From 80ab91e51c0da801ecd40a8e9e55984a45df2f80 Mon Sep 17 00:00:00 2001 From: swordfish-autoresearch Date: Mon, 20 Apr 2026 01:50:16 +0000 Subject: [PATCH 1/2] autoresearch: profile run 20260420T014943Z source SHA: 20ab7f3e9e603f905be9e51b853e15d628b64241 shapes: voice impls: fp16,marlin repeats: 5 GPU: NVIDIA A100-SXM4-80GB marlin SHA: 1f25790bdd49fba53106164a24666dade68d7c90 Headline (8b-b1 marlin): 0.7 TFLOPS --- docs/profiling/INDEX.md | 8 + .../20260420T014943Z/70b-tp2-b1.ncu.csv | 10 + .../20260420T014943Z/70b-tp2-b4.ncu.csv | 10 + .../20260420T014943Z/70b-tp2-b8.ncu.csv | 10 + .../marlin/20260420T014943Z/8b-b1.ncu.csv | 10 + .../marlin/20260420T014943Z/8b-b4.ncu.csv | 10 + .../marlin/20260420T014943Z/8b-b8.ncu.csv | 10 + .../marlin/20260420T014943Z/SUMMARY.md | 26 + .../profiling/marlin/20260420T014943Z/env.txt | 43 + .../marlin/20260420T014943Z/manifest.json | 223 + .../marlin/20260420T014943Z/results.csv | 13 + .../marlin/20260420T014943Z/trace.json | 354751 +++++++++++++++ 12 files changed, 355124 insertions(+) create mode 100644 docs/profiling/INDEX.md create mode 100644 docs/profiling/marlin/20260420T014943Z/70b-tp2-b1.ncu.csv create mode 100644 docs/profiling/marlin/20260420T014943Z/70b-tp2-b4.ncu.csv create mode 100644 docs/profiling/marlin/20260420T014943Z/70b-tp2-b8.ncu.csv create mode 100644 docs/profiling/marlin/20260420T014943Z/8b-b1.ncu.csv create mode 100644 docs/profiling/marlin/20260420T014943Z/8b-b4.ncu.csv create mode 100644 docs/profiling/marlin/20260420T014943Z/8b-b8.ncu.csv create mode 100644 docs/profiling/marlin/20260420T014943Z/SUMMARY.md create mode 100644 docs/profiling/marlin/20260420T014943Z/env.txt create mode 100644 docs/profiling/marlin/20260420T014943Z/manifest.json create mode 100644 docs/profiling/marlin/20260420T014943Z/results.csv create mode 100644 docs/profiling/marlin/20260420T014943Z/trace.json diff --git a/docs/profiling/INDEX.md b/docs/profiling/INDEX.md new file mode 100644 index 0000000..732fc21 --- /dev/null +++ b/docs/profiling/INDEX.md @@ -0,0 +1,8 @@ +# Autoresearch run index + +One row per profiling run produced by the swordfish-autoresearch chart. +Newest first. PR column links to the draft PR carrying the artifacts. + +| timestamp (UTC) | source SHA | shapes | impls | GPU | 8b-b1 marlin TFLOPS | run dir | PR | +|---|---|---|---|---|---|---|---| +| 20260420T014943Z | `20ab7f3` | voice | fp16,marlin | NVIDIA A100-SXM4-80GB | 0.7 | [`20260420T014943Z/`](./marlin/20260420T014943Z/) | _pending_ | diff --git a/docs/profiling/marlin/20260420T014943Z/70b-tp2-b1.ncu.csv b/docs/profiling/marlin/20260420T014943Z/70b-tp2-b1.ncu.csv new file mode 100644 index 0000000..b324fc2 --- /dev/null +++ b/docs/profiling/marlin/20260420T014943Z/70b-tp2-b1.ncu.csv @@ -0,0 +1,10 @@ +==PROF== Connected to process 906 (/usr/bin/python3.10) + +==ERROR== An error was reported by the driver + +==ERROR== Profiling failed because a driver resource was unavailable. Ensure that no other tool (like DCGM) is concurrently collecting profiling data. See https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html#faq for more details. +==ERROR== Failed to profile "distribution_elementwise_grid..." in process 906 +==PROF== Trying to shutdown target application +==ERROR== The application returned an error code (9). +==ERROR== An error occurred while trying to profile. +==WARNING== No kernels were profiled. diff --git a/docs/profiling/marlin/20260420T014943Z/70b-tp2-b4.ncu.csv b/docs/profiling/marlin/20260420T014943Z/70b-tp2-b4.ncu.csv new file mode 100644 index 0000000..5f13bb5 --- /dev/null +++ b/docs/profiling/marlin/20260420T014943Z/70b-tp2-b4.ncu.csv @@ -0,0 +1,10 @@ +==PROF== Connected to process 995 (/usr/bin/python3.10) + +==ERROR== An error was reported by the driver + +==ERROR== Profiling failed because a driver resource was unavailable. Ensure that no other tool (like DCGM) is concurrently collecting profiling data. See https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html#faq for more details. +==ERROR== Failed to profile "distribution_elementwise_grid..." in process 995 +==PROF== Trying to shutdown target application +==ERROR== The application returned an error code (9). +==ERROR== An error occurred while trying to profile. +==WARNING== No kernels were profiled. diff --git a/docs/profiling/marlin/20260420T014943Z/70b-tp2-b8.ncu.csv b/docs/profiling/marlin/20260420T014943Z/70b-tp2-b8.ncu.csv new file mode 100644 index 0000000..9ffe9bb --- /dev/null +++ b/docs/profiling/marlin/20260420T014943Z/70b-tp2-b8.ncu.csv @@ -0,0 +1,10 @@ +==PROF== Connected to process 1084 (/usr/bin/python3.10) + +==ERROR== An error was reported by the driver + +==ERROR== Profiling failed because a driver resource was unavailable. Ensure that no other tool (like DCGM) is concurrently collecting profiling data. See https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html#faq for more details. +==ERROR== Failed to profile "distribution_elementwise_grid..." in process 1084 +==PROF== Trying to shutdown target application +==ERROR== The application returned an error code (9). +==ERROR== An error occurred while trying to profile. +==WARNING== No kernels were profiled. diff --git a/docs/profiling/marlin/20260420T014943Z/8b-b1.ncu.csv b/docs/profiling/marlin/20260420T014943Z/8b-b1.ncu.csv new file mode 100644 index 0000000..abd03e8 --- /dev/null +++ b/docs/profiling/marlin/20260420T014943Z/8b-b1.ncu.csv @@ -0,0 +1,10 @@ +==PROF== Connected to process 639 (/usr/bin/python3.10) + +==ERROR== An error was reported by the driver + +==ERROR== Profiling failed because a driver resource was unavailable. Ensure that no other tool (like DCGM) is concurrently collecting profiling data. See https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html#faq for more details. +==ERROR== Failed to profile "distribution_elementwise_grid..." in process 639 +==PROF== Trying to shutdown target application +==ERROR== The application returned an error code (9). +==ERROR== An error occurred while trying to profile. +==WARNING== No kernels were profiled. diff --git a/docs/profiling/marlin/20260420T014943Z/8b-b4.ncu.csv b/docs/profiling/marlin/20260420T014943Z/8b-b4.ncu.csv new file mode 100644 index 0000000..8d76f1f --- /dev/null +++ b/docs/profiling/marlin/20260420T014943Z/8b-b4.ncu.csv @@ -0,0 +1,10 @@ +==PROF== Connected to process 728 (/usr/bin/python3.10) + +==ERROR== An error was reported by the driver + +==ERROR== Profiling failed because a driver resource was unavailable. Ensure that no other tool (like DCGM) is concurrently collecting profiling data. See https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html#faq for more details. +==ERROR== Failed to profile "distribution_elementwise_grid..." in process 728 +==PROF== Trying to shutdown target application +==ERROR== The application returned an error code (9). +==ERROR== An error occurred while trying to profile. +==WARNING== No kernels were profiled. diff --git a/docs/profiling/marlin/20260420T014943Z/8b-b8.ncu.csv b/docs/profiling/marlin/20260420T014943Z/8b-b8.ncu.csv new file mode 100644 index 0000000..9b0a0f1 --- /dev/null +++ b/docs/profiling/marlin/20260420T014943Z/8b-b8.ncu.csv @@ -0,0 +1,10 @@ +==PROF== Connected to process 817 (/usr/bin/python3.10) + +==ERROR== An error was reported by the driver + +==ERROR== Profiling failed because a driver resource was unavailable. Ensure that no other tool (like DCGM) is concurrently collecting profiling data. See https://docs.nvidia.com/nsight-compute/ProfilingGuide/index.html#faq for more details. +==ERROR== Failed to profile "distribution_elementwise_grid..." in process 817 +==PROF== Trying to shutdown target application +==ERROR== The application returned an error code (9). +==ERROR== An error occurred while trying to profile. +==WARNING== No kernels were profiled. diff --git a/docs/profiling/marlin/20260420T014943Z/SUMMARY.md b/docs/profiling/marlin/20260420T014943Z/SUMMARY.md new file mode 100644 index 0000000..1b280f8 --- /dev/null +++ b/docs/profiling/marlin/20260420T014943Z/SUMMARY.md @@ -0,0 +1,26 @@ +# Autoresearch run `20260420T014943Z` + +- **source SHA:** `20ab7f3` +- **GPU:** NVIDIA A100-SXM4-80GB (cc 8.0, 79.3 GB) +- **CUDA / torch / triton:** 12.4 / 2.4.0a0+07cecf4168.nv24.05 / 3.0.0 +- **shapes:** `voice` **impls:** `fp16,marlin` **repeats:** 5 +- **marlin SHA:** `1f25790bdd49fba53106164a24666dade68d7c90` + +## Results + +| shape | impl | ms_mean | ms_p95 | TFLOPS | speedup vs fp16 | error | +|---|---|---|---|---|---|---| +| 8b-b1 | fp16 | 0.031 | 0.033 | 1.1 | x1.00 | | +| 8b-b1 | marlin | 0.049 | 0.051 | 0.7 | x0.64 | | +| 8b-b4 | fp16 | 0.031 | 0.031 | 4.4 | x1.00 | | +| 8b-b4 | marlin | 0.050 | 0.050 | 2.7 | x0.61 | | +| 8b-b8 | fp16 | 0.031 | 0.032 | 8.6 | x1.00 | | +| 8b-b8 | marlin | 0.050 | 0.050 | 5.4 | x0.63 | | +| 70b-tp2-b1 | fp16 | 0.051 | 0.056 | 1.3 | x1.00 | | +| 70b-tp2-b1 | marlin | 0.049 | 0.050 | 1.4 | x1.02 | | +| 70b-tp2-b4 | fp16 | 0.049 | 0.050 | 5.4 | x1.00 | | +| 70b-tp2-b4 | marlin | 0.066 | 0.133 | 4.1 | x0.75 | | +| 70b-tp2-b8 | fp16 | 0.049 | 0.050 | 10.8 | x1.00 | | +| 70b-tp2-b8 | marlin | 0.049 | 0.049 | 10.9 | x1.01 | | + +![roofline](./roofline.png) diff --git a/docs/profiling/marlin/20260420T014943Z/env.txt b/docs/profiling/marlin/20260420T014943Z/env.txt new file mode 100644 index 0000000..fe16e63 --- /dev/null +++ b/docs/profiling/marlin/20260420T014943Z/env.txt @@ -0,0 +1,43 @@ +=== profile_marlin.sh @ 20260420T014943Z === +--- host --- +Linux swordfish-profile-sf-prof-260420-014607-r5hlj 6.6.126.1-1.azl3 #1 SMP PREEMPT_DYNAMIC Wed Mar 4 05:04:40 UTC 2026 x86_64 x86_64 x86_64 GNU/Linux +--- nvidia-smi --- +Mon Apr 20 01:49:43 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 580.105.08 Driver Version: 580.105.08 CUDA Version: 13.0 | ++-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA A100-SXM4-80GB On | 0000000E:00:00.0 Off | 0 | +| N/A 34C P0 70W / 400W | 0MiB / 81920MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| No running processes found | ++-----------------------------------------------------------------------------------------+ +--- nvcc --- +nvcc: NVIDIA (R) Cuda compiler driver +Copyright (c) 2005-2024 NVIDIA Corporation +Built on Thu_Mar_28_02:18:24_PDT_2024 +Cuda compilation tools, release 12.4, V12.4.131 +Build cuda_12.4.r12.4/compiler.34097967_0 +--- nsys --- +NVIDIA Nsight Systems version 2024.2.1.106-242134037904v0 +--- ncu --- +NVIDIA (R) Nsight Compute Command Line Profiler +Copyright (c) 2018-2024 NVIDIA Corporation +Version 2024.1.1.0 (build 33998838) (public-release) +--- python / torch / triton / marlin --- +python 3.10.12 +torch 2.4.0a0+07cecf4168.nv24.05 cuda 12.4 +triton 3.0.0 +marlin unknown +--- repo SHA --- +20ab7f3e9e603f905be9e51b853e15d628b64241 diff --git a/docs/profiling/marlin/20260420T014943Z/manifest.json b/docs/profiling/marlin/20260420T014943Z/manifest.json new file mode 100644 index 0000000..84afc7d --- /dev/null +++ b/docs/profiling/marlin/20260420T014943Z/manifest.json @@ -0,0 +1,223 @@ +{ + "env": { + "host": "swordfish-profile-sf-prof-260420-014607-r5hlj", + "os": "Linux 6.6.126.1-1.azl3", + "python": "3.10.12", + "torch": "2.4.0a0+07cecf4168.nv24.05", + "cuda_available": true, + "timestamp": "2026-04-20T01:49:49+0000", + "gpu_name": "NVIDIA A100-SXM4-80GB", + "gpu_cc": "8.0", + "gpu_mem_gb": 79.3, + "gpu_sm_count": 108, + "torch_cuda": "12.4", + "cudnn": 90100, + "triton": "3.0.0" + }, + "rows": [ + { + "name": "8b-b1", + "M": 1, + "N": 4096, + "K": 4096, + "group_size": 128, + "priority": 0, + "tag": "llama-3-8b", + "impl": "fp16", + "ms_mean": 0.03149350357055664, + "ms_p50": 0.031086080074310303, + "ms_p95": 0.03273663997650147, + "ms_min": 0.030670719146728517, + "tflops_mean": 1.0654397953795818, + "error": null, + "speedup_vs_fp16": 1.0 + }, + { + "name": "8b-b1", + "M": 1, + "N": 4096, + "K": 4096, + "group_size": 128, + "priority": 0, + "tag": "llama-3-8b", + "impl": "marlin", + "ms_mean": 0.049491456031799316, + "ms_p50": 0.049548802375793455, + "ms_p95": 0.05056960105895996, + "ms_min": 0.04887360095977783, + "tflops_mean": 0.6779843369013141, + "error": null, + "speedup_vs_fp16": 0.6363422314817611 + }, + { + "name": "8b-b4", + "M": 4, + "N": 4096, + "K": 4096, + "group_size": 128, + "priority": 0, + "tag": "llama-3-8b", + "impl": "fp16", + "ms_mean": 0.03054387187957764, + "ms_p50": 0.030387840270996093, + "ms_p95": 0.030972800254821777, + "ms_min": 0.03024319887161255, + "tflops_mean": 4.394260443769775, + "error": null, + "speedup_vs_fp16": 1.0 + }, + { + "name": "8b-b4", + "M": 4, + "N": 4096, + "K": 4096, + "group_size": 128, + "priority": 0, + "tag": "llama-3-8b", + "impl": "marlin", + "ms_mean": 0.04976857566833496, + "ms_p50": 0.04982399940490723, + "ms_p95": 0.05031424045562744, + "ms_min": 0.049259519577026366, + "tflops_mean": 2.6968368332348205, + "error": null, + "speedup_vs_fp16": 0.6137180232588221 + }, + { + "name": "8b-b8", + "M": 8, + "N": 4096, + "K": 4096, + "group_size": 128, + "priority": 0, + "tag": "llama-3-8b", + "impl": "fp16", + "ms_mean": 0.031084928035736082, + "ms_p50": 0.0308134388923645, + "ms_p95": 0.03185983896255493, + "ms_min": 0.03056960105895996, + "tflops_mean": 8.635550183400756, + "error": null, + "speedup_vs_fp16": 1.0 + }, + { + "name": "8b-b8", + "M": 8, + "N": 4096, + "K": 4096, + "group_size": 128, + "priority": 0, + "tag": "llama-3-8b", + "impl": "marlin", + "ms_mean": 0.04965529632568359, + "ms_p50": 0.04962175846099853, + "ms_p95": 0.05010816097259521, + "ms_min": 0.049258241653442385, + "tflops_mean": 5.405978331885517, + "error": null, + "speedup_vs_fp16": 0.626014349644668 + }, + { + "name": "70b-tp2-b1", + "M": 1, + "N": 8192, + "K": 4096, + "group_size": 128, + "priority": 0, + "tag": "llama-3-70b", + "impl": "fp16", + "ms_mean": 0.050512895584106446, + "ms_p50": 0.04921984195709229, + "ms_p95": 0.05594431877136231, + "ms_min": 0.04904191970825195, + "tflops_mean": 1.3285491402539071, + "error": null, + "speedup_vs_fp16": 1.0 + }, + { + "name": "70b-tp2-b1", + "M": 1, + "N": 8192, + "K": 4096, + "group_size": 128, + "priority": 0, + "tag": "llama-3-70b", + "impl": "marlin", + "ms_mean": 0.04932083320617676, + "ms_p50": 0.04902592182159424, + "ms_p95": 0.05009727954864502, + "ms_min": 0.048891520500183104, + "tflops_mean": 1.360659576034809, + "error": null, + "speedup_vs_fp16": 1.0241695506835111 + }, + { + "name": "70b-tp2-b4", + "M": 4, + "N": 8192, + "K": 4096, + "group_size": 128, + "priority": 0, + "tag": "llama-3-70b", + "impl": "fp16", + "ms_mean": 0.049335807800292966, + "ms_p50": 0.04932799816131592, + "ms_p95": 0.04952127933502197, + "ms_min": 0.049187841415405276, + "tflops_mean": 5.440986333630194, + "error": null, + "speedup_vs_fp16": 1.0 + }, + { + "name": "70b-tp2-b4", + "M": 4, + "N": 8192, + "K": 4096, + "group_size": 128, + "priority": 0, + "tag": "llama-3-70b", + "impl": "marlin", + "ms_mean": 0.06605465698242188, + "ms_p50": 0.04948544025421143, + "ms_p95": 0.1330508804321289, + "ms_min": 0.04902656078338623, + "tflops_mean": 4.063838467459374, + "error": null, + "speedup_vs_fp16": 0.7468937097564817 + }, + { + "name": "70b-tp2-b8", + "M": 8, + "N": 8192, + "K": 4096, + "group_size": 128, + "priority": 0, + "tag": "llama-3-70b", + "impl": "fp16", + "ms_mean": 0.049490047454833985, + "ms_p50": 0.04956799983978272, + "ms_p95": 0.0496127986907959, + "ms_min": 0.049237117767333985, + "tflops_mean": 10.848058137142898, + "error": null, + "speedup_vs_fp16": 1.0 + }, + { + "name": "70b-tp2-b8", + "M": 8, + "N": 8192, + "K": 4096, + "group_size": 128, + "priority": 0, + "tag": "llama-3-70b", + "impl": "marlin", + "ms_mean": 0.04924096012115479, + "ms_p50": 0.04936704158782959, + "ms_p95": 0.049479680061340334, + "ms_min": 0.04883967876434326, + "tflops_mean": 10.902933465940905, + "error": null, + "speedup_vs_fp16": 1.0050585393352673 + } + ] +} \ No newline at end of file diff --git a/docs/profiling/marlin/20260420T014943Z/results.csv b/docs/profiling/marlin/20260420T014943Z/results.csv new file mode 100644 index 0000000..a0d196d --- /dev/null +++ b/docs/profiling/marlin/20260420T014943Z/results.csv @@ -0,0 +1,13 @@ +name,impl,M,N,K,group_size,priority,tag,ms_mean,ms_p50,ms_p95,ms_min,tflops_mean,speedup_vs_fp16,error +8b-b1,fp16,1,4096,4096,128,0,llama-3-8b,0.03149350357055664,0.031086080074310303,0.03273663997650147,0.030670719146728517,1.0654397953795818,1.0, +8b-b1,marlin,1,4096,4096,128,0,llama-3-8b,0.049491456031799316,0.049548802375793455,0.05056960105895996,0.04887360095977783,0.6779843369013141,0.6363422314817611, +8b-b4,fp16,4,4096,4096,128,0,llama-3-8b,0.03054387187957764,0.030387840270996093,0.030972800254821777,0.03024319887161255,4.394260443769775,1.0, +8b-b4,marlin,4,4096,4096,128,0,llama-3-8b,0.04976857566833496,0.04982399940490723,0.05031424045562744,0.049259519577026366,2.6968368332348205,0.6137180232588221, +8b-b8,fp16,8,4096,4096,128,0,llama-3-8b,0.031084928035736082,0.0308134388923645,0.03185983896255493,0.03056960105895996,8.635550183400756,1.0, +8b-b8,marlin,8,4096,4096,128,0,llama-3-8b,0.04965529632568359,0.04962175846099853,0.05010816097259521,0.049258241653442385,5.405978331885517,0.626014349644668, +70b-tp2-b1,fp16,1,8192,4096,128,0,llama-3-70b,0.050512895584106446,0.04921984195709229,0.05594431877136231,0.04904191970825195,1.3285491402539071,1.0, +70b-tp2-b1,marlin,1,8192,4096,128,0,llama-3-70b,0.04932083320617676,0.04902592182159424,0.05009727954864502,0.048891520500183104,1.360659576034809,1.0241695506835111, +70b-tp2-b4,fp16,4,8192,4096,128,0,llama-3-70b,0.049335807800292966,0.04932799816131592,0.04952127933502197,0.049187841415405276,5.440986333630194,1.0, +70b-tp2-b4,marlin,4,8192,4096,128,0,llama-3-70b,0.06605465698242188,0.04948544025421143,0.1330508804321289,0.04902656078338623,4.063838467459374,0.7468937097564817, +70b-tp2-b8,fp16,8,8192,4096,128,0,llama-3-70b,0.049490047454833985,0.04956799983978272,0.0496127986907959,0.049237117767333985,10.848058137142898,1.0, +70b-tp2-b8,marlin,8,8192,4096,128,0,llama-3-70b,0.04924096012115479,0.04936704158782959,0.049479680061340334,0.04883967876434326,10.902933465940905,1.0050585393352673, diff --git a/docs/profiling/marlin/20260420T014943Z/trace.json b/docs/profiling/marlin/20260420T014943Z/trace.json new file mode 100644 index 0000000..3a10c7a --- /dev/null +++ b/docs/profiling/marlin/20260420T014943Z/trace.json @@ -0,0 +1,354751 @@ + +{ + "schemaVersion": 1, + "deviceProperties": [ + { + "id": 0, "name": "NVIDIA A100-SXM4-80GB", "totalGlobalMem": 85094825984, + "computeMajor": 8, "computeMinor": 0, + "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048, + "regsPerBlock": 65536, "regsPerMultiprocessor": 65536, "warpSize": 32, + "sharedMemPerBlock": 49152, "sharedMemPerMultiprocessor": 167936, + "numSms": 108, "sharedMemPerBlockOptin": 166912 + } + ], + "with_flops": 1, + "record_shapes": 1, + "traceEvents": [ + { + "ph": "X", "cat": "cpu_op", "name": "aten::randn", "pid": 435, "tid": 435, + "ts": 1776649789511767, "dur": 184994, + "args": { + "External id": 1,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 0 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649789511843, "dur": 183056, + "args": { + "External id": 2,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::normal_", "pid": 435, "tid": 435, + "ts": 1776649789695046, "dur": 1710, + "args": { + "External id": 3,"Record function id": 0, "Concrete Inputs": ["", "0.", "1.", ""], "Input type": ["c10::Half", "Scalar", "Scalar", ""], "Input Dims": [[1, 4096], [], [], []], "Ev Idx": 2 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::randn", "pid": 435, "tid": 435, + "ts": 1776649789696900, "dur": 337, + "args": { + "External id": 4,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "", "5", "", "", "False"], "Input type": ["ScalarList", "", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649789696919, "dur": 280, + "args": { + "External id": 5,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::normal_", "pid": 435, "tid": 435, + "ts": 1776649789697204, "dur": 32, + "args": { + "External id": 6,"Record function id": 0, "Concrete Inputs": ["", "0.", "1.", ""], "Input type": ["c10::Half", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 5 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 435, "tid": 435, + "ts": 1776649789697402, "dur": 10889, + "args": { + "External id": 7,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "double"], "Input Dims": [[4096, 4096], []], "Ev Idx": 6 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649789708510, "dur": 92, + "args": { + "External id": 8,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 4096], []], "Ev Idx": 7 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649789708537, "dur": 65, + "args": { + "External id": 9,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 4096], []], "Ev Idx": 8 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 435, "tid": 435, + "ts": 1776649789708654, "dur": 2811, + "args": { + "External id": 10,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[32, 128, 4096]], "Ev Idx": 9 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649789708678, "dur": 17, + "args": { + "External id": 11,"Record function id": 0, "Concrete Inputs": ["[0]", "5", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 435, "tid": 435, + "ts": 1776649789708709, "dur": 2753, + "args": { + "External id": 12,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[32, 128, 4096], [0]], "Ev Idx": 11 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 435, "tid": 435, + "ts": 1776649789708764, "dur": 13, + "args": { + "External id": 13,"Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]", ""], "Input type": ["c10::Half", "ScalarList", ""], "Input Dims": [[0], [], []], "Ev Idx": 12 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 435, "tid": 435, + "ts": 1776649789711537, "dur": 15561, + "args": { + "External id": 14,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1]", "False"], "Input type": ["c10::Half", "ScalarList", "Scalar"], "Input Dims": [[32, 128, 4096], [], []], "Ev Idx": 13 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649789711602, "dur": 17, + "args": { + "External id": 15,"Record function id": 0, "Concrete Inputs": ["", "[32, 1, 4096]", "[4096, 0, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 14 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 435, "tid": 435, + "ts": 1776649789727284, "dur": 7699, + "args": { + "External id": 16,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "double"], "Input Dims": [[32, 4096], []], "Ev Idx": 15 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 435, "tid": 435, + "ts": 1776649789735124, "dur": 12957, + "args": { + "External id": 17,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "9.9999999999999995e-07", ""], "Input type": ["c10::Half", "Scalar", ""], "Input Dims": [[32, 4096], [], []], "Ev Idx": 16 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649789735166, "dur": 7, + "args": { + "External id": 18,"Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 4096], [], [], [], []], "Ev Idx": 17 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649789748295, "dur": 2, + "args": { + "External id": 19,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 4096], [], [], [], []], "Ev Idx": 18 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649789748347, "dur": 69, + "args": { + "External id": 20,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 4096], []], "Ev Idx": 19 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649789748400, "dur": 9, + "args": { + "External id": 21,"Record function id": 0, "Concrete Inputs": ["", "[32, 1, 4096]", "[4096, 4096, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 20 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 435, "tid": 435, + "ts": 1776649789748430, "dur": 120, + "args": { + "External id": 22,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[32, 128, 4096], [32, 1, 4096]], "Ev Idx": 21 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 435, "tid": 435, + "ts": 1776649789748626, "dur": 8928, + "args": { + "External id": 23,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[32, 128, 4096]], "Ev Idx": 22 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 435, "tid": 435, + "ts": 1776649789757618, "dur": 60, + "args": { + "External id": 24,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "-8", "7"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[32, 128, 4096], [], []], "Ev Idx": 23 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649789757628, "dur": 1, + "args": { + "External id": 25,"Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 128, 4096], [], [], [], []], "Ev Idx": 24 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649789757688, "dur": 8977, + "args": { + "External id": 26,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 128, 4096], [], [], [], []], "Ev Idx": 25 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649789757740, "dur": 8923, + "args": { + "External id": 27,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["c10::Half", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[32, 128, 4096], [], [], [], [], [], []], "Ev Idx": 26 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649789757798, "dur": 17, + "args": { + "External id": 28,"Record function id": 0, "Concrete Inputs": ["[32, 128, 4096]", "[524288, 4096, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 27 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649789757835, "dur": 8818, + "args": { + "External id": 29,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "c10::Half", "Scalar"], "Input Dims": [[32, 128, 4096], [32, 128, 4096], []], "Ev Idx": 28 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 435, "tid": 435, + "ts": 1776649789766826, "dur": 10349, + "args": { + "External id": 30,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["signed char", "long int", "Scalar"], "Input Dims": [[32, 128, 4096], [], []], "Ev Idx": 29 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649789777249, "dur": 158, + "args": { + "External id": 31,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "False", "False", ""], "Input type": ["signed char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 128, 4096], [], [], [], []], "Ev Idx": 30 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649789777255, "dur": 152, + "args": { + "External id": 32,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "", "", "", "False", ""], "Input type": ["signed char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[32, 128, 4096], [], [], [], [], [], []], "Ev Idx": 31 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649789777277, "dur": 26, + "args": { + "External id": 33,"Record function id": 0, "Concrete Inputs": ["[32, 128, 4096]", "[524288, 4096, 1]", "0", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 32 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649789777307, "dur": 99, + "args": { + "External id": 34,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["unsigned char", "signed char", "Scalar"], "Input Dims": [[32, 128, 4096], [32, 128, 4096], []], "Ev Idx": 33 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649789777442, "dur": 20, + "args": { + "External id": 35,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["unsigned char", "ScalarList"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 34 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649789777445, "dur": 17, + "args": { + "External id": 36,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["unsigned char", "ScalarList"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 35 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649789777539, "dur": 86, + "args": { + "External id": 37,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 36 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649789777610, "dur": 3, + "args": { + "External id": 38,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]", "[4096, 1]", "0"], "Input type": ["unsigned char", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 37 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649789777628, "dur": 5, + "args": { + "External id": 39,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "2"], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 38 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649789777630, "dur": 2, + "args": { + "External id": 40,"Record function id": 0, "Concrete Inputs": ["", "[4096, 2048]", "[4096, 2]", "0"], "Input type": ["unsigned char", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 39 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649789777644, "dur": 3, + "args": { + "External id": 41,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 40 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649789777645, "dur": 1, + "args": { + "External id": 42,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]", "[4096, 1]", "0"], "Input type": ["unsigned char", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 41 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649789777648, "dur": 3, + "args": { + "External id": 43,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "1", "9223372036854775807", "2"], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 42 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649789777650, "dur": 1, + "args": { + "External id": 44,"Record function id": 0, "Concrete Inputs": ["", "[4096, 2048]", "[4096, 2]", "1"], "Input type": ["unsigned char", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 43 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__lshift__", "pid": 435, "tid": 435, + "ts": 1776649789777727, "dur": 8247, + "args": { + "External id": 45,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 44 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649789777778, "dur": 11, + "args": { + "External id": 46,"Record function id": 0, "Concrete Inputs": ["[4096, 2048]", "0", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 45 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__or__", "pid": 435, "tid": 435, + "ts": 1776649789786036, "dur": 9837, + "args": { + "External id": 47,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["unsigned char", "unsigned char"], "Input Dims": [[4096, 2048], [4096, 2048]], "Ev Idx": 46 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_or", "pid": 435, "tid": 435, + "ts": 1776649789786054, "dur": 9812, + "args": { + "External id": 48,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["unsigned char", "unsigned char"], "Input Dims": [[4096, 2048], [4096, 2048]], "Ev Idx": 47 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649789795952, "dur": 1, + "args": { + "External id": 49,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], []], "Ev Idx": 48 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__and__", "pid": 435, "tid": 435, + "ts": 1776649789796075, "dur": 136, + "args": { + "External id": 50,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "15"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 49 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_and", "pid": 435, "tid": 435, + "ts": 1776649789796086, "dur": 124, + "args": { + "External id": 51,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "15"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 50 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_and", "pid": 435, "tid": 435, + "ts": 1776649789796122, "dur": 85, + "args": { + "External id": 52,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["unsigned char", "long int"], "Input Dims": [[4096, 2048], []], "Ev Idx": 51 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649789796248, "dur": 74, + "args": { + "External id": 53,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], []], "Ev Idx": 52 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649789796252, "dur": 69, + "args": { + "External id": 54,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["unsigned char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], [], [], []], "Ev Idx": 53 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649789796271, "dur": 13, + "args": { + "External id": 55,"Record function id": 0, "Concrete Inputs": ["[4096, 2048]", "[2048, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 54 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649789796288, "dur": 32, + "args": { + "External id": 56,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "unsigned char", "Scalar"], "Input Dims": [[4096, 2048], [4096, 2048], []], "Ev Idx": 55 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__rshift__", "pid": 435, "tid": 435, + "ts": 1776649789796351, "dur": 76, + "args": { + "External id": 57,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 56 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649789796365, "dur": 9, + "args": { + "External id": 58,"Record function id": 0, "Concrete Inputs": ["[4096, 2048]", "0", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 57 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649789796433, "dur": 27, + "args": { + "External id": 59,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], []], "Ev Idx": 58 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649789796435, "dur": 25, + "args": { + "External id": 60,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["unsigned char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], [], [], []], "Ev Idx": 59 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649789796438, "dur": 6, + "args": { + "External id": 61,"Record function id": 0, "Concrete Inputs": ["[4096, 2048]", "[2048, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 60 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649789796445, "dur": 14, + "args": { + "External id": 62,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "unsigned char", "Scalar"], "Input Dims": [[4096, 2048], [4096, 2048], []], "Ev Idx": 61 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::stack", "pid": 435, "tid": 435, + "ts": 1776649789796569, "dur": 4187, + "args": { + "External id": 63,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["TensorList", "Scalar"], "Input Dims": [[[4096, 2048], [4096, 2048]], []], "Ev Idx": 62 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649789796597, "dur": 9, + "args": { + "External id": 64,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["signed char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 63 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649789796601, "dur": 4, + "args": { + "External id": 65,"Record function id": 0, "Concrete Inputs": ["", "[4096, 2048, 1]", "[2048, 1, 1]", ""], "Input type": ["signed char", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 2048], [], [], []], "Ev Idx": 64 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649789796607, "dur": 2, + "args": { + "External id": 66,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["signed char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 65 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649789796608, "dur": 0, + "args": { + "External id": 67,"Record function id": 0, "Concrete Inputs": ["", "[4096, 2048, 1]", "[2048, 1, 1]", ""], "Input type": ["signed char", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 2048], [], [], []], "Ev Idx": 66 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 435, "tid": 435, + "ts": 1776649789796614, "dur": 4129, + "args": { + "External id": 68,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["TensorList", "Scalar"], "Input Dims": [[[4096, 2048, 1], [4096, 2048, 1]], []], "Ev Idx": 67 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649789800784, "dur": 16, + "args": { + "External id": 69,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["signed char", "ScalarList"], "Input Dims": [[4096, 2048, 2], []], "Ev Idx": 68 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649789800787, "dur": 12, + "args": { + "External id": 70,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["signed char", "ScalarList"], "Input Dims": [[4096, 2048, 2], []], "Ev Idx": 69 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::repeat_interleave", "pid": 435, "tid": 435, + "ts": 1776649789800854, "dur": 219, + "args": { + "External id": 71,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "128", "0", ""], "Input type": ["c10::Half", "Scalar", "Scalar", ""], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 70 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649789800866, "dur": 8, + "args": { + "External id": 72,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 4096], []], "Ev Idx": 71 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649789800871, "dur": 2, + "args": { + "External id": 73,"Record function id": 0, "Concrete Inputs": ["", "[32, 1, 4096]", "[4096, 4096, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 72 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 435, "tid": 435, + "ts": 1776649789800895, "dur": 18, + "args": { + "External id": 74,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]", "False"], "Input type": ["c10::Half", "ScalarList", "Scalar"], "Input Dims": [[32, 1, 4096], [], []], "Ev Idx": 73 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649789800909, "dur": 2, + "args": { + "External id": 75,"Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]", "[4096, 0, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 1, 4096], [], [], []], "Ev Idx": 74 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649789800934, "dur": 122, + "args": { + "External id": 76,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 75 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649789800952, "dur": 18, + "args": { + "External id": 77,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[32, 128, 4096], [], [], [], [], []], "Ev Idx": 76 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649789800956, "dur": 13, + "args": { + "External id": 78,"Record function id": 0, "Concrete Inputs": ["[32, 128, 4096]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 77 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649789800971, "dur": 84, + "args": { + "External id": 79,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[32, 128, 4096], [32, 128, 4096], []], "Ev Idx": 78 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::flatten", "pid": 435, "tid": 435, + "ts": 1776649789801064, "dur": 8, + "args": { + "External id": 80,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[32, 128, 4096], [], []], "Ev Idx": 79 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649789801068, "dur": 3, + "args": { + "External id": 81,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 80 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 435, "tid": 435, + "ts": 1776649789801114, "dur": 54, + "args": { + "External id": 82,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["signed char", "long int", "Scalar"], "Input Dims": [[4096, 4096], [], []], "Ev Idx": 81 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649789801189, "dur": 528, + "args": { + "External id": 83,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["signed char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 82 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649789801191, "dur": 525, + "args": { + "External id": 84,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "", "", "", "False", ""], "Input type": ["signed char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], [], []], "Ev Idx": 83 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649789801195, "dur": 440, + "args": { + "External id": 85,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "[4096, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 84 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649789801639, "dur": 76, + "args": { + "External id": 86,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "signed char", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 85 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 435, "tid": 435, + "ts": 1776649789801738, "dur": 284, + "args": { + "External id": 87,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4096, 4096], [4096, 4096]], "Ev Idx": 86 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789802301, "dur": 175529, + "args": { + "External id": 88,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 87 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789802320, "dur": 175494, + "args": { + "External id": 89,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 88 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789977921, "dur": 114, + "args": { + "External id": 90,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 89 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789977944, "dur": 90, + "args": { + "External id": 91,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 90 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789978042, "dur": 33, + "args": { + "External id": 92,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 91 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789978049, "dur": 25, + "args": { + "External id": 93,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 92 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789978081, "dur": 26, + "args": { + "External id": 94,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 93 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789978083, "dur": 24, + "args": { + "External id": 95,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 94 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789978113, "dur": 25, + "args": { + "External id": 96,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 95 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789978115, "dur": 22, + "args": { + "External id": 97,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 96 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789978143, "dur": 26, + "args": { + "External id": 98,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 97 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789978145, "dur": 23, + "args": { + "External id": 99,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 98 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789978173, "dur": 26, + "args": { + "External id": 100,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 99 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789978176, "dur": 22, + "args": { + "External id": 101,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 100 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789978203, "dur": 26, + "args": { + "External id": 102,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 101 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789978206, "dur": 22, + "args": { + "External id": 103,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 102 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789978233, "dur": 25, + "args": { + "External id": 104,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 103 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789978236, "dur": 22, + "args": { + "External id": 105,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 104 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789978263, "dur": 25, + "args": { + "External id": 106,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 105 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789978265, "dur": 22, + "args": { + "External id": 107,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 106 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789978692, "dur": 42, + "args": { + "External id": 108,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 107 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789978696, "dur": 37, + "args": { + "External id": 109,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 108 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789978741, "dur": 27, + "args": { + "External id": 110,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 109 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789978744, "dur": 23, + "args": { + "External id": 111,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 110 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789978773, "dur": 25, + "args": { + "External id": 112,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 111 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789978775, "dur": 22, + "args": { + "External id": 113,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 112 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789978802, "dur": 25, + "args": { + "External id": 114,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 113 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789978805, "dur": 22, + "args": { + "External id": 115,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 114 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789978831, "dur": 26, + "args": { + "External id": 116,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 115 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789978834, "dur": 22, + "args": { + "External id": 117,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 116 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789978861, "dur": 25, + "args": { + "External id": 118,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 117 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789978863, "dur": 22, + "args": { + "External id": 119,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 118 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789978890, "dur": 31, + "args": { + "External id": 120,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 119 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789978897, "dur": 23, + "args": { + "External id": 121,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 120 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789978925, "dur": 26, + "args": { + "External id": 122,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 121 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789978928, "dur": 22, + "args": { + "External id": 123,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 122 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789978955, "dur": 25, + "args": { + "External id": 124,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 123 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789978958, "dur": 22, + "args": { + "External id": 125,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 124 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789978984, "dur": 25, + "args": { + "External id": 126,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 125 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789978987, "dur": 22, + "args": { + "External id": 127,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 126 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979014, "dur": 24, + "args": { + "External id": 128,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 127 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979016, "dur": 22, + "args": { + "External id": 129,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 128 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979043, "dur": 24, + "args": { + "External id": 130,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 129 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979045, "dur": 22, + "args": { + "External id": 131,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 130 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979071, "dur": 25, + "args": { + "External id": 132,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 131 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979074, "dur": 22, + "args": { + "External id": 133,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 132 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979101, "dur": 29, + "args": { + "External id": 134,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 133 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979103, "dur": 26, + "args": { + "External id": 135,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 134 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979134, "dur": 26, + "args": { + "External id": 136,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 135 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979137, "dur": 22, + "args": { + "External id": 137,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 136 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979164, "dur": 25, + "args": { + "External id": 138,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 137 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979167, "dur": 21, + "args": { + "External id": 139,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 138 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979193, "dur": 25, + "args": { + "External id": 140,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 139 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979196, "dur": 21, + "args": { + "External id": 141,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 140 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979223, "dur": 24, + "args": { + "External id": 142,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 141 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979225, "dur": 21, + "args": { + "External id": 143,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 142 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979251, "dur": 25, + "args": { + "External id": 144,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 143 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979254, "dur": 21, + "args": { + "External id": 145,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 144 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979281, "dur": 24, + "args": { + "External id": 146,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 145 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979283, "dur": 22, + "args": { + "External id": 147,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 146 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979310, "dur": 25, + "args": { + "External id": 148,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 147 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979312, "dur": 22, + "args": { + "External id": 149,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 148 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979339, "dur": 29, + "args": { + "External id": 150,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 149 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979345, "dur": 22, + "args": { + "External id": 151,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 150 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979373, "dur": 33, + "args": { + "External id": 152,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 151 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979375, "dur": 30, + "args": { + "External id": 153,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 152 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979411, "dur": 25, + "args": { + "External id": 154,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 153 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979413, "dur": 22, + "args": { + "External id": 155,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 154 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979440, "dur": 26, + "args": { + "External id": 156,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 155 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979443, "dur": 22, + "args": { + "External id": 157,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 156 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979470, "dur": 25, + "args": { + "External id": 158,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 157 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979472, "dur": 22, + "args": { + "External id": 159,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 158 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979499, "dur": 25, + "args": { + "External id": 160,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 159 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979502, "dur": 22, + "args": { + "External id": 161,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 160 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979529, "dur": 24, + "args": { + "External id": 162,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 161 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979531, "dur": 22, + "args": { + "External id": 163,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 162 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979558, "dur": 34, + "args": { + "External id": 164,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 163 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979568, "dur": 23, + "args": { + "External id": 165,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 164 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979596, "dur": 26, + "args": { + "External id": 166,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 165 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979599, "dur": 22, + "args": { + "External id": 167,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 166 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979626, "dur": 25, + "args": { + "External id": 168,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 167 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979628, "dur": 22, + "args": { + "External id": 169,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 168 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979655, "dur": 25, + "args": { + "External id": 170,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 169 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979657, "dur": 22, + "args": { + "External id": 171,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 170 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979684, "dur": 25, + "args": { + "External id": 172,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 171 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979686, "dur": 22, + "args": { + "External id": 173,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 172 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979713, "dur": 25, + "args": { + "External id": 174,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 173 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979716, "dur": 21, + "args": { + "External id": 175,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 174 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979742, "dur": 25, + "args": { + "External id": 176,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 175 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979745, "dur": 21, + "args": { + "External id": 177,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 176 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979772, "dur": 29, + "args": { + "External id": 178,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 177 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979774, "dur": 27, + "args": { + "External id": 179,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 178 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979806, "dur": 25, + "args": { + "External id": 180,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 179 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979808, "dur": 23, + "args": { + "External id": 181,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 180 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979836, "dur": 24, + "args": { + "External id": 182,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 181 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979838, "dur": 22, + "args": { + "External id": 183,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 182 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979865, "dur": 25, + "args": { + "External id": 184,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 183 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979867, "dur": 22, + "args": { + "External id": 185,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 184 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979894, "dur": 25, + "args": { + "External id": 186,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 185 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979896, "dur": 22, + "args": { + "External id": 187,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 186 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979923, "dur": 25, + "args": { + "External id": 188,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 187 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979925, "dur": 22, + "args": { + "External id": 189,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 188 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979952, "dur": 25, + "args": { + "External id": 190,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 189 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979954, "dur": 22, + "args": { + "External id": 191,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 190 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789979981, "dur": 25, + "args": { + "External id": 192,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 191 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789979983, "dur": 22, + "args": { + "External id": 193,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 192 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980010, "dur": 29, + "args": { + "External id": 194,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 193 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980016, "dur": 23, + "args": { + "External id": 195,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 194 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980044, "dur": 25, + "args": { + "External id": 196,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 195 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980046, "dur": 22, + "args": { + "External id": 197,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 196 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980073, "dur": 25, + "args": { + "External id": 198,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 197 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980076, "dur": 21, + "args": { + "External id": 199,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 198 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980102, "dur": 25, + "args": { + "External id": 200,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 199 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980105, "dur": 22, + "args": { + "External id": 201,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 200 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980132, "dur": 25, + "args": { + "External id": 202,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 201 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980134, "dur": 22, + "args": { + "External id": 203,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 202 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980161, "dur": 25, + "args": { + "External id": 204,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 203 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980163, "dur": 22, + "args": { + "External id": 205,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 204 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980190, "dur": 25, + "args": { + "External id": 206,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 205 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980192, "dur": 22, + "args": { + "External id": 207,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 206 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980323, "dur": 37, + "args": { + "External id": 208,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 207 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980327, "dur": 32, + "args": { + "External id": 209,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 208 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980367, "dur": 26, + "args": { + "External id": 210,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 209 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980369, "dur": 23, + "args": { + "External id": 211,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 210 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980398, "dur": 25, + "args": { + "External id": 212,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 211 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980400, "dur": 22, + "args": { + "External id": 213,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 212 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980428, "dur": 25, + "args": { + "External id": 214,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 213 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980430, "dur": 22, + "args": { + "External id": 215,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 214 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980457, "dur": 25, + "args": { + "External id": 216,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 215 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980459, "dur": 23, + "args": { + "External id": 217,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 216 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980486, "dur": 26, + "args": { + "External id": 218,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 217 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980489, "dur": 22, + "args": { + "External id": 219,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 218 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980516, "dur": 25, + "args": { + "External id": 220,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 219 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980518, "dur": 22, + "args": { + "External id": 221,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 220 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980545, "dur": 30, + "args": { + "External id": 222,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 221 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980552, "dur": 22, + "args": { + "External id": 223,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 222 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980579, "dur": 26, + "args": { + "External id": 224,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 223 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980582, "dur": 22, + "args": { + "External id": 225,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 224 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980609, "dur": 25, + "args": { + "External id": 226,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 225 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980611, "dur": 22, + "args": { + "External id": 227,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 226 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980695, "dur": 33, + "args": { + "External id": 228,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 227 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980698, "dur": 29, + "args": { + "External id": 229,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 228 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980733, "dur": 26, + "args": { + "External id": 230,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 229 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980736, "dur": 22, + "args": { + "External id": 231,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 230 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980763, "dur": 26, + "args": { + "External id": 232,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 231 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980766, "dur": 22, + "args": { + "External id": 233,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 232 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980793, "dur": 31, + "args": { + "External id": 234,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 233 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980800, "dur": 23, + "args": { + "External id": 235,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 234 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980829, "dur": 25, + "args": { + "External id": 236,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 235 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980831, "dur": 22, + "args": { + "External id": 237,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 236 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980858, "dur": 25, + "args": { + "External id": 238,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 237 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980860, "dur": 22, + "args": { + "External id": 239,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 238 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980887, "dur": 25, + "args": { + "External id": 240,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 239 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980889, "dur": 22, + "args": { + "External id": 241,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 240 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980916, "dur": 25, + "args": { + "External id": 242,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 241 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980918, "dur": 22, + "args": { + "External id": 243,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 242 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980945, "dur": 25, + "args": { + "External id": 244,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 243 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980948, "dur": 22, + "args": { + "External id": 245,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 244 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789980975, "dur": 24, + "args": { + "External id": 246,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 245 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789980977, "dur": 22, + "args": { + "External id": 247,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 246 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981003, "dur": 29, + "args": { + "External id": 248,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 247 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981006, "dur": 26, + "args": { + "External id": 249,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 248 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981037, "dur": 26, + "args": { + "External id": 250,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 249 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981039, "dur": 23, + "args": { + "External id": 251,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 250 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981067, "dur": 25, + "args": { + "External id": 252,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 251 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981069, "dur": 22, + "args": { + "External id": 253,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 252 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981096, "dur": 25, + "args": { + "External id": 254,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 253 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981098, "dur": 22, + "args": { + "External id": 255,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 254 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981125, "dur": 25, + "args": { + "External id": 256,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 255 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981128, "dur": 22, + "args": { + "External id": 257,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 256 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981155, "dur": 25, + "args": { + "External id": 258,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 257 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981157, "dur": 22, + "args": { + "External id": 259,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 258 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981184, "dur": 25, + "args": { + "External id": 260,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 259 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981186, "dur": 22, + "args": { + "External id": 261,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 260 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981214, "dur": 25, + "args": { + "External id": 262,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 261 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981216, "dur": 22, + "args": { + "External id": 263,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 262 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981243, "dur": 35, + "args": { + "External id": 264,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 263 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981254, "dur": 23, + "args": { + "External id": 265,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 264 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981283, "dur": 25, + "args": { + "External id": 266,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 265 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981285, "dur": 22, + "args": { + "External id": 267,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 266 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981313, "dur": 25, + "args": { + "External id": 268,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 267 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981315, "dur": 22, + "args": { + "External id": 269,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 268 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981342, "dur": 25, + "args": { + "External id": 270,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 269 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981344, "dur": 22, + "args": { + "External id": 271,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 270 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981371, "dur": 25, + "args": { + "External id": 272,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 271 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981374, "dur": 21, + "args": { + "External id": 273,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 272 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981400, "dur": 25, + "args": { + "External id": 274,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 273 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981402, "dur": 22, + "args": { + "External id": 275,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 274 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981429, "dur": 25, + "args": { + "External id": 276,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 275 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981432, "dur": 21, + "args": { + "External id": 277,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 276 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981458, "dur": 29, + "args": { + "External id": 278,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 277 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981464, "dur": 22, + "args": { + "External id": 279,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 278 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981492, "dur": 25, + "args": { + "External id": 280,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 279 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981494, "dur": 22, + "args": { + "External id": 281,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 280 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981521, "dur": 25, + "args": { + "External id": 282,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 281 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981524, "dur": 21, + "args": { + "External id": 283,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 282 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981550, "dur": 25, + "args": { + "External id": 284,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 283 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981553, "dur": 21, + "args": { + "External id": 285,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 284 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981579, "dur": 25, + "args": { + "External id": 286,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 285 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981581, "dur": 22, + "args": { + "External id": 287,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 286 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981612, "dur": 26, + "args": { + "External id": 288,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 287 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981615, "dur": 22, + "args": { + "External id": 289,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 288 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981643, "dur": 26, + "args": { + "External id": 290,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 289 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981645, "dur": 23, + "args": { + "External id": 291,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 290 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981673, "dur": 34, + "args": { + "External id": 292,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 291 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981681, "dur": 25, + "args": { + "External id": 293,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 292 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981712, "dur": 25, + "args": { + "External id": 294,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 293 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981714, "dur": 23, + "args": { + "External id": 295,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 294 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981742, "dur": 25, + "args": { + "External id": 296,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 295 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981744, "dur": 22, + "args": { + "External id": 297,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 296 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981771, "dur": 25, + "args": { + "External id": 298,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 297 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981773, "dur": 23, + "args": { + "External id": 299,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 298 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981801, "dur": 24, + "args": { + "External id": 300,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 299 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981803, "dur": 22, + "args": { + "External id": 301,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 300 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981830, "dur": 25, + "args": { + "External id": 302,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 301 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981832, "dur": 22, + "args": { + "External id": 303,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 302 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981859, "dur": 25, + "args": { + "External id": 304,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 303 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981861, "dur": 22, + "args": { + "External id": 305,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 304 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981888, "dur": 25, + "args": { + "External id": 306,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 305 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981890, "dur": 22, + "args": { + "External id": 307,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 306 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981918, "dur": 29, + "args": { + "External id": 308,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 307 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981923, "dur": 23, + "args": { + "External id": 309,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 308 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981951, "dur": 25, + "args": { + "External id": 310,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 309 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981953, "dur": 23, + "args": { + "External id": 311,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 310 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789981981, "dur": 25, + "args": { + "External id": 312,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 311 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789981983, "dur": 22, + "args": { + "External id": 313,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 312 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982010, "dur": 25, + "args": { + "External id": 314,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 313 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982012, "dur": 22, + "args": { + "External id": 315,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 314 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982039, "dur": 25, + "args": { + "External id": 316,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 315 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982042, "dur": 22, + "args": { + "External id": 317,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 316 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982069, "dur": 24, + "args": { + "External id": 318,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 317 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982071, "dur": 22, + "args": { + "External id": 319,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 318 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982098, "dur": 24, + "args": { + "External id": 320,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 319 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982100, "dur": 22, + "args": { + "External id": 321,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 320 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982127, "dur": 29, + "args": { + "External id": 322,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 321 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982133, "dur": 22, + "args": { + "External id": 323,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 322 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982161, "dur": 25, + "args": { + "External id": 324,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 323 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982163, "dur": 22, + "args": { + "External id": 325,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 324 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982190, "dur": 25, + "args": { + "External id": 326,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 325 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982192, "dur": 22, + "args": { + "External id": 327,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 326 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982280, "dur": 33, + "args": { + "External id": 328,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 327 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982284, "dur": 28, + "args": { + "External id": 329,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 328 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982319, "dur": 25, + "args": { + "External id": 330,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 329 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982321, "dur": 23, + "args": { + "External id": 331,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 330 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982349, "dur": 25, + "args": { + "External id": 332,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 331 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982351, "dur": 22, + "args": { + "External id": 333,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 332 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982378, "dur": 25, + "args": { + "External id": 334,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 333 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982381, "dur": 21, + "args": { + "External id": 335,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 334 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982407, "dur": 31, + "args": { + "External id": 336,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 335 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982415, "dur": 23, + "args": { + "External id": 337,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 336 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982443, "dur": 25, + "args": { + "External id": 338,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 337 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982445, "dur": 23, + "args": { + "External id": 339,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 338 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982473, "dur": 25, + "args": { + "External id": 340,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 339 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982475, "dur": 22, + "args": { + "External id": 341,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 340 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982502, "dur": 25, + "args": { + "External id": 342,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 341 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982504, "dur": 22, + "args": { + "External id": 343,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 342 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982531, "dur": 25, + "args": { + "External id": 344,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 343 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982534, "dur": 22, + "args": { + "External id": 345,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 344 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982561, "dur": 25, + "args": { + "External id": 346,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 345 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982563, "dur": 22, + "args": { + "External id": 347,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 346 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982644, "dur": 32, + "args": { + "External id": 348,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 347 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982647, "dur": 28, + "args": { + "External id": 349,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 348 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982681, "dur": 31, + "args": { + "External id": 350,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 349 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982688, "dur": 24, + "args": { + "External id": 351,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 350 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982717, "dur": 25, + "args": { + "External id": 352,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 351 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982719, "dur": 23, + "args": { + "External id": 353,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 352 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982747, "dur": 25, + "args": { + "External id": 354,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 353 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982749, "dur": 22, + "args": { + "External id": 355,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 354 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982776, "dur": 25, + "args": { + "External id": 356,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 355 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982779, "dur": 22, + "args": { + "External id": 357,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 356 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982806, "dur": 25, + "args": { + "External id": 358,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 357 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982808, "dur": 22, + "args": { + "External id": 359,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 358 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982835, "dur": 25, + "args": { + "External id": 360,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 359 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982837, "dur": 22, + "args": { + "External id": 361,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 360 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982864, "dur": 25, + "args": { + "External id": 362,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 361 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982866, "dur": 22, + "args": { + "External id": 363,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 362 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982893, "dur": 29, + "args": { + "External id": 364,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 363 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982899, "dur": 22, + "args": { + "External id": 365,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 364 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982927, "dur": 25, + "args": { + "External id": 366,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 365 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982929, "dur": 22, + "args": { + "External id": 367,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 366 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982956, "dur": 24, + "args": { + "External id": 368,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 367 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982958, "dur": 22, + "args": { + "External id": 369,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 368 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789982985, "dur": 24, + "args": { + "External id": 370,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 369 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789982987, "dur": 21, + "args": { + "External id": 371,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 370 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983013, "dur": 25, + "args": { + "External id": 372,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 371 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983016, "dur": 21, + "args": { + "External id": 373,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 372 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983042, "dur": 25, + "args": { + "External id": 374,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 373 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983044, "dur": 22, + "args": { + "External id": 375,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 374 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983071, "dur": 25, + "args": { + "External id": 376,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 375 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983073, "dur": 22, + "args": { + "External id": 377,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 376 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983100, "dur": 32, + "args": { + "External id": 378,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 377 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983108, "dur": 23, + "args": { + "External id": 379,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 378 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983136, "dur": 26, + "args": { + "External id": 380,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 379 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983139, "dur": 22, + "args": { + "External id": 381,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 380 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983166, "dur": 24, + "args": { + "External id": 382,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 381 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983168, "dur": 22, + "args": { + "External id": 383,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 382 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983195, "dur": 24, + "args": { + "External id": 384,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 383 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983197, "dur": 22, + "args": { + "External id": 385,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 384 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983224, "dur": 24, + "args": { + "External id": 386,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 385 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983226, "dur": 22, + "args": { + "External id": 387,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 386 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983253, "dur": 24, + "args": { + "External id": 388,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 387 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983255, "dur": 21, + "args": { + "External id": 389,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 388 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983281, "dur": 25, + "args": { + "External id": 390,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 389 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983283, "dur": 22, + "args": { + "External id": 391,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 390 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983310, "dur": 28, + "args": { + "External id": 392,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 391 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983312, "dur": 26, + "args": { + "External id": 393,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 392 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983343, "dur": 25, + "args": { + "External id": 394,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 393 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983345, "dur": 23, + "args": { + "External id": 395,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 394 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983373, "dur": 24, + "args": { + "External id": 396,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 395 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983375, "dur": 22, + "args": { + "External id": 397,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 396 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983402, "dur": 24, + "args": { + "External id": 398,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 397 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983404, "dur": 22, + "args": { + "External id": 399,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 398 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983431, "dur": 24, + "args": { + "External id": 400,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 399 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983433, "dur": 22, + "args": { + "External id": 401,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 400 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983460, "dur": 25, + "args": { + "External id": 402,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 401 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983462, "dur": 22, + "args": { + "External id": 403,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 402 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983489, "dur": 24, + "args": { + "External id": 404,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 403 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983491, "dur": 21, + "args": { + "External id": 405,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 404 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983517, "dur": 25, + "args": { + "External id": 406,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 405 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983519, "dur": 22, + "args": { + "External id": 407,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 406 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983546, "dur": 29, + "args": { + "External id": 408,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 407 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983552, "dur": 22, + "args": { + "External id": 409,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 408 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983579, "dur": 25, + "args": { + "External id": 410,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 409 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983581, "dur": 22, + "args": { + "External id": 411,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 410 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983608, "dur": 25, + "args": { + "External id": 412,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 411 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983610, "dur": 22, + "args": { + "External id": 413,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 412 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983637, "dur": 25, + "args": { + "External id": 414,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 413 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983639, "dur": 22, + "args": { + "External id": 415,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 414 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983666, "dur": 25, + "args": { + "External id": 416,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 415 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983668, "dur": 22, + "args": { + "External id": 417,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 416 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983695, "dur": 25, + "args": { + "External id": 418,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 417 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983698, "dur": 21, + "args": { + "External id": 419,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 418 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983724, "dur": 32, + "args": { + "External id": 420,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 419 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983732, "dur": 23, + "args": { + "External id": 421,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 420 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983761, "dur": 28, + "args": { + "External id": 422,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 421 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983766, "dur": 22, + "args": { + "External id": 423,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 422 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983794, "dur": 25, + "args": { + "External id": 424,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 423 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983796, "dur": 22, + "args": { + "External id": 425,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 424 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983823, "dur": 24, + "args": { + "External id": 426,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 425 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983825, "dur": 22, + "args": { + "External id": 427,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 426 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983852, "dur": 24, + "args": { + "External id": 428,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 427 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983854, "dur": 21, + "args": { + "External id": 429,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 428 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983880, "dur": 25, + "args": { + "External id": 430,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 429 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983882, "dur": 22, + "args": { + "External id": 431,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 430 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983909, "dur": 25, + "args": { + "External id": 432,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 431 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983911, "dur": 22, + "args": { + "External id": 433,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 432 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983938, "dur": 24, + "args": { + "External id": 434,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 433 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983940, "dur": 22, + "args": { + "External id": 435,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 434 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789983967, "dur": 29, + "args": { + "External id": 436,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 435 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789983969, "dur": 26, + "args": { + "External id": 437,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 436 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984001, "dur": 25, + "args": { + "External id": 438,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 437 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984003, "dur": 23, + "args": { + "External id": 439,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 438 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984031, "dur": 25, + "args": { + "External id": 440,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 439 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984033, "dur": 22, + "args": { + "External id": 441,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 440 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984060, "dur": 25, + "args": { + "External id": 442,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 441 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984063, "dur": 21, + "args": { + "External id": 443,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 442 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984089, "dur": 25, + "args": { + "External id": 444,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 443 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984092, "dur": 22, + "args": { + "External id": 445,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 444 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984119, "dur": 24, + "args": { + "External id": 446,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 445 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984121, "dur": 22, + "args": { + "External id": 447,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 446 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984207, "dur": 54, + "args": { + "External id": 448,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 447 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984211, "dur": 49, + "args": { + "External id": 449,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 448 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984268, "dur": 32, + "args": { + "External id": 450,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 449 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984271, "dur": 28, + "args": { + "External id": 451,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 450 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984305, "dur": 26, + "args": { + "External id": 452,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 451 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984308, "dur": 22, + "args": { + "External id": 453,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 452 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984335, "dur": 26, + "args": { + "External id": 454,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 453 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984338, "dur": 22, + "args": { + "External id": 455,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 454 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984365, "dur": 25, + "args": { + "External id": 456,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 455 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984367, "dur": 22, + "args": { + "External id": 457,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 456 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984395, "dur": 25, + "args": { + "External id": 458,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 457 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984397, "dur": 22, + "args": { + "External id": 459,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 458 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984424, "dur": 25, + "args": { + "External id": 460,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 459 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984427, "dur": 21, + "args": { + "External id": 461,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 460 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984454, "dur": 25, + "args": { + "External id": 462,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 461 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984456, "dur": 22, + "args": { + "External id": 463,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 462 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984483, "dur": 25, + "args": { + "External id": 464,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 463 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984485, "dur": 22, + "args": { + "External id": 465,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 464 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984512, "dur": 30, + "args": { + "External id": 466,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 465 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984518, "dur": 23, + "args": { + "External id": 467,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 466 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984600, "dur": 33, + "args": { + "External id": 468,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 467 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984604, "dur": 28, + "args": { + "External id": 469,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 468 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984638, "dur": 26, + "args": { + "External id": 470,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 469 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984641, "dur": 22, + "args": { + "External id": 471,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 470 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984668, "dur": 25, + "args": { + "External id": 472,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 471 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984671, "dur": 22, + "args": { + "External id": 473,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 472 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984698, "dur": 24, + "args": { + "External id": 474,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 473 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984700, "dur": 22, + "args": { + "External id": 475,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 474 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984727, "dur": 25, + "args": { + "External id": 476,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 475 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984729, "dur": 22, + "args": { + "External id": 477,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 476 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984756, "dur": 30, + "args": { + "External id": 478,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 477 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984758, "dur": 27, + "args": { + "External id": 479,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 478 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984791, "dur": 25, + "args": { + "External id": 480,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 479 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984793, "dur": 22, + "args": { + "External id": 481,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 480 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984820, "dur": 25, + "args": { + "External id": 482,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 481 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984823, "dur": 22, + "args": { + "External id": 483,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 482 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984849, "dur": 25, + "args": { + "External id": 484,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 483 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984852, "dur": 22, + "args": { + "External id": 485,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 484 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984878, "dur": 25, + "args": { + "External id": 486,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 485 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984881, "dur": 22, + "args": { + "External id": 487,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 486 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984908, "dur": 24, + "args": { + "External id": 488,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 487 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984910, "dur": 22, + "args": { + "External id": 489,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 488 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984937, "dur": 24, + "args": { + "External id": 490,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 489 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984939, "dur": 22, + "args": { + "External id": 491,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 490 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984966, "dur": 24, + "args": { + "External id": 492,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 491 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789984968, "dur": 21, + "args": { + "External id": 493,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 492 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789984994, "dur": 30, + "args": { + "External id": 494,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 493 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985000, "dur": 23, + "args": { + "External id": 495,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 494 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985029, "dur": 24, + "args": { + "External id": 496,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 495 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985031, "dur": 22, + "args": { + "External id": 497,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 496 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985058, "dur": 25, + "args": { + "External id": 498,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 497 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985060, "dur": 22, + "args": { + "External id": 499,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 498 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985087, "dur": 25, + "args": { + "External id": 500,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 499 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985089, "dur": 22, + "args": { + "External id": 501,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 500 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985117, "dur": 24, + "args": { + "External id": 502,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 501 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985119, "dur": 22, + "args": { + "External id": 503,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 502 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985146, "dur": 24, + "args": { + "External id": 504,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 503 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985148, "dur": 22, + "args": { + "External id": 505,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 504 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985175, "dur": 25, + "args": { + "External id": 506,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 505 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985177, "dur": 22, + "args": { + "External id": 507,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 506 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985204, "dur": 29, + "args": { + "External id": 508,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 507 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985210, "dur": 22, + "args": { + "External id": 509,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 508 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985238, "dur": 25, + "args": { + "External id": 510,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 509 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985240, "dur": 22, + "args": { + "External id": 511,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 510 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985267, "dur": 100, + "args": { + "External id": 512,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 511 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985342, "dur": 24, + "args": { + "External id": 513,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 512 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985372, "dur": 26, + "args": { + "External id": 514,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 513 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985375, "dur": 22, + "args": { + "External id": 515,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 514 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985402, "dur": 25, + "args": { + "External id": 516,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 515 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985404, "dur": 22, + "args": { + "External id": 517,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 516 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985431, "dur": 25, + "args": { + "External id": 518,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 517 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985433, "dur": 22, + "args": { + "External id": 519,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 518 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985460, "dur": 24, + "args": { + "External id": 520,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 519 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985462, "dur": 22, + "args": { + "External id": 521,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 520 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985489, "dur": 29, + "args": { + "External id": 522,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 521 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985494, "dur": 23, + "args": { + "External id": 523,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 522 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985522, "dur": 25, + "args": { + "External id": 524,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 523 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985525, "dur": 22, + "args": { + "External id": 525,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 524 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985552, "dur": 24, + "args": { + "External id": 526,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 525 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985554, "dur": 22, + "args": { + "External id": 527,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 526 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985581, "dur": 24, + "args": { + "External id": 528,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 527 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985583, "dur": 22, + "args": { + "External id": 529,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 528 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985609, "dur": 25, + "args": { + "External id": 530,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 529 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985612, "dur": 21, + "args": { + "External id": 531,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 530 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985638, "dur": 25, + "args": { + "External id": 532,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 531 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985641, "dur": 21, + "args": { + "External id": 533,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 532 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985667, "dur": 25, + "args": { + "External id": 534,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 533 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985669, "dur": 22, + "args": { + "External id": 535,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 534 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985696, "dur": 29, + "args": { + "External id": 536,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 535 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985698, "dur": 26, + "args": { + "External id": 537,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 536 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985730, "dur": 25, + "args": { + "External id": 538,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 537 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985732, "dur": 23, + "args": { + "External id": 539,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 538 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985760, "dur": 24, + "args": { + "External id": 540,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 539 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985762, "dur": 22, + "args": { + "External id": 541,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 540 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985789, "dur": 24, + "args": { + "External id": 542,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 541 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985791, "dur": 22, + "args": { + "External id": 543,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 542 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985817, "dur": 25, + "args": { + "External id": 544,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 543 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985820, "dur": 21, + "args": { + "External id": 545,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 544 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985846, "dur": 25, + "args": { + "External id": 546,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 545 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985848, "dur": 22, + "args": { + "External id": 547,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 546 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985875, "dur": 31, + "args": { + "External id": 548,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 547 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985883, "dur": 23, + "args": { + "External id": 549,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 548 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985911, "dur": 25, + "args": { + "External id": 550,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 549 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985913, "dur": 22, + "args": { + "External id": 551,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 550 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985940, "dur": 28, + "args": { + "External id": 552,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 551 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985946, "dur": 22, + "args": { + "External id": 553,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 552 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789985997, "dur": 26, + "args": { + "External id": 554,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 553 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789985999, "dur": 23, + "args": { + "External id": 555,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 554 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986027, "dur": 25, + "args": { + "External id": 556,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 555 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986030, "dur": 22, + "args": { + "External id": 557,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 556 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986057, "dur": 25, + "args": { + "External id": 558,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 557 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986059, "dur": 22, + "args": { + "External id": 559,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 558 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986086, "dur": 25, + "args": { + "External id": 560,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 559 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986088, "dur": 22, + "args": { + "External id": 561,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 560 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986115, "dur": 25, + "args": { + "External id": 562,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 561 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986117, "dur": 22, + "args": { + "External id": 563,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 562 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986144, "dur": 25, + "args": { + "External id": 564,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 563 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986147, "dur": 21, + "args": { + "External id": 565,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 564 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986173, "dur": 29, + "args": { + "External id": 566,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 565 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986179, "dur": 22, + "args": { + "External id": 567,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 566 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986267, "dur": 34, + "args": { + "External id": 568,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 567 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986271, "dur": 29, + "args": { + "External id": 569,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 568 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986306, "dur": 27, + "args": { + "External id": 570,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 569 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986309, "dur": 23, + "args": { + "External id": 571,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 570 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986337, "dur": 26, + "args": { + "External id": 572,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 571 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986340, "dur": 22, + "args": { + "External id": 573,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 572 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986367, "dur": 25, + "args": { + "External id": 574,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 573 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986369, "dur": 22, + "args": { + "External id": 575,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 574 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986396, "dur": 25, + "args": { + "External id": 576,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 575 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986398, "dur": 22, + "args": { + "External id": 577,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 576 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986425, "dur": 25, + "args": { + "External id": 578,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 577 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986427, "dur": 22, + "args": { + "External id": 579,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 578 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986454, "dur": 25, + "args": { + "External id": 580,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 579 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986456, "dur": 22, + "args": { + "External id": 581,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 580 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986484, "dur": 30, + "args": { + "External id": 582,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 581 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986491, "dur": 23, + "args": { + "External id": 583,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 582 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986519, "dur": 25, + "args": { + "External id": 584,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 583 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986521, "dur": 23, + "args": { + "External id": 585,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 584 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986549, "dur": 24, + "args": { + "External id": 586,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 585 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986551, "dur": 22, + "args": { + "External id": 587,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 586 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986633, "dur": 31, + "args": { + "External id": 588,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 587 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986636, "dur": 27, + "args": { + "External id": 589,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 588 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986670, "dur": 26, + "args": { + "External id": 590,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 589 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986672, "dur": 23, + "args": { + "External id": 591,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 590 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986701, "dur": 25, + "args": { + "External id": 592,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 591 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986703, "dur": 22, + "args": { + "External id": 593,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 592 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986730, "dur": 25, + "args": { + "External id": 594,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 593 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986732, "dur": 23, + "args": { + "External id": 595,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 594 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986760, "dur": 31, + "args": { + "External id": 596,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 595 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986767, "dur": 23, + "args": { + "External id": 597,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 596 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986796, "dur": 25, + "args": { + "External id": 598,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 597 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986798, "dur": 22, + "args": { + "External id": 599,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 598 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986825, "dur": 25, + "args": { + "External id": 600,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 599 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986828, "dur": 21, + "args": { + "External id": 601,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 600 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986854, "dur": 25, + "args": { + "External id": 602,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 601 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986856, "dur": 22, + "args": { + "External id": 603,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 602 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986883, "dur": 25, + "args": { + "External id": 604,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 603 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986885, "dur": 22, + "args": { + "External id": 605,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 604 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986912, "dur": 25, + "args": { + "External id": 606,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 605 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986914, "dur": 22, + "args": { + "External id": 607,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 606 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986941, "dur": 25, + "args": { + "External id": 608,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 607 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986943, "dur": 22, + "args": { + "External id": 609,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 608 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789986970, "dur": 34, + "args": { + "External id": 610,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 609 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789986980, "dur": 23, + "args": { + "External id": 611,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 610 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987009, "dur": 25, + "args": { + "External id": 612,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 611 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987011, "dur": 22, + "args": { + "External id": 613,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 612 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987038, "dur": 26, + "args": { + "External id": 614,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 613 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987041, "dur": 22, + "args": { + "External id": 615,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 614 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987068, "dur": 25, + "args": { + "External id": 616,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 615 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987070, "dur": 22, + "args": { + "External id": 617,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 616 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987097, "dur": 25, + "args": { + "External id": 618,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 617 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987099, "dur": 22, + "args": { + "External id": 619,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 618 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987126, "dur": 25, + "args": { + "External id": 620,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 619 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987128, "dur": 22, + "args": { + "External id": 621,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 620 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987155, "dur": 25, + "args": { + "External id": 622,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 621 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987157, "dur": 22, + "args": { + "External id": 623,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 622 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987184, "dur": 29, + "args": { + "External id": 624,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 623 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987186, "dur": 27, + "args": { + "External id": 625,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 624 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987218, "dur": 25, + "args": { + "External id": 626,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 625 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987220, "dur": 22, + "args": { + "External id": 627,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 626 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987247, "dur": 33, + "args": { + "External id": 628,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 627 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987256, "dur": 23, + "args": { + "External id": 629,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 628 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987285, "dur": 24, + "args": { + "External id": 630,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 629 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987287, "dur": 22, + "args": { + "External id": 631,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 630 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987314, "dur": 24, + "args": { + "External id": 632,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 631 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987316, "dur": 21, + "args": { + "External id": 633,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 632 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987342, "dur": 25, + "args": { + "External id": 634,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 633 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987345, "dur": 21, + "args": { + "External id": 635,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 634 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987371, "dur": 25, + "args": { + "External id": 636,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 635 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987374, "dur": 22, + "args": { + "External id": 637,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 636 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987401, "dur": 28, + "args": { + "External id": 638,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 637 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987403, "dur": 26, + "args": { + "External id": 639,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 638 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987434, "dur": 26, + "args": { + "External id": 640,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 639 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987436, "dur": 23, + "args": { + "External id": 641,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 640 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987464, "dur": 25, + "args": { + "External id": 642,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 641 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987466, "dur": 22, + "args": { + "External id": 643,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 642 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987493, "dur": 25, + "args": { + "External id": 644,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 643 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987495, "dur": 22, + "args": { + "External id": 645,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 644 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987522, "dur": 25, + "args": { + "External id": 646,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 645 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987525, "dur": 22, + "args": { + "External id": 647,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 646 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987552, "dur": 24, + "args": { + "External id": 648,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 647 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987554, "dur": 22, + "args": { + "External id": 649,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 648 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987581, "dur": 24, + "args": { + "External id": 650,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 649 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987583, "dur": 22, + "args": { + "External id": 651,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 650 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987609, "dur": 25, + "args": { + "External id": 652,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 651 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987612, "dur": 21, + "args": { + "External id": 653,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 652 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987639, "dur": 29, + "args": { + "External id": 654,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 653 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987644, "dur": 23, + "args": { + "External id": 655,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 654 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987672, "dur": 25, + "args": { + "External id": 656,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 655 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987674, "dur": 23, + "args": { + "External id": 657,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 656 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987702, "dur": 24, + "args": { + "External id": 658,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 657 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987704, "dur": 22, + "args": { + "External id": 659,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 658 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987731, "dur": 24, + "args": { + "External id": 660,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 659 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987733, "dur": 22, + "args": { + "External id": 661,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 660 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987760, "dur": 25, + "args": { + "External id": 662,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 661 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987762, "dur": 22, + "args": { + "External id": 663,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 662 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987789, "dur": 25, + "args": { + "External id": 664,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 663 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987791, "dur": 22, + "args": { + "External id": 665,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 664 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987818, "dur": 24, + "args": { + "External id": 666,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 665 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987820, "dur": 22, + "args": { + "External id": 667,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 666 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987847, "dur": 29, + "args": { + "External id": 668,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 667 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987853, "dur": 22, + "args": { + "External id": 669,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 668 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987880, "dur": 25, + "args": { + "External id": 670,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 669 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987882, "dur": 23, + "args": { + "External id": 671,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 670 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987909, "dur": 25, + "args": { + "External id": 672,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 671 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987912, "dur": 21, + "args": { + "External id": 673,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 672 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987938, "dur": 25, + "args": { + "External id": 674,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 673 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987941, "dur": 21, + "args": { + "External id": 675,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 674 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789987967, "dur": 32, + "args": { + "External id": 676,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 675 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789987975, "dur": 23, + "args": { + "External id": 677,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 676 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789988003, "dur": 26, + "args": { + "External id": 678,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 677 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789988006, "dur": 22, + "args": { + "External id": 679,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 678 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789988033, "dur": 25, + "args": { + "External id": 680,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 679 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789988036, "dur": 21, + "args": { + "External id": 681,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 680 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789988062, "dur": 29, + "args": { + "External id": 682,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 681 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789988065, "dur": 25, + "args": { + "External id": 683,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 682 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789988096, "dur": 25, + "args": { + "External id": 684,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 683 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789988098, "dur": 22, + "args": { + "External id": 685,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 684 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649789988125, "dur": 25, + "args": { + "External id": 686,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 685 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649789988128, "dur": 22, + "args": { + "External id": 687,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 4096]], "Ev Idx": 686 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 435, "tid": 435, + "ts": 1776649789990003, "dur": 11, + "args": { + "External id": 688,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Dims": [[1024]], "Ev Idx": 687 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__and__", "pid": 435, "tid": 435, + "ts": 1776649789990975, "dur": 103, + "args": { + "External id": 689,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "15"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 688 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_and", "pid": 435, "tid": 435, + "ts": 1776649789990979, "dur": 98, + "args": { + "External id": 690,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "15"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 689 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_and", "pid": 435, "tid": 435, + "ts": 1776649789990994, "dur": 79, + "args": { + "External id": 691,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["unsigned char", "long int"], "Input Dims": [[4096, 2048], []], "Ev Idx": 690 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649789991089, "dur": 64, + "args": { + "External id": 692,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], []], "Ev Idx": 691 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649789991094, "dur": 59, + "args": { + "External id": 693,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["unsigned char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], [], [], []], "Ev Idx": 692 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649789991108, "dur": 11, + "args": { + "External id": 694,"Record function id": 0, "Concrete Inputs": ["[4096, 2048]", "[2048, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 693 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649789991122, "dur": 30, + "args": { + "External id": 695,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "unsigned char", "Scalar"], "Input Dims": [[4096, 2048], [4096, 2048], []], "Ev Idx": 694 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__rshift__", "pid": 435, "tid": 435, + "ts": 1776649789991163, "dur": 39, + "args": { + "External id": 696,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 695 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649789991175, "dur": 7, + "args": { + "External id": 697,"Record function id": 0, "Concrete Inputs": ["[4096, 2048]", "0", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 696 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649789991206, "dur": 23, + "args": { + "External id": 698,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], []], "Ev Idx": 697 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649789991207, "dur": 22, + "args": { + "External id": 699,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["unsigned char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], [], [], []], "Ev Idx": 698 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649789991209, "dur": 6, + "args": { + "External id": 700,"Record function id": 0, "Concrete Inputs": ["[4096, 2048]", "[2048, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 699 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649789991216, "dur": 12, + "args": { + "External id": 701,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "unsigned char", "Scalar"], "Input Dims": [[4096, 2048], [4096, 2048], []], "Ev Idx": 700 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::stack", "pid": 435, "tid": 435, + "ts": 1776649789991245, "dur": 64, + "args": { + "External id": 702,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["TensorList", "Scalar"], "Input Dims": [[[4096, 2048], [4096, 2048]], []], "Ev Idx": 701 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649789991254, "dur": 8, + "args": { + "External id": 703,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["signed char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 702 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649789991258, "dur": 4, + "args": { + "External id": 704,"Record function id": 0, "Concrete Inputs": ["", "[4096, 2048, 1]", "[2048, 1, 1]", ""], "Input type": ["signed char", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 2048], [], [], []], "Ev Idx": 703 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649789991263, "dur": 2, + "args": { + "External id": 705,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["signed char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 704 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649789991264, "dur": 1, + "args": { + "External id": 706,"Record function id": 0, "Concrete Inputs": ["", "[4096, 2048, 1]", "[2048, 1, 1]", ""], "Input type": ["signed char", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 2048], [], [], []], "Ev Idx": 705 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 435, "tid": 435, + "ts": 1776649789991267, "dur": 40, + "args": { + "External id": 707,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["TensorList", "Scalar"], "Input Dims": [[[4096, 2048, 1], [4096, 2048, 1]], []], "Ev Idx": 706 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649789991322, "dur": 13, + "args": { + "External id": 708,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["signed char", "ScalarList"], "Input Dims": [[4096, 2048, 2], []], "Ev Idx": 707 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649789991324, "dur": 10, + "args": { + "External id": 709,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["signed char", "ScalarList"], "Input Dims": [[4096, 2048, 2], []], "Ev Idx": 708 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::repeat_interleave", "pid": 435, "tid": 435, + "ts": 1776649789991344, "dur": 74, + "args": { + "External id": 710,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "128", "0", ""], "Input type": ["c10::Half", "Scalar", "Scalar", ""], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 709 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649789991347, "dur": 4, + "args": { + "External id": 711,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 4096], []], "Ev Idx": 710 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649789991350, "dur": 0, + "args": { + "External id": 712,"Record function id": 0, "Concrete Inputs": ["", "[32, 1, 4096]", "[4096, 4096, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 711 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 435, "tid": 435, + "ts": 1776649789991355, "dur": 12, + "args": { + "External id": 713,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]", "False"], "Input type": ["c10::Half", "ScalarList", "Scalar"], "Input Dims": [[32, 1, 4096], [], []], "Ev Idx": 712 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649789991362, "dur": 3, + "args": { + "External id": 714,"Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]", "[4096, 0, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 1, 4096], [], [], []], "Ev Idx": 713 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649789991370, "dur": 40, + "args": { + "External id": 715,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 714 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649789991376, "dur": 11, + "args": { + "External id": 716,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[32, 128, 4096], [], [], [], [], []], "Ev Idx": 715 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649789991380, "dur": 6, + "args": { + "External id": 717,"Record function id": 0, "Concrete Inputs": ["[32, 128, 4096]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 716 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649789991388, "dur": 21, + "args": { + "External id": 718,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[32, 128, 4096], [32, 128, 4096], []], "Ev Idx": 717 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::flatten", "pid": 435, "tid": 435, + "ts": 1776649789991413, "dur": 5, + "args": { + "External id": 719,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[32, 128, 4096], [], []], "Ev Idx": 718 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649789991415, "dur": 2, + "args": { + "External id": 720,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 719 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 435, "tid": 435, + "ts": 1776649789991429, "dur": 32, + "args": { + "External id": 721,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["signed char", "long int", "Scalar"], "Input Dims": [[4096, 4096], [], []], "Ev Idx": 720 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649789991467, "dur": 27, + "args": { + "External id": 722,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["signed char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 721 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649789991468, "dur": 25, + "args": { + "External id": 723,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "", "", "", "False", ""], "Input type": ["signed char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], [], []], "Ev Idx": 722 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649789991471, "dur": 6, + "args": { + "External id": 724,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "[4096, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 723 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649789991478, "dur": 15, + "args": { + "External id": 725,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "signed char", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 724 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 435, "tid": 435, + "ts": 1776649789991503, "dur": 432, + "args": { + "External id": 726,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4096, 4096], [4096, 4096]], "Ev Idx": 725 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 435, "tid": 435, + "ts": 1776649789991984, "dur": 19, + "args": { + "External id": 727,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[4096, 4096]], "Ev Idx": 726 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 435, "tid": 435, + "ts": 1776649789991993, "dur": 8, + "args": { + "External id": 728,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[4096, 4096], [], []], "Ev Idx": 727 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649789991999, "dur": 1, + "args": { + "External id": 729,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]", "[1, 4096]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 728 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 435, "tid": 435, + "ts": 1776649789992029, "dur": 37, + "args": { + "External id": 730,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 729 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649789992030, "dur": 35, + "args": { + "External id": 731,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 730 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649789992033, "dur": 10, + "args": { + "External id": 732,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[4096, 4096], [], [], [], [], []], "Ev Idx": 731 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649789992035, "dur": 7, + "args": { + "External id": 733,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 732 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649789992044, "dur": 21, + "args": { + "External id": 734,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 733 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649789992426, "dur": 23, + "args": { + "External id": 735,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "", "", "", "False", ""], "Input type": ["ScalarList", "", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 734 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 435, "tid": 435, + "ts": 1776649789992506, "dur": 11, + "args": { + "External id": 736,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Dims": [[4096, 4096]], "Ev Idx": 735 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 435, "tid": 435, + "ts": 1776649789992511, "dur": 5, + "args": { + "External id": 737,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Dims": [[4096, 4096]], "Ev Idx": 736 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::uniform_", "pid": 435, "tid": 435, + "ts": 1776649789992881, "dur": 107929, + "args": { + "External id": 738,"Record function id": 0, "Concrete Inputs": ["", "-0.015624999999999998", "0.015624999999999998", ""], "Input type": ["float", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 737 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790101199, "dur": 279239, + "args": { + "External id": 739,"Record function id": 0, "Concrete Inputs": ["", "", "5", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], []], "Ev Idx": 738 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790101206, "dur": 279229, + "args": { + "External id": 740,"Record function id": 0, "Concrete Inputs": ["", "5", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], [], []], "Ev Idx": 739 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790101231, "dur": 61, + "args": { + "External id": 741,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "[4096, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 740 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790101296, "dur": 279131, + "args": { + "External id": 742,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "float", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 741 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790101318, "dur": 183549, + "args": { + "External id": 743,"Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 742 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790101320, "dur": 183544, + "args": { + "External id": 744,"Record function id": 0, "Concrete Inputs": ["", "5", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], [], []], "Ev Idx": 743 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790101322, "dur": 47, + "args": { + "External id": 745,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "[4096, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 744 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790101370, "dur": 183475, + "args": { + "External id": 746,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "float", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 745 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::expand_as", "pid": 435, "tid": 435, + "ts": 1776649790284953, "dur": 37, + "args": { + "External id": 747,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4096, 4096], [4096, 4096]], "Ev Idx": 746 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 435, "tid": 435, + "ts": 1776649790284970, "dur": 20, + "args": { + "External id": 748,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]", "False"], "Input type": ["c10::Half", "ScalarList", "Scalar"], "Input Dims": [[4096, 4096], [], []], "Ev Idx": 747 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790284981, "dur": 8, + "args": { + "External id": 749,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]", "[4096, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 748 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790284997, "dur": 91784, + "args": { + "External id": 750,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 749 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_has_compatible_shallow_copy_type", "pid": 435, "tid": 435, + "ts": 1776649790380621, "dur": 3, + "args": { + "External id": 751,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "c10::Half"], "Input Dims": [[4096, 4096], [4096, 4096]], "Ev Idx": 750 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_has_compatible_shallow_copy_type", "pid": 435, "tid": 435, + "ts": 1776649790380685, "dur": 0, + "args": { + "External id": 752,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "c10::Half"], "Input Dims": [[4096, 4096], [4096, 4096]], "Ev Idx": 751 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790388423, "dur": 87, + "args": { + "External id": 753,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 752 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790388632, "dur": 37, + "args": { + "External id": 754,"Record function id": 0, "Concrete Inputs": ["[256, 8192]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 753 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790388690, "dur": 9, + "args": { + "External id": 755,"Record function id": 0, "Concrete Inputs": ["[32, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 754 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790388777, "dur": 25, + "args": { + "External id": 756,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 755 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790388781, "dur": 4, + "args": { + "External id": 757,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 756 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790388793, "dur": 9, + "args": { + "External id": 758,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 757 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790388873, "dur": 2494, + "args": { + "External id": 759,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 8192], [], [], [], [], [], [], []], "Ev Idx": 758 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790388877, "dur": 2489, + "args": { + "External id": 760,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[256, 8192], [], [], [], [], [], []], "Ev Idx": 759 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790388886, "dur": 26, + "args": { + "External id": 761,"Record function id": 0, "Concrete Inputs": ["[256, 8192]", "[8192, 1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 760 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790388914, "dur": 2451, + "args": { + "External id": 762,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 8192], [256, 8192], []], "Ev Idx": 761 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790391486, "dur": 137, + "args": { + "External id": 763,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[32, 4096], [], [], [], [], [], [], []], "Ev Idx": 762 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790391488, "dur": 135, + "args": { + "External id": 764,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[32, 4096], [], [], [], [], [], []], "Ev Idx": 763 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790391492, "dur": 10, + "args": { + "External id": 765,"Record function id": 0, "Concrete Inputs": ["[32, 4096]", "[4096, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 764 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790391503, "dur": 119, + "args": { + "External id": 766,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[32, 4096], [32, 4096], []], "Ev Idx": 765 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790391631, "dur": 32, + "args": { + "External id": 767,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[512], [], [], [], [], [], [], []], "Ev Idx": 766 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790391632, "dur": 31, + "args": { + "External id": 768,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[512], [], [], [], [], [], []], "Ev Idx": 767 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790391634, "dur": 7, + "args": { + "External id": 769,"Record function id": 0, "Concrete Inputs": ["[512]", "[1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 768 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790391642, "dur": 21, + "args": { + "External id": 770,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[512], [512], []], "Ev Idx": 769 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 435, "tid": 435, + "ts": 1776649790391680, "dur": 24, + "args": { + "External id": 771,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[32, 4096]], "Ev Idx": 770 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 435, "tid": 435, + "ts": 1776649790391686, "dur": 13, + "args": { + "External id": 772,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[32, 4096], [], []], "Ev Idx": 771 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790391694, "dur": 5, + "args": { + "External id": 773,"Record function id": 0, "Concrete Inputs": ["", "[4096, 32]", "[1, 4096]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 772 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 435, "tid": 435, + "ts": 1776649790391708, "dur": 4, + "args": { + "External id": 774,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[4096, 4096]], "Ev Idx": 773 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 435, "tid": 435, + "ts": 1776649790391709, "dur": 2, + "args": { + "External id": 775,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[4096, 4096], [], []], "Ev Idx": 774 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790391711, "dur": 0, + "args": { + "External id": 776,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]", "[1, 4096]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 775 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790391721, "dur": 47, + "args": { + "External id": 777,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 128, 4096]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 4096], []], "Ev Idx": 776 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 435, "tid": 435, + "ts": 1776649790391752, "dur": 16, + "args": { + "External id": 778,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]", "[128, 1, 4096]"], "Input type": ["c10::Half", "ScalarList", "ScalarList"], "Input Dims": [[4096, 4096], [], []], "Ev Idx": 777 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 435, "tid": 435, + "ts": 1776649790391817, "dur": 16, + "args": { + "External id": 779,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1, 0, 2]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 778 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790391831, "dur": 1, + "args": { + "External id": 780,"Record function id": 0, "Concrete Inputs": ["", "[128, 32, 4096]", "[1, 128, 4096]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 128, 4096], [], [], []], "Ev Idx": 779 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790391837, "dur": 91, + "args": { + "External id": 781,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[128, -1]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[128, 32, 4096], []], "Ev Idx": 780 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649790391840, "dur": 65, + "args": { + "External id": 782,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[128, 32, 4096], []], "Ev Idx": 781 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649790391846, "dur": 16, + "args": { + "External id": 783,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[128, 32, 4096], [], [], [], [], []], "Ev Idx": 782 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790391850, "dur": 11, + "args": { + "External id": 784,"Record function id": 0, "Concrete Inputs": ["[128, 32, 4096]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 783 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790391864, "dur": 40, + "args": { + "External id": 785,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[128, 32, 4096], [128, 32, 4096], []], "Ev Idx": 784 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 435, "tid": 435, + "ts": 1776649790391915, "dur": 12, + "args": { + "External id": 786,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[128, 131072]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[128, 32, 4096], []], "Ev Idx": 785 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790391936, "dur": 33, + "args": { + "External id": 787,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1, -1]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 32], []], "Ev Idx": 786 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649790391938, "dur": 28, + "args": { + "External id": 788,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[4096, 32], []], "Ev Idx": 787 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649790391940, "dur": 9, + "args": { + "External id": 789,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[4096, 32], [], [], [], [], []], "Ev Idx": 788 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790391941, "dur": 7, + "args": { + "External id": 790,"Record function id": 0, "Concrete Inputs": ["[4096, 32]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 789 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790391950, "dur": 16, + "args": { + "External id": 791,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[4096, 32], [4096, 32], []], "Ev Idx": 790 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 435, "tid": 435, + "ts": 1776649790391968, "dur": 1, + "args": { + "External id": 792,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1, 131072]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 32], []], "Ev Idx": 791 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 435, "tid": 435, + "ts": 1776649790391978, "dur": 35, + "args": { + "External id": 793,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[128, 131072], [1, 131072]], "Ev Idx": 792 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 435, "tid": 435, + "ts": 1776649790392017, "dur": 489, + "args": { + "External id": 794,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[128, 131072]], "Ev Idx": 793 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790392542, "dur": 388, + "args": { + "External id": 795,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[128, 131072], [], [], [], []], "Ev Idx": 794 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790392544, "dur": 386, + "args": { + "External id": 796,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "", "", "", "False", ""], "Input type": ["c10::Half", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[128, 131072], [], [], [], [], [], []], "Ev Idx": 795 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790392548, "dur": 230, + "args": { + "External id": 797,"Record function id": 0, "Concrete Inputs": ["[128, 131072]", "[131072, 1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 796 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790392782, "dur": 146, + "args": { + "External id": 798,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "c10::Half", "Scalar"], "Input Dims": [[128, 131072], [128, 131072], []], "Ev Idx": 797 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 435, "tid": 435, + "ts": 1776649790392967, "dur": 99, + "args": { + "External id": 799,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["int", "long int", "Scalar"], "Input Dims": [[128, 131072], [], []], "Ev Idx": 798 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 435, "tid": 435, + "ts": 1776649790393114, "dur": 327, + "args": { + "External id": 800,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "15"], "Input type": ["int", "Scalar", "Scalar"], "Input Dims": [[128, 131072], [], []], "Ev Idx": 799 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790393120, "dur": 1, + "args": { + "External id": 801,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[128, 131072], [], [], [], []], "Ev Idx": 800 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790393458, "dur": 8, + "args": { + "External id": 802,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[128, -1, 4096]"], "Input type": ["int", "ScalarList"], "Input Dims": [[128, 131072], []], "Ev Idx": 801 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790393460, "dur": 6, + "args": { + "External id": 803,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[128, -1, 4096]"], "Input type": ["int", "ScalarList"], "Input Dims": [[128, 131072], []], "Ev Idx": 802 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 435, "tid": 435, + "ts": 1776649790393470, "dur": 20, + "args": { + "External id": 804,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1, 0, 2]"], "Input type": ["int", "ScalarList"], "Input Dims": [[128, 32, 4096], []], "Ev Idx": 803 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790393487, "dur": 1, + "args": { + "External id": 805,"Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]", "[4096, 131072, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Dims": [[128, 32, 4096], [], [], []], "Ev Idx": 804 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790393494, "dur": 93, + "args": { + "External id": 806,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["int", "ScalarList"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 805 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649790393498, "dur": 83, + "args": { + "External id": 807,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 806 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649790393500, "dur": 11, + "args": { + "External id": 808,"Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "0"], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[32, 128, 4096], [], [], [], [], []], "Ev Idx": 807 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790393502, "dur": 8, + "args": { + "External id": 809,"Record function id": 0, "Concrete Inputs": ["[32, 128, 4096]", "3", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 808 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790393513, "dur": 67, + "args": { + "External id": 810,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[32, 128, 4096], [32, 128, 4096], []], "Ev Idx": 809 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 435, "tid": 435, + "ts": 1776649790393584, "dur": 2, + "args": { + "External id": 811,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["int", "ScalarList"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 810 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790393598, "dur": 3, + "args": { + "External id": 812,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 64]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[1, 131072], []], "Ev Idx": 811 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790393599, "dur": 2, + "args": { + "External id": 813,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 64]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[1, 131072], []], "Ev Idx": 812 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649790393614, "dur": 11, + "args": { + "External id": 814,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[2048, 64], [], [], [], []], "Ev Idx": 813 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790393621, "dur": 1, + "args": { + "External id": 815,"Record function id": 0, "Concrete Inputs": ["", "[2048, 64]", "[64, 1]", "0"], "Input type": ["c10::Half", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[2048, 64], [], [], []], "Ev Idx": 814 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790393693, "dur": 15, + "args": { + "External id": 816,"Record function id": 0, "Concrete Inputs": ["[64]", "4", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 815 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790393717, "dur": 43, + "args": { + "External id": 817,"Record function id": 0, "Concrete Inputs": ["", "", "4", "False", "False", ""], "Input type": ["long int", "", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[64], [], [], [], [], []], "Ev Idx": 816 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790393719, "dur": 41, + "args": { + "External id": 818,"Record function id": 0, "Concrete Inputs": ["", "4", "", "", "", "False", ""], "Input type": ["long int", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[64], [], [], [], [], [], []], "Ev Idx": 817 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790393721, "dur": 9, + "args": { + "External id": 819,"Record function id": 0, "Concrete Inputs": ["[64]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 818 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790393732, "dur": 27, + "args": { + "External id": 820,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Dims": [[64], [64], []], "Ev Idx": 819 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 435, "tid": 435, + "ts": 1776649790393764, "dur": 1, + "args": { + "External id": 821,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Dims": [[64]], "Ev Idx": 820 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::index", "pid": 435, "tid": 435, + "ts": 1776649790393783, "dur": 16409, + "args": { + "External id": 822,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", ""], "Input Dims": [[2048, 64], []], "Ev Idx": 821 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790393811, "dur": 1, + "args": { + "External id": 823,"Record function id": 0, "Concrete Inputs": ["", "[2048, 64]", "[64, 0]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[2048, 64], [], [], []], "Ev Idx": 822 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790393814, "dur": 2, + "args": { + "External id": 824,"Record function id": 0, "Concrete Inputs": ["", "[1, 64]"], "Input type": ["long int", "ScalarList"], "Input Dims": [[64], []], "Ev Idx": 823 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790393815, "dur": 1, + "args": { + "External id": 825,"Record function id": 0, "Concrete Inputs": ["", "[1, 64]"], "Input type": ["long int", "ScalarList"], "Input Dims": [[64], []], "Ev Idx": 824 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790410306, "dur": 26, + "args": { + "External id": 826,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 4096]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[2048, 64], []], "Ev Idx": 825 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790410311, "dur": 20, + "args": { + "External id": 827,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 4096]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[2048, 64], []], "Ev Idx": 826 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790410341, "dur": 4, + "args": { + "External id": 828,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 16, 256, 16]"], "Input type": ["int", "ScalarList"], "Input Dims": [[4096, 4096], []], "Ev Idx": 827 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790410342, "dur": 3, + "args": { + "External id": 829,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 16, 256, 16]"], "Input type": ["int", "ScalarList"], "Input Dims": [[4096, 4096], []], "Ev Idx": 828 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 435, "tid": 435, + "ts": 1776649790410352, "dur": 26, + "args": { + "External id": 830,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[0, 2, 1, 3]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 16, 256, 16], []], "Ev Idx": 829 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790410370, "dur": 6, + "args": { + "External id": 831,"Record function id": 0, "Concrete Inputs": ["", "[256, 256, 16, 16]", "[65536, 16, 4096, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Dims": [[256, 16, 256, 16], [], [], []], "Ev Idx": 830 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790410383, "dur": 116, + "args": { + "External id": 832,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 65536]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 256, 16, 16], []], "Ev Idx": 831 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649790410395, "dur": 96, + "args": { + "External id": 833,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[256, 256, 16, 16], []], "Ev Idx": 832 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649790410402, "dur": 31, + "args": { + "External id": 834,"Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "0"], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[256, 256, 16, 16], [], [], [], [], []], "Ev Idx": 833 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790410408, "dur": 24, + "args": { + "External id": 835,"Record function id": 0, "Concrete Inputs": ["[256, 256, 16, 16]", "3", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 834 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790410436, "dur": 54, + "args": { + "External id": 836,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 256, 16, 16], [256, 256, 16, 16], []], "Ev Idx": 835 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 435, "tid": 435, + "ts": 1776649790410495, "dur": 4, + "args": { + "External id": 837,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 65536]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 256, 16, 16], []], "Ev Idx": 836 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790410514, "dur": 3, + "args": { + "External id": 838,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 65536], []], "Ev Idx": 837 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790410516, "dur": 1, + "args": { + "External id": 839,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 65536], []], "Ev Idx": 838 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649790410536, "dur": 11, + "args": { + "External id": 840,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[16384, 1024], [], [], [], []], "Ev Idx": 839 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790410543, "dur": 1, + "args": { + "External id": 841,"Record function id": 0, "Concrete Inputs": ["", "[16384, 1024]", "[1024, 1]", "0"], "Input type": ["int", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[16384, 1024], [], [], []], "Ev Idx": 840 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::index", "pid": 435, "tid": 435, + "ts": 1776649790410553, "dur": 197, + "args": { + "External id": 842,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["int", ""], "Input Dims": [[16384, 1024], []], "Ev Idx": 841 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790410565, "dur": 68, + "args": { + "External id": 843,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "False", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Dims": [[1024], [], [], [], [], [], [], []], "Ev Idx": 842 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790410577, "dur": 55, + "args": { + "External id": 844,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Dims": [[1024], [], [], [], [], [], []], "Ev Idx": 843 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790410584, "dur": 9, + "args": { + "External id": 845,"Record function id": 0, "Concrete Inputs": ["[1024]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 844 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790410595, "dur": 37, + "args": { + "External id": 846,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Dims": [[1024], [1024], []], "Ev Idx": 845 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790410637, "dur": 12, + "args": { + "External id": 847,"Record function id": 0, "Concrete Inputs": ["", "[16384, 1024]", "[1024, 0]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Dims": [[16384, 1024], [], [], []], "Ev Idx": 846 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790410651, "dur": 2, + "args": { + "External id": 848,"Record function id": 0, "Concrete Inputs": ["", "[1, 1024]"], "Input type": ["long int", "ScalarList"], "Input Dims": [[1024], []], "Ev Idx": 847 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790410652, "dur": 0, + "args": { + "External id": 849,"Record function id": 0, "Concrete Inputs": ["", "[1, 1024]"], "Input type": ["long int", "ScalarList"], "Input Dims": [[1024], []], "Ev Idx": 848 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790410765, "dur": 4, + "args": { + "External id": 850,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 65536]"], "Input type": ["int", "ScalarList"], "Input Dims": [[16384, 1024], []], "Ev Idx": 849 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790410766, "dur": 3, + "args": { + "External id": 851,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 65536]"], "Input type": ["int", "ScalarList"], "Input Dims": [[16384, 1024], []], "Ev Idx": 850 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790410898, "dur": 45479, + "args": { + "External id": 852,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 65536], [], [], [], [], [], [], []], "Ev Idx": 851 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790410900, "dur": 45477, + "args": { + "External id": 853,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[256, 65536], [], [], [], [], [], []], "Ev Idx": 852 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790410904, "dur": 23, + "args": { + "External id": 854,"Record function id": 0, "Concrete Inputs": ["[256, 65536]", "[65536, 1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 853 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790410928, "dur": 45447, + "args": { + "External id": 855,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 65536], [256, 65536], []], "Ev Idx": 854 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 435, "tid": 435, + "ts": 1776649790456402, "dur": 10, + "args": { + "External id": 856,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 65536]], "Ev Idx": 855 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 435, "tid": 435, + "ts": 1776649790456405, "dur": 6, + "args": { + "External id": 857,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 65536]], "Ev Idx": 856 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790456413, "dur": 1, + "args": { + "External id": 858,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 65536], [], [], [], [], [], [], []], "Ev Idx": 857 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::resolve_conj", "pid": 435, "tid": 435, + "ts": 1776649790456423, "dur": 1, + "args": { + "External id": 859,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 65536]], "Ev Idx": 858 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::resolve_neg", "pid": 435, "tid": 435, + "ts": 1776649790456429, "dur": 0, + "args": { + "External id": 860,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 65536]], "Ev Idx": 859 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 435, "tid": 435, + "ts": 1776649790527799, "dur": 8, + "args": { + "External id": 861,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 8192]], "Ev Idx": 860 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790527841, "dur": 833, + "args": { + "External id": 862,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 8192], [], [], [], [], [], [], []], "Ev Idx": 861 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790527846, "dur": 827, + "args": { + "External id": 863,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[256, 8192], [], [], [], [], [], []], "Ev Idx": 862 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790527891, "dur": 60, + "args": { + "External id": 864,"Record function id": 0, "Concrete Inputs": ["[256, 8192]", "[8192, 1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 863 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790527956, "dur": 715, + "args": { + "External id": 865,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 8192], [256, 8192], []], "Ev Idx": 864 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790528725, "dur": 0, + "args": { + "External id": 866,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 8192], [], [], [], [], [], [], []], "Ev Idx": 865 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649790528756, "dur": 27, + "args": { + "External id": 867,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[256, 8192], [], [], [], []], "Ev Idx": 866 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790528770, "dur": 5, + "args": { + "External id": 868,"Record function id": 0, "Concrete Inputs": ["", "[256, 8192]", "[8192, 1]", "0"], "Input type": ["int", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[256, 8192], [], [], []], "Ev Idx": 867 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649790528784, "dur": 8, + "args": { + "External id": 869,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[256, 8192], [], [], [], []], "Ev Idx": 868 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790528789, "dur": 1, + "args": { + "External id": 870,"Record function id": 0, "Concrete Inputs": ["", "[256, 8192]", "[8192, 1]", "0"], "Input type": ["int", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[256, 8192], [], [], []], "Ev Idx": 869 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790528795, "dur": 52, + "args": { + "External id": 871,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 8192], [256, 8192], []], "Ev Idx": 870 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790528859, "dur": 0, + "args": { + "External id": 872,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[32, 4096], [], [], [], [], [], [], []], "Ev Idx": 871 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649790528865, "dur": 5, + "args": { + "External id": 873,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[32, 4096], [], [], [], []], "Ev Idx": 872 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790528867, "dur": 1, + "args": { + "External id": 874,"Record function id": 0, "Concrete Inputs": ["", "[32, 4096]", "[4096, 1]", "0"], "Input type": ["c10::Half", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 873 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649790528871, "dur": 3, + "args": { + "External id": 875,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[32, 4096], [], [], [], []], "Ev Idx": 874 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790528873, "dur": 0, + "args": { + "External id": 876,"Record function id": 0, "Concrete Inputs": ["", "[32, 4096]", "[4096, 1]", "0"], "Input type": ["c10::Half", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 875 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790528876, "dur": 19, + "args": { + "External id": 877,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[32, 4096], [32, 4096], []], "Ev Idx": 876 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790529484, "dur": 19, + "args": { + "External id": 878,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 877 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790529515, "dur": 1560, + "args": { + "External id": 879,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 878 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790529518, "dur": 6, + "args": { + "External id": 880,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 879 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790529527, "dur": 1547, + "args": { + "External id": 881,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 880 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790529547, "dur": 1524, + "args": { + "External id": 882,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 881 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790531365, "dur": 12, + "args": { + "External id": 883,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 882 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790531384, "dur": 31, + "args": { + "External id": 884,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 883 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790531386, "dur": 5, + "args": { + "External id": 885,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 884 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790531393, "dur": 21, + "args": { + "External id": 886,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 885 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790531395, "dur": 18, + "args": { + "External id": 887,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 886 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790531439, "dur": 5, + "args": { + "External id": 888,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 887 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790531449, "dur": 22, + "args": { + "External id": 889,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 888 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790531450, "dur": 4, + "args": { + "External id": 890,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 889 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790531456, "dur": 15, + "args": { + "External id": 891,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 890 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790531457, "dur": 12, + "args": { + "External id": 892,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 891 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790531490, "dur": 5, + "args": { + "External id": 893,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 892 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790531499, "dur": 22, + "args": { + "External id": 894,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 893 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790531501, "dur": 4, + "args": { + "External id": 895,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 894 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790531506, "dur": 15, + "args": { + "External id": 896,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 895 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790531508, "dur": 11, + "args": { + "External id": 897,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 896 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790531540, "dur": 5, + "args": { + "External id": 898,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 897 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790531549, "dur": 22, + "args": { + "External id": 899,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 898 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790531550, "dur": 5, + "args": { + "External id": 900,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 899 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790531556, "dur": 14, + "args": { + "External id": 901,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 900 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790531558, "dur": 11, + "args": { + "External id": 902,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 901 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790531589, "dur": 5, + "args": { + "External id": 903,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 902 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790531598, "dur": 21, + "args": { + "External id": 904,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 903 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790531599, "dur": 5, + "args": { + "External id": 905,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 904 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790531605, "dur": 14, + "args": { + "External id": 906,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 905 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790531607, "dur": 11, + "args": { + "External id": 907,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 906 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790531637, "dur": 5, + "args": { + "External id": 908,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 907 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790531646, "dur": 22, + "args": { + "External id": 909,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 908 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790531647, "dur": 5, + "args": { + "External id": 910,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 909 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790531653, "dur": 14, + "args": { + "External id": 911,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 910 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790531655, "dur": 11, + "args": { + "External id": 912,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 911 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790531686, "dur": 5, + "args": { + "External id": 913,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 912 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790531695, "dur": 21, + "args": { + "External id": 914,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 913 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790531696, "dur": 5, + "args": { + "External id": 915,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 914 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790531702, "dur": 14, + "args": { + "External id": 916,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 915 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790531704, "dur": 11, + "args": { + "External id": 917,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 916 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790531734, "dur": 5, + "args": { + "External id": 918,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 917 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790531743, "dur": 22, + "args": { + "External id": 919,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 918 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790531744, "dur": 5, + "args": { + "External id": 920,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 919 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790531750, "dur": 14, + "args": { + "External id": 921,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 920 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790531752, "dur": 11, + "args": { + "External id": 922,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 921 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790531783, "dur": 5, + "args": { + "External id": 923,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 922 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790531792, "dur": 20, + "args": { + "External id": 924,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 923 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790531793, "dur": 4, + "args": { + "External id": 925,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 924 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790531798, "dur": 14, + "args": { + "External id": 926,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 925 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790531800, "dur": 11, + "args": { + "External id": 927,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 926 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790531953, "dur": 7, + "args": { + "External id": 928,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 927 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790531965, "dur": 27, + "args": { + "External id": 929,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 928 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790531967, "dur": 4, + "args": { + "External id": 930,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 929 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790531972, "dur": 19, + "args": { + "External id": 931,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 930 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790531974, "dur": 16, + "args": { + "External id": 932,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 931 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532014, "dur": 5, + "args": { + "External id": 933,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 932 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790532023, "dur": 22, + "args": { + "External id": 934,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 933 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532025, "dur": 4, + "args": { + "External id": 935,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 934 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790532030, "dur": 15, + "args": { + "External id": 936,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 935 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790532032, "dur": 11, + "args": { + "External id": 937,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 936 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532064, "dur": 5, + "args": { + "External id": 938,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 937 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790532073, "dur": 22, + "args": { + "External id": 939,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 938 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532075, "dur": 4, + "args": { + "External id": 940,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 939 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790532080, "dur": 14, + "args": { + "External id": 941,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 940 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790532082, "dur": 11, + "args": { + "External id": 942,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 941 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532113, "dur": 5, + "args": { + "External id": 943,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 942 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790532122, "dur": 21, + "args": { + "External id": 944,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 943 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532123, "dur": 5, + "args": { + "External id": 945,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 944 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790532129, "dur": 14, + "args": { + "External id": 946,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 945 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790532131, "dur": 11, + "args": { + "External id": 947,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 946 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532161, "dur": 6, + "args": { + "External id": 948,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 947 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790532171, "dur": 21, + "args": { + "External id": 949,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 948 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532172, "dur": 4, + "args": { + "External id": 950,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 949 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790532177, "dur": 15, + "args": { + "External id": 951,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 950 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790532179, "dur": 11, + "args": { + "External id": 952,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 951 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532210, "dur": 5, + "args": { + "External id": 953,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 952 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790532264, "dur": 25, + "args": { + "External id": 954,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 953 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532265, "dur": 5, + "args": { + "External id": 955,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 954 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790532272, "dur": 16, + "args": { + "External id": 956,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 955 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790532274, "dur": 13, + "args": { + "External id": 957,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 956 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532309, "dur": 5, + "args": { + "External id": 958,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 957 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790532319, "dur": 22, + "args": { + "External id": 959,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 958 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532320, "dur": 4, + "args": { + "External id": 960,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 959 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790532326, "dur": 14, + "args": { + "External id": 961,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 960 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790532327, "dur": 12, + "args": { + "External id": 962,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 961 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532359, "dur": 5, + "args": { + "External id": 963,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 962 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790532368, "dur": 21, + "args": { + "External id": 964,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 963 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532369, "dur": 4, + "args": { + "External id": 965,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 964 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790532375, "dur": 14, + "args": { + "External id": 966,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 965 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790532376, "dur": 12, + "args": { + "External id": 967,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 966 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532407, "dur": 5, + "args": { + "External id": 968,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 967 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790532416, "dur": 21, + "args": { + "External id": 969,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 968 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532417, "dur": 5, + "args": { + "External id": 970,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 969 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790532423, "dur": 14, + "args": { + "External id": 971,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 970 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790532425, "dur": 11, + "args": { + "External id": 972,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 971 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532456, "dur": 5, + "args": { + "External id": 973,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 972 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790532465, "dur": 21, + "args": { + "External id": 974,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 973 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532467, "dur": 4, + "args": { + "External id": 975,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 974 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790532472, "dur": 14, + "args": { + "External id": 976,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 975 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790532474, "dur": 11, + "args": { + "External id": 977,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 976 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532505, "dur": 5, + "args": { + "External id": 978,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 977 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790532514, "dur": 21, + "args": { + "External id": 979,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 978 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532515, "dur": 5, + "args": { + "External id": 980,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 979 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790532521, "dur": 14, + "args": { + "External id": 981,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 980 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790532522, "dur": 11, + "args": { + "External id": 982,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 981 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532553, "dur": 5, + "args": { + "External id": 983,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 982 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790532562, "dur": 21, + "args": { + "External id": 984,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 983 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532563, "dur": 5, + "args": { + "External id": 985,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 984 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790532569, "dur": 14, + "args": { + "External id": 986,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 985 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790532570, "dur": 11, + "args": { + "External id": 987,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 986 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532601, "dur": 5, + "args": { + "External id": 988,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 987 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790532610, "dur": 21, + "args": { + "External id": 989,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 988 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532611, "dur": 5, + "args": { + "External id": 990,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 989 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790532617, "dur": 14, + "args": { + "External id": 991,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 990 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790532619, "dur": 10, + "args": { + "External id": 992,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 991 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532649, "dur": 5, + "args": { + "External id": 993,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 992 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790532658, "dur": 21, + "args": { + "External id": 994,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 993 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532659, "dur": 4, + "args": { + "External id": 995,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 994 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790532665, "dur": 14, + "args": { + "External id": 996,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 995 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790532666, "dur": 11, + "args": { + "External id": 997,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 996 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532697, "dur": 5, + "args": { + "External id": 998,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 997 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790532706, "dur": 21, + "args": { + "External id": 999,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 998 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532707, "dur": 5, + "args": { + "External id": 1000,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 999 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790532713, "dur": 14, + "args": { + "External id": 1001,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1000 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790532715, "dur": 11, + "args": { + "External id": 1002,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1001 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532745, "dur": 6, + "args": { + "External id": 1003,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1002 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790532754, "dur": 22, + "args": { + "External id": 1004,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1003 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532756, "dur": 4, + "args": { + "External id": 1005,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1004 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790532761, "dur": 14, + "args": { + "External id": 1006,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1005 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790532763, "dur": 11, + "args": { + "External id": 1007,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1006 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532794, "dur": 5, + "args": { + "External id": 1008,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1007 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790532803, "dur": 21, + "args": { + "External id": 1009,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1008 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532804, "dur": 5, + "args": { + "External id": 1010,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1009 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790532810, "dur": 14, + "args": { + "External id": 1011,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1010 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790532812, "dur": 11, + "args": { + "External id": 1012,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1011 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532842, "dur": 5, + "args": { + "External id": 1013,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1012 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790532851, "dur": 22, + "args": { + "External id": 1014,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1013 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532853, "dur": 4, + "args": { + "External id": 1015,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1014 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790532858, "dur": 14, + "args": { + "External id": 1016,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1015 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790532860, "dur": 11, + "args": { + "External id": 1017,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1016 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532890, "dur": 5, + "args": { + "External id": 1018,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1017 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790532899, "dur": 23, + "args": { + "External id": 1019,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1018 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532901, "dur": 4, + "args": { + "External id": 1020,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1019 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790532908, "dur": 14, + "args": { + "External id": 1021,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1020 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790532909, "dur": 12, + "args": { + "External id": 1022,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1021 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532940, "dur": 5, + "args": { + "External id": 1023,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1022 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790532950, "dur": 31, + "args": { + "External id": 1024,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1023 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532960, "dur": 5, + "args": { + "External id": 1025,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1024 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790532966, "dur": 14, + "args": { + "External id": 1026,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1025 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790532968, "dur": 11, + "args": { + "External id": 1027,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1026 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790532999, "dur": 5, + "args": { + "External id": 1028,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1027 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790533009, "dur": 21, + "args": { + "External id": 1029,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1028 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533010, "dur": 4, + "args": { + "External id": 1030,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1029 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790533015, "dur": 15, + "args": { + "External id": 1031,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1030 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790533017, "dur": 11, + "args": { + "External id": 1032,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1031 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533048, "dur": 5, + "args": { + "External id": 1033,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1032 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790533057, "dur": 21, + "args": { + "External id": 1034,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1033 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533058, "dur": 5, + "args": { + "External id": 1035,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1034 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790533064, "dur": 14, + "args": { + "External id": 1036,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1035 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790533066, "dur": 11, + "args": { + "External id": 1037,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1036 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533096, "dur": 5, + "args": { + "External id": 1038,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1037 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790533105, "dur": 21, + "args": { + "External id": 1039,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1038 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533107, "dur": 4, + "args": { + "External id": 1040,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1039 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790533112, "dur": 14, + "args": { + "External id": 1041,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1040 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790533114, "dur": 11, + "args": { + "External id": 1042,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1041 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533144, "dur": 5, + "args": { + "External id": 1043,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1042 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790533153, "dur": 21, + "args": { + "External id": 1044,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1043 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533154, "dur": 5, + "args": { + "External id": 1045,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1044 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790533160, "dur": 14, + "args": { + "External id": 1046,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1045 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790533162, "dur": 11, + "args": { + "External id": 1047,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1046 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533192, "dur": 5, + "args": { + "External id": 1048,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1047 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790533201, "dur": 21, + "args": { + "External id": 1049,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1048 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533202, "dur": 5, + "args": { + "External id": 1050,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1049 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790533208, "dur": 14, + "args": { + "External id": 1051,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1050 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790533210, "dur": 11, + "args": { + "External id": 1052,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1051 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533240, "dur": 5, + "args": { + "External id": 1053,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1052 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790533249, "dur": 21, + "args": { + "External id": 1054,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1053 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533250, "dur": 5, + "args": { + "External id": 1055,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1054 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790533256, "dur": 14, + "args": { + "External id": 1056,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1055 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790533258, "dur": 11, + "args": { + "External id": 1057,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1056 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533288, "dur": 5, + "args": { + "External id": 1058,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1057 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790533297, "dur": 22, + "args": { + "External id": 1059,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1058 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533299, "dur": 4, + "args": { + "External id": 1060,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1059 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790533304, "dur": 15, + "args": { + "External id": 1061,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1060 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790533306, "dur": 11, + "args": { + "External id": 1062,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1061 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533337, "dur": 5, + "args": { + "External id": 1063,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1062 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790533346, "dur": 22, + "args": { + "External id": 1064,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1063 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533347, "dur": 5, + "args": { + "External id": 1065,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1064 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790533353, "dur": 15, + "args": { + "External id": 1066,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1065 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790533355, "dur": 11, + "args": { + "External id": 1067,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1066 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533386, "dur": 5, + "args": { + "External id": 1068,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1067 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790533395, "dur": 21, + "args": { + "External id": 1069,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1068 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533396, "dur": 5, + "args": { + "External id": 1070,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1069 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790533402, "dur": 14, + "args": { + "External id": 1071,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1070 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790533404, "dur": 11, + "args": { + "External id": 1072,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1071 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533434, "dur": 5, + "args": { + "External id": 1073,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1072 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790533443, "dur": 21, + "args": { + "External id": 1074,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1073 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533444, "dur": 5, + "args": { + "External id": 1075,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1074 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790533450, "dur": 14, + "args": { + "External id": 1076,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1075 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790533452, "dur": 11, + "args": { + "External id": 1077,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1076 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533482, "dur": 5, + "args": { + "External id": 1078,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1077 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790533491, "dur": 21, + "args": { + "External id": 1079,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1078 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533492, "dur": 5, + "args": { + "External id": 1080,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1079 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790533498, "dur": 14, + "args": { + "External id": 1081,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1080 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790533500, "dur": 11, + "args": { + "External id": 1082,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1081 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533530, "dur": 5, + "args": { + "External id": 1083,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1082 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790533539, "dur": 21, + "args": { + "External id": 1084,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1083 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533540, "dur": 4, + "args": { + "External id": 1085,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1084 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790533546, "dur": 14, + "args": { + "External id": 1086,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1085 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790533547, "dur": 12, + "args": { + "External id": 1087,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1086 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533578, "dur": 5, + "args": { + "External id": 1088,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1087 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790533587, "dur": 21, + "args": { + "External id": 1089,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1088 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533588, "dur": 5, + "args": { + "External id": 1090,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1089 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790533594, "dur": 14, + "args": { + "External id": 1091,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1090 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790533596, "dur": 11, + "args": { + "External id": 1092,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1091 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533626, "dur": 6, + "args": { + "External id": 1093,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1092 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790533636, "dur": 21, + "args": { + "External id": 1094,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1093 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533637, "dur": 4, + "args": { + "External id": 1095,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1094 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790533643, "dur": 13, + "args": { + "External id": 1096,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1095 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790533644, "dur": 11, + "args": { + "External id": 1097,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1096 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533675, "dur": 5, + "args": { + "External id": 1098,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1097 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790533684, "dur": 21, + "args": { + "External id": 1099,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1098 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533685, "dur": 5, + "args": { + "External id": 1100,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1099 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790533691, "dur": 14, + "args": { + "External id": 1101,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1100 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790533692, "dur": 11, + "args": { + "External id": 1102,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1101 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533723, "dur": 5, + "args": { + "External id": 1103,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1102 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790533732, "dur": 21, + "args": { + "External id": 1104,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1103 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533733, "dur": 5, + "args": { + "External id": 1105,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1104 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790533739, "dur": 14, + "args": { + "External id": 1106,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1105 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790533741, "dur": 11, + "args": { + "External id": 1107,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1106 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533771, "dur": 5, + "args": { + "External id": 1108,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1107 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790533780, "dur": 21, + "args": { + "External id": 1109,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1108 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533781, "dur": 5, + "args": { + "External id": 1110,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1109 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790533787, "dur": 14, + "args": { + "External id": 1111,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1110 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790533788, "dur": 11, + "args": { + "External id": 1112,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1111 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533819, "dur": 5, + "args": { + "External id": 1113,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1112 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790533828, "dur": 21, + "args": { + "External id": 1114,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1113 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533829, "dur": 4, + "args": { + "External id": 1115,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1114 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790533835, "dur": 14, + "args": { + "External id": 1116,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1115 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790533836, "dur": 11, + "args": { + "External id": 1117,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1116 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533869, "dur": 5, + "args": { + "External id": 1118,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1117 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790533879, "dur": 21, + "args": { + "External id": 1119,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1118 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533880, "dur": 4, + "args": { + "External id": 1120,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1119 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790533885, "dur": 14, + "args": { + "External id": 1121,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1120 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790533887, "dur": 11, + "args": { + "External id": 1122,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1121 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533918, "dur": 5, + "args": { + "External id": 1123,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1122 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790533927, "dur": 21, + "args": { + "External id": 1124,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1123 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533928, "dur": 4, + "args": { + "External id": 1125,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1124 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790533934, "dur": 13, + "args": { + "External id": 1126,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1125 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790533935, "dur": 11, + "args": { + "External id": 1127,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1126 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533966, "dur": 5, + "args": { + "External id": 1128,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1127 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790533975, "dur": 21, + "args": { + "External id": 1129,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1128 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790533976, "dur": 4, + "args": { + "External id": 1130,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1129 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790533982, "dur": 14, + "args": { + "External id": 1131,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1130 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790533983, "dur": 11, + "args": { + "External id": 1132,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1131 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534014, "dur": 5, + "args": { + "External id": 1133,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1132 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790534023, "dur": 21, + "args": { + "External id": 1134,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1133 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534024, "dur": 4, + "args": { + "External id": 1135,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1134 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790534029, "dur": 14, + "args": { + "External id": 1136,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1135 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790534031, "dur": 11, + "args": { + "External id": 1137,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1136 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534062, "dur": 5, + "args": { + "External id": 1138,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1137 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790534071, "dur": 21, + "args": { + "External id": 1139,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1138 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534072, "dur": 4, + "args": { + "External id": 1140,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1139 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790534078, "dur": 14, + "args": { + "External id": 1141,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1140 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790534079, "dur": 11, + "args": { + "External id": 1142,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1141 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534109, "dur": 5, + "args": { + "External id": 1143,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1142 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790534118, "dur": 21, + "args": { + "External id": 1144,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1143 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534119, "dur": 5, + "args": { + "External id": 1145,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1144 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790534125, "dur": 14, + "args": { + "External id": 1146,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1145 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790534127, "dur": 11, + "args": { + "External id": 1147,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1146 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534157, "dur": 5, + "args": { + "External id": 1148,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1147 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790534166, "dur": 21, + "args": { + "External id": 1149,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1148 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534167, "dur": 5, + "args": { + "External id": 1150,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1149 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790534173, "dur": 14, + "args": { + "External id": 1151,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1150 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790534175, "dur": 11, + "args": { + "External id": 1152,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1151 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534205, "dur": 5, + "args": { + "External id": 1153,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1152 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790534214, "dur": 21, + "args": { + "External id": 1154,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1153 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534215, "dur": 4, + "args": { + "External id": 1155,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1154 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790534220, "dur": 14, + "args": { + "External id": 1156,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1155 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790534222, "dur": 11, + "args": { + "External id": 1157,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1156 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534252, "dur": 5, + "args": { + "External id": 1158,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1157 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790534261, "dur": 21, + "args": { + "External id": 1159,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1158 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534263, "dur": 4, + "args": { + "External id": 1160,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1159 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790534268, "dur": 14, + "args": { + "External id": 1161,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1160 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790534270, "dur": 11, + "args": { + "External id": 1162,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1161 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534300, "dur": 5, + "args": { + "External id": 1163,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1162 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790534309, "dur": 21, + "args": { + "External id": 1164,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1163 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534311, "dur": 4, + "args": { + "External id": 1165,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1164 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790534316, "dur": 14, + "args": { + "External id": 1166,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1165 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790534318, "dur": 11, + "args": { + "External id": 1167,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1166 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534348, "dur": 5, + "args": { + "External id": 1168,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1167 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790534357, "dur": 22, + "args": { + "External id": 1169,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1168 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534358, "dur": 5, + "args": { + "External id": 1170,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1169 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790534364, "dur": 14, + "args": { + "External id": 1171,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1170 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790534366, "dur": 11, + "args": { + "External id": 1172,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1171 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534396, "dur": 5, + "args": { + "External id": 1173,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1172 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790534405, "dur": 21, + "args": { + "External id": 1174,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1173 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534406, "dur": 5, + "args": { + "External id": 1175,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1174 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790534412, "dur": 14, + "args": { + "External id": 1176,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1175 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790534414, "dur": 11, + "args": { + "External id": 1177,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1176 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534516, "dur": 7, + "args": { + "External id": 1178,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1177 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790534528, "dur": 26, + "args": { + "External id": 1179,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1178 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534530, "dur": 4, + "args": { + "External id": 1180,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1179 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790534536, "dur": 17, + "args": { + "External id": 1181,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1180 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790534537, "dur": 15, + "args": { + "External id": 1182,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1181 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534575, "dur": 5, + "args": { + "External id": 1183,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1182 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790534584, "dur": 22, + "args": { + "External id": 1184,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1183 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534586, "dur": 4, + "args": { + "External id": 1185,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1184 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790534591, "dur": 15, + "args": { + "External id": 1186,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1185 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790534593, "dur": 11, + "args": { + "External id": 1187,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1186 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534625, "dur": 6, + "args": { + "External id": 1188,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1187 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790534635, "dur": 22, + "args": { + "External id": 1189,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1188 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534636, "dur": 5, + "args": { + "External id": 1190,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1189 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790534642, "dur": 14, + "args": { + "External id": 1191,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1190 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790534644, "dur": 11, + "args": { + "External id": 1192,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1191 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534675, "dur": 5, + "args": { + "External id": 1193,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1192 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790534684, "dur": 21, + "args": { + "External id": 1194,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1193 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534685, "dur": 5, + "args": { + "External id": 1195,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1194 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790534691, "dur": 14, + "args": { + "External id": 1196,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1195 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790534693, "dur": 11, + "args": { + "External id": 1197,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1196 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534724, "dur": 5, + "args": { + "External id": 1198,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1197 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790534733, "dur": 21, + "args": { + "External id": 1199,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1198 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534734, "dur": 4, + "args": { + "External id": 1200,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1199 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790534740, "dur": 13, + "args": { + "External id": 1201,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1200 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790534741, "dur": 11, + "args": { + "External id": 1202,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1201 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534772, "dur": 5, + "args": { + "External id": 1203,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1202 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790534781, "dur": 21, + "args": { + "External id": 1204,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1203 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534782, "dur": 4, + "args": { + "External id": 1205,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1204 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790534788, "dur": 14, + "args": { + "External id": 1206,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1205 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790534789, "dur": 11, + "args": { + "External id": 1207,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1206 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534820, "dur": 5, + "args": { + "External id": 1208,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1207 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790534829, "dur": 21, + "args": { + "External id": 1209,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1208 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534830, "dur": 5, + "args": { + "External id": 1210,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1209 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790534836, "dur": 14, + "args": { + "External id": 1211,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1210 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790534838, "dur": 11, + "args": { + "External id": 1212,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1211 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534868, "dur": 5, + "args": { + "External id": 1213,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1212 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790534877, "dur": 21, + "args": { + "External id": 1214,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1213 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534879, "dur": 4, + "args": { + "External id": 1215,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1214 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790534884, "dur": 14, + "args": { + "External id": 1216,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1215 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790534886, "dur": 11, + "args": { + "External id": 1217,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1216 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534917, "dur": 5, + "args": { + "External id": 1218,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1217 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790534926, "dur": 21, + "args": { + "External id": 1219,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1218 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534927, "dur": 5, + "args": { + "External id": 1220,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1219 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790534933, "dur": 14, + "args": { + "External id": 1221,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1220 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790534934, "dur": 12, + "args": { + "External id": 1222,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1221 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534965, "dur": 5, + "args": { + "External id": 1223,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1222 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790534974, "dur": 21, + "args": { + "External id": 1224,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1223 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790534975, "dur": 5, + "args": { + "External id": 1225,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1224 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790534981, "dur": 14, + "args": { + "External id": 1226,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1225 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790534982, "dur": 12, + "args": { + "External id": 1227,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1226 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535065, "dur": 6, + "args": { + "External id": 1228,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1227 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790535076, "dur": 25, + "args": { + "External id": 1229,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1228 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535077, "dur": 5, + "args": { + "External id": 1230,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1229 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790535083, "dur": 17, + "args": { + "External id": 1231,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1230 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790535085, "dur": 14, + "args": { + "External id": 1232,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1231 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535121, "dur": 6, + "args": { + "External id": 1233,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1232 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790535131, "dur": 21, + "args": { + "External id": 1234,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1233 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535132, "dur": 4, + "args": { + "External id": 1235,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1234 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790535138, "dur": 14, + "args": { + "External id": 1236,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1235 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790535139, "dur": 11, + "args": { + "External id": 1237,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1236 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535170, "dur": 6, + "args": { + "External id": 1238,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1237 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790535180, "dur": 21, + "args": { + "External id": 1239,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1238 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535181, "dur": 4, + "args": { + "External id": 1240,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1239 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790535187, "dur": 14, + "args": { + "External id": 1241,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1240 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790535188, "dur": 12, + "args": { + "External id": 1242,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1241 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535219, "dur": 6, + "args": { + "External id": 1243,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1242 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790535229, "dur": 21, + "args": { + "External id": 1244,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1243 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535230, "dur": 4, + "args": { + "External id": 1245,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1244 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790535236, "dur": 14, + "args": { + "External id": 1246,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1245 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790535237, "dur": 12, + "args": { + "External id": 1247,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1246 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535269, "dur": 5, + "args": { + "External id": 1248,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1247 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790535278, "dur": 21, + "args": { + "External id": 1249,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1248 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535279, "dur": 4, + "args": { + "External id": 1250,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1249 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790535284, "dur": 15, + "args": { + "External id": 1251,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1250 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790535286, "dur": 11, + "args": { + "External id": 1252,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1251 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535317, "dur": 5, + "args": { + "External id": 1253,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1252 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790535326, "dur": 22, + "args": { + "External id": 1254,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1253 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535327, "dur": 5, + "args": { + "External id": 1255,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1254 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790535333, "dur": 14, + "args": { + "External id": 1256,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1255 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790535335, "dur": 11, + "args": { + "External id": 1257,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1256 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535365, "dur": 6, + "args": { + "External id": 1258,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1257 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790535375, "dur": 21, + "args": { + "External id": 1259,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1258 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535376, "dur": 4, + "args": { + "External id": 1260,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1259 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790535381, "dur": 15, + "args": { + "External id": 1261,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1260 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790535383, "dur": 11, + "args": { + "External id": 1262,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1261 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535415, "dur": 5, + "args": { + "External id": 1263,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1262 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790535424, "dur": 21, + "args": { + "External id": 1264,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1263 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535425, "dur": 4, + "args": { + "External id": 1265,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1264 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790535431, "dur": 14, + "args": { + "External id": 1266,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1265 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790535432, "dur": 12, + "args": { + "External id": 1267,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1266 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535463, "dur": 6, + "args": { + "External id": 1268,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1267 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790535473, "dur": 22, + "args": { + "External id": 1269,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1268 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535474, "dur": 5, + "args": { + "External id": 1270,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1269 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790535480, "dur": 14, + "args": { + "External id": 1271,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1270 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790535482, "dur": 11, + "args": { + "External id": 1272,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1271 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535513, "dur": 5, + "args": { + "External id": 1273,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1272 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790535522, "dur": 21, + "args": { + "External id": 1274,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1273 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535523, "dur": 4, + "args": { + "External id": 1275,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1274 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790535529, "dur": 14, + "args": { + "External id": 1276,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1275 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790535531, "dur": 11, + "args": { + "External id": 1277,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1276 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535561, "dur": 6, + "args": { + "External id": 1278,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1277 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790535570, "dur": 22, + "args": { + "External id": 1279,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1278 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535572, "dur": 4, + "args": { + "External id": 1280,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1279 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790535577, "dur": 15, + "args": { + "External id": 1281,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1280 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790535579, "dur": 11, + "args": { + "External id": 1282,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1281 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535610, "dur": 6, + "args": { + "External id": 1283,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1282 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790535620, "dur": 21, + "args": { + "External id": 1284,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1283 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535621, "dur": 4, + "args": { + "External id": 1285,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1284 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790535627, "dur": 14, + "args": { + "External id": 1286,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1285 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790535628, "dur": 11, + "args": { + "External id": 1287,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1286 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535659, "dur": 5, + "args": { + "External id": 1288,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1287 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790535668, "dur": 21, + "args": { + "External id": 1289,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1288 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535669, "dur": 4, + "args": { + "External id": 1290,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1289 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790535675, "dur": 14, + "args": { + "External id": 1291,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1290 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790535676, "dur": 11, + "args": { + "External id": 1292,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1291 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535707, "dur": 5, + "args": { + "External id": 1293,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1292 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790535716, "dur": 21, + "args": { + "External id": 1294,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1293 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535717, "dur": 4, + "args": { + "External id": 1295,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1294 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790535723, "dur": 13, + "args": { + "External id": 1296,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1295 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790535724, "dur": 11, + "args": { + "External id": 1297,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1296 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535755, "dur": 5, + "args": { + "External id": 1298,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1297 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790535764, "dur": 21, + "args": { + "External id": 1299,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1298 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535765, "dur": 4, + "args": { + "External id": 1300,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1299 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790535771, "dur": 14, + "args": { + "External id": 1301,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1300 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790535772, "dur": 12, + "args": { + "External id": 1302,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1301 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535803, "dur": 6, + "args": { + "External id": 1303,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1302 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790535812, "dur": 22, + "args": { + "External id": 1304,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1303 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535814, "dur": 4, + "args": { + "External id": 1305,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1304 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790535819, "dur": 14, + "args": { + "External id": 1306,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1305 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790535821, "dur": 11, + "args": { + "External id": 1307,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1306 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535852, "dur": 5, + "args": { + "External id": 1308,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1307 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790535861, "dur": 21, + "args": { + "External id": 1309,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1308 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535862, "dur": 4, + "args": { + "External id": 1310,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1309 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790535867, "dur": 14, + "args": { + "External id": 1311,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1310 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790535869, "dur": 11, + "args": { + "External id": 1312,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1311 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535899, "dur": 6, + "args": { + "External id": 1313,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1312 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790535908, "dur": 22, + "args": { + "External id": 1314,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1313 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535910, "dur": 4, + "args": { + "External id": 1315,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1314 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790535915, "dur": 14, + "args": { + "External id": 1316,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1315 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790535917, "dur": 11, + "args": { + "External id": 1317,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1316 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535947, "dur": 5, + "args": { + "External id": 1318,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1317 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790535956, "dur": 22, + "args": { + "External id": 1319,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1318 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535958, "dur": 4, + "args": { + "External id": 1320,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1319 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790535963, "dur": 14, + "args": { + "External id": 1321,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1320 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790535965, "dur": 11, + "args": { + "External id": 1322,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1321 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790535996, "dur": 5, + "args": { + "External id": 1323,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1322 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790536005, "dur": 21, + "args": { + "External id": 1324,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1323 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536006, "dur": 4, + "args": { + "External id": 1325,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1324 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790536012, "dur": 13, + "args": { + "External id": 1326,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1325 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790536013, "dur": 11, + "args": { + "External id": 1327,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1326 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536043, "dur": 5, + "args": { + "External id": 1328,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1327 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790536052, "dur": 22, + "args": { + "External id": 1329,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1328 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536053, "dur": 5, + "args": { + "External id": 1330,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1329 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790536059, "dur": 14, + "args": { + "External id": 1331,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1330 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790536061, "dur": 11, + "args": { + "External id": 1332,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1331 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536092, "dur": 5, + "args": { + "External id": 1333,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1332 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790536101, "dur": 21, + "args": { + "External id": 1334,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1333 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536102, "dur": 4, + "args": { + "External id": 1335,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1334 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790536107, "dur": 14, + "args": { + "External id": 1336,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1335 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790536109, "dur": 11, + "args": { + "External id": 1337,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1336 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536139, "dur": 5, + "args": { + "External id": 1338,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1337 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790536148, "dur": 21, + "args": { + "External id": 1339,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1338 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536149, "dur": 5, + "args": { + "External id": 1340,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1339 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790536155, "dur": 14, + "args": { + "External id": 1341,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1340 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790536157, "dur": 11, + "args": { + "External id": 1342,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1341 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536187, "dur": 5, + "args": { + "External id": 1343,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1342 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790536196, "dur": 41, + "args": { + "External id": 1344,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1343 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536197, "dur": 5, + "args": { + "External id": 1345,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1344 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790536203, "dur": 34, + "args": { + "External id": 1346,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1345 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790536205, "dur": 30, + "args": { + "External id": 1347,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1346 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536259, "dur": 6, + "args": { + "External id": 1348,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1347 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790536269, "dur": 22, + "args": { + "External id": 1349,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1348 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536270, "dur": 5, + "args": { + "External id": 1350,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1349 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790536276, "dur": 15, + "args": { + "External id": 1351,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1350 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790536278, "dur": 12, + "args": { + "External id": 1352,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1351 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536310, "dur": 5, + "args": { + "External id": 1353,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1352 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790536319, "dur": 21, + "args": { + "External id": 1354,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1353 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536320, "dur": 4, + "args": { + "External id": 1355,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1354 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790536326, "dur": 14, + "args": { + "External id": 1356,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1355 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790536327, "dur": 12, + "args": { + "External id": 1357,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1356 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536358, "dur": 5, + "args": { + "External id": 1358,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1357 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790536367, "dur": 21, + "args": { + "External id": 1359,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1358 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536368, "dur": 5, + "args": { + "External id": 1360,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1359 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790536374, "dur": 14, + "args": { + "External id": 1361,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1360 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790536376, "dur": 11, + "args": { + "External id": 1362,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1361 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536407, "dur": 5, + "args": { + "External id": 1363,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1362 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790536416, "dur": 21, + "args": { + "External id": 1364,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1363 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536417, "dur": 4, + "args": { + "External id": 1365,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1364 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790536422, "dur": 14, + "args": { + "External id": 1366,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1365 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790536424, "dur": 11, + "args": { + "External id": 1367,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1366 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536454, "dur": 6, + "args": { + "External id": 1368,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1367 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790536464, "dur": 21, + "args": { + "External id": 1369,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1368 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536465, "dur": 4, + "args": { + "External id": 1370,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1369 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790536470, "dur": 14, + "args": { + "External id": 1371,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1370 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790536472, "dur": 11, + "args": { + "External id": 1372,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1371 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536503, "dur": 5, + "args": { + "External id": 1373,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1372 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790536512, "dur": 21, + "args": { + "External id": 1374,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1373 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536513, "dur": 4, + "args": { + "External id": 1375,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1374 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790536519, "dur": 13, + "args": { + "External id": 1376,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1375 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790536520, "dur": 11, + "args": { + "External id": 1377,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1376 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536551, "dur": 5, + "args": { + "External id": 1378,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1377 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790536560, "dur": 21, + "args": { + "External id": 1379,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1378 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536561, "dur": 4, + "args": { + "External id": 1380,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1379 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790536567, "dur": 14, + "args": { + "External id": 1381,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1380 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790536568, "dur": 12, + "args": { + "External id": 1382,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1381 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536599, "dur": 5, + "args": { + "External id": 1383,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1382 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790536608, "dur": 21, + "args": { + "External id": 1384,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1383 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536609, "dur": 5, + "args": { + "External id": 1385,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1384 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790536615, "dur": 14, + "args": { + "External id": 1386,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1385 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790536617, "dur": 11, + "args": { + "External id": 1387,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1386 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536647, "dur": 6, + "args": { + "External id": 1388,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1387 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790536657, "dur": 21, + "args": { + "External id": 1389,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1388 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536658, "dur": 4, + "args": { + "External id": 1390,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1389 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790536663, "dur": 14, + "args": { + "External id": 1391,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1390 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790536665, "dur": 11, + "args": { + "External id": 1392,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1391 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536695, "dur": 6, + "args": { + "External id": 1393,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1392 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790536705, "dur": 21, + "args": { + "External id": 1394,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1393 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536706, "dur": 4, + "args": { + "External id": 1395,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1394 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790536711, "dur": 14, + "args": { + "External id": 1396,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1395 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790536713, "dur": 11, + "args": { + "External id": 1397,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1396 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536743, "dur": 5, + "args": { + "External id": 1398,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1397 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790536752, "dur": 22, + "args": { + "External id": 1399,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1398 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536754, "dur": 4, + "args": { + "External id": 1400,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1399 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790536759, "dur": 14, + "args": { + "External id": 1401,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1400 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790536761, "dur": 11, + "args": { + "External id": 1402,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1401 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536791, "dur": 5, + "args": { + "External id": 1403,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1402 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790536800, "dur": 21, + "args": { + "External id": 1404,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1403 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536801, "dur": 5, + "args": { + "External id": 1405,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1404 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790536807, "dur": 14, + "args": { + "External id": 1406,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1405 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790536809, "dur": 11, + "args": { + "External id": 1407,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1406 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536839, "dur": 5, + "args": { + "External id": 1408,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1407 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790536848, "dur": 21, + "args": { + "External id": 1409,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1408 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536849, "dur": 5, + "args": { + "External id": 1410,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1409 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790536855, "dur": 14, + "args": { + "External id": 1411,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1410 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790536857, "dur": 11, + "args": { + "External id": 1412,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1411 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536887, "dur": 6, + "args": { + "External id": 1413,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1412 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790536897, "dur": 21, + "args": { + "External id": 1414,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1413 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536898, "dur": 4, + "args": { + "External id": 1415,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1414 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790536904, "dur": 14, + "args": { + "External id": 1416,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1415 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790536905, "dur": 11, + "args": { + "External id": 1417,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1416 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536935, "dur": 6, + "args": { + "External id": 1418,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1417 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790536944, "dur": 22, + "args": { + "External id": 1419,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1418 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536946, "dur": 4, + "args": { + "External id": 1420,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1419 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790536951, "dur": 15, + "args": { + "External id": 1421,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1420 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790536953, "dur": 11, + "args": { + "External id": 1422,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1421 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536984, "dur": 5, + "args": { + "External id": 1423,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1422 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790536993, "dur": 21, + "args": { + "External id": 1424,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1423 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790536994, "dur": 4, + "args": { + "External id": 1425,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1424 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790537000, "dur": 13, + "args": { + "External id": 1426,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1425 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790537001, "dur": 11, + "args": { + "External id": 1427,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1426 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537032, "dur": 5, + "args": { + "External id": 1428,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1427 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790537041, "dur": 21, + "args": { + "External id": 1429,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1428 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537042, "dur": 5, + "args": { + "External id": 1430,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1429 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790537048, "dur": 14, + "args": { + "External id": 1431,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1430 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790537049, "dur": 11, + "args": { + "External id": 1432,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1431 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537080, "dur": 5, + "args": { + "External id": 1433,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1432 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790537089, "dur": 21, + "args": { + "External id": 1434,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1433 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537090, "dur": 5, + "args": { + "External id": 1435,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1434 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790537096, "dur": 14, + "args": { + "External id": 1436,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1435 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790537098, "dur": 11, + "args": { + "External id": 1437,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1436 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537128, "dur": 5, + "args": { + "External id": 1438,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1437 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790537137, "dur": 21, + "args": { + "External id": 1439,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1438 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537138, "dur": 4, + "args": { + "External id": 1440,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1439 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790537144, "dur": 13, + "args": { + "External id": 1441,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1440 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790537145, "dur": 11, + "args": { + "External id": 1442,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1441 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537176, "dur": 5, + "args": { + "External id": 1443,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1442 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790537185, "dur": 21, + "args": { + "External id": 1444,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1443 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537186, "dur": 5, + "args": { + "External id": 1445,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1444 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790537192, "dur": 14, + "args": { + "External id": 1446,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1445 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790537193, "dur": 11, + "args": { + "External id": 1447,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1446 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537224, "dur": 5, + "args": { + "External id": 1448,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1447 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790537233, "dur": 21, + "args": { + "External id": 1449,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1448 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537234, "dur": 4, + "args": { + "External id": 1450,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1449 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790537240, "dur": 13, + "args": { + "External id": 1451,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1450 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790537241, "dur": 11, + "args": { + "External id": 1452,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1451 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537272, "dur": 5, + "args": { + "External id": 1453,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1452 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790537281, "dur": 21, + "args": { + "External id": 1454,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1453 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537282, "dur": 4, + "args": { + "External id": 1455,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1454 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790537288, "dur": 14, + "args": { + "External id": 1456,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1455 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790537289, "dur": 11, + "args": { + "External id": 1457,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1456 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537320, "dur": 5, + "args": { + "External id": 1458,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1457 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790537329, "dur": 21, + "args": { + "External id": 1459,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1458 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537330, "dur": 4, + "args": { + "External id": 1460,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1459 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790537336, "dur": 14, + "args": { + "External id": 1461,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1460 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790537337, "dur": 11, + "args": { + "External id": 1462,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1461 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537368, "dur": 5, + "args": { + "External id": 1463,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1462 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790537377, "dur": 21, + "args": { + "External id": 1464,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1463 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537378, "dur": 5, + "args": { + "External id": 1465,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1464 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790537384, "dur": 14, + "args": { + "External id": 1466,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1465 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790537386, "dur": 11, + "args": { + "External id": 1467,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1466 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537416, "dur": 5, + "args": { + "External id": 1468,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1467 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790537425, "dur": 20, + "args": { + "External id": 1469,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1468 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537426, "dur": 4, + "args": { + "External id": 1470,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1469 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790537431, "dur": 14, + "args": { + "External id": 1471,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1470 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790537433, "dur": 11, + "args": { + "External id": 1472,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1471 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537463, "dur": 5, + "args": { + "External id": 1473,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1472 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790537472, "dur": 21, + "args": { + "External id": 1474,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1473 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537474, "dur": 4, + "args": { + "External id": 1475,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1474 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790537479, "dur": 14, + "args": { + "External id": 1476,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1475 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790537481, "dur": 11, + "args": { + "External id": 1477,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1476 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537573, "dur": 6, + "args": { + "External id": 1478,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1477 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790537584, "dur": 25, + "args": { + "External id": 1479,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1478 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537585, "dur": 5, + "args": { + "External id": 1480,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1479 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790537591, "dur": 17, + "args": { + "External id": 1481,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1480 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790537593, "dur": 14, + "args": { + "External id": 1482,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1481 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537630, "dur": 5, + "args": { + "External id": 1483,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1482 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790537639, "dur": 28, + "args": { + "External id": 1484,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1483 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537640, "dur": 10, + "args": { + "External id": 1485,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1484 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790537652, "dur": 15, + "args": { + "External id": 1486,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1485 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790537653, "dur": 13, + "args": { + "External id": 1487,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1486 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537687, "dur": 5, + "args": { + "External id": 1488,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1487 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790537696, "dur": 21, + "args": { + "External id": 1489,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1488 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537697, "dur": 4, + "args": { + "External id": 1490,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1489 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790537703, "dur": 14, + "args": { + "External id": 1491,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1490 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790537704, "dur": 11, + "args": { + "External id": 1492,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1491 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537735, "dur": 5, + "args": { + "External id": 1493,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1492 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790537744, "dur": 21, + "args": { + "External id": 1494,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1493 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537745, "dur": 4, + "args": { + "External id": 1495,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1494 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790537751, "dur": 14, + "args": { + "External id": 1496,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1495 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790537752, "dur": 11, + "args": { + "External id": 1497,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1496 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537782, "dur": 5, + "args": { + "External id": 1498,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1497 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790537791, "dur": 21, + "args": { + "External id": 1499,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1498 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537792, "dur": 5, + "args": { + "External id": 1500,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1499 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790537798, "dur": 14, + "args": { + "External id": 1501,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1500 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790537800, "dur": 11, + "args": { + "External id": 1502,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1501 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537830, "dur": 5, + "args": { + "External id": 1503,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1502 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790537839, "dur": 21, + "args": { + "External id": 1504,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1503 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537840, "dur": 4, + "args": { + "External id": 1505,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1504 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790537845, "dur": 14, + "args": { + "External id": 1506,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1505 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790537847, "dur": 11, + "args": { + "External id": 1507,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1506 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537877, "dur": 5, + "args": { + "External id": 1508,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1507 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790537886, "dur": 21, + "args": { + "External id": 1509,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1508 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537887, "dur": 4, + "args": { + "External id": 1510,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1509 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790537893, "dur": 13, + "args": { + "External id": 1511,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1510 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790537894, "dur": 11, + "args": { + "External id": 1512,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1511 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537924, "dur": 6, + "args": { + "External id": 1513,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1512 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790537933, "dur": 21, + "args": { + "External id": 1514,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1513 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537934, "dur": 5, + "args": { + "External id": 1515,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1514 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790537940, "dur": 14, + "args": { + "External id": 1516,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1515 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790537942, "dur": 11, + "args": { + "External id": 1517,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1516 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537972, "dur": 5, + "args": { + "External id": 1518,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1517 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790537980, "dur": 22, + "args": { + "External id": 1519,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1518 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790537982, "dur": 4, + "args": { + "External id": 1520,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1519 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790537987, "dur": 14, + "args": { + "External id": 1521,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1520 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790537989, "dur": 11, + "args": { + "External id": 1522,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1521 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538019, "dur": 5, + "args": { + "External id": 1523,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1522 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790538028, "dur": 21, + "args": { + "External id": 1524,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1523 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538029, "dur": 5, + "args": { + "External id": 1525,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1524 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790538035, "dur": 14, + "args": { + "External id": 1526,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1525 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790538037, "dur": 10, + "args": { + "External id": 1527,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1526 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538117, "dur": 6, + "args": { + "External id": 1528,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1527 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790538128, "dur": 24, + "args": { + "External id": 1529,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1528 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538129, "dur": 4, + "args": { + "External id": 1530,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1529 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790538135, "dur": 16, + "args": { + "External id": 1531,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1530 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790538136, "dur": 14, + "args": { + "External id": 1532,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1531 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538172, "dur": 5, + "args": { + "External id": 1533,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1532 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790538181, "dur": 31, + "args": { + "External id": 1534,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1533 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538182, "dur": 4, + "args": { + "External id": 1535,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1534 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790538188, "dur": 23, + "args": { + "External id": 1536,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1535 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790538199, "dur": 11, + "args": { + "External id": 1537,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1536 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538230, "dur": 5, + "args": { + "External id": 1538,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1537 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790538239, "dur": 22, + "args": { + "External id": 1539,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1538 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538240, "dur": 5, + "args": { + "External id": 1540,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1539 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790538246, "dur": 14, + "args": { + "External id": 1541,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1540 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790538248, "dur": 11, + "args": { + "External id": 1542,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1541 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538278, "dur": 5, + "args": { + "External id": 1543,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1542 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790538287, "dur": 21, + "args": { + "External id": 1544,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1543 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538288, "dur": 5, + "args": { + "External id": 1545,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1544 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790538294, "dur": 14, + "args": { + "External id": 1546,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1545 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790538296, "dur": 10, + "args": { + "External id": 1547,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1546 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538326, "dur": 5, + "args": { + "External id": 1548,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1547 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790538334, "dur": 21, + "args": { + "External id": 1549,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1548 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538336, "dur": 4, + "args": { + "External id": 1550,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1549 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790538341, "dur": 14, + "args": { + "External id": 1551,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1550 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790538343, "dur": 11, + "args": { + "External id": 1552,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1551 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538373, "dur": 5, + "args": { + "External id": 1553,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1552 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790538381, "dur": 21, + "args": { + "External id": 1554,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1553 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538382, "dur": 5, + "args": { + "External id": 1555,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1554 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790538388, "dur": 14, + "args": { + "External id": 1556,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1555 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790538390, "dur": 11, + "args": { + "External id": 1557,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1556 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538420, "dur": 5, + "args": { + "External id": 1558,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1557 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790538429, "dur": 21, + "args": { + "External id": 1559,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1558 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538430, "dur": 4, + "args": { + "External id": 1560,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1559 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790538436, "dur": 14, + "args": { + "External id": 1561,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1560 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790538437, "dur": 12, + "args": { + "External id": 1562,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1561 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538468, "dur": 5, + "args": { + "External id": 1563,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1562 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790538477, "dur": 21, + "args": { + "External id": 1564,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1563 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538478, "dur": 4, + "args": { + "External id": 1565,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1564 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790538483, "dur": 14, + "args": { + "External id": 1566,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1565 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790538485, "dur": 11, + "args": { + "External id": 1567,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1566 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538515, "dur": 5, + "args": { + "External id": 1568,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1567 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790538524, "dur": 21, + "args": { + "External id": 1569,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1568 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538525, "dur": 4, + "args": { + "External id": 1570,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1569 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790538531, "dur": 14, + "args": { + "External id": 1571,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1570 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790538532, "dur": 12, + "args": { + "External id": 1572,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1571 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538563, "dur": 5, + "args": { + "External id": 1573,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1572 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790538571, "dur": 21, + "args": { + "External id": 1574,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1573 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538573, "dur": 4, + "args": { + "External id": 1575,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1574 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790538578, "dur": 14, + "args": { + "External id": 1576,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1575 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790538580, "dur": 11, + "args": { + "External id": 1577,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1576 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538610, "dur": 5, + "args": { + "External id": 1578,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1577 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790538619, "dur": 20, + "args": { + "External id": 1579,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1578 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538620, "dur": 4, + "args": { + "External id": 1580,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1579 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790538626, "dur": 13, + "args": { + "External id": 1581,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1580 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790538627, "dur": 11, + "args": { + "External id": 1582,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1581 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538657, "dur": 5, + "args": { + "External id": 1583,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1582 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790538666, "dur": 21, + "args": { + "External id": 1584,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1583 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538667, "dur": 5, + "args": { + "External id": 1585,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1584 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790538673, "dur": 14, + "args": { + "External id": 1586,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1585 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790538675, "dur": 10, + "args": { + "External id": 1587,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1586 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538704, "dur": 5, + "args": { + "External id": 1588,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1587 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790538713, "dur": 21, + "args": { + "External id": 1589,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1588 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538714, "dur": 4, + "args": { + "External id": 1590,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1589 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790538720, "dur": 13, + "args": { + "External id": 1591,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1590 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790538721, "dur": 11, + "args": { + "External id": 1592,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1591 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538751, "dur": 5, + "args": { + "External id": 1593,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1592 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790538760, "dur": 21, + "args": { + "External id": 1594,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1593 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538761, "dur": 4, + "args": { + "External id": 1595,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1594 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790538767, "dur": 13, + "args": { + "External id": 1596,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1595 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790538768, "dur": 11, + "args": { + "External id": 1597,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1596 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538798, "dur": 5, + "args": { + "External id": 1598,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1597 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790538807, "dur": 21, + "args": { + "External id": 1599,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1598 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538808, "dur": 4, + "args": { + "External id": 1600,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1599 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790538814, "dur": 13, + "args": { + "External id": 1601,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1600 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790538815, "dur": 11, + "args": { + "External id": 1602,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1601 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538845, "dur": 5, + "args": { + "External id": 1603,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1602 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790538854, "dur": 21, + "args": { + "External id": 1604,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1603 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538855, "dur": 4, + "args": { + "External id": 1605,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1604 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790538860, "dur": 14, + "args": { + "External id": 1606,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1605 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790538862, "dur": 11, + "args": { + "External id": 1607,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1606 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538892, "dur": 5, + "args": { + "External id": 1608,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1607 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790538901, "dur": 21, + "args": { + "External id": 1609,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1608 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538902, "dur": 5, + "args": { + "External id": 1610,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1609 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790538908, "dur": 13, + "args": { + "External id": 1611,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1610 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790538910, "dur": 10, + "args": { + "External id": 1612,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1611 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538939, "dur": 5, + "args": { + "External id": 1613,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1612 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790538948, "dur": 21, + "args": { + "External id": 1614,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1613 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538949, "dur": 4, + "args": { + "External id": 1615,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1614 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790538955, "dur": 13, + "args": { + "External id": 1616,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1615 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790538956, "dur": 11, + "args": { + "External id": 1617,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1616 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538986, "dur": 5, + "args": { + "External id": 1618,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1617 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790538995, "dur": 20, + "args": { + "External id": 1619,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1618 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790538996, "dur": 4, + "args": { + "External id": 1620,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1619 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790539002, "dur": 13, + "args": { + "External id": 1621,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1620 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790539003, "dur": 11, + "args": { + "External id": 1622,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1621 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539033, "dur": 5, + "args": { + "External id": 1623,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1622 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790539041, "dur": 21, + "args": { + "External id": 1624,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1623 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539043, "dur": 4, + "args": { + "External id": 1625,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1624 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790539048, "dur": 14, + "args": { + "External id": 1626,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1625 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790539050, "dur": 11, + "args": { + "External id": 1627,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1626 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539080, "dur": 5, + "args": { + "External id": 1628,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1627 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790539090, "dur": 21, + "args": { + "External id": 1629,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1628 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539091, "dur": 5, + "args": { + "External id": 1630,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1629 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790539097, "dur": 13, + "args": { + "External id": 1631,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1630 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790539098, "dur": 11, + "args": { + "External id": 1632,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1631 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539128, "dur": 5, + "args": { + "External id": 1633,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1632 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790539137, "dur": 21, + "args": { + "External id": 1634,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1633 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539138, "dur": 5, + "args": { + "External id": 1635,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1634 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790539144, "dur": 14, + "args": { + "External id": 1636,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1635 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790539146, "dur": 10, + "args": { + "External id": 1637,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1636 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539175, "dur": 5, + "args": { + "External id": 1638,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1637 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790539184, "dur": 21, + "args": { + "External id": 1639,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1638 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539185, "dur": 4, + "args": { + "External id": 1640,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1639 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790539191, "dur": 13, + "args": { + "External id": 1641,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1640 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790539192, "dur": 11, + "args": { + "External id": 1642,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1641 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539222, "dur": 5, + "args": { + "External id": 1643,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1642 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790539231, "dur": 21, + "args": { + "External id": 1644,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1643 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539232, "dur": 4, + "args": { + "External id": 1645,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1644 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790539237, "dur": 14, + "args": { + "External id": 1646,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1645 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790539239, "dur": 11, + "args": { + "External id": 1647,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1646 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539269, "dur": 5, + "args": { + "External id": 1648,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1647 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790539278, "dur": 21, + "args": { + "External id": 1649,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1648 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539279, "dur": 5, + "args": { + "External id": 1650,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1649 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790539285, "dur": 13, + "args": { + "External id": 1651,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1650 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790539286, "dur": 11, + "args": { + "External id": 1652,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1651 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539316, "dur": 5, + "args": { + "External id": 1653,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1652 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790539325, "dur": 21, + "args": { + "External id": 1654,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1653 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539326, "dur": 5, + "args": { + "External id": 1655,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1654 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790539332, "dur": 14, + "args": { + "External id": 1656,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1655 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790539333, "dur": 11, + "args": { + "External id": 1657,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1656 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539364, "dur": 5, + "args": { + "External id": 1658,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1657 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790539372, "dur": 21, + "args": { + "External id": 1659,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1658 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539374, "dur": 4, + "args": { + "External id": 1660,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1659 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790539379, "dur": 14, + "args": { + "External id": 1661,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1660 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790539381, "dur": 11, + "args": { + "External id": 1662,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1661 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539410, "dur": 5, + "args": { + "External id": 1663,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1662 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790539419, "dur": 21, + "args": { + "External id": 1664,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1663 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539420, "dur": 5, + "args": { + "External id": 1665,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1664 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790539426, "dur": 14, + "args": { + "External id": 1666,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1665 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790539428, "dur": 11, + "args": { + "External id": 1667,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1666 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539458, "dur": 5, + "args": { + "External id": 1668,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1667 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790539467, "dur": 20, + "args": { + "External id": 1669,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1668 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539468, "dur": 4, + "args": { + "External id": 1670,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1669 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790539473, "dur": 14, + "args": { + "External id": 1671,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1670 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790539475, "dur": 11, + "args": { + "External id": 1672,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1671 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539505, "dur": 5, + "args": { + "External id": 1673,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1672 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790539514, "dur": 20, + "args": { + "External id": 1674,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1673 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539515, "dur": 4, + "args": { + "External id": 1675,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1674 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790539520, "dur": 14, + "args": { + "External id": 1676,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1675 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790539522, "dur": 11, + "args": { + "External id": 1677,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1676 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539552, "dur": 5, + "args": { + "External id": 1678,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1677 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790539561, "dur": 21, + "args": { + "External id": 1679,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1678 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539562, "dur": 5, + "args": { + "External id": 1680,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1679 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790539568, "dur": 14, + "args": { + "External id": 1681,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1680 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790539570, "dur": 11, + "args": { + "External id": 1682,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1681 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539600, "dur": 5, + "args": { + "External id": 1683,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1682 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790539609, "dur": 21, + "args": { + "External id": 1684,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1683 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539610, "dur": 4, + "args": { + "External id": 1685,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1684 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790539616, "dur": 13, + "args": { + "External id": 1686,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1685 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790539617, "dur": 11, + "args": { + "External id": 1687,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1686 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539647, "dur": 5, + "args": { + "External id": 1688,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1687 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790539656, "dur": 21, + "args": { + "External id": 1689,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1688 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539657, "dur": 5, + "args": { + "External id": 1690,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1689 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790539663, "dur": 14, + "args": { + "External id": 1691,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1690 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790539665, "dur": 10, + "args": { + "External id": 1692,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1691 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539694, "dur": 5, + "args": { + "External id": 1693,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1692 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790539703, "dur": 21, + "args": { + "External id": 1694,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1693 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539705, "dur": 4, + "args": { + "External id": 1695,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1694 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790539710, "dur": 14, + "args": { + "External id": 1696,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1695 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790539712, "dur": 11, + "args": { + "External id": 1697,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1696 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539742, "dur": 5, + "args": { + "External id": 1698,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1697 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790539751, "dur": 20, + "args": { + "External id": 1699,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1698 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539752, "dur": 4, + "args": { + "External id": 1700,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1699 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790539757, "dur": 14, + "args": { + "External id": 1701,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1700 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790539759, "dur": 11, + "args": { + "External id": 1702,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1701 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539789, "dur": 5, + "args": { + "External id": 1703,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1702 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790539798, "dur": 21, + "args": { + "External id": 1704,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1703 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539799, "dur": 4, + "args": { + "External id": 1705,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1704 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790539804, "dur": 14, + "args": { + "External id": 1706,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1705 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790539806, "dur": 11, + "args": { + "External id": 1707,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1706 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539836, "dur": 5, + "args": { + "External id": 1708,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1707 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790539845, "dur": 21, + "args": { + "External id": 1709,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1708 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539846, "dur": 4, + "args": { + "External id": 1710,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1709 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790539851, "dur": 14, + "args": { + "External id": 1711,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1710 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790539853, "dur": 11, + "args": { + "External id": 1712,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1711 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539883, "dur": 4, + "args": { + "External id": 1713,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1712 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790539891, "dur": 22, + "args": { + "External id": 1714,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1713 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539892, "dur": 5, + "args": { + "External id": 1715,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1714 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790539898, "dur": 14, + "args": { + "External id": 1716,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1715 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790539900, "dur": 11, + "args": { + "External id": 1717,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1716 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539930, "dur": 5, + "args": { + "External id": 1718,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1717 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790539939, "dur": 21, + "args": { + "External id": 1719,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1718 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539940, "dur": 5, + "args": { + "External id": 1720,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1719 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790539946, "dur": 14, + "args": { + "External id": 1721,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1720 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790539947, "dur": 11, + "args": { + "External id": 1722,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1721 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539977, "dur": 5, + "args": { + "External id": 1723,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1722 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790539986, "dur": 21, + "args": { + "External id": 1724,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1723 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790539987, "dur": 5, + "args": { + "External id": 1725,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1724 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790539993, "dur": 14, + "args": { + "External id": 1726,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1725 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790539995, "dur": 11, + "args": { + "External id": 1727,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1726 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540025, "dur": 5, + "args": { + "External id": 1728,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1727 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790540034, "dur": 21, + "args": { + "External id": 1729,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1728 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540035, "dur": 5, + "args": { + "External id": 1730,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1729 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790540041, "dur": 14, + "args": { + "External id": 1731,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1730 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790540043, "dur": 11, + "args": { + "External id": 1732,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1731 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540073, "dur": 5, + "args": { + "External id": 1733,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1732 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790540082, "dur": 21, + "args": { + "External id": 1734,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1733 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540083, "dur": 4, + "args": { + "External id": 1735,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1734 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790540089, "dur": 14, + "args": { + "External id": 1736,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1735 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790540090, "dur": 11, + "args": { + "External id": 1737,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1736 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540121, "dur": 5, + "args": { + "External id": 1738,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1737 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790540130, "dur": 21, + "args": { + "External id": 1739,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1738 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540131, "dur": 4, + "args": { + "External id": 1740,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1739 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790540137, "dur": 14, + "args": { + "External id": 1741,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1740 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790540138, "dur": 11, + "args": { + "External id": 1742,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1741 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540169, "dur": 5, + "args": { + "External id": 1743,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1742 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790540178, "dur": 21, + "args": { + "External id": 1744,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1743 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540179, "dur": 4, + "args": { + "External id": 1745,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1744 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790540184, "dur": 14, + "args": { + "External id": 1746,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1745 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790540186, "dur": 11, + "args": { + "External id": 1747,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1746 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540236, "dur": 6, + "args": { + "External id": 1748,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1747 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790540247, "dur": 23, + "args": { + "External id": 1749,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1748 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540248, "dur": 5, + "args": { + "External id": 1750,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1749 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790540254, "dur": 16, + "args": { + "External id": 1751,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1750 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790540256, "dur": 13, + "args": { + "External id": 1752,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1751 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540290, "dur": 5, + "args": { + "External id": 1753,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1752 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790540299, "dur": 21, + "args": { + "External id": 1754,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1753 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540300, "dur": 5, + "args": { + "External id": 1755,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1754 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790540306, "dur": 14, + "args": { + "External id": 1756,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1755 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790540307, "dur": 12, + "args": { + "External id": 1757,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1756 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540338, "dur": 6, + "args": { + "External id": 1758,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1757 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790540347, "dur": 22, + "args": { + "External id": 1759,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1758 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540349, "dur": 4, + "args": { + "External id": 1760,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1759 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790540354, "dur": 14, + "args": { + "External id": 1761,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1760 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790540356, "dur": 11, + "args": { + "External id": 1762,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1761 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540386, "dur": 5, + "args": { + "External id": 1763,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1762 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790540395, "dur": 21, + "args": { + "External id": 1764,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1763 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540397, "dur": 4, + "args": { + "External id": 1765,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1764 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790540402, "dur": 14, + "args": { + "External id": 1766,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1765 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790540404, "dur": 11, + "args": { + "External id": 1767,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1766 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540434, "dur": 5, + "args": { + "External id": 1768,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1767 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790540443, "dur": 22, + "args": { + "External id": 1769,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1768 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540445, "dur": 4, + "args": { + "External id": 1770,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1769 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790540450, "dur": 14, + "args": { + "External id": 1771,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1770 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790540452, "dur": 11, + "args": { + "External id": 1772,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1771 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540482, "dur": 5, + "args": { + "External id": 1773,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1772 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790540491, "dur": 21, + "args": { + "External id": 1774,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1773 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540492, "dur": 5, + "args": { + "External id": 1775,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1774 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790540498, "dur": 13, + "args": { + "External id": 1776,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1775 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790540499, "dur": 11, + "args": { + "External id": 1777,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1776 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540590, "dur": 7, + "args": { + "External id": 1778,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1777 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790540602, "dur": 25, + "args": { + "External id": 1779,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1778 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540603, "dur": 5, + "args": { + "External id": 1780,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1779 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790540609, "dur": 17, + "args": { + "External id": 1781,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1780 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790540611, "dur": 14, + "args": { + "External id": 1782,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1781 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540648, "dur": 5, + "args": { + "External id": 1783,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1782 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790540657, "dur": 22, + "args": { + "External id": 1784,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1783 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540659, "dur": 4, + "args": { + "External id": 1785,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1784 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790540664, "dur": 15, + "args": { + "External id": 1786,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1785 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790540666, "dur": 11, + "args": { + "External id": 1787,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1786 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540698, "dur": 5, + "args": { + "External id": 1788,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1787 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790540707, "dur": 21, + "args": { + "External id": 1789,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1788 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540708, "dur": 5, + "args": { + "External id": 1790,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1789 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790540714, "dur": 14, + "args": { + "External id": 1791,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1790 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790540716, "dur": 11, + "args": { + "External id": 1792,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1791 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540747, "dur": 5, + "args": { + "External id": 1793,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1792 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790540756, "dur": 22, + "args": { + "External id": 1794,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1793 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540758, "dur": 4, + "args": { + "External id": 1795,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1794 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790540763, "dur": 15, + "args": { + "External id": 1796,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1795 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790540765, "dur": 11, + "args": { + "External id": 1797,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1796 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540796, "dur": 5, + "args": { + "External id": 1798,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1797 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790540805, "dur": 21, + "args": { + "External id": 1799,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1798 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540806, "dur": 5, + "args": { + "External id": 1800,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1799 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790540812, "dur": 14, + "args": { + "External id": 1801,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1800 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790540813, "dur": 11, + "args": { + "External id": 1802,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1801 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540844, "dur": 5, + "args": { + "External id": 1803,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1802 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790540853, "dur": 21, + "args": { + "External id": 1804,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1803 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540854, "dur": 5, + "args": { + "External id": 1805,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1804 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790540860, "dur": 14, + "args": { + "External id": 1806,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1805 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790540861, "dur": 12, + "args": { + "External id": 1807,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1806 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540892, "dur": 5, + "args": { + "External id": 1808,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1807 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790540901, "dur": 21, + "args": { + "External id": 1809,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1808 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540902, "dur": 5, + "args": { + "External id": 1810,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1809 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790540908, "dur": 14, + "args": { + "External id": 1811,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1810 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790540909, "dur": 12, + "args": { + "External id": 1812,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1811 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540940, "dur": 6, + "args": { + "External id": 1813,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1812 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790540950, "dur": 21, + "args": { + "External id": 1814,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1813 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540951, "dur": 4, + "args": { + "External id": 1815,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1814 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790540956, "dur": 14, + "args": { + "External id": 1816,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1815 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790540958, "dur": 11, + "args": { + "External id": 1817,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1816 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790540991, "dur": 5, + "args": { + "External id": 1818,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1817 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790541000, "dur": 22, + "args": { + "External id": 1819,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1818 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541001, "dur": 5, + "args": { + "External id": 1820,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1819 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790541007, "dur": 14, + "args": { + "External id": 1821,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1820 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790541009, "dur": 11, + "args": { + "External id": 1822,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1821 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541040, "dur": 5, + "args": { + "External id": 1823,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1822 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790541049, "dur": 21, + "args": { + "External id": 1824,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1823 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541050, "dur": 4, + "args": { + "External id": 1825,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1824 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790541056, "dur": 14, + "args": { + "External id": 1826,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1825 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790541057, "dur": 11, + "args": { + "External id": 1827,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1826 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541138, "dur": 6, + "args": { + "External id": 1828,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1827 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790541149, "dur": 24, + "args": { + "External id": 1829,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1828 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541150, "dur": 5, + "args": { + "External id": 1830,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1829 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790541156, "dur": 17, + "args": { + "External id": 1831,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1830 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790541158, "dur": 14, + "args": { + "External id": 1832,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1831 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541194, "dur": 5, + "args": { + "External id": 1833,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1832 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790541203, "dur": 22, + "args": { + "External id": 1834,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1833 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541204, "dur": 5, + "args": { + "External id": 1835,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1834 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790541210, "dur": 15, + "args": { + "External id": 1836,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1835 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790541212, "dur": 11, + "args": { + "External id": 1837,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1836 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541243, "dur": 5, + "args": { + "External id": 1838,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1837 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790541252, "dur": 22, + "args": { + "External id": 1839,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1838 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541253, "dur": 5, + "args": { + "External id": 1840,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1839 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790541259, "dur": 15, + "args": { + "External id": 1841,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1840 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790541261, "dur": 11, + "args": { + "External id": 1842,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1841 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541292, "dur": 6, + "args": { + "External id": 1843,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1842 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790541302, "dur": 21, + "args": { + "External id": 1844,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1843 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541303, "dur": 4, + "args": { + "External id": 1845,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1844 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790541308, "dur": 15, + "args": { + "External id": 1846,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1845 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790541310, "dur": 11, + "args": { + "External id": 1847,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1846 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541341, "dur": 5, + "args": { + "External id": 1848,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1847 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790541350, "dur": 21, + "args": { + "External id": 1849,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1848 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541351, "dur": 4, + "args": { + "External id": 1850,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1849 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790541357, "dur": 14, + "args": { + "External id": 1851,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1850 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790541358, "dur": 11, + "args": { + "External id": 1852,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1851 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541389, "dur": 5, + "args": { + "External id": 1853,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1852 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790541398, "dur": 21, + "args": { + "External id": 1854,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1853 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541399, "dur": 5, + "args": { + "External id": 1855,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1854 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790541405, "dur": 14, + "args": { + "External id": 1856,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1855 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790541407, "dur": 11, + "args": { + "External id": 1857,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1856 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541438, "dur": 5, + "args": { + "External id": 1858,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1857 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790541447, "dur": 21, + "args": { + "External id": 1859,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1858 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541448, "dur": 5, + "args": { + "External id": 1860,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1859 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790541454, "dur": 14, + "args": { + "External id": 1861,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1860 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790541455, "dur": 12, + "args": { + "External id": 1862,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1861 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541486, "dur": 5, + "args": { + "External id": 1863,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1862 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790541495, "dur": 21, + "args": { + "External id": 1864,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1863 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541496, "dur": 5, + "args": { + "External id": 1865,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1864 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790541502, "dur": 14, + "args": { + "External id": 1866,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1865 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790541503, "dur": 12, + "args": { + "External id": 1867,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1866 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541535, "dur": 5, + "args": { + "External id": 1868,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1867 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790541544, "dur": 21, + "args": { + "External id": 1869,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1868 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541545, "dur": 4, + "args": { + "External id": 1870,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1869 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790541550, "dur": 14, + "args": { + "External id": 1871,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1870 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790541552, "dur": 11, + "args": { + "External id": 1872,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1871 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541583, "dur": 5, + "args": { + "External id": 1873,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1872 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790541592, "dur": 21, + "args": { + "External id": 1874,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1873 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541593, "dur": 4, + "args": { + "External id": 1875,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1874 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790541598, "dur": 15, + "args": { + "External id": 1876,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1875 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790541600, "dur": 12, + "args": { + "External id": 1877,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1876 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541631, "dur": 5, + "args": { + "External id": 1878,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1877 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790541640, "dur": 21, + "args": { + "External id": 1879,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1878 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541641, "dur": 4, + "args": { + "External id": 1880,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1879 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790541646, "dur": 14, + "args": { + "External id": 1881,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1880 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790541648, "dur": 11, + "args": { + "External id": 1882,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1881 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541679, "dur": 5, + "args": { + "External id": 1883,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1882 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790541688, "dur": 22, + "args": { + "External id": 1884,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1883 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541690, "dur": 4, + "args": { + "External id": 1885,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1884 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790541695, "dur": 14, + "args": { + "External id": 1886,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1885 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790541697, "dur": 11, + "args": { + "External id": 1887,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1886 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541728, "dur": 5, + "args": { + "External id": 1888,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1887 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790541737, "dur": 21, + "args": { + "External id": 1889,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1888 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541738, "dur": 4, + "args": { + "External id": 1890,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1889 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790541743, "dur": 15, + "args": { + "External id": 1891,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1890 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790541745, "dur": 11, + "args": { + "External id": 1892,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1891 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541776, "dur": 5, + "args": { + "External id": 1893,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1892 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790541784, "dur": 22, + "args": { + "External id": 1894,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1893 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541786, "dur": 4, + "args": { + "External id": 1895,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1894 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790541791, "dur": 14, + "args": { + "External id": 1896,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1895 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790541793, "dur": 11, + "args": { + "External id": 1897,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1896 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541824, "dur": 5, + "args": { + "External id": 1898,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1897 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790541833, "dur": 21, + "args": { + "External id": 1899,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1898 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541834, "dur": 4, + "args": { + "External id": 1900,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1899 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790541839, "dur": 14, + "args": { + "External id": 1901,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1900 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790541841, "dur": 11, + "args": { + "External id": 1902,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1901 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541872, "dur": 5, + "args": { + "External id": 1903,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1902 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790541880, "dur": 22, + "args": { + "External id": 1904,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1903 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541882, "dur": 4, + "args": { + "External id": 1905,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1904 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790541887, "dur": 14, + "args": { + "External id": 1906,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1905 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790541889, "dur": 11, + "args": { + "External id": 1907,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1906 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541919, "dur": 5, + "args": { + "External id": 1908,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1907 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790541928, "dur": 21, + "args": { + "External id": 1909,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1908 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541929, "dur": 5, + "args": { + "External id": 1910,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1909 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790541935, "dur": 14, + "args": { + "External id": 1911,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1910 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790541937, "dur": 11, + "args": { + "External id": 1912,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1911 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541967, "dur": 5, + "args": { + "External id": 1913,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1912 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790541976, "dur": 21, + "args": { + "External id": 1914,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1913 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790541977, "dur": 5, + "args": { + "External id": 1915,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1914 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790541983, "dur": 14, + "args": { + "External id": 1916,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1915 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790541984, "dur": 11, + "args": { + "External id": 1917,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1916 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542015, "dur": 5, + "args": { + "External id": 1918,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1917 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790542024, "dur": 21, + "args": { + "External id": 1919,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1918 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542025, "dur": 4, + "args": { + "External id": 1920,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1919 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790542030, "dur": 15, + "args": { + "External id": 1921,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1920 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790542032, "dur": 11, + "args": { + "External id": 1922,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1921 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542063, "dur": 5, + "args": { + "External id": 1923,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1922 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790542072, "dur": 20, + "args": { + "External id": 1924,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1923 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542073, "dur": 4, + "args": { + "External id": 1925,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1924 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790542078, "dur": 14, + "args": { + "External id": 1926,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1925 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790542080, "dur": 11, + "args": { + "External id": 1927,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1926 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542110, "dur": 5, + "args": { + "External id": 1928,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1927 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790542119, "dur": 21, + "args": { + "External id": 1929,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1928 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542120, "dur": 5, + "args": { + "External id": 1930,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1929 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790542126, "dur": 14, + "args": { + "External id": 1931,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1930 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790542128, "dur": 11, + "args": { + "External id": 1932,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1931 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542158, "dur": 5, + "args": { + "External id": 1933,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1932 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790542167, "dur": 21, + "args": { + "External id": 1934,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1933 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542168, "dur": 4, + "args": { + "External id": 1935,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1934 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790542174, "dur": 14, + "args": { + "External id": 1936,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1935 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790542175, "dur": 11, + "args": { + "External id": 1937,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1936 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542206, "dur": 5, + "args": { + "External id": 1938,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1937 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790542215, "dur": 20, + "args": { + "External id": 1939,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1938 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542216, "dur": 4, + "args": { + "External id": 1940,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1939 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790542221, "dur": 14, + "args": { + "External id": 1941,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1940 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790542223, "dur": 11, + "args": { + "External id": 1942,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1941 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542253, "dur": 5, + "args": { + "External id": 1943,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1942 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790542262, "dur": 21, + "args": { + "External id": 1944,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1943 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542263, "dur": 5, + "args": { + "External id": 1945,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1944 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790542269, "dur": 14, + "args": { + "External id": 1946,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1945 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790542270, "dur": 11, + "args": { + "External id": 1947,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1946 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542301, "dur": 5, + "args": { + "External id": 1948,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1947 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790542310, "dur": 21, + "args": { + "External id": 1949,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1948 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542311, "dur": 4, + "args": { + "External id": 1950,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1949 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790542316, "dur": 15, + "args": { + "External id": 1951,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1950 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790542318, "dur": 11, + "args": { + "External id": 1952,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1951 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542348, "dur": 5, + "args": { + "External id": 1953,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1952 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790542357, "dur": 21, + "args": { + "External id": 1954,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1953 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542358, "dur": 5, + "args": { + "External id": 1955,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1954 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790542364, "dur": 14, + "args": { + "External id": 1956,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1955 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790542366, "dur": 11, + "args": { + "External id": 1957,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1956 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542396, "dur": 5, + "args": { + "External id": 1958,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1957 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790542405, "dur": 21, + "args": { + "External id": 1959,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1958 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542406, "dur": 4, + "args": { + "External id": 1960,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1959 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790542411, "dur": 14, + "args": { + "External id": 1961,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1960 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790542413, "dur": 11, + "args": { + "External id": 1962,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1961 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542443, "dur": 5, + "args": { + "External id": 1963,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1962 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790542452, "dur": 21, + "args": { + "External id": 1964,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1963 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542453, "dur": 5, + "args": { + "External id": 1965,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1964 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790542459, "dur": 14, + "args": { + "External id": 1966,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1965 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790542461, "dur": 10, + "args": { + "External id": 1967,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1966 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542491, "dur": 5, + "args": { + "External id": 1968,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1967 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790542500, "dur": 21, + "args": { + "External id": 1969,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1968 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542501, "dur": 4, + "args": { + "External id": 1970,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1969 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790542506, "dur": 15, + "args": { + "External id": 1971,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1970 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790542508, "dur": 11, + "args": { + "External id": 1972,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1971 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542539, "dur": 5, + "args": { + "External id": 1973,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1972 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790542547, "dur": 22, + "args": { + "External id": 1974,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1973 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542549, "dur": 4, + "args": { + "External id": 1975,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1974 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790542554, "dur": 14, + "args": { + "External id": 1976,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1975 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790542556, "dur": 11, + "args": { + "External id": 1977,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1976 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542586, "dur": 6, + "args": { + "External id": 1978,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1977 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790542595, "dur": 21, + "args": { + "External id": 1979,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1978 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542597, "dur": 4, + "args": { + "External id": 1980,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1979 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790542602, "dur": 14, + "args": { + "External id": 1981,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1980 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790542604, "dur": 11, + "args": { + "External id": 1982,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1981 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542634, "dur": 5, + "args": { + "External id": 1983,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1982 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790542643, "dur": 21, + "args": { + "External id": 1984,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1983 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542644, "dur": 4, + "args": { + "External id": 1985,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1984 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790542650, "dur": 13, + "args": { + "External id": 1986,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1985 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790542651, "dur": 11, + "args": { + "External id": 1987,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1986 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542681, "dur": 5, + "args": { + "External id": 1988,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1987 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790542690, "dur": 21, + "args": { + "External id": 1989,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1988 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542691, "dur": 5, + "args": { + "External id": 1990,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1989 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790542697, "dur": 14, + "args": { + "External id": 1991,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1990 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790542699, "dur": 10, + "args": { + "External id": 1992,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1991 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542729, "dur": 5, + "args": { + "External id": 1993,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1992 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790542738, "dur": 21, + "args": { + "External id": 1994,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1993 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542739, "dur": 4, + "args": { + "External id": 1995,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1994 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790542745, "dur": 14, + "args": { + "External id": 1996,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 1995 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790542746, "dur": 11, + "args": { + "External id": 1997,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 1996 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542777, "dur": 5, + "args": { + "External id": 1998,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1997 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790542786, "dur": 21, + "args": { + "External id": 1999,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 1998 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542787, "dur": 4, + "args": { + "External id": 2000,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1999 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790542793, "dur": 13, + "args": { + "External id": 2001,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2000 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790542794, "dur": 11, + "args": { + "External id": 2002,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2001 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542824, "dur": 5, + "args": { + "External id": 2003,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2002 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790542833, "dur": 21, + "args": { + "External id": 2004,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2003 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542834, "dur": 5, + "args": { + "External id": 2005,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2004 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790542840, "dur": 14, + "args": { + "External id": 2006,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2005 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790542842, "dur": 11, + "args": { + "External id": 2007,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2006 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542872, "dur": 5, + "args": { + "External id": 2008,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2007 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790542881, "dur": 21, + "args": { + "External id": 2009,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2008 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542882, "dur": 5, + "args": { + "External id": 2010,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2009 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790542888, "dur": 14, + "args": { + "External id": 2011,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2010 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790542890, "dur": 11, + "args": { + "External id": 2012,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2011 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542920, "dur": 5, + "args": { + "External id": 2013,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2012 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790542929, "dur": 21, + "args": { + "External id": 2014,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2013 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542930, "dur": 5, + "args": { + "External id": 2015,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2014 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790542936, "dur": 14, + "args": { + "External id": 2016,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2015 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790542937, "dur": 12, + "args": { + "External id": 2017,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2016 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542968, "dur": 5, + "args": { + "External id": 2018,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2017 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790542977, "dur": 21, + "args": { + "External id": 2019,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2018 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790542978, "dur": 5, + "args": { + "External id": 2020,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2019 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790542984, "dur": 14, + "args": { + "External id": 2021,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2020 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790542986, "dur": 11, + "args": { + "External id": 2022,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2021 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543016, "dur": 5, + "args": { + "External id": 2023,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2022 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790543025, "dur": 21, + "args": { + "External id": 2024,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2023 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543026, "dur": 5, + "args": { + "External id": 2025,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2024 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790543032, "dur": 14, + "args": { + "External id": 2026,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2025 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790543033, "dur": 12, + "args": { + "External id": 2027,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2026 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543064, "dur": 5, + "args": { + "External id": 2028,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2027 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790543073, "dur": 21, + "args": { + "External id": 2029,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2028 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543074, "dur": 4, + "args": { + "External id": 2030,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2029 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790543080, "dur": 14, + "args": { + "External id": 2031,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2030 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790543081, "dur": 11, + "args": { + "External id": 2032,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2031 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543112, "dur": 5, + "args": { + "External id": 2033,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2032 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790543121, "dur": 21, + "args": { + "External id": 2034,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2033 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543122, "dur": 4, + "args": { + "External id": 2035,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2034 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790543128, "dur": 14, + "args": { + "External id": 2036,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2035 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790543129, "dur": 11, + "args": { + "External id": 2037,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2036 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543160, "dur": 6, + "args": { + "External id": 2038,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2037 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790543170, "dur": 20, + "args": { + "External id": 2039,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2038 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543171, "dur": 4, + "args": { + "External id": 2040,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2039 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790543176, "dur": 14, + "args": { + "External id": 2041,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2040 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790543178, "dur": 11, + "args": { + "External id": 2042,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2041 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543208, "dur": 5, + "args": { + "External id": 2043,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2042 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790543217, "dur": 21, + "args": { + "External id": 2044,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2043 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543218, "dur": 5, + "args": { + "External id": 2045,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2044 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790543224, "dur": 14, + "args": { + "External id": 2046,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2045 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790543225, "dur": 11, + "args": { + "External id": 2047,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2046 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543255, "dur": 6, + "args": { + "External id": 2048,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2047 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790543273, "dur": 22, + "args": { + "External id": 2049,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2048 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543275, "dur": 4, + "args": { + "External id": 2050,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2049 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790543280, "dur": 14, + "args": { + "External id": 2051,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2050 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790543282, "dur": 11, + "args": { + "External id": 2052,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2051 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543313, "dur": 5, + "args": { + "External id": 2053,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2052 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790543322, "dur": 21, + "args": { + "External id": 2054,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2053 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543323, "dur": 4, + "args": { + "External id": 2055,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2054 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790543329, "dur": 13, + "args": { + "External id": 2056,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2055 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790543330, "dur": 11, + "args": { + "External id": 2057,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2056 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543360, "dur": 5, + "args": { + "External id": 2058,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2057 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790543369, "dur": 21, + "args": { + "External id": 2059,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2058 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543370, "dur": 5, + "args": { + "External id": 2060,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2059 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790543376, "dur": 14, + "args": { + "External id": 2061,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2060 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790543378, "dur": 11, + "args": { + "External id": 2062,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2061 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543408, "dur": 5, + "args": { + "External id": 2063,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2062 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790543417, "dur": 21, + "args": { + "External id": 2064,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2063 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543418, "dur": 5, + "args": { + "External id": 2065,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2064 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790543424, "dur": 14, + "args": { + "External id": 2066,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2065 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790543426, "dur": 11, + "args": { + "External id": 2067,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2066 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543456, "dur": 5, + "args": { + "External id": 2068,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2067 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790543465, "dur": 21, + "args": { + "External id": 2069,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2068 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543466, "dur": 4, + "args": { + "External id": 2070,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2069 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790543471, "dur": 14, + "args": { + "External id": 2071,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2070 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790543473, "dur": 11, + "args": { + "External id": 2072,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2071 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543504, "dur": 5, + "args": { + "External id": 2073,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2072 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790543513, "dur": 21, + "args": { + "External id": 2074,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2073 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543514, "dur": 4, + "args": { + "External id": 2075,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2074 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790543519, "dur": 14, + "args": { + "External id": 2076,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2075 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790543521, "dur": 11, + "args": { + "External id": 2077,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2076 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543612, "dur": 6, + "args": { + "External id": 2078,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2077 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790543623, "dur": 25, + "args": { + "External id": 2079,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2078 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543624, "dur": 5, + "args": { + "External id": 2080,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2079 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790543630, "dur": 18, + "args": { + "External id": 2081,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2080 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790543632, "dur": 14, + "args": { + "External id": 2082,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2081 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543669, "dur": 5, + "args": { + "External id": 2083,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2082 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790543678, "dur": 22, + "args": { + "External id": 2084,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2083 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543680, "dur": 4, + "args": { + "External id": 2085,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2084 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790543685, "dur": 15, + "args": { + "External id": 2086,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2085 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790543687, "dur": 12, + "args": { + "External id": 2087,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2086 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543719, "dur": 5, + "args": { + "External id": 2088,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2087 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790543728, "dur": 22, + "args": { + "External id": 2089,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2088 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543730, "dur": 4, + "args": { + "External id": 2090,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2089 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790543735, "dur": 15, + "args": { + "External id": 2091,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2090 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790543737, "dur": 11, + "args": { + "External id": 2092,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2091 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543768, "dur": 5, + "args": { + "External id": 2093,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2092 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790543777, "dur": 22, + "args": { + "External id": 2094,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2093 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543779, "dur": 4, + "args": { + "External id": 2095,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2094 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790543784, "dur": 15, + "args": { + "External id": 2096,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2095 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790543786, "dur": 11, + "args": { + "External id": 2097,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2096 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543817, "dur": 5, + "args": { + "External id": 2098,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2097 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790543826, "dur": 21, + "args": { + "External id": 2099,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2098 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543827, "dur": 4, + "args": { + "External id": 2100,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2099 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790543833, "dur": 14, + "args": { + "External id": 2101,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2100 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790543834, "dur": 12, + "args": { + "External id": 2102,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2101 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543865, "dur": 5, + "args": { + "External id": 2103,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2102 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790543874, "dur": 21, + "args": { + "External id": 2104,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2103 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543875, "dur": 5, + "args": { + "External id": 2105,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2104 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790543881, "dur": 14, + "args": { + "External id": 2106,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2105 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790543883, "dur": 11, + "args": { + "External id": 2107,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2106 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543913, "dur": 5, + "args": { + "External id": 2108,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2107 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790543922, "dur": 21, + "args": { + "External id": 2109,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2108 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543923, "dur": 5, + "args": { + "External id": 2110,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2109 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790543929, "dur": 14, + "args": { + "External id": 2111,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2110 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790543931, "dur": 11, + "args": { + "External id": 2112,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2111 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543961, "dur": 5, + "args": { + "External id": 2113,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2112 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790543970, "dur": 21, + "args": { + "External id": 2114,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2113 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790543971, "dur": 5, + "args": { + "External id": 2115,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2114 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790543977, "dur": 14, + "args": { + "External id": 2116,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2115 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790543979, "dur": 11, + "args": { + "External id": 2117,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2116 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544010, "dur": 5, + "args": { + "External id": 2118,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2117 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790544019, "dur": 21, + "args": { + "External id": 2119,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2118 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544020, "dur": 4, + "args": { + "External id": 2120,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2119 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790544025, "dur": 14, + "args": { + "External id": 2121,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2120 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790544027, "dur": 11, + "args": { + "External id": 2122,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2121 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544057, "dur": 6, + "args": { + "External id": 2123,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2122 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790544067, "dur": 21, + "args": { + "External id": 2124,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2123 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544068, "dur": 4, + "args": { + "External id": 2125,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2124 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790544073, "dur": 14, + "args": { + "External id": 2126,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2125 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790544075, "dur": 11, + "args": { + "External id": 2127,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2126 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544155, "dur": 7, + "args": { + "External id": 2128,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2127 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790544167, "dur": 24, + "args": { + "External id": 2129,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2128 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544168, "dur": 4, + "args": { + "External id": 2130,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2129 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790544174, "dur": 17, + "args": { + "External id": 2131,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2130 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790544176, "dur": 13, + "args": { + "External id": 2132,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2131 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544212, "dur": 24, + "args": { + "External id": 2133,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2132 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790544242, "dur": 24, + "args": { + "External id": 2134,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2133 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544243, "dur": 5, + "args": { + "External id": 2135,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2134 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790544249, "dur": 17, + "args": { + "External id": 2136,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2135 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790544251, "dur": 13, + "args": { + "External id": 2137,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2136 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544286, "dur": 5, + "args": { + "External id": 2138,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2137 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790544295, "dur": 24, + "args": { + "External id": 2139,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2138 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544298, "dur": 5, + "args": { + "External id": 2140,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2139 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790544304, "dur": 14, + "args": { + "External id": 2141,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2140 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790544306, "dur": 11, + "args": { + "External id": 2142,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2141 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544337, "dur": 5, + "args": { + "External id": 2143,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2142 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790544346, "dur": 22, + "args": { + "External id": 2144,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2143 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544348, "dur": 4, + "args": { + "External id": 2145,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2144 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790544353, "dur": 14, + "args": { + "External id": 2146,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2145 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790544355, "dur": 11, + "args": { + "External id": 2147,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2146 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544388, "dur": 6, + "args": { + "External id": 2148,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2147 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790544398, "dur": 23, + "args": { + "External id": 2149,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2148 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544399, "dur": 6, + "args": { + "External id": 2150,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2149 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790544406, "dur": 15, + "args": { + "External id": 2151,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2150 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790544408, "dur": 12, + "args": { + "External id": 2152,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2151 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544439, "dur": 5, + "args": { + "External id": 2153,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2152 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790544448, "dur": 21, + "args": { + "External id": 2154,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2153 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544449, "dur": 5, + "args": { + "External id": 2155,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2154 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790544455, "dur": 14, + "args": { + "External id": 2156,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2155 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790544456, "dur": 12, + "args": { + "External id": 2157,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2156 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544487, "dur": 5, + "args": { + "External id": 2158,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2157 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790544496, "dur": 22, + "args": { + "External id": 2159,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2158 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544497, "dur": 5, + "args": { + "External id": 2160,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2159 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790544503, "dur": 14, + "args": { + "External id": 2161,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2160 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790544504, "dur": 12, + "args": { + "External id": 2162,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2161 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544535, "dur": 5, + "args": { + "External id": 2163,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2162 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790544544, "dur": 22, + "args": { + "External id": 2164,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2163 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544545, "dur": 5, + "args": { + "External id": 2165,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2164 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790544551, "dur": 14, + "args": { + "External id": 2166,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2165 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790544553, "dur": 11, + "args": { + "External id": 2167,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2166 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544584, "dur": 5, + "args": { + "External id": 2168,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2167 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790544593, "dur": 21, + "args": { + "External id": 2169,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2168 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544594, "dur": 4, + "args": { + "External id": 2170,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2169 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790544599, "dur": 15, + "args": { + "External id": 2171,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2170 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790544601, "dur": 11, + "args": { + "External id": 2172,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2171 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544632, "dur": 5, + "args": { + "External id": 2173,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2172 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790544641, "dur": 21, + "args": { + "External id": 2174,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2173 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544642, "dur": 4, + "args": { + "External id": 2175,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2174 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790544647, "dur": 14, + "args": { + "External id": 2176,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2175 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790544649, "dur": 11, + "args": { + "External id": 2177,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2176 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544680, "dur": 5, + "args": { + "External id": 2178,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2177 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790544689, "dur": 21, + "args": { + "External id": 2179,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2178 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544690, "dur": 4, + "args": { + "External id": 2180,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2179 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790544696, "dur": 14, + "args": { + "External id": 2181,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2180 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790544697, "dur": 11, + "args": { + "External id": 2182,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2181 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544728, "dur": 5, + "args": { + "External id": 2183,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2182 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790544737, "dur": 21, + "args": { + "External id": 2184,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2183 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544738, "dur": 5, + "args": { + "External id": 2185,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2184 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790544744, "dur": 14, + "args": { + "External id": 2186,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2185 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790544745, "dur": 12, + "args": { + "External id": 2187,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2186 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544776, "dur": 5, + "args": { + "External id": 2188,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2187 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790544785, "dur": 21, + "args": { + "External id": 2189,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2188 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544786, "dur": 5, + "args": { + "External id": 2190,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2189 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790544792, "dur": 14, + "args": { + "External id": 2191,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2190 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790544793, "dur": 11, + "args": { + "External id": 2192,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2191 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544824, "dur": 5, + "args": { + "External id": 2193,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2192 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790544833, "dur": 21, + "args": { + "External id": 2194,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2193 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544834, "dur": 5, + "args": { + "External id": 2195,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2194 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790544840, "dur": 14, + "args": { + "External id": 2196,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2195 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790544842, "dur": 11, + "args": { + "External id": 2197,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2196 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544872, "dur": 5, + "args": { + "External id": 2198,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2197 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790544881, "dur": 21, + "args": { + "External id": 2199,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2198 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544882, "dur": 5, + "args": { + "External id": 2200,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2199 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790544888, "dur": 14, + "args": { + "External id": 2201,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2200 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790544890, "dur": 10, + "args": { + "External id": 2202,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2201 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544920, "dur": 5, + "args": { + "External id": 2203,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2202 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790544929, "dur": 21, + "args": { + "External id": 2204,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2203 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544930, "dur": 4, + "args": { + "External id": 2205,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2204 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790544936, "dur": 14, + "args": { + "External id": 2206,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2205 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790544937, "dur": 11, + "args": { + "External id": 2207,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2206 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544968, "dur": 5, + "args": { + "External id": 2208,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2207 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790544977, "dur": 21, + "args": { + "External id": 2209,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2208 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790544978, "dur": 4, + "args": { + "External id": 2210,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2209 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790544983, "dur": 14, + "args": { + "External id": 2211,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2210 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790544985, "dur": 11, + "args": { + "External id": 2212,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2211 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545016, "dur": 5, + "args": { + "External id": 2213,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2212 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790545025, "dur": 21, + "args": { + "External id": 2214,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2213 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545026, "dur": 4, + "args": { + "External id": 2215,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2214 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790545031, "dur": 14, + "args": { + "External id": 2216,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2215 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790545033, "dur": 11, + "args": { + "External id": 2217,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2216 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545063, "dur": 5, + "args": { + "External id": 2218,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2217 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790545072, "dur": 22, + "args": { + "External id": 2219,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2218 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545073, "dur": 5, + "args": { + "External id": 2220,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2219 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790545079, "dur": 14, + "args": { + "External id": 2221,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2220 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790545081, "dur": 11, + "args": { + "External id": 2222,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2221 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545111, "dur": 5, + "args": { + "External id": 2223,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2222 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790545120, "dur": 22, + "args": { + "External id": 2224,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2223 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545122, "dur": 4, + "args": { + "External id": 2225,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2224 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790545127, "dur": 14, + "args": { + "External id": 2226,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2225 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790545129, "dur": 11, + "args": { + "External id": 2227,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2226 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545160, "dur": 5, + "args": { + "External id": 2228,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2227 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790545169, "dur": 21, + "args": { + "External id": 2229,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2228 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545170, "dur": 4, + "args": { + "External id": 2230,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2229 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790545175, "dur": 14, + "args": { + "External id": 2231,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2230 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790545177, "dur": 11, + "args": { + "External id": 2232,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2231 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545208, "dur": 5, + "args": { + "External id": 2233,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2232 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790545217, "dur": 21, + "args": { + "External id": 2234,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2233 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545218, "dur": 4, + "args": { + "External id": 2235,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2234 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790545224, "dur": 14, + "args": { + "External id": 2236,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2235 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790545225, "dur": 11, + "args": { + "External id": 2237,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2236 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545256, "dur": 5, + "args": { + "External id": 2238,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2237 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790545265, "dur": 21, + "args": { + "External id": 2239,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2238 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545266, "dur": 4, + "args": { + "External id": 2240,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2239 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790545271, "dur": 15, + "args": { + "External id": 2241,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2240 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790545273, "dur": 11, + "args": { + "External id": 2242,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2241 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545303, "dur": 6, + "args": { + "External id": 2243,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2242 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790545313, "dur": 21, + "args": { + "External id": 2244,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2243 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545314, "dur": 4, + "args": { + "External id": 2245,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2244 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790545319, "dur": 14, + "args": { + "External id": 2246,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2245 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790545321, "dur": 11, + "args": { + "External id": 2247,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2246 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545352, "dur": 5, + "args": { + "External id": 2248,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2247 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790545361, "dur": 21, + "args": { + "External id": 2249,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2248 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545362, "dur": 4, + "args": { + "External id": 2250,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2249 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790545368, "dur": 13, + "args": { + "External id": 2251,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2250 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790545369, "dur": 11, + "args": { + "External id": 2252,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2251 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545399, "dur": 5, + "args": { + "External id": 2253,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2252 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790545408, "dur": 21, + "args": { + "External id": 2254,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2253 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545409, "dur": 5, + "args": { + "External id": 2255,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2254 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790545415, "dur": 14, + "args": { + "External id": 2256,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2255 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790545417, "dur": 11, + "args": { + "External id": 2257,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2256 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545447, "dur": 5, + "args": { + "External id": 2258,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2257 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790545456, "dur": 21, + "args": { + "External id": 2259,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2258 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545457, "dur": 5, + "args": { + "External id": 2260,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2259 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790545463, "dur": 14, + "args": { + "External id": 2261,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2260 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790545464, "dur": 12, + "args": { + "External id": 2262,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2261 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545495, "dur": 5, + "args": { + "External id": 2263,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2262 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790545504, "dur": 21, + "args": { + "External id": 2264,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2263 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545505, "dur": 5, + "args": { + "External id": 2265,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2264 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790545511, "dur": 14, + "args": { + "External id": 2266,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2265 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790545512, "dur": 12, + "args": { + "External id": 2267,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2266 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545543, "dur": 5, + "args": { + "External id": 2268,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2267 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790545552, "dur": 21, + "args": { + "External id": 2269,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2268 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545553, "dur": 5, + "args": { + "External id": 2270,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2269 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790545559, "dur": 14, + "args": { + "External id": 2271,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2270 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790545560, "dur": 12, + "args": { + "External id": 2272,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2271 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545591, "dur": 5, + "args": { + "External id": 2273,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2272 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790545600, "dur": 25, + "args": { + "External id": 2274,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2273 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545601, "dur": 5, + "args": { + "External id": 2275,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2274 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790545607, "dur": 17, + "args": { + "External id": 2276,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2275 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790545612, "dur": 11, + "args": { + "External id": 2277,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2276 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545643, "dur": 5, + "args": { + "External id": 2278,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2277 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790545652, "dur": 21, + "args": { + "External id": 2279,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2278 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545653, "dur": 4, + "args": { + "External id": 2280,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2279 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790545659, "dur": 14, + "args": { + "External id": 2281,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2280 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790545660, "dur": 11, + "args": { + "External id": 2282,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2281 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545691, "dur": 5, + "args": { + "External id": 2283,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2282 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790545700, "dur": 21, + "args": { + "External id": 2284,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2283 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545701, "dur": 4, + "args": { + "External id": 2285,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2284 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790545707, "dur": 13, + "args": { + "External id": 2286,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2285 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790545708, "dur": 11, + "args": { + "External id": 2287,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2286 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545739, "dur": 5, + "args": { + "External id": 2288,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2287 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790545748, "dur": 21, + "args": { + "External id": 2289,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2288 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545749, "dur": 4, + "args": { + "External id": 2290,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2289 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790545755, "dur": 13, + "args": { + "External id": 2291,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2290 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790545756, "dur": 11, + "args": { + "External id": 2292,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2291 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545787, "dur": 5, + "args": { + "External id": 2293,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2292 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790545796, "dur": 21, + "args": { + "External id": 2294,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2293 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545797, "dur": 5, + "args": { + "External id": 2295,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2294 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790545803, "dur": 14, + "args": { + "External id": 2296,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2295 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790545805, "dur": 11, + "args": { + "External id": 2297,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2296 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545835, "dur": 5, + "args": { + "External id": 2298,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2297 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790545844, "dur": 22, + "args": { + "External id": 2299,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2298 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545846, "dur": 4, + "args": { + "External id": 2300,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2299 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790545852, "dur": 14, + "args": { + "External id": 2301,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2300 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790545853, "dur": 12, + "args": { + "External id": 2302,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2301 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545884, "dur": 5, + "args": { + "External id": 2303,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2302 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790545893, "dur": 21, + "args": { + "External id": 2304,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2303 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545894, "dur": 5, + "args": { + "External id": 2305,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2304 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790545900, "dur": 14, + "args": { + "External id": 2306,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2305 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790545901, "dur": 12, + "args": { + "External id": 2307,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2306 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545932, "dur": 5, + "args": { + "External id": 2308,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2307 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790545941, "dur": 21, + "args": { + "External id": 2309,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2308 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545942, "dur": 5, + "args": { + "External id": 2310,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2309 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790545948, "dur": 14, + "args": { + "External id": 2311,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2310 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790545950, "dur": 11, + "args": { + "External id": 2312,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2311 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545980, "dur": 5, + "args": { + "External id": 2313,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2312 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790545989, "dur": 21, + "args": { + "External id": 2314,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2313 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790545990, "dur": 5, + "args": { + "External id": 2315,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2314 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790545996, "dur": 14, + "args": { + "External id": 2316,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2315 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790545998, "dur": 11, + "args": { + "External id": 2317,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2316 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546028, "dur": 5, + "args": { + "External id": 2318,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2317 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790546037, "dur": 22, + "args": { + "External id": 2319,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2318 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546038, "dur": 5, + "args": { + "External id": 2320,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2319 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790546044, "dur": 15, + "args": { + "External id": 2321,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2320 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790546046, "dur": 11, + "args": { + "External id": 2322,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2321 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546076, "dur": 6, + "args": { + "External id": 2323,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2322 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790546085, "dur": 22, + "args": { + "External id": 2324,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2323 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546087, "dur": 4, + "args": { + "External id": 2325,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2324 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790546092, "dur": 14, + "args": { + "External id": 2326,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2325 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790546094, "dur": 11, + "args": { + "External id": 2327,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2326 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546125, "dur": 5, + "args": { + "External id": 2328,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2327 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790546134, "dur": 21, + "args": { + "External id": 2329,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2328 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546135, "dur": 4, + "args": { + "External id": 2330,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2329 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790546140, "dur": 15, + "args": { + "External id": 2331,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2330 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790546142, "dur": 11, + "args": { + "External id": 2332,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2331 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546173, "dur": 5, + "args": { + "External id": 2333,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2332 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790546182, "dur": 21, + "args": { + "External id": 2334,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2333 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546183, "dur": 4, + "args": { + "External id": 2335,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2334 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790546188, "dur": 14, + "args": { + "External id": 2336,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2335 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790546190, "dur": 11, + "args": { + "External id": 2337,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2336 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546220, "dur": 6, + "args": { + "External id": 2338,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2337 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790546229, "dur": 21, + "args": { + "External id": 2339,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2338 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546231, "dur": 4, + "args": { + "External id": 2340,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2339 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790546236, "dur": 14, + "args": { + "External id": 2341,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2340 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790546238, "dur": 11, + "args": { + "External id": 2342,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2341 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546268, "dur": 5, + "args": { + "External id": 2343,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2342 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790546277, "dur": 21, + "args": { + "External id": 2344,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2343 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546278, "dur": 5, + "args": { + "External id": 2345,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2344 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790546284, "dur": 14, + "args": { + "External id": 2346,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2345 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790546285, "dur": 12, + "args": { + "External id": 2347,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2346 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546316, "dur": 5, + "args": { + "External id": 2348,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2347 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790546325, "dur": 21, + "args": { + "External id": 2349,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2348 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546326, "dur": 5, + "args": { + "External id": 2350,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2349 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790546332, "dur": 14, + "args": { + "External id": 2351,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2350 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790546334, "dur": 11, + "args": { + "External id": 2352,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2351 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546364, "dur": 5, + "args": { + "External id": 2353,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2352 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790546373, "dur": 22, + "args": { + "External id": 2354,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2353 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546375, "dur": 4, + "args": { + "External id": 2355,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2354 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790546380, "dur": 14, + "args": { + "External id": 2356,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2355 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790546382, "dur": 11, + "args": { + "External id": 2357,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2356 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546413, "dur": 5, + "args": { + "External id": 2358,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2357 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790546422, "dur": 21, + "args": { + "External id": 2359,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2358 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546423, "dur": 4, + "args": { + "External id": 2360,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2359 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790546428, "dur": 14, + "args": { + "External id": 2361,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2360 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790546430, "dur": 11, + "args": { + "External id": 2362,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2361 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546460, "dur": 6, + "args": { + "External id": 2363,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2362 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790546469, "dur": 22, + "args": { + "External id": 2364,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2363 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546471, "dur": 4, + "args": { + "External id": 2365,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2364 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790546476, "dur": 14, + "args": { + "External id": 2366,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2365 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790546478, "dur": 11, + "args": { + "External id": 2367,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2366 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546509, "dur": 5, + "args": { + "External id": 2368,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2367 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790546518, "dur": 21, + "args": { + "External id": 2369,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2368 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546519, "dur": 4, + "args": { + "External id": 2370,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2369 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790546524, "dur": 14, + "args": { + "External id": 2371,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2370 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790546526, "dur": 11, + "args": { + "External id": 2372,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2371 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546556, "dur": 6, + "args": { + "External id": 2373,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2372 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790546565, "dur": 21, + "args": { + "External id": 2374,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2373 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546567, "dur": 4, + "args": { + "External id": 2375,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2374 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790546572, "dur": 14, + "args": { + "External id": 2376,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 2375 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790546574, "dur": 11, + "args": { + "External id": 2377,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 2376 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::randn", "pid": 435, "tid": 435, + "ts": 1776649790546758, "dur": 56, + "args": { + "External id": 2378,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 2377 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546763, "dur": 8, + "args": { + "External id": 2379,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2378 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::normal_", "pid": 435, "tid": 435, + "ts": 1776649790546776, "dur": 38, + "args": { + "External id": 2380,"Record function id": 0, "Concrete Inputs": ["", "0.", "1.", ""], "Input type": ["c10::Half", "Scalar", "Scalar", ""], "Input Dims": [[4, 4096], [], [], []], "Ev Idx": 2379 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::randn", "pid": 435, "tid": 435, + "ts": 1776649790546838, "dur": 30, + "args": { + "External id": 2381,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "", "5", "", "", "False"], "Input type": ["ScalarList", "", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2380 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546842, "dur": 6, + "args": { + "External id": 2382,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2381 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::normal_", "pid": 435, "tid": 435, + "ts": 1776649790546849, "dur": 18, + "args": { + "External id": 2383,"Record function id": 0, "Concrete Inputs": ["", "0.", "1.", ""], "Input type": ["c10::Half", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 2382 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 435, "tid": 435, + "ts": 1776649790546896, "dur": 43, + "args": { + "External id": 2384,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "double"], "Input Dims": [[4096, 4096], []], "Ev Idx": 2383 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790546954, "dur": 16, + "args": { + "External id": 2385,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 4096], []], "Ev Idx": 2384 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790546956, "dur": 14, + "args": { + "External id": 2386,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 4096], []], "Ev Idx": 2385 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 435, "tid": 435, + "ts": 1776649790546975, "dur": 49, + "args": { + "External id": 2387,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[32, 128, 4096]], "Ev Idx": 2386 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790546981, "dur": 5, + "args": { + "External id": 2388,"Record function id": 0, "Concrete Inputs": ["[0]", "5", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2387 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 435, "tid": 435, + "ts": 1776649790546989, "dur": 34, + "args": { + "External id": 2389,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[32, 128, 4096], [0]], "Ev Idx": 2388 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 435, "tid": 435, + "ts": 1776649790546998, "dur": 8, + "args": { + "External id": 2390,"Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]", ""], "Input type": ["c10::Half", "ScalarList", ""], "Input Dims": [[0], [], []], "Ev Idx": 2389 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 435, "tid": 435, + "ts": 1776649790547033, "dur": 45, + "args": { + "External id": 2391,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1]", "False"], "Input type": ["c10::Half", "ScalarList", "Scalar"], "Input Dims": [[32, 128, 4096], [], []], "Ev Idx": 2390 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790547050, "dur": 3, + "args": { + "External id": 2392,"Record function id": 0, "Concrete Inputs": ["", "[32, 1, 4096]", "[4096, 0, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 2391 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 435, "tid": 435, + "ts": 1776649790547089, "dur": 27, + "args": { + "External id": 2393,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "double"], "Input Dims": [[32, 4096], []], "Ev Idx": 2392 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 435, "tid": 435, + "ts": 1776649790547126, "dur": 31, + "args": { + "External id": 2394,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "9.9999999999999995e-07", ""], "Input type": ["c10::Half", "Scalar", ""], "Input Dims": [[32, 4096], [], []], "Ev Idx": 2393 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790547132, "dur": 0, + "args": { + "External id": 2395,"Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 4096], [], [], [], []], "Ev Idx": 2394 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790547164, "dur": 0, + "args": { + "External id": 2396,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 4096], [], [], [], []], "Ev Idx": 2395 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649790547170, "dur": 7, + "args": { + "External id": 2397,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 4096], []], "Ev Idx": 2396 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790547175, "dur": 1, + "args": { + "External id": 2398,"Record function id": 0, "Concrete Inputs": ["", "[32, 1, 4096]", "[4096, 4096, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 2397 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 435, "tid": 435, + "ts": 1776649790547180, "dur": 25, + "args": { + "External id": 2399,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[32, 128, 4096], [32, 1, 4096]], "Ev Idx": 2398 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 435, "tid": 435, + "ts": 1776649790547210, "dur": 23, + "args": { + "External id": 2400,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[32, 128, 4096]], "Ev Idx": 2399 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 435, "tid": 435, + "ts": 1776649790547239, "dur": 16, + "args": { + "External id": 2401,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "-8", "7"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[32, 128, 4096], [], []], "Ev Idx": 2400 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790547240, "dur": 0, + "args": { + "External id": 2402,"Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 128, 4096], [], [], [], []], "Ev Idx": 2401 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790547260, "dur": 44, + "args": { + "External id": 2403,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 128, 4096], [], [], [], []], "Ev Idx": 2402 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790547262, "dur": 42, + "args": { + "External id": 2404,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["c10::Half", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[32, 128, 4096], [], [], [], [], [], []], "Ev Idx": 2403 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790547270, "dur": 7, + "args": { + "External id": 2405,"Record function id": 0, "Concrete Inputs": ["[32, 128, 4096]", "[524288, 4096, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2404 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790547280, "dur": 23, + "args": { + "External id": 2406,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "c10::Half", "Scalar"], "Input Dims": [[32, 128, 4096], [32, 128, 4096], []], "Ev Idx": 2405 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 435, "tid": 435, + "ts": 1776649790547316, "dur": 29, + "args": { + "External id": 2407,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["signed char", "long int", "Scalar"], "Input Dims": [[32, 128, 4096], [], []], "Ev Idx": 2406 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790547350, "dur": 26, + "args": { + "External id": 2408,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "False", "False", ""], "Input type": ["signed char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 128, 4096], [], [], [], []], "Ev Idx": 2407 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790547351, "dur": 24, + "args": { + "External id": 2409,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "", "", "", "False", ""], "Input type": ["signed char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[32, 128, 4096], [], [], [], [], [], []], "Ev Idx": 2408 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790547354, "dur": 5, + "args": { + "External id": 2410,"Record function id": 0, "Concrete Inputs": ["[32, 128, 4096]", "[524288, 4096, 1]", "0", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2409 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790547360, "dur": 15, + "args": { + "External id": 2411,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["unsigned char", "signed char", "Scalar"], "Input Dims": [[32, 128, 4096], [32, 128, 4096], []], "Ev Idx": 2410 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790547383, "dur": 5, + "args": { + "External id": 2412,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["unsigned char", "ScalarList"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 2411 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790547384, "dur": 3, + "args": { + "External id": 2413,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["unsigned char", "ScalarList"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 2412 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649790547403, "dur": 11, + "args": { + "External id": 2414,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 2413 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790547410, "dur": 1, + "args": { + "External id": 2415,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]", "[4096, 1]", "0"], "Input type": ["unsigned char", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 2414 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649790547416, "dur": 5, + "args": { + "External id": 2416,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "2"], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 2415 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790547418, "dur": 2, + "args": { + "External id": 2417,"Record function id": 0, "Concrete Inputs": ["", "[4096, 2048]", "[4096, 2]", "0"], "Input type": ["unsigned char", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 2416 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649790547427, "dur": 3, + "args": { + "External id": 2418,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 2417 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790547429, "dur": 0, + "args": { + "External id": 2419,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]", "[4096, 1]", "0"], "Input type": ["unsigned char", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 2418 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649790547432, "dur": 2, + "args": { + "External id": 2420,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "1", "9223372036854775807", "2"], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 2419 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790547433, "dur": 1, + "args": { + "External id": 2421,"Record function id": 0, "Concrete Inputs": ["", "[4096, 2048]", "[4096, 2]", "1"], "Input type": ["unsigned char", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 2420 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__lshift__", "pid": 435, "tid": 435, + "ts": 1776649790547442, "dur": 43, + "args": { + "External id": 2422,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 2421 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790547455, "dur": 6, + "args": { + "External id": 2423,"Record function id": 0, "Concrete Inputs": ["[4096, 2048]", "0", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2422 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__or__", "pid": 435, "tid": 435, + "ts": 1776649790547491, "dur": 29, + "args": { + "External id": 2424,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["unsigned char", "unsigned char"], "Input Dims": [[4096, 2048], [4096, 2048]], "Ev Idx": 2423 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_or", "pid": 435, "tid": 435, + "ts": 1776649790547495, "dur": 24, + "args": { + "External id": 2425,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["unsigned char", "unsigned char"], "Input Dims": [[4096, 2048], [4096, 2048]], "Ev Idx": 2424 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790547525, "dur": 1, + "args": { + "External id": 2426,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], []], "Ev Idx": 2425 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__and__", "pid": 435, "tid": 435, + "ts": 1776649790547543, "dur": 32, + "args": { + "External id": 2427,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "15"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 2426 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_and", "pid": 435, "tid": 435, + "ts": 1776649790547545, "dur": 30, + "args": { + "External id": 2428,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "15"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 2427 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_and", "pid": 435, "tid": 435, + "ts": 1776649790547550, "dur": 23, + "args": { + "External id": 2429,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["unsigned char", "long int"], "Input Dims": [[4096, 2048], []], "Ev Idx": 2428 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790547579, "dur": 26, + "args": { + "External id": 2430,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], []], "Ev Idx": 2429 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790547581, "dur": 24, + "args": { + "External id": 2431,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["unsigned char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], [], [], []], "Ev Idx": 2430 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790547584, "dur": 6, + "args": { + "External id": 2432,"Record function id": 0, "Concrete Inputs": ["[4096, 2048]", "[2048, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2431 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790547591, "dur": 13, + "args": { + "External id": 2433,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "unsigned char", "Scalar"], "Input Dims": [[4096, 2048], [4096, 2048], []], "Ev Idx": 2432 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__rshift__", "pid": 435, "tid": 435, + "ts": 1776649790547612, "dur": 29, + "args": { + "External id": 2434,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 2433 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790547618, "dur": 5, + "args": { + "External id": 2435,"Record function id": 0, "Concrete Inputs": ["[4096, 2048]", "0", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2434 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790547644, "dur": 31, + "args": { + "External id": 2436,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], []], "Ev Idx": 2435 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790547645, "dur": 29, + "args": { + "External id": 2437,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["unsigned char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], [], [], []], "Ev Idx": 2436 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790547654, "dur": 5, + "args": { + "External id": 2438,"Record function id": 0, "Concrete Inputs": ["[4096, 2048]", "[2048, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2437 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790547661, "dur": 12, + "args": { + "External id": 2439,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "unsigned char", "Scalar"], "Input Dims": [[4096, 2048], [4096, 2048], []], "Ev Idx": 2438 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::stack", "pid": 435, "tid": 435, + "ts": 1776649790547689, "dur": 56, + "args": { + "External id": 2440,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["TensorList", "Scalar"], "Input Dims": [[[4096, 2048], [4096, 2048]], []], "Ev Idx": 2439 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649790547694, "dur": 3, + "args": { + "External id": 2441,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["signed char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 2440 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790547696, "dur": 1, + "args": { + "External id": 2442,"Record function id": 0, "Concrete Inputs": ["", "[4096, 2048, 1]", "[2048, 1, 1]", ""], "Input type": ["signed char", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 2048], [], [], []], "Ev Idx": 2441 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649790547698, "dur": 2, + "args": { + "External id": 2443,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["signed char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 2442 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790547700, "dur": 0, + "args": { + "External id": 2444,"Record function id": 0, "Concrete Inputs": ["", "[4096, 2048, 1]", "[2048, 1, 1]", ""], "Input type": ["signed char", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 2048], [], [], []], "Ev Idx": 2443 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 435, "tid": 435, + "ts": 1776649790547703, "dur": 40, + "args": { + "External id": 2445,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["TensorList", "Scalar"], "Input Dims": [[[4096, 2048, 1], [4096, 2048, 1]], []], "Ev Idx": 2444 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790547752, "dur": 4, + "args": { + "External id": 2446,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["signed char", "ScalarList"], "Input Dims": [[4096, 2048, 2], []], "Ev Idx": 2445 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790547753, "dur": 3, + "args": { + "External id": 2447,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["signed char", "ScalarList"], "Input Dims": [[4096, 2048, 2], []], "Ev Idx": 2446 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::repeat_interleave", "pid": 435, "tid": 435, + "ts": 1776649790547764, "dur": 66, + "args": { + "External id": 2448,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "128", "0", ""], "Input type": ["c10::Half", "Scalar", "Scalar", ""], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 2447 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649790547767, "dur": 4, + "args": { + "External id": 2449,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 4096], []], "Ev Idx": 2448 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790547769, "dur": 1, + "args": { + "External id": 2450,"Record function id": 0, "Concrete Inputs": ["", "[32, 1, 4096]", "[4096, 4096, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 2449 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 435, "tid": 435, + "ts": 1776649790547773, "dur": 8, + "args": { + "External id": 2451,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]", "False"], "Input type": ["c10::Half", "ScalarList", "Scalar"], "Input Dims": [[32, 1, 4096], [], []], "Ev Idx": 2450 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790547779, "dur": 0, + "args": { + "External id": 2452,"Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]", "[4096, 0, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 1, 4096], [], [], []], "Ev Idx": 2451 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649790547783, "dur": 38, + "args": { + "External id": 2453,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 2452 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649790547788, "dur": 11, + "args": { + "External id": 2454,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[32, 128, 4096], [], [], [], [], []], "Ev Idx": 2453 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790547792, "dur": 7, + "args": { + "External id": 2455,"Record function id": 0, "Concrete Inputs": ["[32, 128, 4096]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2454 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790547801, "dur": 20, + "args": { + "External id": 2456,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[32, 128, 4096], [32, 128, 4096], []], "Ev Idx": 2455 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::flatten", "pid": 435, "tid": 435, + "ts": 1776649790547824, "dur": 5, + "args": { + "External id": 2457,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[32, 128, 4096], [], []], "Ev Idx": 2456 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790547827, "dur": 2, + "args": { + "External id": 2458,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 2457 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 435, "tid": 435, + "ts": 1776649790547838, "dur": 22, + "args": { + "External id": 2459,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["signed char", "long int", "Scalar"], "Input Dims": [[4096, 4096], [], []], "Ev Idx": 2458 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790547866, "dur": 25, + "args": { + "External id": 2460,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["signed char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 2459 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790547867, "dur": 24, + "args": { + "External id": 2461,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "", "", "", "False", ""], "Input type": ["signed char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], [], []], "Ev Idx": 2460 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790547870, "dur": 5, + "args": { + "External id": 2462,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "[4096, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2461 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790547877, "dur": 13, + "args": { + "External id": 2463,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "signed char", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 2462 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 435, "tid": 435, + "ts": 1776649790547898, "dur": 19, + "args": { + "External id": 2464,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4096, 4096], [4096, 4096]], "Ev Idx": 2463 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790547990, "dur": 179, + "args": { + "External id": 2465,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2464 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790547997, "dur": 170, + "args": { + "External id": 2466,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2465 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790548177, "dur": 31, + "args": { + "External id": 2467,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2466 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790548180, "dur": 27, + "args": { + "External id": 2468,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2467 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790548213, "dur": 49, + "args": { + "External id": 2469,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2468 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790548234, "dur": 27, + "args": { + "External id": 2470,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2469 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790548268, "dur": 27, + "args": { + "External id": 2471,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2470 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790548271, "dur": 23, + "args": { + "External id": 2472,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2471 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790548299, "dur": 26, + "args": { + "External id": 2473,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2472 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790548302, "dur": 22, + "args": { + "External id": 2474,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2473 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790548330, "dur": 26, + "args": { + "External id": 2475,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2474 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790548332, "dur": 23, + "args": { + "External id": 2476,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2475 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790548360, "dur": 26, + "args": { + "External id": 2477,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2476 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790548362, "dur": 23, + "args": { + "External id": 2478,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2477 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790548390, "dur": 25, + "args": { + "External id": 2479,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2478 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790548393, "dur": 22, + "args": { + "External id": 2480,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2479 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790548420, "dur": 25, + "args": { + "External id": 2481,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2480 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790548422, "dur": 22, + "args": { + "External id": 2482,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2481 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790548450, "dur": 25, + "args": { + "External id": 2483,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2482 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790548452, "dur": 22, + "args": { + "External id": 2484,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2483 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790548541, "dur": 35, + "args": { + "External id": 2485,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2484 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790548544, "dur": 31, + "args": { + "External id": 2486,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2485 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790548582, "dur": 28, + "args": { + "External id": 2487,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2486 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790548585, "dur": 24, + "args": { + "External id": 2488,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2487 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790548614, "dur": 26, + "args": { + "External id": 2489,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2488 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790548617, "dur": 23, + "args": { + "External id": 2490,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2489 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790548645, "dur": 25, + "args": { + "External id": 2491,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2490 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790548647, "dur": 23, + "args": { + "External id": 2492,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2491 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790548675, "dur": 25, + "args": { + "External id": 2493,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2492 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790548677, "dur": 22, + "args": { + "External id": 2494,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2493 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790548705, "dur": 25, + "args": { + "External id": 2495,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2494 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790548707, "dur": 23, + "args": { + "External id": 2496,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2495 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790548735, "dur": 25, + "args": { + "External id": 2497,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2496 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790548737, "dur": 22, + "args": { + "External id": 2498,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2497 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790548764, "dur": 26, + "args": { + "External id": 2499,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2498 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790548767, "dur": 22, + "args": { + "External id": 2500,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2499 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790548794, "dur": 25, + "args": { + "External id": 2501,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2500 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790548796, "dur": 23, + "args": { + "External id": 2502,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2501 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790548824, "dur": 25, + "args": { + "External id": 2503,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2502 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790548826, "dur": 22, + "args": { + "External id": 2504,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2503 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790548853, "dur": 25, + "args": { + "External id": 2505,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2504 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790548855, "dur": 22, + "args": { + "External id": 2506,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2505 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790548883, "dur": 25, + "args": { + "External id": 2507,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2506 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790548885, "dur": 22, + "args": { + "External id": 2508,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2507 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790548912, "dur": 25, + "args": { + "External id": 2509,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2508 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790548914, "dur": 22, + "args": { + "External id": 2510,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2509 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790548941, "dur": 26, + "args": { + "External id": 2511,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2510 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790548944, "dur": 22, + "args": { + "External id": 2512,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2511 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790548971, "dur": 25, + "args": { + "External id": 2513,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2512 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790548973, "dur": 22, + "args": { + "External id": 2514,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2513 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549000, "dur": 25, + "args": { + "External id": 2515,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2514 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549002, "dur": 22, + "args": { + "External id": 2516,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2515 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549029, "dur": 25, + "args": { + "External id": 2517,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2516 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549031, "dur": 23, + "args": { + "External id": 2518,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2517 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549059, "dur": 25, + "args": { + "External id": 2519,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2518 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549061, "dur": 22, + "args": { + "External id": 2520,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2519 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549088, "dur": 25, + "args": { + "External id": 2521,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2520 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549091, "dur": 22, + "args": { + "External id": 2522,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2521 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549118, "dur": 25, + "args": { + "External id": 2523,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2522 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549120, "dur": 22, + "args": { + "External id": 2524,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2523 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549147, "dur": 25, + "args": { + "External id": 2525,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2524 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549149, "dur": 23, + "args": { + "External id": 2526,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2525 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549177, "dur": 26, + "args": { + "External id": 2527,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2526 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549180, "dur": 22, + "args": { + "External id": 2528,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2527 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549207, "dur": 26, + "args": { + "External id": 2529,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2528 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549210, "dur": 22, + "args": { + "External id": 2530,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2529 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549237, "dur": 25, + "args": { + "External id": 2531,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2530 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549239, "dur": 23, + "args": { + "External id": 2532,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2531 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549267, "dur": 26, + "args": { + "External id": 2533,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2532 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549270, "dur": 22, + "args": { + "External id": 2534,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2533 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549297, "dur": 26, + "args": { + "External id": 2535,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2534 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549300, "dur": 22, + "args": { + "External id": 2536,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2535 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549327, "dur": 25, + "args": { + "External id": 2537,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2536 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549329, "dur": 23, + "args": { + "External id": 2538,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2537 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549357, "dur": 25, + "args": { + "External id": 2539,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2538 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549359, "dur": 22, + "args": { + "External id": 2540,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2539 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549386, "dur": 25, + "args": { + "External id": 2541,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2540 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549388, "dur": 23, + "args": { + "External id": 2542,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2541 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549416, "dur": 25, + "args": { + "External id": 2543,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2542 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549418, "dur": 22, + "args": { + "External id": 2544,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2543 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549445, "dur": 25, + "args": { + "External id": 2545,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2544 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549447, "dur": 22, + "args": { + "External id": 2546,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2545 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549474, "dur": 25, + "args": { + "External id": 2547,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2546 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549477, "dur": 22, + "args": { + "External id": 2548,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2547 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549504, "dur": 24, + "args": { + "External id": 2549,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2548 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549506, "dur": 22, + "args": { + "External id": 2550,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2549 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549533, "dur": 25, + "args": { + "External id": 2551,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2550 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549535, "dur": 22, + "args": { + "External id": 2552,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2551 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549562, "dur": 25, + "args": { + "External id": 2553,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2552 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549564, "dur": 22, + "args": { + "External id": 2554,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2553 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549591, "dur": 26, + "args": { + "External id": 2555,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2554 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549594, "dur": 22, + "args": { + "External id": 2556,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2555 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549621, "dur": 25, + "args": { + "External id": 2557,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2556 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549623, "dur": 22, + "args": { + "External id": 2558,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2557 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549650, "dur": 26, + "args": { + "External id": 2559,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2558 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549653, "dur": 22, + "args": { + "External id": 2560,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2559 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549689, "dur": 26, + "args": { + "External id": 2561,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2560 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549691, "dur": 23, + "args": { + "External id": 2562,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2561 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549719, "dur": 26, + "args": { + "External id": 2563,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2562 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549722, "dur": 22, + "args": { + "External id": 2564,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2563 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549749, "dur": 26, + "args": { + "External id": 2565,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2564 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549751, "dur": 23, + "args": { + "External id": 2566,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2565 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549779, "dur": 25, + "args": { + "External id": 2567,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2566 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549781, "dur": 23, + "args": { + "External id": 2568,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2567 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549809, "dur": 25, + "args": { + "External id": 2569,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2568 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549811, "dur": 22, + "args": { + "External id": 2570,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2569 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549838, "dur": 25, + "args": { + "External id": 2571,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2570 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549840, "dur": 23, + "args": { + "External id": 2572,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2571 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549868, "dur": 25, + "args": { + "External id": 2573,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2572 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549870, "dur": 22, + "args": { + "External id": 2574,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2573 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549897, "dur": 26, + "args": { + "External id": 2575,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2574 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549900, "dur": 22, + "args": { + "External id": 2576,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2575 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549927, "dur": 26, + "args": { + "External id": 2577,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2576 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549930, "dur": 22, + "args": { + "External id": 2578,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2577 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549957, "dur": 25, + "args": { + "External id": 2579,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2578 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549959, "dur": 23, + "args": { + "External id": 2580,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2579 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790549987, "dur": 25, + "args": { + "External id": 2581,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2580 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790549989, "dur": 22, + "args": { + "External id": 2582,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2581 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550016, "dur": 26, + "args": { + "External id": 2583,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2582 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550019, "dur": 22, + "args": { + "External id": 2584,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2583 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550107, "dur": 35, + "args": { + "External id": 2585,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2584 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550111, "dur": 30, + "args": { + "External id": 2586,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2585 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550147, "dur": 28, + "args": { + "External id": 2587,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2586 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550150, "dur": 24, + "args": { + "External id": 2588,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2587 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550179, "dur": 27, + "args": { + "External id": 2589,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2588 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550181, "dur": 24, + "args": { + "External id": 2590,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2589 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550210, "dur": 25, + "args": { + "External id": 2591,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2590 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550212, "dur": 23, + "args": { + "External id": 2592,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2591 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550240, "dur": 26, + "args": { + "External id": 2593,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2592 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550242, "dur": 23, + "args": { + "External id": 2594,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2593 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550270, "dur": 25, + "args": { + "External id": 2595,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2594 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550272, "dur": 23, + "args": { + "External id": 2596,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2595 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550300, "dur": 26, + "args": { + "External id": 2597,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2596 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550302, "dur": 23, + "args": { + "External id": 2598,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2597 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550330, "dur": 25, + "args": { + "External id": 2599,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2598 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550332, "dur": 23, + "args": { + "External id": 2600,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2599 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550360, "dur": 25, + "args": { + "External id": 2601,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2600 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550362, "dur": 23, + "args": { + "External id": 2602,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2601 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550390, "dur": 25, + "args": { + "External id": 2603,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2602 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550392, "dur": 22, + "args": { + "External id": 2604,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2603 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550474, "dur": 34, + "args": { + "External id": 2605,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2604 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550478, "dur": 29, + "args": { + "External id": 2606,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2605 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550513, "dur": 27, + "args": { + "External id": 2607,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2606 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550516, "dur": 23, + "args": { + "External id": 2608,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2607 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550544, "dur": 26, + "args": { + "External id": 2609,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2608 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550547, "dur": 23, + "args": { + "External id": 2610,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2609 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550575, "dur": 26, + "args": { + "External id": 2611,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2610 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550577, "dur": 23, + "args": { + "External id": 2612,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2611 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550605, "dur": 26, + "args": { + "External id": 2613,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2612 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550607, "dur": 23, + "args": { + "External id": 2614,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2613 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550635, "dur": 26, + "args": { + "External id": 2615,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2614 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550637, "dur": 23, + "args": { + "External id": 2616,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2615 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550665, "dur": 25, + "args": { + "External id": 2617,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2616 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550667, "dur": 23, + "args": { + "External id": 2618,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2617 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550694, "dur": 26, + "args": { + "External id": 2619,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2618 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550697, "dur": 22, + "args": { + "External id": 2620,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2619 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550724, "dur": 26, + "args": { + "External id": 2621,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2620 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550727, "dur": 22, + "args": { + "External id": 2622,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2621 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550754, "dur": 25, + "args": { + "External id": 2623,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2622 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550756, "dur": 22, + "args": { + "External id": 2624,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2623 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550783, "dur": 26, + "args": { + "External id": 2625,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2624 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550785, "dur": 23, + "args": { + "External id": 2626,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2625 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550813, "dur": 25, + "args": { + "External id": 2627,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2626 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550815, "dur": 22, + "args": { + "External id": 2628,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2627 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550842, "dur": 26, + "args": { + "External id": 2629,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2628 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550845, "dur": 22, + "args": { + "External id": 2630,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2629 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550872, "dur": 26, + "args": { + "External id": 2631,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2630 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550875, "dur": 22, + "args": { + "External id": 2632,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2631 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550902, "dur": 26, + "args": { + "External id": 2633,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2632 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550905, "dur": 22, + "args": { + "External id": 2634,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2633 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550932, "dur": 25, + "args": { + "External id": 2635,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2634 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550934, "dur": 23, + "args": { + "External id": 2636,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2635 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550962, "dur": 25, + "args": { + "External id": 2637,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2636 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550964, "dur": 22, + "args": { + "External id": 2638,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2637 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790550991, "dur": 26, + "args": { + "External id": 2639,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2638 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790550994, "dur": 22, + "args": { + "External id": 2640,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2639 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551021, "dur": 25, + "args": { + "External id": 2641,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2640 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551023, "dur": 22, + "args": { + "External id": 2642,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2641 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551050, "dur": 25, + "args": { + "External id": 2643,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2642 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551052, "dur": 23, + "args": { + "External id": 2644,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2643 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551080, "dur": 25, + "args": { + "External id": 2645,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2644 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551082, "dur": 22, + "args": { + "External id": 2646,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2645 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551109, "dur": 26, + "args": { + "External id": 2647,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2646 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551111, "dur": 23, + "args": { + "External id": 2648,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2647 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551139, "dur": 25, + "args": { + "External id": 2649,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2648 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551142, "dur": 22, + "args": { + "External id": 2650,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2649 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551169, "dur": 25, + "args": { + "External id": 2651,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2650 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551171, "dur": 22, + "args": { + "External id": 2652,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2651 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551199, "dur": 25, + "args": { + "External id": 2653,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2652 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551201, "dur": 22, + "args": { + "External id": 2654,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2653 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551228, "dur": 25, + "args": { + "External id": 2655,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2654 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551230, "dur": 22, + "args": { + "External id": 2656,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2655 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551257, "dur": 25, + "args": { + "External id": 2657,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2656 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551259, "dur": 23, + "args": { + "External id": 2658,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2657 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551287, "dur": 25, + "args": { + "External id": 2659,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2658 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551289, "dur": 22, + "args": { + "External id": 2660,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2659 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551316, "dur": 26, + "args": { + "External id": 2661,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2660 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551319, "dur": 22, + "args": { + "External id": 2662,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2661 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551347, "dur": 25, + "args": { + "External id": 2663,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2662 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551349, "dur": 22, + "args": { + "External id": 2664,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2663 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551376, "dur": 24, + "args": { + "External id": 2665,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2664 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551378, "dur": 22, + "args": { + "External id": 2666,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2665 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551405, "dur": 25, + "args": { + "External id": 2667,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2666 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551407, "dur": 23, + "args": { + "External id": 2668,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2667 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551435, "dur": 25, + "args": { + "External id": 2669,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2668 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551437, "dur": 22, + "args": { + "External id": 2670,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2669 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551464, "dur": 25, + "args": { + "External id": 2671,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2670 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551466, "dur": 22, + "args": { + "External id": 2672,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2671 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551493, "dur": 25, + "args": { + "External id": 2673,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2672 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551495, "dur": 22, + "args": { + "External id": 2674,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2673 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551522, "dur": 26, + "args": { + "External id": 2675,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2674 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551525, "dur": 22, + "args": { + "External id": 2676,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2675 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551552, "dur": 25, + "args": { + "External id": 2677,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2676 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551554, "dur": 23, + "args": { + "External id": 2678,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2677 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551581, "dur": 25, + "args": { + "External id": 2679,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2678 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551584, "dur": 22, + "args": { + "External id": 2680,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2679 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551611, "dur": 25, + "args": { + "External id": 2681,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2680 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551613, "dur": 22, + "args": { + "External id": 2682,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2681 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551640, "dur": 25, + "args": { + "External id": 2683,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2682 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551642, "dur": 22, + "args": { + "External id": 2684,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2683 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551669, "dur": 25, + "args": { + "External id": 2685,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2684 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551671, "dur": 23, + "args": { + "External id": 2686,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2685 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551699, "dur": 25, + "args": { + "External id": 2687,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2686 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551701, "dur": 22, + "args": { + "External id": 2688,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2687 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551728, "dur": 25, + "args": { + "External id": 2689,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2688 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551730, "dur": 23, + "args": { + "External id": 2690,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2689 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551757, "dur": 25, + "args": { + "External id": 2691,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2690 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551760, "dur": 22, + "args": { + "External id": 2692,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2691 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551787, "dur": 25, + "args": { + "External id": 2693,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2692 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551789, "dur": 22, + "args": { + "External id": 2694,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2693 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551816, "dur": 25, + "args": { + "External id": 2695,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2694 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551818, "dur": 23, + "args": { + "External id": 2696,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2695 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551846, "dur": 25, + "args": { + "External id": 2697,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2696 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551848, "dur": 22, + "args": { + "External id": 2698,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2697 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551875, "dur": 26, + "args": { + "External id": 2699,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2698 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551878, "dur": 22, + "args": { + "External id": 2700,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2699 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551905, "dur": 25, + "args": { + "External id": 2701,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2700 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551907, "dur": 22, + "args": { + "External id": 2702,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2701 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790551934, "dur": 26, + "args": { + "External id": 2703,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2702 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790551937, "dur": 22, + "args": { + "External id": 2704,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2703 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552023, "dur": 34, + "args": { + "External id": 2705,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2704 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552027, "dur": 29, + "args": { + "External id": 2706,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2705 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552063, "dur": 27, + "args": { + "External id": 2707,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2706 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552066, "dur": 24, + "args": { + "External id": 2708,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2707 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552095, "dur": 26, + "args": { + "External id": 2709,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2708 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552097, "dur": 23, + "args": { + "External id": 2710,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2709 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552125, "dur": 26, + "args": { + "External id": 2711,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2710 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552127, "dur": 24, + "args": { + "External id": 2712,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2711 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552156, "dur": 26, + "args": { + "External id": 2713,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2712 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552158, "dur": 23, + "args": { + "External id": 2714,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2713 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552186, "dur": 26, + "args": { + "External id": 2715,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2714 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552188, "dur": 23, + "args": { + "External id": 2716,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2715 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552235, "dur": 30, + "args": { + "External id": 2717,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2716 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552238, "dur": 26, + "args": { + "External id": 2718,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2717 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552270, "dur": 25, + "args": { + "External id": 2719,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2718 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552272, "dur": 23, + "args": { + "External id": 2720,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2719 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552300, "dur": 26, + "args": { + "External id": 2721,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2720 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552302, "dur": 23, + "args": { + "External id": 2722,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2721 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552330, "dur": 26, + "args": { + "External id": 2723,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2722 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552333, "dur": 23, + "args": { + "External id": 2724,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2723 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552414, "dur": 34, + "args": { + "External id": 2725,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2724 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552417, "dur": 30, + "args": { + "External id": 2726,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2725 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552453, "dur": 27, + "args": { + "External id": 2727,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2726 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552456, "dur": 23, + "args": { + "External id": 2728,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2727 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552485, "dur": 25, + "args": { + "External id": 2729,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2728 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552487, "dur": 23, + "args": { + "External id": 2730,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2729 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552515, "dur": 25, + "args": { + "External id": 2731,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2730 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552517, "dur": 23, + "args": { + "External id": 2732,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2731 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552545, "dur": 26, + "args": { + "External id": 2733,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2732 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552547, "dur": 23, + "args": { + "External id": 2734,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2733 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552576, "dur": 25, + "args": { + "External id": 2735,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2734 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552578, "dur": 22, + "args": { + "External id": 2736,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2735 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552606, "dur": 25, + "args": { + "External id": 2737,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2736 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552608, "dur": 23, + "args": { + "External id": 2738,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2737 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552636, "dur": 26, + "args": { + "External id": 2739,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2738 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552638, "dur": 23, + "args": { + "External id": 2740,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2739 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552666, "dur": 26, + "args": { + "External id": 2741,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2740 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552669, "dur": 22, + "args": { + "External id": 2742,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2741 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552696, "dur": 26, + "args": { + "External id": 2743,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2742 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552699, "dur": 22, + "args": { + "External id": 2744,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2743 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552726, "dur": 26, + "args": { + "External id": 2745,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2744 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552729, "dur": 22, + "args": { + "External id": 2746,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2745 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552757, "dur": 26, + "args": { + "External id": 2747,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2746 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552759, "dur": 23, + "args": { + "External id": 2748,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2747 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552787, "dur": 26, + "args": { + "External id": 2749,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2748 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552790, "dur": 22, + "args": { + "External id": 2750,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2749 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552818, "dur": 25, + "args": { + "External id": 2751,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2750 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552820, "dur": 23, + "args": { + "External id": 2752,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2751 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552848, "dur": 26, + "args": { + "External id": 2753,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2752 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552850, "dur": 23, + "args": { + "External id": 2754,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2753 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552878, "dur": 26, + "args": { + "External id": 2755,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2754 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552880, "dur": 24, + "args": { + "External id": 2756,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2755 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552909, "dur": 25, + "args": { + "External id": 2757,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2756 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552911, "dur": 23, + "args": { + "External id": 2758,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2757 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552939, "dur": 25, + "args": { + "External id": 2759,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2758 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552941, "dur": 23, + "args": { + "External id": 2760,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2759 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552969, "dur": 26, + "args": { + "External id": 2761,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2760 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790552971, "dur": 23, + "args": { + "External id": 2762,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2761 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790552999, "dur": 26, + "args": { + "External id": 2763,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2762 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553001, "dur": 23, + "args": { + "External id": 2764,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2763 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553029, "dur": 26, + "args": { + "External id": 2765,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2764 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553031, "dur": 23, + "args": { + "External id": 2766,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2765 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553059, "dur": 26, + "args": { + "External id": 2767,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2766 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553061, "dur": 23, + "args": { + "External id": 2768,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2767 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553089, "dur": 26, + "args": { + "External id": 2769,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2768 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553092, "dur": 22, + "args": { + "External id": 2770,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2769 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553119, "dur": 26, + "args": { + "External id": 2771,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2770 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553122, "dur": 22, + "args": { + "External id": 2772,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2771 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553149, "dur": 26, + "args": { + "External id": 2773,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2772 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553152, "dur": 23, + "args": { + "External id": 2774,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2773 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553180, "dur": 25, + "args": { + "External id": 2775,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2774 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553182, "dur": 22, + "args": { + "External id": 2776,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2775 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553209, "dur": 26, + "args": { + "External id": 2777,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2776 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553211, "dur": 23, + "args": { + "External id": 2778,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2777 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553239, "dur": 26, + "args": { + "External id": 2779,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2778 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553241, "dur": 23, + "args": { + "External id": 2780,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2779 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553269, "dur": 26, + "args": { + "External id": 2781,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2780 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553272, "dur": 22, + "args": { + "External id": 2782,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2781 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553299, "dur": 26, + "args": { + "External id": 2783,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2782 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553302, "dur": 22, + "args": { + "External id": 2784,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2783 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553330, "dur": 25, + "args": { + "External id": 2785,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2784 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553332, "dur": 22, + "args": { + "External id": 2786,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2785 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553359, "dur": 26, + "args": { + "External id": 2787,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2786 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553362, "dur": 22, + "args": { + "External id": 2788,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2787 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553389, "dur": 27, + "args": { + "External id": 2789,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2788 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553392, "dur": 23, + "args": { + "External id": 2790,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2789 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553420, "dur": 26, + "args": { + "External id": 2791,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2790 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553422, "dur": 23, + "args": { + "External id": 2792,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2791 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553450, "dur": 26, + "args": { + "External id": 2793,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2792 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553453, "dur": 22, + "args": { + "External id": 2794,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2793 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553481, "dur": 26, + "args": { + "External id": 2795,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2794 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553483, "dur": 23, + "args": { + "External id": 2796,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2795 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553511, "dur": 26, + "args": { + "External id": 2797,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2796 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553513, "dur": 23, + "args": { + "External id": 2798,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2797 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553541, "dur": 26, + "args": { + "External id": 2799,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2798 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553543, "dur": 23, + "args": { + "External id": 2800,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2799 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553571, "dur": 26, + "args": { + "External id": 2801,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2800 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553573, "dur": 23, + "args": { + "External id": 2802,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2801 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553601, "dur": 26, + "args": { + "External id": 2803,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2802 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553603, "dur": 23, + "args": { + "External id": 2804,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2803 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553631, "dur": 26, + "args": { + "External id": 2805,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2804 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553633, "dur": 23, + "args": { + "External id": 2806,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2805 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553661, "dur": 26, + "args": { + "External id": 2807,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2806 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553663, "dur": 23, + "args": { + "External id": 2808,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2807 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553691, "dur": 26, + "args": { + "External id": 2809,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2808 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553694, "dur": 22, + "args": { + "External id": 2810,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2809 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553721, "dur": 26, + "args": { + "External id": 2811,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2810 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553724, "dur": 22, + "args": { + "External id": 2812,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2811 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553752, "dur": 25, + "args": { + "External id": 2813,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2812 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553754, "dur": 22, + "args": { + "External id": 2814,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2813 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553782, "dur": 25, + "args": { + "External id": 2815,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2814 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553784, "dur": 22, + "args": { + "External id": 2816,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2815 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553812, "dur": 25, + "args": { + "External id": 2817,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2816 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553814, "dur": 22, + "args": { + "External id": 2818,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2817 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553841, "dur": 26, + "args": { + "External id": 2819,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2818 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553844, "dur": 22, + "args": { + "External id": 2820,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2819 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553871, "dur": 26, + "args": { + "External id": 2821,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2820 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553874, "dur": 22, + "args": { + "External id": 2822,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2821 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553902, "dur": 25, + "args": { + "External id": 2823,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2822 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553904, "dur": 23, + "args": { + "External id": 2824,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2823 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790553991, "dur": 34, + "args": { + "External id": 2825,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2824 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790553994, "dur": 30, + "args": { + "External id": 2826,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2825 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554030, "dur": 28, + "args": { + "External id": 2827,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2826 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554033, "dur": 24, + "args": { + "External id": 2828,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2827 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554062, "dur": 27, + "args": { + "External id": 2829,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2828 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554065, "dur": 23, + "args": { + "External id": 2830,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2829 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554093, "dur": 26, + "args": { + "External id": 2831,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2830 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554095, "dur": 23, + "args": { + "External id": 2832,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2831 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554124, "dur": 26, + "args": { + "External id": 2833,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2832 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554126, "dur": 23, + "args": { + "External id": 2834,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2833 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554154, "dur": 26, + "args": { + "External id": 2835,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2834 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554156, "dur": 23, + "args": { + "External id": 2836,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2835 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554184, "dur": 26, + "args": { + "External id": 2837,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2836 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554187, "dur": 23, + "args": { + "External id": 2838,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2837 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554215, "dur": 25, + "args": { + "External id": 2839,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2838 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554217, "dur": 23, + "args": { + "External id": 2840,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2839 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554245, "dur": 25, + "args": { + "External id": 2841,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2840 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554247, "dur": 22, + "args": { + "External id": 2842,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2841 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554275, "dur": 25, + "args": { + "External id": 2843,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2842 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554277, "dur": 22, + "args": { + "External id": 2844,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2843 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554359, "dur": 33, + "args": { + "External id": 2845,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2844 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554362, "dur": 29, + "args": { + "External id": 2846,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2845 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554397, "dur": 26, + "args": { + "External id": 2847,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2846 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554400, "dur": 23, + "args": { + "External id": 2848,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2847 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554428, "dur": 26, + "args": { + "External id": 2849,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2848 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554430, "dur": 23, + "args": { + "External id": 2850,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2849 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554458, "dur": 26, + "args": { + "External id": 2851,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2850 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554461, "dur": 22, + "args": { + "External id": 2852,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2851 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554489, "dur": 25, + "args": { + "External id": 2853,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2852 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554491, "dur": 22, + "args": { + "External id": 2854,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2853 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554518, "dur": 26, + "args": { + "External id": 2855,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2854 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554520, "dur": 23, + "args": { + "External id": 2856,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2855 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554548, "dur": 25, + "args": { + "External id": 2857,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2856 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554550, "dur": 23, + "args": { + "External id": 2858,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2857 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554578, "dur": 25, + "args": { + "External id": 2859,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2858 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554580, "dur": 22, + "args": { + "External id": 2860,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2859 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554607, "dur": 25, + "args": { + "External id": 2861,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2860 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554609, "dur": 23, + "args": { + "External id": 2862,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2861 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554637, "dur": 25, + "args": { + "External id": 2863,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2862 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554639, "dur": 23, + "args": { + "External id": 2864,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2863 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554667, "dur": 25, + "args": { + "External id": 2865,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2864 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554669, "dur": 22, + "args": { + "External id": 2866,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2865 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554697, "dur": 25, + "args": { + "External id": 2867,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2866 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554699, "dur": 22, + "args": { + "External id": 2868,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2867 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554727, "dur": 25, + "args": { + "External id": 2869,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2868 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554729, "dur": 23, + "args": { + "External id": 2870,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2869 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554757, "dur": 25, + "args": { + "External id": 2871,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2870 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554759, "dur": 22, + "args": { + "External id": 2872,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2871 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554787, "dur": 25, + "args": { + "External id": 2873,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2872 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554789, "dur": 22, + "args": { + "External id": 2874,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2873 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554816, "dur": 26, + "args": { + "External id": 2875,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2874 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554819, "dur": 22, + "args": { + "External id": 2876,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2875 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554846, "dur": 25, + "args": { + "External id": 2877,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2876 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554848, "dur": 23, + "args": { + "External id": 2878,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2877 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554876, "dur": 25, + "args": { + "External id": 2879,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2878 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554878, "dur": 22, + "args": { + "External id": 2880,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2879 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554905, "dur": 25, + "args": { + "External id": 2881,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2880 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554907, "dur": 23, + "args": { + "External id": 2882,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2881 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554935, "dur": 25, + "args": { + "External id": 2883,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2882 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554937, "dur": 22, + "args": { + "External id": 2884,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2883 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554964, "dur": 26, + "args": { + "External id": 2885,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2884 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554967, "dur": 22, + "args": { + "External id": 2886,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2885 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790554995, "dur": 25, + "args": { + "External id": 2887,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2886 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790554997, "dur": 22, + "args": { + "External id": 2888,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2887 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555025, "dur": 25, + "args": { + "External id": 2889,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2888 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555027, "dur": 22, + "args": { + "External id": 2890,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2889 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555054, "dur": 25, + "args": { + "External id": 2891,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2890 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555056, "dur": 23, + "args": { + "External id": 2892,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2891 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555084, "dur": 25, + "args": { + "External id": 2893,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2892 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555086, "dur": 22, + "args": { + "External id": 2894,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2893 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555113, "dur": 25, + "args": { + "External id": 2895,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2894 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555115, "dur": 22, + "args": { + "External id": 2896,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2895 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555142, "dur": 25, + "args": { + "External id": 2897,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2896 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555144, "dur": 22, + "args": { + "External id": 2898,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2897 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555171, "dur": 25, + "args": { + "External id": 2899,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2898 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555173, "dur": 23, + "args": { + "External id": 2900,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2899 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555200, "dur": 26, + "args": { + "External id": 2901,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2900 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555203, "dur": 22, + "args": { + "External id": 2902,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2901 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555230, "dur": 25, + "args": { + "External id": 2903,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2902 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555232, "dur": 22, + "args": { + "External id": 2904,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2903 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555259, "dur": 25, + "args": { + "External id": 2905,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2904 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555261, "dur": 22, + "args": { + "External id": 2906,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2905 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555288, "dur": 25, + "args": { + "External id": 2907,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2906 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555290, "dur": 22, + "args": { + "External id": 2908,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2907 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555317, "dur": 30, + "args": { + "External id": 2909,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2908 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555323, "dur": 23, + "args": { + "External id": 2910,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2909 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555351, "dur": 25, + "args": { + "External id": 2911,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2910 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555353, "dur": 23, + "args": { + "External id": 2912,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2911 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555381, "dur": 24, + "args": { + "External id": 2913,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2912 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555383, "dur": 22, + "args": { + "External id": 2914,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2913 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555410, "dur": 25, + "args": { + "External id": 2915,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2914 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555412, "dur": 22, + "args": { + "External id": 2916,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2915 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555439, "dur": 26, + "args": { + "External id": 2917,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2916 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555442, "dur": 22, + "args": { + "External id": 2918,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2917 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555469, "dur": 25, + "args": { + "External id": 2919,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2918 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555471, "dur": 22, + "args": { + "External id": 2920,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2919 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555498, "dur": 25, + "args": { + "External id": 2921,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2920 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555500, "dur": 22, + "args": { + "External id": 2922,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2921 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555527, "dur": 26, + "args": { + "External id": 2923,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2922 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555529, "dur": 23, + "args": { + "External id": 2924,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2923 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555557, "dur": 25, + "args": { + "External id": 2925,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2924 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555559, "dur": 22, + "args": { + "External id": 2926,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2925 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555586, "dur": 25, + "args": { + "External id": 2927,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2926 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555588, "dur": 22, + "args": { + "External id": 2928,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2927 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555615, "dur": 25, + "args": { + "External id": 2929,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2928 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555617, "dur": 22, + "args": { + "External id": 2930,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2929 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555644, "dur": 25, + "args": { + "External id": 2931,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2930 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555646, "dur": 22, + "args": { + "External id": 2932,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2931 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555673, "dur": 25, + "args": { + "External id": 2933,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2932 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555675, "dur": 22, + "args": { + "External id": 2934,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2933 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555702, "dur": 25, + "args": { + "External id": 2935,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2934 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555704, "dur": 22, + "args": { + "External id": 2936,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2935 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555731, "dur": 26, + "args": { + "External id": 2937,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2936 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555733, "dur": 23, + "args": { + "External id": 2938,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2937 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555761, "dur": 25, + "args": { + "External id": 2939,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2938 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555763, "dur": 22, + "args": { + "External id": 2940,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2939 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555790, "dur": 25, + "args": { + "External id": 2941,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2940 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555792, "dur": 22, + "args": { + "External id": 2942,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2941 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555819, "dur": 25, + "args": { + "External id": 2943,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2942 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555821, "dur": 22, + "args": { + "External id": 2944,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2943 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555905, "dur": 34, + "args": { + "External id": 2945,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2944 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555909, "dur": 29, + "args": { + "External id": 2946,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2945 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555945, "dur": 26, + "args": { + "External id": 2947,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2946 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555947, "dur": 24, + "args": { + "External id": 2948,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2947 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790555976, "dur": 26, + "args": { + "External id": 2949,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2948 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790555978, "dur": 23, + "args": { + "External id": 2950,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2949 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556006, "dur": 26, + "args": { + "External id": 2951,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2950 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556009, "dur": 22, + "args": { + "External id": 2952,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2951 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556036, "dur": 26, + "args": { + "External id": 2953,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2952 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556038, "dur": 23, + "args": { + "External id": 2954,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2953 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556066, "dur": 25, + "args": { + "External id": 2955,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2954 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556068, "dur": 22, + "args": { + "External id": 2956,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2955 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556095, "dur": 25, + "args": { + "External id": 2957,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2956 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556098, "dur": 22, + "args": { + "External id": 2958,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2957 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556125, "dur": 24, + "args": { + "External id": 2959,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2958 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556127, "dur": 22, + "args": { + "External id": 2960,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2959 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556154, "dur": 25, + "args": { + "External id": 2961,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2960 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556156, "dur": 22, + "args": { + "External id": 2962,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2961 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556183, "dur": 25, + "args": { + "External id": 2963,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2962 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556185, "dur": 22, + "args": { + "External id": 2964,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2963 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556281, "dur": 35, + "args": { + "External id": 2965,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2964 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556284, "dur": 31, + "args": { + "External id": 2966,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2965 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556322, "dur": 26, + "args": { + "External id": 2967,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2966 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556324, "dur": 23, + "args": { + "External id": 2968,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2967 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556352, "dur": 25, + "args": { + "External id": 2969,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2968 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556355, "dur": 22, + "args": { + "External id": 2970,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2969 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556382, "dur": 24, + "args": { + "External id": 2971,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2970 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556384, "dur": 22, + "args": { + "External id": 2972,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2971 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556411, "dur": 25, + "args": { + "External id": 2973,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2972 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556413, "dur": 22, + "args": { + "External id": 2974,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2973 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556440, "dur": 25, + "args": { + "External id": 2975,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2974 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556442, "dur": 22, + "args": { + "External id": 2976,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2975 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556469, "dur": 25, + "args": { + "External id": 2977,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2976 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556471, "dur": 22, + "args": { + "External id": 2978,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2977 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556498, "dur": 25, + "args": { + "External id": 2979,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2978 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556501, "dur": 22, + "args": { + "External id": 2980,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2979 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556528, "dur": 24, + "args": { + "External id": 2981,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2980 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556530, "dur": 22, + "args": { + "External id": 2982,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2981 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556557, "dur": 25, + "args": { + "External id": 2983,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2982 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556559, "dur": 22, + "args": { + "External id": 2984,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2983 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556586, "dur": 25, + "args": { + "External id": 2985,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2984 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556588, "dur": 22, + "args": { + "External id": 2986,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2985 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556615, "dur": 25, + "args": { + "External id": 2987,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2986 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556617, "dur": 22, + "args": { + "External id": 2988,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2987 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556645, "dur": 24, + "args": { + "External id": 2989,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2988 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556647, "dur": 22, + "args": { + "External id": 2990,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2989 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556674, "dur": 24, + "args": { + "External id": 2991,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2990 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556676, "dur": 21, + "args": { + "External id": 2992,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2991 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556702, "dur": 25, + "args": { + "External id": 2993,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2992 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556705, "dur": 21, + "args": { + "External id": 2994,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2993 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556731, "dur": 25, + "args": { + "External id": 2995,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2994 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556734, "dur": 21, + "args": { + "External id": 2996,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2995 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556760, "dur": 25, + "args": { + "External id": 2997,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2996 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556763, "dur": 21, + "args": { + "External id": 2998,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2997 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556789, "dur": 25, + "args": { + "External id": 2999,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2998 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556792, "dur": 21, + "args": { + "External id": 3000,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 2999 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556818, "dur": 25, + "args": { + "External id": 3001,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3000 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556821, "dur": 22, + "args": { + "External id": 3002,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3001 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556848, "dur": 24, + "args": { + "External id": 3003,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3002 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556850, "dur": 22, + "args": { + "External id": 3004,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3003 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556877, "dur": 24, + "args": { + "External id": 3005,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3004 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556879, "dur": 22, + "args": { + "External id": 3006,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3005 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556906, "dur": 25, + "args": { + "External id": 3007,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3006 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556908, "dur": 22, + "args": { + "External id": 3008,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3007 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556935, "dur": 25, + "args": { + "External id": 3009,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3008 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556937, "dur": 23, + "args": { + "External id": 3010,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3009 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556965, "dur": 24, + "args": { + "External id": 3011,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3010 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556967, "dur": 22, + "args": { + "External id": 3012,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3011 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790556994, "dur": 24, + "args": { + "External id": 3013,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3012 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790556996, "dur": 22, + "args": { + "External id": 3014,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3013 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790557023, "dur": 25, + "args": { + "External id": 3015,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3014 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790557025, "dur": 22, + "args": { + "External id": 3016,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3015 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790557052, "dur": 25, + "args": { + "External id": 3017,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3016 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790557054, "dur": 22, + "args": { + "External id": 3018,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3017 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790557081, "dur": 25, + "args": { + "External id": 3019,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3018 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790557083, "dur": 22, + "args": { + "External id": 3020,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3019 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790557110, "dur": 25, + "args": { + "External id": 3021,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3020 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790557112, "dur": 22, + "args": { + "External id": 3022,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3021 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790557139, "dur": 25, + "args": { + "External id": 3023,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3022 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790557142, "dur": 22, + "args": { + "External id": 3024,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3023 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790557168, "dur": 25, + "args": { + "External id": 3025,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3024 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790557171, "dur": 22, + "args": { + "External id": 3026,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3025 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790557197, "dur": 25, + "args": { + "External id": 3027,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3026 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790557200, "dur": 22, + "args": { + "External id": 3028,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3027 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790557227, "dur": 25, + "args": { + "External id": 3029,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3028 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790557229, "dur": 22, + "args": { + "External id": 3030,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3029 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790557256, "dur": 25, + "args": { + "External id": 3031,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3030 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790557258, "dur": 22, + "args": { + "External id": 3032,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3031 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790557285, "dur": 25, + "args": { + "External id": 3033,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3032 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790557287, "dur": 22, + "args": { + "External id": 3034,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3033 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790557314, "dur": 25, + "args": { + "External id": 3035,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3034 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790557316, "dur": 22, + "args": { + "External id": 3036,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3035 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790557343, "dur": 25, + "args": { + "External id": 3037,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3036 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790557345, "dur": 23, + "args": { + "External id": 3038,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3037 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790557373, "dur": 24, + "args": { + "External id": 3039,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3038 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790557375, "dur": 22, + "args": { + "External id": 3040,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3039 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790557402, "dur": 25, + "args": { + "External id": 3041,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3040 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790557404, "dur": 22, + "args": { + "External id": 3042,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3041 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790557431, "dur": 25, + "args": { + "External id": 3043,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3042 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790557434, "dur": 22, + "args": { + "External id": 3044,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3043 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790557461, "dur": 26, + "args": { + "External id": 3045,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3044 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790557464, "dur": 23, + "args": { + "External id": 3046,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3045 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790557492, "dur": 25, + "args": { + "External id": 3047,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3046 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790557494, "dur": 22, + "args": { + "External id": 3048,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3047 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790557521, "dur": 25, + "args": { + "External id": 3049,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3048 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790557523, "dur": 22, + "args": { + "External id": 3050,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3049 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790557550, "dur": 25, + "args": { + "External id": 3051,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3050 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790557552, "dur": 22, + "args": { + "External id": 3052,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3051 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790557579, "dur": 25, + "args": { + "External id": 3053,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3052 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790557581, "dur": 22, + "args": { + "External id": 3054,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3053 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790557608, "dur": 25, + "args": { + "External id": 3055,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3054 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790557611, "dur": 21, + "args": { + "External id": 3056,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3055 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790557637, "dur": 32, + "args": { + "External id": 3057,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3056 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790557640, "dur": 28, + "args": { + "External id": 3058,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3057 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790557674, "dur": 26, + "args": { + "External id": 3059,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3058 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790557676, "dur": 23, + "args": { + "External id": 3060,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3059 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790557704, "dur": 26, + "args": { + "External id": 3061,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3060 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790557706, "dur": 23, + "args": { + "External id": 3062,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3061 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790557734, "dur": 25, + "args": { + "External id": 3063,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3062 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790557736, "dur": 22, + "args": { + "External id": 3064,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 4096]], "Ev Idx": 3063 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__and__", "pid": 435, "tid": 435, + "ts": 1776649790557859, "dur": 39, + "args": { + "External id": 3065,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "15"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 3064 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_and", "pid": 435, "tid": 435, + "ts": 1776649790557861, "dur": 37, + "args": { + "External id": 3066,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "15"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 3065 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_and", "pid": 435, "tid": 435, + "ts": 1776649790557867, "dur": 29, + "args": { + "External id": 3067,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["unsigned char", "long int"], "Input Dims": [[4096, 2048], []], "Ev Idx": 3066 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790557903, "dur": 32, + "args": { + "External id": 3068,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], []], "Ev Idx": 3067 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790557905, "dur": 29, + "args": { + "External id": 3069,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["unsigned char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], [], [], []], "Ev Idx": 3068 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790557910, "dur": 7, + "args": { + "External id": 3070,"Record function id": 0, "Concrete Inputs": ["[4096, 2048]", "[2048, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3069 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790557918, "dur": 16, + "args": { + "External id": 3071,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "unsigned char", "Scalar"], "Input Dims": [[4096, 2048], [4096, 2048], []], "Ev Idx": 3070 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__rshift__", "pid": 435, "tid": 435, + "ts": 1776649790557942, "dur": 36, + "args": { + "External id": 3072,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 3071 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790557958, "dur": 5, + "args": { + "External id": 3073,"Record function id": 0, "Concrete Inputs": ["[4096, 2048]", "0", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3072 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790557981, "dur": 23, + "args": { + "External id": 3074,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], []], "Ev Idx": 3073 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790557982, "dur": 21, + "args": { + "External id": 3075,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["unsigned char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], [], [], []], "Ev Idx": 3074 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790557985, "dur": 5, + "args": { + "External id": 3076,"Record function id": 0, "Concrete Inputs": ["[4096, 2048]", "[2048, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3075 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790557991, "dur": 12, + "args": { + "External id": 3077,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "unsigned char", "Scalar"], "Input Dims": [[4096, 2048], [4096, 2048], []], "Ev Idx": 3076 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::stack", "pid": 435, "tid": 435, + "ts": 1776649790558015, "dur": 40, + "args": { + "External id": 3078,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["TensorList", "Scalar"], "Input Dims": [[[4096, 2048], [4096, 2048]], []], "Ev Idx": 3077 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649790558018, "dur": 5, + "args": { + "External id": 3079,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["signed char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 3078 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790558021, "dur": 1, + "args": { + "External id": 3080,"Record function id": 0, "Concrete Inputs": ["", "[4096, 2048, 1]", "[2048, 1, 1]", ""], "Input type": ["signed char", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 2048], [], [], []], "Ev Idx": 3079 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649790558024, "dur": 1, + "args": { + "External id": 3081,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["signed char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 3080 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790558025, "dur": 0, + "args": { + "External id": 3082,"Record function id": 0, "Concrete Inputs": ["", "[4096, 2048, 1]", "[2048, 1, 1]", ""], "Input type": ["signed char", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 2048], [], [], []], "Ev Idx": 3081 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 435, "tid": 435, + "ts": 1776649790558027, "dur": 26, + "args": { + "External id": 3083,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["TensorList", "Scalar"], "Input Dims": [[[4096, 2048, 1], [4096, 2048, 1]], []], "Ev Idx": 3082 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790558062, "dur": 5, + "args": { + "External id": 3084,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["signed char", "ScalarList"], "Input Dims": [[4096, 2048, 2], []], "Ev Idx": 3083 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790558063, "dur": 4, + "args": { + "External id": 3085,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["signed char", "ScalarList"], "Input Dims": [[4096, 2048, 2], []], "Ev Idx": 3084 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::repeat_interleave", "pid": 435, "tid": 435, + "ts": 1776649790558072, "dur": 52, + "args": { + "External id": 3086,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "128", "0", ""], "Input type": ["c10::Half", "Scalar", "Scalar", ""], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 3085 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649790558075, "dur": 3, + "args": { + "External id": 3087,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 4096], []], "Ev Idx": 3086 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790558077, "dur": 0, + "args": { + "External id": 3088,"Record function id": 0, "Concrete Inputs": ["", "[32, 1, 4096]", "[4096, 4096, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 3087 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 435, "tid": 435, + "ts": 1776649790558080, "dur": 5, + "args": { + "External id": 3089,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]", "False"], "Input type": ["c10::Half", "ScalarList", "Scalar"], "Input Dims": [[32, 1, 4096], [], []], "Ev Idx": 3088 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790558083, "dur": 1, + "args": { + "External id": 3090,"Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]", "[4096, 0, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 1, 4096], [], [], []], "Ev Idx": 3089 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649790558087, "dur": 31, + "args": { + "External id": 3091,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 3090 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649790558089, "dur": 9, + "args": { + "External id": 3092,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[32, 128, 4096], [], [], [], [], []], "Ev Idx": 3091 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790558091, "dur": 7, + "args": { + "External id": 3093,"Record function id": 0, "Concrete Inputs": ["[32, 128, 4096]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3092 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790558100, "dur": 17, + "args": { + "External id": 3094,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[32, 128, 4096], [32, 128, 4096], []], "Ev Idx": 3093 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::flatten", "pid": 435, "tid": 435, + "ts": 1776649790558120, "dur": 4, + "args": { + "External id": 3095,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[32, 128, 4096], [], []], "Ev Idx": 3094 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790558122, "dur": 1, + "args": { + "External id": 3096,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 3095 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 435, "tid": 435, + "ts": 1776649790558132, "dur": 22, + "args": { + "External id": 3097,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["signed char", "long int", "Scalar"], "Input Dims": [[4096, 4096], [], []], "Ev Idx": 3096 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790558159, "dur": 24, + "args": { + "External id": 3098,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["signed char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 3097 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790558160, "dur": 23, + "args": { + "External id": 3099,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "", "", "", "False", ""], "Input type": ["signed char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], [], []], "Ev Idx": 3098 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790558163, "dur": 5, + "args": { + "External id": 3100,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "[4096, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3099 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790558169, "dur": 13, + "args": { + "External id": 3101,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "signed char", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 3100 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 435, "tid": 435, + "ts": 1776649790558190, "dur": 19, + "args": { + "External id": 3102,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4096, 4096], [4096, 4096]], "Ev Idx": 3101 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 435, "tid": 435, + "ts": 1776649790558221, "dur": 12, + "args": { + "External id": 3103,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[4096, 4096]], "Ev Idx": 3102 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 435, "tid": 435, + "ts": 1776649790558226, "dur": 6, + "args": { + "External id": 3104,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[4096, 4096], [], []], "Ev Idx": 3103 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790558230, "dur": 1, + "args": { + "External id": 3105,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]", "[1, 4096]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 3104 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 435, "tid": 435, + "ts": 1776649790558238, "dur": 28, + "args": { + "External id": 3106,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 3105 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649790558240, "dur": 25, + "args": { + "External id": 3107,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 3106 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649790558242, "dur": 7, + "args": { + "External id": 3108,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[4096, 4096], [], [], [], [], []], "Ev Idx": 3107 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790558243, "dur": 5, + "args": { + "External id": 3109,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3108 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790558250, "dur": 15, + "args": { + "External id": 3110,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 3109 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790558315, "dur": 49, + "args": { + "External id": 3111,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "", "", "", "False", ""], "Input type": ["ScalarList", "", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3110 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 435, "tid": 435, + "ts": 1776649790558375, "dur": 7, + "args": { + "External id": 3112,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Dims": [[4096, 4096]], "Ev Idx": 3111 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 435, "tid": 435, + "ts": 1776649790558378, "dur": 4, + "args": { + "External id": 3113,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Dims": [[4096, 4096]], "Ev Idx": 3112 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::uniform_", "pid": 435, "tid": 435, + "ts": 1776649790558447, "dur": 106213, + "args": { + "External id": 3114,"Record function id": 0, "Concrete Inputs": ["", "-0.015624999999999998", "0.015624999999999998", ""], "Input type": ["float", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 3113 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790664815, "dur": 19330, + "args": { + "External id": 3115,"Record function id": 0, "Concrete Inputs": ["", "", "5", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], []], "Ev Idx": 3114 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790664821, "dur": 19322, + "args": { + "External id": 3116,"Record function id": 0, "Concrete Inputs": ["", "5", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], [], []], "Ev Idx": 3115 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790664847, "dur": 56, + "args": { + "External id": 3117,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "[4096, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3116 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790664908, "dur": 19226, + "args": { + "External id": 3118,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "float", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 3117 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790664925, "dur": 11559, + "args": { + "External id": 3119,"Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 3118 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790664927, "dur": 11556, + "args": { + "External id": 3120,"Record function id": 0, "Concrete Inputs": ["", "5", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], [], []], "Ev Idx": 3119 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790664929, "dur": 50, + "args": { + "External id": 3121,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "[4096, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3120 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790664981, "dur": 11500, + "args": { + "External id": 3122,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "float", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 3121 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::expand_as", "pid": 435, "tid": 435, + "ts": 1776649790676489, "dur": 18, + "args": { + "External id": 3123,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4096, 4096], [4096, 4096]], "Ev Idx": 3122 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 435, "tid": 435, + "ts": 1776649790676494, "dur": 13, + "args": { + "External id": 3124,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]", "False"], "Input type": ["c10::Half", "ScalarList", "Scalar"], "Input Dims": [[4096, 4096], [], []], "Ev Idx": 3123 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790676501, "dur": 5, + "args": { + "External id": 3125,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]", "[4096, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 3124 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790676510, "dur": 3665, + "args": { + "External id": 3126,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 3125 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_has_compatible_shallow_copy_type", "pid": 435, "tid": 435, + "ts": 1776649790772333, "dur": 4, + "args": { + "External id": 3127,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "c10::Half"], "Input Dims": [[4096, 4096], [4096, 4096]], "Ev Idx": 3126 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_has_compatible_shallow_copy_type", "pid": 435, "tid": 435, + "ts": 1776649790772383, "dur": 0, + "args": { + "External id": 3128,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "c10::Half"], "Input Dims": [[4096, 4096], [4096, 4096]], "Ev Idx": 3127 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790780151, "dur": 386, + "args": { + "External id": 3129,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 3128 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790780635, "dur": 37, + "args": { + "External id": 3130,"Record function id": 0, "Concrete Inputs": ["[256, 8192]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3129 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790780692, "dur": 2, + "args": { + "External id": 3131,"Record function id": 0, "Concrete Inputs": ["[32, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3130 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790780708, "dur": 11, + "args": { + "External id": 3132,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3131 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790780712, "dur": 1, + "args": { + "External id": 3133,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3132 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790780716, "dur": 2, + "args": { + "External id": 3134,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3133 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790780753, "dur": 2509, + "args": { + "External id": 3135,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 8192], [], [], [], [], [], [], []], "Ev Idx": 3134 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790780757, "dur": 2504, + "args": { + "External id": 3136,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[256, 8192], [], [], [], [], [], []], "Ev Idx": 3135 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790780768, "dur": 22, + "args": { + "External id": 3137,"Record function id": 0, "Concrete Inputs": ["[256, 8192]", "[8192, 1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3136 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790780793, "dur": 2467, + "args": { + "External id": 3138,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 8192], [256, 8192], []], "Ev Idx": 3137 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790783272, "dur": 73, + "args": { + "External id": 3139,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[32, 4096], [], [], [], [], [], [], []], "Ev Idx": 3138 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790783273, "dur": 72, + "args": { + "External id": 3140,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[32, 4096], [], [], [], [], [], []], "Ev Idx": 3139 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790783276, "dur": 8, + "args": { + "External id": 3141,"Record function id": 0, "Concrete Inputs": ["[32, 4096]", "[4096, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3140 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790783285, "dur": 59, + "args": { + "External id": 3142,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[32, 4096], [32, 4096], []], "Ev Idx": 3141 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790783351, "dur": 32, + "args": { + "External id": 3143,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[512], [], [], [], [], [], [], []], "Ev Idx": 3142 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790783352, "dur": 31, + "args": { + "External id": 3144,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[512], [], [], [], [], [], []], "Ev Idx": 3143 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790783355, "dur": 6, + "args": { + "External id": 3145,"Record function id": 0, "Concrete Inputs": ["[512]", "[1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3144 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790783362, "dur": 21, + "args": { + "External id": 3146,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[512], [512], []], "Ev Idx": 3145 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 435, "tid": 435, + "ts": 1776649790783398, "dur": 24, + "args": { + "External id": 3147,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[32, 4096]], "Ev Idx": 3146 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 435, "tid": 435, + "ts": 1776649790783404, "dur": 14, + "args": { + "External id": 3148,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[32, 4096], [], []], "Ev Idx": 3147 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790783412, "dur": 5, + "args": { + "External id": 3149,"Record function id": 0, "Concrete Inputs": ["", "[4096, 32]", "[1, 4096]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 3148 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 435, "tid": 435, + "ts": 1776649790783426, "dur": 4, + "args": { + "External id": 3150,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[4096, 4096]], "Ev Idx": 3149 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 435, "tid": 435, + "ts": 1776649790783427, "dur": 2, + "args": { + "External id": 3151,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[4096, 4096], [], []], "Ev Idx": 3150 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790783429, "dur": 0, + "args": { + "External id": 3152,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]", "[1, 4096]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 3151 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790783439, "dur": 24, + "args": { + "External id": 3153,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 128, 4096]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 4096], []], "Ev Idx": 3152 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 435, "tid": 435, + "ts": 1776649790783456, "dur": 7, + "args": { + "External id": 3154,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]", "[128, 1, 4096]"], "Input type": ["c10::Half", "ScalarList", "ScalarList"], "Input Dims": [[4096, 4096], [], []], "Ev Idx": 3153 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 435, "tid": 435, + "ts": 1776649790783471, "dur": 8, + "args": { + "External id": 3155,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1, 0, 2]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 3154 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790783477, "dur": 1, + "args": { + "External id": 3156,"Record function id": 0, "Concrete Inputs": ["", "[128, 32, 4096]", "[1, 128, 4096]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 128, 4096], [], [], []], "Ev Idx": 3155 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790783482, "dur": 75, + "args": { + "External id": 3157,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[128, -1]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[128, 32, 4096], []], "Ev Idx": 3156 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649790783485, "dur": 62, + "args": { + "External id": 3158,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[128, 32, 4096], []], "Ev Idx": 3157 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649790783490, "dur": 14, + "args": { + "External id": 3159,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[128, 32, 4096], [], [], [], [], []], "Ev Idx": 3158 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790783494, "dur": 10, + "args": { + "External id": 3160,"Record function id": 0, "Concrete Inputs": ["[128, 32, 4096]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3159 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790783506, "dur": 40, + "args": { + "External id": 3161,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[128, 32, 4096], [128, 32, 4096], []], "Ev Idx": 3160 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 435, "tid": 435, + "ts": 1776649790783550, "dur": 6, + "args": { + "External id": 3162,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[128, 131072]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[128, 32, 4096], []], "Ev Idx": 3161 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790783565, "dur": 30, + "args": { + "External id": 3163,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1, -1]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 32], []], "Ev Idx": 3162 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649790783567, "dur": 25, + "args": { + "External id": 3164,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[4096, 32], []], "Ev Idx": 3163 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649790783568, "dur": 8, + "args": { + "External id": 3165,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[4096, 32], [], [], [], [], []], "Ev Idx": 3164 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790783569, "dur": 6, + "args": { + "External id": 3166,"Record function id": 0, "Concrete Inputs": ["[4096, 32]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3165 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790783577, "dur": 14, + "args": { + "External id": 3167,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[4096, 32], [4096, 32], []], "Ev Idx": 3166 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 435, "tid": 435, + "ts": 1776649790783594, "dur": 1, + "args": { + "External id": 3168,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1, 131072]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 32], []], "Ev Idx": 3167 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 435, "tid": 435, + "ts": 1776649790783604, "dur": 37, + "args": { + "External id": 3169,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[128, 131072], [1, 131072]], "Ev Idx": 3168 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 435, "tid": 435, + "ts": 1776649790783647, "dur": 24, + "args": { + "External id": 3170,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[128, 131072]], "Ev Idx": 3169 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790783679, "dur": 29, + "args": { + "External id": 3171,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[128, 131072], [], [], [], []], "Ev Idx": 3170 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790783681, "dur": 27, + "args": { + "External id": 3172,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "", "", "", "False", ""], "Input type": ["c10::Half", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[128, 131072], [], [], [], [], [], []], "Ev Idx": 3171 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790783684, "dur": 5, + "args": { + "External id": 3173,"Record function id": 0, "Concrete Inputs": ["[128, 131072]", "[131072, 1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3172 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790783690, "dur": 17, + "args": { + "External id": 3174,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "c10::Half", "Scalar"], "Input Dims": [[128, 131072], [128, 131072], []], "Ev Idx": 3173 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 435, "tid": 435, + "ts": 1776649790783723, "dur": 29, + "args": { + "External id": 3175,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["int", "long int", "Scalar"], "Input Dims": [[128, 131072], [], []], "Ev Idx": 3174 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 435, "tid": 435, + "ts": 1776649790783762, "dur": 32, + "args": { + "External id": 3176,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "15"], "Input type": ["int", "Scalar", "Scalar"], "Input Dims": [[128, 131072], [], []], "Ev Idx": 3175 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790783767, "dur": 1, + "args": { + "External id": 3177,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[128, 131072], [], [], [], []], "Ev Idx": 3176 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790783802, "dur": 8, + "args": { + "External id": 3178,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[128, -1, 4096]"], "Input type": ["int", "ScalarList"], "Input Dims": [[128, 131072], []], "Ev Idx": 3177 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790783804, "dur": 5, + "args": { + "External id": 3179,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[128, -1, 4096]"], "Input type": ["int", "ScalarList"], "Input Dims": [[128, 131072], []], "Ev Idx": 3178 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 435, "tid": 435, + "ts": 1776649790783813, "dur": 5, + "args": { + "External id": 3180,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1, 0, 2]"], "Input type": ["int", "ScalarList"], "Input Dims": [[128, 32, 4096], []], "Ev Idx": 3179 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790783816, "dur": 1, + "args": { + "External id": 3181,"Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]", "[4096, 131072, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Dims": [[128, 32, 4096], [], [], []], "Ev Idx": 3180 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790783821, "dur": 35, + "args": { + "External id": 3182,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["int", "ScalarList"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 3181 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649790783824, "dur": 29, + "args": { + "External id": 3183,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 3182 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649790783825, "dur": 9, + "args": { + "External id": 3184,"Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "0"], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[32, 128, 4096], [], [], [], [], []], "Ev Idx": 3183 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790783827, "dur": 6, + "args": { + "External id": 3185,"Record function id": 0, "Concrete Inputs": ["[32, 128, 4096]", "3", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3184 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790783835, "dur": 17, + "args": { + "External id": 3186,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[32, 128, 4096], [32, 128, 4096], []], "Ev Idx": 3185 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 435, "tid": 435, + "ts": 1776649790783855, "dur": 1, + "args": { + "External id": 3187,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["int", "ScalarList"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 3186 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790783864, "dur": 3, + "args": { + "External id": 3188,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 64]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[1, 131072], []], "Ev Idx": 3187 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790783865, "dur": 2, + "args": { + "External id": 3189,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 64]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[1, 131072], []], "Ev Idx": 3188 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649790783880, "dur": 10, + "args": { + "External id": 3190,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[2048, 64], [], [], [], []], "Ev Idx": 3189 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790783886, "dur": 1, + "args": { + "External id": 3191,"Record function id": 0, "Concrete Inputs": ["", "[2048, 64]", "[64, 1]", "0"], "Input type": ["c10::Half", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[2048, 64], [], [], []], "Ev Idx": 3190 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790783907, "dur": 1, + "args": { + "External id": 3192,"Record function id": 0, "Concrete Inputs": ["[64]", "4", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3191 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790783917, "dur": 404, + "args": { + "External id": 3193,"Record function id": 0, "Concrete Inputs": ["", "", "4", "False", "False", ""], "Input type": ["long int", "", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[64], [], [], [], [], []], "Ev Idx": 3192 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790783918, "dur": 403, + "args": { + "External id": 3194,"Record function id": 0, "Concrete Inputs": ["", "4", "", "", "", "False", ""], "Input type": ["long int", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[64], [], [], [], [], [], []], "Ev Idx": 3193 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790783920, "dur": 7, + "args": { + "External id": 3195,"Record function id": 0, "Concrete Inputs": ["[64]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3194 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790783928, "dur": 392, + "args": { + "External id": 3196,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Dims": [[64], [64], []], "Ev Idx": 3195 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 435, "tid": 435, + "ts": 1776649790784326, "dur": 1, + "args": { + "External id": 3197,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Dims": [[64]], "Ev Idx": 3196 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::index", "pid": 435, "tid": 435, + "ts": 1776649790784333, "dur": 50, + "args": { + "External id": 3198,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", ""], "Input Dims": [[2048, 64], []], "Ev Idx": 3197 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790784345, "dur": 1, + "args": { + "External id": 3199,"Record function id": 0, "Concrete Inputs": ["", "[2048, 64]", "[64, 0]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[2048, 64], [], [], []], "Ev Idx": 3198 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790784348, "dur": 2, + "args": { + "External id": 3200,"Record function id": 0, "Concrete Inputs": ["", "[1, 64]"], "Input type": ["long int", "ScalarList"], "Input Dims": [[64], []], "Ev Idx": 3199 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790784349, "dur": 1, + "args": { + "External id": 3201,"Record function id": 0, "Concrete Inputs": ["", "[1, 64]"], "Input type": ["long int", "ScalarList"], "Input Dims": [[64], []], "Ev Idx": 3200 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790784396, "dur": 4, + "args": { + "External id": 3202,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 4096]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[2048, 64], []], "Ev Idx": 3201 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790784397, "dur": 3, + "args": { + "External id": 3203,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 4096]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[2048, 64], []], "Ev Idx": 3202 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790784404, "dur": 2, + "args": { + "External id": 3204,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 16, 256, 16]"], "Input type": ["int", "ScalarList"], "Input Dims": [[4096, 4096], []], "Ev Idx": 3203 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790784405, "dur": 1, + "args": { + "External id": 3205,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 16, 256, 16]"], "Input type": ["int", "ScalarList"], "Input Dims": [[4096, 4096], []], "Ev Idx": 3204 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 435, "tid": 435, + "ts": 1776649790784409, "dur": 5, + "args": { + "External id": 3206,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[0, 2, 1, 3]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 16, 256, 16], []], "Ev Idx": 3205 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790784412, "dur": 1, + "args": { + "External id": 3207,"Record function id": 0, "Concrete Inputs": ["", "[256, 256, 16, 16]", "[65536, 16, 4096, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Dims": [[256, 16, 256, 16], [], [], []], "Ev Idx": 3206 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790784417, "dur": 36, + "args": { + "External id": 3208,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 65536]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 256, 16, 16], []], "Ev Idx": 3207 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649790784420, "dur": 29, + "args": { + "External id": 3209,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[256, 256, 16, 16], []], "Ev Idx": 3208 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649790784421, "dur": 9, + "args": { + "External id": 3210,"Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "0"], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[256, 256, 16, 16], [], [], [], [], []], "Ev Idx": 3209 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790784423, "dur": 7, + "args": { + "External id": 3211,"Record function id": 0, "Concrete Inputs": ["[256, 256, 16, 16]", "3", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3210 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790784432, "dur": 16, + "args": { + "External id": 3212,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 256, 16, 16], [256, 256, 16, 16], []], "Ev Idx": 3211 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 435, "tid": 435, + "ts": 1776649790784451, "dur": 2, + "args": { + "External id": 3213,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 65536]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 256, 16, 16], []], "Ev Idx": 3212 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790784463, "dur": 3, + "args": { + "External id": 3214,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 65536], []], "Ev Idx": 3213 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790784464, "dur": 1, + "args": { + "External id": 3215,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 65536], []], "Ev Idx": 3214 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649790784471, "dur": 5, + "args": { + "External id": 3216,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[16384, 1024], [], [], [], []], "Ev Idx": 3215 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790784474, "dur": 1, + "args": { + "External id": 3217,"Record function id": 0, "Concrete Inputs": ["", "[16384, 1024]", "[1024, 1]", "0"], "Input type": ["int", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[16384, 1024], [], [], []], "Ev Idx": 3216 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::index", "pid": 435, "tid": 435, + "ts": 1776649790784478, "dur": 124, + "args": { + "External id": 3218,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["int", ""], "Input Dims": [[16384, 1024], []], "Ev Idx": 3217 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790784481, "dur": 91, + "args": { + "External id": 3219,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "False", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Dims": [[1024], [], [], [], [], [], [], []], "Ev Idx": 3218 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790784484, "dur": 87, + "args": { + "External id": 3220,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Dims": [[1024], [], [], [], [], [], []], "Ev Idx": 3219 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790784487, "dur": 6, + "args": { + "External id": 3221,"Record function id": 0, "Concrete Inputs": ["[1024]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3220 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790784495, "dur": 76, + "args": { + "External id": 3222,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Dims": [[1024], [1024], []], "Ev Idx": 3221 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790784574, "dur": 1, + "args": { + "External id": 3223,"Record function id": 0, "Concrete Inputs": ["", "[16384, 1024]", "[1024, 0]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Dims": [[16384, 1024], [], [], []], "Ev Idx": 3222 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790784576, "dur": 2, + "args": { + "External id": 3224,"Record function id": 0, "Concrete Inputs": ["", "[1, 1024]"], "Input type": ["long int", "ScalarList"], "Input Dims": [[1024], []], "Ev Idx": 3223 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790784577, "dur": 1, + "args": { + "External id": 3225,"Record function id": 0, "Concrete Inputs": ["", "[1, 1024]"], "Input type": ["long int", "ScalarList"], "Input Dims": [[1024], []], "Ev Idx": 3224 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790784613, "dur": 3, + "args": { + "External id": 3226,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 65536]"], "Input type": ["int", "ScalarList"], "Input Dims": [[16384, 1024], []], "Ev Idx": 3225 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790784614, "dur": 2, + "args": { + "External id": 3227,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 65536]"], "Input type": ["int", "ScalarList"], "Input Dims": [[16384, 1024], []], "Ev Idx": 3226 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790798573, "dur": 45305, + "args": { + "External id": 3228,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 65536], [], [], [], [], [], [], []], "Ev Idx": 3227 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790798575, "dur": 45302, + "args": { + "External id": 3229,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[256, 65536], [], [], [], [], [], []], "Ev Idx": 3228 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790798581, "dur": 31, + "args": { + "External id": 3230,"Record function id": 0, "Concrete Inputs": ["[256, 65536]", "[65536, 1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3229 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790798614, "dur": 45257, + "args": { + "External id": 3231,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 65536], [256, 65536], []], "Ev Idx": 3230 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 435, "tid": 435, + "ts": 1776649790843912, "dur": 20, + "args": { + "External id": 3232,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 65536]], "Ev Idx": 3231 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 435, "tid": 435, + "ts": 1776649790843917, "dur": 15, + "args": { + "External id": 3233,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 65536]], "Ev Idx": 3232 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790843934, "dur": 0, + "args": { + "External id": 3234,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 65536], [], [], [], [], [], [], []], "Ev Idx": 3233 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::resolve_conj", "pid": 435, "tid": 435, + "ts": 1776649790843938, "dur": 1, + "args": { + "External id": 3235,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 65536]], "Ev Idx": 3234 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::resolve_neg", "pid": 435, "tid": 435, + "ts": 1776649790843940, "dur": 1, + "args": { + "External id": 3236,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 65536]], "Ev Idx": 3235 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 435, "tid": 435, + "ts": 1776649790903962, "dur": 9, + "args": { + "External id": 3237,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 8192]], "Ev Idx": 3236 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790904002, "dur": 808, + "args": { + "External id": 3238,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 8192], [], [], [], [], [], [], []], "Ev Idx": 3237 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790904007, "dur": 802, + "args": { + "External id": 3239,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[256, 8192], [], [], [], [], [], []], "Ev Idx": 3238 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790904027, "dur": 61, + "args": { + "External id": 3240,"Record function id": 0, "Concrete Inputs": ["[256, 8192]", "[8192, 1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3239 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790904092, "dur": 715, + "args": { + "External id": 3241,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 8192], [256, 8192], []], "Ev Idx": 3240 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790906021, "dur": 1, + "args": { + "External id": 3242,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 8192], [], [], [], [], [], [], []], "Ev Idx": 3241 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649790906054, "dur": 28, + "args": { + "External id": 3243,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[256, 8192], [], [], [], []], "Ev Idx": 3242 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790906069, "dur": 5, + "args": { + "External id": 3244,"Record function id": 0, "Concrete Inputs": ["", "[256, 8192]", "[8192, 1]", "0"], "Input type": ["int", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[256, 8192], [], [], []], "Ev Idx": 3243 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649790906084, "dur": 7, + "args": { + "External id": 3245,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[256, 8192], [], [], [], []], "Ev Idx": 3244 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790906089, "dur": 0, + "args": { + "External id": 3246,"Record function id": 0, "Concrete Inputs": ["", "[256, 8192]", "[8192, 1]", "0"], "Input type": ["int", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[256, 8192], [], [], []], "Ev Idx": 3245 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790906095, "dur": 58, + "args": { + "External id": 3247,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 8192], [256, 8192], []], "Ev Idx": 3246 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790906165, "dur": 1, + "args": { + "External id": 3248,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[32, 4096], [], [], [], [], [], [], []], "Ev Idx": 3247 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649790906171, "dur": 5, + "args": { + "External id": 3249,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[32, 4096], [], [], [], []], "Ev Idx": 3248 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790906174, "dur": 1, + "args": { + "External id": 3250,"Record function id": 0, "Concrete Inputs": ["", "[32, 4096]", "[4096, 1]", "0"], "Input type": ["c10::Half", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 3249 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649790906178, "dur": 3, + "args": { + "External id": 3251,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[32, 4096], [], [], [], []], "Ev Idx": 3250 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790906179, "dur": 1, + "args": { + "External id": 3252,"Record function id": 0, "Concrete Inputs": ["", "[32, 4096]", "[4096, 1]", "0"], "Input type": ["c10::Half", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 3251 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790906182, "dur": 19, + "args": { + "External id": 3253,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[32, 4096], [32, 4096], []], "Ev Idx": 3252 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790906853, "dur": 20, + "args": { + "External id": 3254,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3253 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790906886, "dur": 61, + "args": { + "External id": 3255,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3254 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790906892, "dur": 6, + "args": { + "External id": 3256,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3255 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790906901, "dur": 46, + "args": { + "External id": 3257,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3256 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790906906, "dur": 39, + "args": { + "External id": 3258,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3257 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790906990, "dur": 6, + "args": { + "External id": 3259,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3258 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790907001, "dur": 23, + "args": { + "External id": 3260,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3259 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907002, "dur": 5, + "args": { + "External id": 3261,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3260 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790907008, "dur": 16, + "args": { + "External id": 3262,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3261 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790907010, "dur": 12, + "args": { + "External id": 3263,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3262 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907044, "dur": 5, + "args": { + "External id": 3264,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3263 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790907054, "dur": 22, + "args": { + "External id": 3265,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3264 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907055, "dur": 4, + "args": { + "External id": 3266,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3265 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790907061, "dur": 14, + "args": { + "External id": 3267,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3266 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790907063, "dur": 11, + "args": { + "External id": 3268,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3267 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907096, "dur": 5, + "args": { + "External id": 3269,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3268 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790907105, "dur": 22, + "args": { + "External id": 3270,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3269 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907106, "dur": 5, + "args": { + "External id": 3271,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3270 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790907112, "dur": 15, + "args": { + "External id": 3272,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3271 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790907114, "dur": 12, + "args": { + "External id": 3273,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3272 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907146, "dur": 5, + "args": { + "External id": 3274,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3273 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790907155, "dur": 22, + "args": { + "External id": 3275,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3274 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907156, "dur": 5, + "args": { + "External id": 3276,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3275 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790907162, "dur": 14, + "args": { + "External id": 3277,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3276 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790907164, "dur": 11, + "args": { + "External id": 3278,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3277 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907195, "dur": 6, + "args": { + "External id": 3279,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3278 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790907205, "dur": 21, + "args": { + "External id": 3280,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3279 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907206, "dur": 4, + "args": { + "External id": 3281,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3280 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790907212, "dur": 14, + "args": { + "External id": 3282,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3281 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790907213, "dur": 12, + "args": { + "External id": 3283,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3282 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907244, "dur": 6, + "args": { + "External id": 3284,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3283 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790907254, "dur": 21, + "args": { + "External id": 3285,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3284 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907255, "dur": 4, + "args": { + "External id": 3286,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3285 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790907261, "dur": 14, + "args": { + "External id": 3287,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3286 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790907262, "dur": 12, + "args": { + "External id": 3288,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3287 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907294, "dur": 5, + "args": { + "External id": 3289,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3288 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790907303, "dur": 22, + "args": { + "External id": 3290,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3289 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907304, "dur": 5, + "args": { + "External id": 3291,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3290 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790907310, "dur": 14, + "args": { + "External id": 3292,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3291 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790907312, "dur": 11, + "args": { + "External id": 3293,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3292 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907342, "dur": 6, + "args": { + "External id": 3294,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3293 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790907352, "dur": 21, + "args": { + "External id": 3295,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3294 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907353, "dur": 4, + "args": { + "External id": 3296,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3295 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790907359, "dur": 14, + "args": { + "External id": 3297,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3296 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790907360, "dur": 11, + "args": { + "External id": 3298,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3297 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907391, "dur": 6, + "args": { + "External id": 3299,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3298 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790907401, "dur": 21, + "args": { + "External id": 3300,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3299 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907402, "dur": 4, + "args": { + "External id": 3301,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3300 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790907407, "dur": 15, + "args": { + "External id": 3302,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3301 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790907409, "dur": 11, + "args": { + "External id": 3303,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3302 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907553, "dur": 6, + "args": { + "External id": 3304,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3303 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790907565, "dur": 25, + "args": { + "External id": 3305,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3304 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907566, "dur": 4, + "args": { + "External id": 3306,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3305 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790907572, "dur": 18, + "args": { + "External id": 3307,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3306 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790907574, "dur": 15, + "args": { + "External id": 3308,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3307 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907611, "dur": 6, + "args": { + "External id": 3309,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3308 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790907621, "dur": 22, + "args": { + "External id": 3310,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3309 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907622, "dur": 5, + "args": { + "External id": 3311,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3310 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790907628, "dur": 14, + "args": { + "External id": 3312,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3311 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790907630, "dur": 11, + "args": { + "External id": 3313,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3312 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907661, "dur": 5, + "args": { + "External id": 3314,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3313 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790907670, "dur": 22, + "args": { + "External id": 3315,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3314 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907671, "dur": 5, + "args": { + "External id": 3316,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3315 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790907677, "dur": 15, + "args": { + "External id": 3317,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3316 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790907679, "dur": 12, + "args": { + "External id": 3318,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3317 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907710, "dur": 6, + "args": { + "External id": 3319,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3318 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790907720, "dur": 21, + "args": { + "External id": 3320,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3319 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907721, "dur": 4, + "args": { + "External id": 3321,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3320 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790907727, "dur": 14, + "args": { + "External id": 3322,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3321 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790907728, "dur": 11, + "args": { + "External id": 3323,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3322 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907759, "dur": 5, + "args": { + "External id": 3324,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3323 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790907768, "dur": 22, + "args": { + "External id": 3325,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3324 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907769, "dur": 5, + "args": { + "External id": 3326,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3325 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790907775, "dur": 14, + "args": { + "External id": 3327,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3326 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790907777, "dur": 11, + "args": { + "External id": 3328,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3327 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907808, "dur": 5, + "args": { + "External id": 3329,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3328 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790907817, "dur": 22, + "args": { + "External id": 3330,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3329 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907819, "dur": 4, + "args": { + "External id": 3331,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3330 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790907824, "dur": 15, + "args": { + "External id": 3332,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3331 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790907826, "dur": 11, + "args": { + "External id": 3333,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3332 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907857, "dur": 5, + "args": { + "External id": 3334,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3333 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790907866, "dur": 21, + "args": { + "External id": 3335,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3334 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907867, "dur": 5, + "args": { + "External id": 3336,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3335 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790907873, "dur": 14, + "args": { + "External id": 3337,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3336 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790907875, "dur": 11, + "args": { + "External id": 3338,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3337 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907906, "dur": 5, + "args": { + "External id": 3339,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3338 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790907915, "dur": 21, + "args": { + "External id": 3340,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3339 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907916, "dur": 5, + "args": { + "External id": 3341,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3340 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790907922, "dur": 14, + "args": { + "External id": 3342,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3341 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790907923, "dur": 11, + "args": { + "External id": 3343,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3342 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907954, "dur": 5, + "args": { + "External id": 3344,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3343 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790907963, "dur": 21, + "args": { + "External id": 3345,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3344 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790907964, "dur": 5, + "args": { + "External id": 3346,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3345 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790907970, "dur": 14, + "args": { + "External id": 3347,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3346 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790907972, "dur": 11, + "args": { + "External id": 3348,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3347 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908002, "dur": 6, + "args": { + "External id": 3349,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3348 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790908012, "dur": 21, + "args": { + "External id": 3350,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3349 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908013, "dur": 4, + "args": { + "External id": 3351,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3350 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790908019, "dur": 13, + "args": { + "External id": 3352,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3351 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790908020, "dur": 11, + "args": { + "External id": 3353,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3352 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908051, "dur": 5, + "args": { + "External id": 3354,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3353 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790908060, "dur": 21, + "args": { + "External id": 3355,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3354 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908061, "dur": 5, + "args": { + "External id": 3356,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3355 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790908067, "dur": 14, + "args": { + "External id": 3357,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3356 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790908069, "dur": 11, + "args": { + "External id": 3358,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3357 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908099, "dur": 6, + "args": { + "External id": 3359,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3358 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790908109, "dur": 21, + "args": { + "External id": 3360,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3359 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908110, "dur": 5, + "args": { + "External id": 3361,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3360 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790908116, "dur": 14, + "args": { + "External id": 3362,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3361 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790908118, "dur": 10, + "args": { + "External id": 3363,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3362 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908148, "dur": 5, + "args": { + "External id": 3364,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3363 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790908157, "dur": 21, + "args": { + "External id": 3365,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3364 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908158, "dur": 5, + "args": { + "External id": 3366,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3365 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790908164, "dur": 14, + "args": { + "External id": 3367,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3366 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790908166, "dur": 11, + "args": { + "External id": 3368,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3367 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908196, "dur": 5, + "args": { + "External id": 3369,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3368 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790908205, "dur": 59, + "args": { + "External id": 3370,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3369 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908206, "dur": 5, + "args": { + "External id": 3371,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3370 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790908212, "dur": 52, + "args": { + "External id": 3372,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3371 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790908213, "dur": 49, + "args": { + "External id": 3373,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3372 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908288, "dur": 6, + "args": { + "External id": 3374,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3373 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790908298, "dur": 22, + "args": { + "External id": 3375,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3374 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908299, "dur": 4, + "args": { + "External id": 3376,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3375 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790908305, "dur": 15, + "args": { + "External id": 3377,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3376 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790908307, "dur": 12, + "args": { + "External id": 3378,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3377 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908339, "dur": 5, + "args": { + "External id": 3379,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3378 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790908348, "dur": 21, + "args": { + "External id": 3380,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3379 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908349, "dur": 5, + "args": { + "External id": 3381,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3380 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790908355, "dur": 14, + "args": { + "External id": 3382,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3381 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790908357, "dur": 11, + "args": { + "External id": 3383,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3382 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908388, "dur": 5, + "args": { + "External id": 3384,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3383 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790908397, "dur": 21, + "args": { + "External id": 3385,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3384 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908398, "dur": 4, + "args": { + "External id": 3386,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3385 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790908404, "dur": 14, + "args": { + "External id": 3387,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3386 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790908405, "dur": 11, + "args": { + "External id": 3388,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3387 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908436, "dur": 5, + "args": { + "External id": 3389,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3388 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790908445, "dur": 21, + "args": { + "External id": 3390,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3389 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908446, "dur": 5, + "args": { + "External id": 3391,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3390 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790908452, "dur": 14, + "args": { + "External id": 3392,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3391 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790908454, "dur": 10, + "args": { + "External id": 3393,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3392 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908484, "dur": 5, + "args": { + "External id": 3394,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3393 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790908494, "dur": 21, + "args": { + "External id": 3395,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3394 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908495, "dur": 4, + "args": { + "External id": 3396,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3395 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790908500, "dur": 15, + "args": { + "External id": 3397,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3396 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790908502, "dur": 11, + "args": { + "External id": 3398,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3397 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908533, "dur": 5, + "args": { + "External id": 3399,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3398 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790908542, "dur": 22, + "args": { + "External id": 3400,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3399 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908544, "dur": 4, + "args": { + "External id": 3401,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3400 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790908549, "dur": 14, + "args": { + "External id": 3402,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3401 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790908551, "dur": 11, + "args": { + "External id": 3403,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3402 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908581, "dur": 5, + "args": { + "External id": 3404,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3403 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790908590, "dur": 22, + "args": { + "External id": 3405,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3404 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908591, "dur": 5, + "args": { + "External id": 3406,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3405 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790908597, "dur": 14, + "args": { + "External id": 3407,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3406 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790908599, "dur": 11, + "args": { + "External id": 3408,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3407 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908630, "dur": 5, + "args": { + "External id": 3409,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3408 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790908639, "dur": 21, + "args": { + "External id": 3410,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3409 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908641, "dur": 4, + "args": { + "External id": 3411,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3410 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790908646, "dur": 14, + "args": { + "External id": 3412,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3411 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790908648, "dur": 11, + "args": { + "External id": 3413,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3412 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908679, "dur": 5, + "args": { + "External id": 3414,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3413 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790908688, "dur": 21, + "args": { + "External id": 3415,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3414 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908689, "dur": 4, + "args": { + "External id": 3416,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3415 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790908695, "dur": 13, + "args": { + "External id": 3417,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3416 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790908696, "dur": 11, + "args": { + "External id": 3418,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3417 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908727, "dur": 5, + "args": { + "External id": 3419,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3418 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790908736, "dur": 22, + "args": { + "External id": 3420,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3419 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908737, "dur": 5, + "args": { + "External id": 3421,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3420 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790908743, "dur": 14, + "args": { + "External id": 3422,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3421 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790908745, "dur": 11, + "args": { + "External id": 3423,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3422 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908776, "dur": 5, + "args": { + "External id": 3424,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3423 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790908784, "dur": 21, + "args": { + "External id": 3425,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3424 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908786, "dur": 4, + "args": { + "External id": 3426,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3425 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790908791, "dur": 14, + "args": { + "External id": 3427,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3426 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790908793, "dur": 11, + "args": { + "External id": 3428,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3427 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908824, "dur": 5, + "args": { + "External id": 3429,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3428 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790908833, "dur": 20, + "args": { + "External id": 3430,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3429 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908834, "dur": 4, + "args": { + "External id": 3431,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3430 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790908840, "dur": 13, + "args": { + "External id": 3432,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3431 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790908841, "dur": 11, + "args": { + "External id": 3433,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3432 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908871, "dur": 5, + "args": { + "External id": 3434,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3433 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790908880, "dur": 21, + "args": { + "External id": 3435,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3434 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908881, "dur": 5, + "args": { + "External id": 3436,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3435 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790908887, "dur": 14, + "args": { + "External id": 3437,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3436 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790908889, "dur": 11, + "args": { + "External id": 3438,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3437 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908919, "dur": 5, + "args": { + "External id": 3439,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3438 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790908928, "dur": 21, + "args": { + "External id": 3440,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3439 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908929, "dur": 5, + "args": { + "External id": 3441,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3440 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790908935, "dur": 13, + "args": { + "External id": 3442,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3441 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790908937, "dur": 10, + "args": { + "External id": 3443,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3442 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908967, "dur": 5, + "args": { + "External id": 3444,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3443 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790908976, "dur": 21, + "args": { + "External id": 3445,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3444 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790908977, "dur": 4, + "args": { + "External id": 3446,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3445 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790908982, "dur": 14, + "args": { + "External id": 3447,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3446 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790908984, "dur": 11, + "args": { + "External id": 3448,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3447 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909015, "dur": 5, + "args": { + "External id": 3449,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3448 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790909024, "dur": 21, + "args": { + "External id": 3450,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3449 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909025, "dur": 4, + "args": { + "External id": 3451,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3450 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790909031, "dur": 13, + "args": { + "External id": 3452,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3451 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790909032, "dur": 11, + "args": { + "External id": 3453,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3452 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909062, "dur": 5, + "args": { + "External id": 3454,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3453 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790909071, "dur": 22, + "args": { + "External id": 3455,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3454 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909073, "dur": 4, + "args": { + "External id": 3456,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3455 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790909078, "dur": 14, + "args": { + "External id": 3457,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3456 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790909080, "dur": 11, + "args": { + "External id": 3458,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3457 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909110, "dur": 5, + "args": { + "External id": 3459,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3458 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790909119, "dur": 22, + "args": { + "External id": 3460,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3459 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909120, "dur": 5, + "args": { + "External id": 3461,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3460 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790909126, "dur": 14, + "args": { + "External id": 3462,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3461 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790909128, "dur": 11, + "args": { + "External id": 3463,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3462 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909158, "dur": 6, + "args": { + "External id": 3464,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3463 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790909168, "dur": 21, + "args": { + "External id": 3465,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3464 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909169, "dur": 4, + "args": { + "External id": 3466,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3465 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790909174, "dur": 14, + "args": { + "External id": 3467,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3466 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790909176, "dur": 11, + "args": { + "External id": 3468,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3467 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909207, "dur": 5, + "args": { + "External id": 3469,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3468 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790909216, "dur": 21, + "args": { + "External id": 3470,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3469 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909217, "dur": 4, + "args": { + "External id": 3471,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3470 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790909223, "dur": 13, + "args": { + "External id": 3472,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3471 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790909224, "dur": 11, + "args": { + "External id": 3473,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3472 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909255, "dur": 5, + "args": { + "External id": 3474,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3473 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790909264, "dur": 21, + "args": { + "External id": 3475,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3474 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909265, "dur": 4, + "args": { + "External id": 3476,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3475 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790909271, "dur": 13, + "args": { + "External id": 3477,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3476 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790909272, "dur": 11, + "args": { + "External id": 3478,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3477 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909310, "dur": 5, + "args": { + "External id": 3479,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3478 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790909319, "dur": 23, + "args": { + "External id": 3480,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3479 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909321, "dur": 4, + "args": { + "External id": 3481,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3480 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790909327, "dur": 14, + "args": { + "External id": 3482,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3481 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790909328, "dur": 12, + "args": { + "External id": 3483,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3482 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909360, "dur": 5, + "args": { + "External id": 3484,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3483 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790909369, "dur": 21, + "args": { + "External id": 3485,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3484 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909370, "dur": 5, + "args": { + "External id": 3486,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3485 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790909376, "dur": 14, + "args": { + "External id": 3487,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3486 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790909377, "dur": 12, + "args": { + "External id": 3488,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3487 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909408, "dur": 6, + "args": { + "External id": 3489,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3488 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790909417, "dur": 22, + "args": { + "External id": 3490,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3489 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909419, "dur": 4, + "args": { + "External id": 3491,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3490 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790909424, "dur": 14, + "args": { + "External id": 3492,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3491 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790909426, "dur": 11, + "args": { + "External id": 3493,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3492 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909457, "dur": 5, + "args": { + "External id": 3494,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3493 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790909466, "dur": 21, + "args": { + "External id": 3495,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3494 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909467, "dur": 5, + "args": { + "External id": 3496,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3495 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790909473, "dur": 14, + "args": { + "External id": 3497,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3496 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790909474, "dur": 11, + "args": { + "External id": 3498,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3497 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909505, "dur": 5, + "args": { + "External id": 3499,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3498 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790909514, "dur": 21, + "args": { + "External id": 3500,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3499 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909515, "dur": 4, + "args": { + "External id": 3501,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3500 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790909521, "dur": 14, + "args": { + "External id": 3502,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3501 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790909522, "dur": 11, + "args": { + "External id": 3503,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3502 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909553, "dur": 5, + "args": { + "External id": 3504,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3503 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790909562, "dur": 21, + "args": { + "External id": 3505,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3504 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909564, "dur": 4, + "args": { + "External id": 3506,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3505 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790909569, "dur": 14, + "args": { + "External id": 3507,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3506 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790909571, "dur": 11, + "args": { + "External id": 3508,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3507 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909601, "dur": 5, + "args": { + "External id": 3509,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3508 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790909610, "dur": 21, + "args": { + "External id": 3510,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3509 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909611, "dur": 5, + "args": { + "External id": 3511,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3510 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790909617, "dur": 14, + "args": { + "External id": 3512,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3511 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790909619, "dur": 11, + "args": { + "External id": 3513,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3512 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909650, "dur": 5, + "args": { + "External id": 3514,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3513 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790909659, "dur": 20, + "args": { + "External id": 3515,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3514 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909660, "dur": 4, + "args": { + "External id": 3516,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3515 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790909665, "dur": 14, + "args": { + "External id": 3517,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3516 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790909667, "dur": 11, + "args": { + "External id": 3518,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3517 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909697, "dur": 6, + "args": { + "External id": 3519,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3518 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790909707, "dur": 21, + "args": { + "External id": 3520,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3519 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909708, "dur": 4, + "args": { + "External id": 3521,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3520 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790909713, "dur": 14, + "args": { + "External id": 3522,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3521 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790909715, "dur": 11, + "args": { + "External id": 3523,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3522 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909746, "dur": 5, + "args": { + "External id": 3524,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3523 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790909755, "dur": 21, + "args": { + "External id": 3525,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3524 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909756, "dur": 5, + "args": { + "External id": 3526,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3525 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790909762, "dur": 13, + "args": { + "External id": 3527,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3526 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790909763, "dur": 11, + "args": { + "External id": 3528,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3527 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909794, "dur": 5, + "args": { + "External id": 3529,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3528 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790909803, "dur": 21, + "args": { + "External id": 3530,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3529 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909804, "dur": 5, + "args": { + "External id": 3531,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3530 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790909810, "dur": 14, + "args": { + "External id": 3532,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3531 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790909812, "dur": 11, + "args": { + "External id": 3533,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3532 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909842, "dur": 5, + "args": { + "External id": 3534,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3533 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790909851, "dur": 21, + "args": { + "External id": 3535,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3534 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909852, "dur": 4, + "args": { + "External id": 3536,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3535 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790909858, "dur": 14, + "args": { + "External id": 3537,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3536 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790909859, "dur": 11, + "args": { + "External id": 3538,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3537 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909890, "dur": 5, + "args": { + "External id": 3539,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3538 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790909899, "dur": 21, + "args": { + "External id": 3540,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3539 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909900, "dur": 5, + "args": { + "External id": 3541,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3540 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790909906, "dur": 14, + "args": { + "External id": 3542,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3541 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790909907, "dur": 11, + "args": { + "External id": 3543,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3542 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909938, "dur": 5, + "args": { + "External id": 3544,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3543 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790909947, "dur": 21, + "args": { + "External id": 3545,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3544 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909948, "dur": 4, + "args": { + "External id": 3546,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3545 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790909954, "dur": 14, + "args": { + "External id": 3547,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3546 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790909955, "dur": 11, + "args": { + "External id": 3548,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3547 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909986, "dur": 5, + "args": { + "External id": 3549,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3548 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790909995, "dur": 21, + "args": { + "External id": 3550,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3549 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790909996, "dur": 4, + "args": { + "External id": 3551,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3550 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790910002, "dur": 13, + "args": { + "External id": 3552,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3551 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790910003, "dur": 11, + "args": { + "External id": 3553,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3552 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910104, "dur": 7, + "args": { + "External id": 3554,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3553 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790910116, "dur": 25, + "args": { + "External id": 3555,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3554 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910117, "dur": 5, + "args": { + "External id": 3556,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3555 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790910123, "dur": 18, + "args": { + "External id": 3557,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3556 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790910125, "dur": 14, + "args": { + "External id": 3558,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3557 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910162, "dur": 5, + "args": { + "External id": 3559,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3558 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790910171, "dur": 22, + "args": { + "External id": 3560,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3559 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910173, "dur": 4, + "args": { + "External id": 3561,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3560 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790910179, "dur": 14, + "args": { + "External id": 3562,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3561 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790910180, "dur": 12, + "args": { + "External id": 3563,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3562 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910212, "dur": 5, + "args": { + "External id": 3564,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3563 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790910222, "dur": 21, + "args": { + "External id": 3565,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3564 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910223, "dur": 5, + "args": { + "External id": 3566,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3565 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790910229, "dur": 14, + "args": { + "External id": 3567,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3566 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790910230, "dur": 12, + "args": { + "External id": 3568,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3567 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910261, "dur": 6, + "args": { + "External id": 3569,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3568 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790910271, "dur": 21, + "args": { + "External id": 3570,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3569 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910272, "dur": 5, + "args": { + "External id": 3571,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3570 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790910278, "dur": 14, + "args": { + "External id": 3572,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3571 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790910279, "dur": 11, + "args": { + "External id": 3573,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3572 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910310, "dur": 5, + "args": { + "External id": 3574,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3573 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790910320, "dur": 21, + "args": { + "External id": 3575,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3574 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910321, "dur": 4, + "args": { + "External id": 3576,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3575 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790910326, "dur": 14, + "args": { + "External id": 3577,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3576 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790910328, "dur": 11, + "args": { + "External id": 3578,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3577 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910359, "dur": 5, + "args": { + "External id": 3579,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3578 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790910368, "dur": 21, + "args": { + "External id": 3580,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3579 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910369, "dur": 4, + "args": { + "External id": 3581,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3580 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790910375, "dur": 14, + "args": { + "External id": 3582,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3581 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790910376, "dur": 11, + "args": { + "External id": 3583,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3582 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910407, "dur": 5, + "args": { + "External id": 3584,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3583 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790910426, "dur": 21, + "args": { + "External id": 3585,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3584 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910427, "dur": 5, + "args": { + "External id": 3586,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3585 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790910433, "dur": 14, + "args": { + "External id": 3587,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3586 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790910435, "dur": 11, + "args": { + "External id": 3588,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3587 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910466, "dur": 6, + "args": { + "External id": 3589,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3588 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790910476, "dur": 21, + "args": { + "External id": 3590,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3589 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910477, "dur": 4, + "args": { + "External id": 3591,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3590 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790910482, "dur": 14, + "args": { + "External id": 3592,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3591 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790910484, "dur": 11, + "args": { + "External id": 3593,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3592 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910514, "dur": 5, + "args": { + "External id": 3594,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3593 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790910523, "dur": 21, + "args": { + "External id": 3595,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3594 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910525, "dur": 4, + "args": { + "External id": 3596,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3595 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790910530, "dur": 14, + "args": { + "External id": 3597,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3596 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790910532, "dur": 11, + "args": { + "External id": 3598,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3597 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910562, "dur": 6, + "args": { + "External id": 3599,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3598 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790910572, "dur": 21, + "args": { + "External id": 3600,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3599 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910573, "dur": 4, + "args": { + "External id": 3601,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3600 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790910579, "dur": 14, + "args": { + "External id": 3602,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3601 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790910580, "dur": 11, + "args": { + "External id": 3603,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3602 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910661, "dur": 6, + "args": { + "External id": 3604,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3603 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790910672, "dur": 25, + "args": { + "External id": 3605,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3604 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910674, "dur": 4, + "args": { + "External id": 3606,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3605 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790910680, "dur": 17, + "args": { + "External id": 3607,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3606 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790910682, "dur": 14, + "args": { + "External id": 3608,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3607 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910719, "dur": 5, + "args": { + "External id": 3609,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3608 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790910728, "dur": 22, + "args": { + "External id": 3610,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3609 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910729, "dur": 5, + "args": { + "External id": 3611,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3610 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790910735, "dur": 15, + "args": { + "External id": 3612,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3611 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790910737, "dur": 11, + "args": { + "External id": 3613,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3612 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910769, "dur": 5, + "args": { + "External id": 3614,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3613 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790910778, "dur": 22, + "args": { + "External id": 3615,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3614 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910779, "dur": 5, + "args": { + "External id": 3616,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3615 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790910785, "dur": 14, + "args": { + "External id": 3617,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3616 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790910787, "dur": 11, + "args": { + "External id": 3618,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3617 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910818, "dur": 5, + "args": { + "External id": 3619,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3618 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790910827, "dur": 21, + "args": { + "External id": 3620,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3619 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910828, "dur": 5, + "args": { + "External id": 3621,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3620 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790910834, "dur": 14, + "args": { + "External id": 3622,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3621 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790910835, "dur": 11, + "args": { + "External id": 3623,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3622 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910866, "dur": 6, + "args": { + "External id": 3624,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3623 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790910876, "dur": 21, + "args": { + "External id": 3625,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3624 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910877, "dur": 4, + "args": { + "External id": 3626,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3625 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790910882, "dur": 14, + "args": { + "External id": 3627,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3626 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790910884, "dur": 11, + "args": { + "External id": 3628,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3627 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910915, "dur": 5, + "args": { + "External id": 3629,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3628 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790910924, "dur": 21, + "args": { + "External id": 3630,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3629 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910926, "dur": 4, + "args": { + "External id": 3631,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3630 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790910931, "dur": 14, + "args": { + "External id": 3632,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3631 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790910933, "dur": 11, + "args": { + "External id": 3633,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3632 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910963, "dur": 5, + "args": { + "External id": 3634,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3633 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790910972, "dur": 21, + "args": { + "External id": 3635,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3634 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790910974, "dur": 4, + "args": { + "External id": 3636,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3635 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790910979, "dur": 14, + "args": { + "External id": 3637,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3636 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790910981, "dur": 11, + "args": { + "External id": 3638,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3637 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911012, "dur": 5, + "args": { + "External id": 3639,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3638 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790911021, "dur": 22, + "args": { + "External id": 3640,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3639 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911023, "dur": 4, + "args": { + "External id": 3641,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3640 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790911028, "dur": 15, + "args": { + "External id": 3642,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3641 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790911030, "dur": 11, + "args": { + "External id": 3643,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3642 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911061, "dur": 5, + "args": { + "External id": 3644,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3643 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790911071, "dur": 21, + "args": { + "External id": 3645,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3644 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911072, "dur": 4, + "args": { + "External id": 3646,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3645 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790911077, "dur": 14, + "args": { + "External id": 3647,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3646 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790911079, "dur": 11, + "args": { + "External id": 3648,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3647 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911110, "dur": 5, + "args": { + "External id": 3649,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3648 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790911119, "dur": 21, + "args": { + "External id": 3650,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3649 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911120, "dur": 4, + "args": { + "External id": 3651,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3650 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790911126, "dur": 14, + "args": { + "External id": 3652,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3651 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790911127, "dur": 11, + "args": { + "External id": 3653,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3652 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911158, "dur": 5, + "args": { + "External id": 3654,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3653 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790911167, "dur": 21, + "args": { + "External id": 3655,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3654 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911168, "dur": 5, + "args": { + "External id": 3656,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3655 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790911174, "dur": 14, + "args": { + "External id": 3657,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3656 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790911176, "dur": 10, + "args": { + "External id": 3658,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3657 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911206, "dur": 6, + "args": { + "External id": 3659,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3658 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790911216, "dur": 21, + "args": { + "External id": 3660,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3659 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911217, "dur": 4, + "args": { + "External id": 3661,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3660 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790911222, "dur": 14, + "args": { + "External id": 3662,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3661 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790911224, "dur": 11, + "args": { + "External id": 3663,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3662 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911255, "dur": 5, + "args": { + "External id": 3664,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3663 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790911265, "dur": 21, + "args": { + "External id": 3665,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3664 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911266, "dur": 4, + "args": { + "External id": 3666,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3665 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790911271, "dur": 15, + "args": { + "External id": 3667,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3666 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790911273, "dur": 11, + "args": { + "External id": 3668,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3667 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911304, "dur": 5, + "args": { + "External id": 3669,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3668 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790911313, "dur": 22, + "args": { + "External id": 3670,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3669 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911314, "dur": 5, + "args": { + "External id": 3671,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3670 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790911320, "dur": 14, + "args": { + "External id": 3672,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3671 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790911322, "dur": 11, + "args": { + "External id": 3673,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3672 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911352, "dur": 6, + "args": { + "External id": 3674,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3673 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790911362, "dur": 21, + "args": { + "External id": 3675,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3674 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911363, "dur": 4, + "args": { + "External id": 3676,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3675 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790911368, "dur": 15, + "args": { + "External id": 3677,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3676 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790911370, "dur": 11, + "args": { + "External id": 3678,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3677 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911401, "dur": 5, + "args": { + "External id": 3679,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3678 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790911410, "dur": 21, + "args": { + "External id": 3680,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3679 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911411, "dur": 5, + "args": { + "External id": 3681,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3680 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790911417, "dur": 14, + "args": { + "External id": 3682,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3681 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790911419, "dur": 11, + "args": { + "External id": 3683,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3682 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911449, "dur": 6, + "args": { + "External id": 3684,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3683 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790911459, "dur": 21, + "args": { + "External id": 3685,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3684 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911460, "dur": 4, + "args": { + "External id": 3686,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3685 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790911465, "dur": 14, + "args": { + "External id": 3687,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3686 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790911467, "dur": 11, + "args": { + "External id": 3688,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3687 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911498, "dur": 5, + "args": { + "External id": 3689,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3688 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790911507, "dur": 21, + "args": { + "External id": 3690,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3689 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911508, "dur": 5, + "args": { + "External id": 3691,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3690 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790911514, "dur": 14, + "args": { + "External id": 3692,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3691 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790911515, "dur": 12, + "args": { + "External id": 3693,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3692 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911546, "dur": 6, + "args": { + "External id": 3694,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3693 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790911556, "dur": 21, + "args": { + "External id": 3695,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3694 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911557, "dur": 4, + "args": { + "External id": 3696,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3695 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790911562, "dur": 15, + "args": { + "External id": 3697,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3696 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790911564, "dur": 11, + "args": { + "External id": 3698,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3697 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911595, "dur": 5, + "args": { + "External id": 3699,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3698 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790911604, "dur": 21, + "args": { + "External id": 3700,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3699 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911605, "dur": 5, + "args": { + "External id": 3701,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3700 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790911611, "dur": 14, + "args": { + "External id": 3702,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3701 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790911612, "dur": 11, + "args": { + "External id": 3703,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3702 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911643, "dur": 5, + "args": { + "External id": 3704,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3703 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790911652, "dur": 21, + "args": { + "External id": 3705,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3704 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911653, "dur": 4, + "args": { + "External id": 3706,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3705 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790911659, "dur": 14, + "args": { + "External id": 3707,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3706 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790911660, "dur": 11, + "args": { + "External id": 3708,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3707 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911691, "dur": 5, + "args": { + "External id": 3709,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3708 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790911700, "dur": 21, + "args": { + "External id": 3710,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3709 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911701, "dur": 4, + "args": { + "External id": 3711,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3710 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790911707, "dur": 14, + "args": { + "External id": 3712,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3711 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790911708, "dur": 11, + "args": { + "External id": 3713,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3712 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911739, "dur": 5, + "args": { + "External id": 3714,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3713 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790911748, "dur": 21, + "args": { + "External id": 3715,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3714 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911749, "dur": 5, + "args": { + "External id": 3716,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3715 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790911755, "dur": 14, + "args": { + "External id": 3717,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3716 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790911756, "dur": 12, + "args": { + "External id": 3718,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3717 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911787, "dur": 5, + "args": { + "External id": 3719,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3718 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790911796, "dur": 21, + "args": { + "External id": 3720,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3719 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911798, "dur": 4, + "args": { + "External id": 3721,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3720 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790911803, "dur": 14, + "args": { + "External id": 3722,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3721 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790911805, "dur": 11, + "args": { + "External id": 3723,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3722 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911836, "dur": 5, + "args": { + "External id": 3724,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3723 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790911845, "dur": 21, + "args": { + "External id": 3725,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3724 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911846, "dur": 5, + "args": { + "External id": 3726,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3725 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790911852, "dur": 14, + "args": { + "External id": 3727,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3726 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790911853, "dur": 11, + "args": { + "External id": 3728,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3727 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911884, "dur": 5, + "args": { + "External id": 3729,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3728 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790911893, "dur": 22, + "args": { + "External id": 3730,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3729 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911895, "dur": 4, + "args": { + "External id": 3731,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3730 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790911900, "dur": 14, + "args": { + "External id": 3732,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3731 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790911902, "dur": 11, + "args": { + "External id": 3733,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3732 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911933, "dur": 5, + "args": { + "External id": 3734,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3733 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790911942, "dur": 21, + "args": { + "External id": 3735,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3734 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911943, "dur": 4, + "args": { + "External id": 3736,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3735 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790911949, "dur": 14, + "args": { + "External id": 3737,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3736 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790911950, "dur": 11, + "args": { + "External id": 3738,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3737 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911981, "dur": 5, + "args": { + "External id": 3739,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3738 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790911990, "dur": 21, + "args": { + "External id": 3740,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3739 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790911991, "dur": 4, + "args": { + "External id": 3741,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3740 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790911997, "dur": 14, + "args": { + "External id": 3742,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3741 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790911998, "dur": 12, + "args": { + "External id": 3743,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3742 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912029, "dur": 5, + "args": { + "External id": 3744,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3743 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790912038, "dur": 21, + "args": { + "External id": 3745,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3744 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912040, "dur": 4, + "args": { + "External id": 3746,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3745 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790912045, "dur": 14, + "args": { + "External id": 3747,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3746 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790912047, "dur": 11, + "args": { + "External id": 3748,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3747 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912077, "dur": 5, + "args": { + "External id": 3749,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3748 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790912086, "dur": 22, + "args": { + "External id": 3750,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3749 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912087, "dur": 5, + "args": { + "External id": 3751,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3750 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790912093, "dur": 14, + "args": { + "External id": 3752,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3751 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790912095, "dur": 11, + "args": { + "External id": 3753,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3752 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912126, "dur": 5, + "args": { + "External id": 3754,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3753 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790912135, "dur": 22, + "args": { + "External id": 3755,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3754 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912136, "dur": 5, + "args": { + "External id": 3756,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3755 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790912142, "dur": 14, + "args": { + "External id": 3757,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3756 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790912144, "dur": 11, + "args": { + "External id": 3758,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3757 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912175, "dur": 5, + "args": { + "External id": 3759,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3758 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790912184, "dur": 21, + "args": { + "External id": 3760,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3759 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912185, "dur": 5, + "args": { + "External id": 3761,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3760 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790912191, "dur": 14, + "args": { + "External id": 3762,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3761 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790912192, "dur": 11, + "args": { + "External id": 3763,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3762 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912243, "dur": 6, + "args": { + "External id": 3764,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3763 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790912254, "dur": 23, + "args": { + "External id": 3765,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3764 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912255, "dur": 4, + "args": { + "External id": 3766,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3765 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790912261, "dur": 15, + "args": { + "External id": 3767,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3766 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790912262, "dur": 13, + "args": { + "External id": 3768,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3767 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912298, "dur": 5, + "args": { + "External id": 3769,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3768 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790912308, "dur": 21, + "args": { + "External id": 3770,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3769 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912309, "dur": 5, + "args": { + "External id": 3771,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3770 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790912315, "dur": 14, + "args": { + "External id": 3772,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3771 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790912317, "dur": 11, + "args": { + "External id": 3773,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3772 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912348, "dur": 5, + "args": { + "External id": 3774,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3773 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790912357, "dur": 21, + "args": { + "External id": 3775,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3774 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912358, "dur": 4, + "args": { + "External id": 3776,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3775 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790912364, "dur": 14, + "args": { + "External id": 3777,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3776 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790912365, "dur": 11, + "args": { + "External id": 3778,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3777 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912396, "dur": 5, + "args": { + "External id": 3779,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3778 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790912405, "dur": 22, + "args": { + "External id": 3780,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3779 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912406, "dur": 5, + "args": { + "External id": 3781,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3780 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790912412, "dur": 14, + "args": { + "External id": 3782,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3781 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790912414, "dur": 11, + "args": { + "External id": 3783,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3782 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912444, "dur": 6, + "args": { + "External id": 3784,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3783 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790912454, "dur": 21, + "args": { + "External id": 3785,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3784 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912455, "dur": 5, + "args": { + "External id": 3786,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3785 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790912461, "dur": 14, + "args": { + "External id": 3787,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3786 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790912463, "dur": 11, + "args": { + "External id": 3788,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3787 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912494, "dur": 5, + "args": { + "External id": 3789,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3788 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790912503, "dur": 21, + "args": { + "External id": 3790,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3789 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912504, "dur": 4, + "args": { + "External id": 3791,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3790 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790912510, "dur": 14, + "args": { + "External id": 3792,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3791 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790912511, "dur": 12, + "args": { + "External id": 3793,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3792 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912542, "dur": 5, + "args": { + "External id": 3794,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3793 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790912551, "dur": 21, + "args": { + "External id": 3795,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3794 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912552, "dur": 5, + "args": { + "External id": 3796,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3795 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790912558, "dur": 14, + "args": { + "External id": 3797,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3796 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790912560, "dur": 11, + "args": { + "External id": 3798,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3797 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912590, "dur": 5, + "args": { + "External id": 3799,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3798 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790912599, "dur": 21, + "args": { + "External id": 3800,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3799 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912600, "dur": 5, + "args": { + "External id": 3801,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3800 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790912606, "dur": 14, + "args": { + "External id": 3802,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3801 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790912608, "dur": 10, + "args": { + "External id": 3803,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3802 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912638, "dur": 5, + "args": { + "External id": 3804,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3803 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790912647, "dur": 22, + "args": { + "External id": 3805,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3804 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912648, "dur": 5, + "args": { + "External id": 3806,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3805 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790912654, "dur": 14, + "args": { + "External id": 3807,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3806 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790912656, "dur": 11, + "args": { + "External id": 3808,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3807 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912687, "dur": 5, + "args": { + "External id": 3809,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3808 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790912696, "dur": 21, + "args": { + "External id": 3810,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3809 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912697, "dur": 4, + "args": { + "External id": 3811,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3810 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790912703, "dur": 14, + "args": { + "External id": 3812,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3811 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790912704, "dur": 12, + "args": { + "External id": 3813,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3812 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912735, "dur": 5, + "args": { + "External id": 3814,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3813 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790912744, "dur": 21, + "args": { + "External id": 3815,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3814 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912745, "dur": 5, + "args": { + "External id": 3816,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3815 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790912751, "dur": 14, + "args": { + "External id": 3817,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3816 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790912753, "dur": 11, + "args": { + "External id": 3818,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3817 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912783, "dur": 5, + "args": { + "External id": 3819,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3818 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790912792, "dur": 22, + "args": { + "External id": 3820,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3819 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912793, "dur": 5, + "args": { + "External id": 3821,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3820 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790912799, "dur": 14, + "args": { + "External id": 3822,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3821 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790912801, "dur": 11, + "args": { + "External id": 3823,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3822 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912832, "dur": 5, + "args": { + "External id": 3824,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3823 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790912840, "dur": 22, + "args": { + "External id": 3825,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3824 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912842, "dur": 4, + "args": { + "External id": 3826,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3825 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790912847, "dur": 14, + "args": { + "External id": 3827,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3826 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790912849, "dur": 11, + "args": { + "External id": 3828,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3827 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912879, "dur": 5, + "args": { + "External id": 3829,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3828 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790912888, "dur": 21, + "args": { + "External id": 3830,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3829 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912889, "dur": 5, + "args": { + "External id": 3831,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3830 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790912895, "dur": 14, + "args": { + "External id": 3832,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3831 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790912897, "dur": 11, + "args": { + "External id": 3833,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3832 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912927, "dur": 5, + "args": { + "External id": 3834,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3833 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790912936, "dur": 22, + "args": { + "External id": 3835,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3834 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912938, "dur": 4, + "args": { + "External id": 3836,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3835 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790912943, "dur": 14, + "args": { + "External id": 3837,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3836 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790912945, "dur": 11, + "args": { + "External id": 3838,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3837 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912976, "dur": 5, + "args": { + "External id": 3839,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3838 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790912985, "dur": 21, + "args": { + "External id": 3840,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3839 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790912986, "dur": 5, + "args": { + "External id": 3841,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3840 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790912992, "dur": 14, + "args": { + "External id": 3842,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3841 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790912994, "dur": 11, + "args": { + "External id": 3843,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3842 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913024, "dur": 5, + "args": { + "External id": 3844,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3843 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790913033, "dur": 21, + "args": { + "External id": 3845,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3844 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913034, "dur": 5, + "args": { + "External id": 3846,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3845 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790913040, "dur": 14, + "args": { + "External id": 3847,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3846 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790913041, "dur": 12, + "args": { + "External id": 3848,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3847 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913072, "dur": 6, + "args": { + "External id": 3849,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3848 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790913082, "dur": 21, + "args": { + "External id": 3850,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3849 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913083, "dur": 4, + "args": { + "External id": 3851,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3850 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790913088, "dur": 15, + "args": { + "External id": 3852,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3851 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790913090, "dur": 11, + "args": { + "External id": 3853,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3852 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913182, "dur": 7, + "args": { + "External id": 3854,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3853 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790913194, "dur": 26, + "args": { + "External id": 3855,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3854 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913195, "dur": 5, + "args": { + "External id": 3856,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3855 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790913201, "dur": 19, + "args": { + "External id": 3857,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3856 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790913203, "dur": 15, + "args": { + "External id": 3858,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3857 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913241, "dur": 6, + "args": { + "External id": 3859,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3858 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790913251, "dur": 22, + "args": { + "External id": 3860,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3859 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913252, "dur": 5, + "args": { + "External id": 3861,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3860 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790913258, "dur": 15, + "args": { + "External id": 3862,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3861 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790913259, "dur": 12, + "args": { + "External id": 3863,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3862 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913292, "dur": 5, + "args": { + "External id": 3864,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3863 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790913301, "dur": 21, + "args": { + "External id": 3865,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3864 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913302, "dur": 5, + "args": { + "External id": 3866,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3865 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790913308, "dur": 14, + "args": { + "External id": 3867,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3866 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790913309, "dur": 12, + "args": { + "External id": 3868,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3867 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913341, "dur": 5, + "args": { + "External id": 3869,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3868 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790913350, "dur": 21, + "args": { + "External id": 3870,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3869 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913351, "dur": 4, + "args": { + "External id": 3871,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3870 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790913357, "dur": 14, + "args": { + "External id": 3872,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3871 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790913358, "dur": 11, + "args": { + "External id": 3873,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3872 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913389, "dur": 5, + "args": { + "External id": 3874,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3873 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790913398, "dur": 22, + "args": { + "External id": 3875,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3874 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913399, "dur": 5, + "args": { + "External id": 3876,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3875 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790913405, "dur": 14, + "args": { + "External id": 3877,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3876 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790913407, "dur": 11, + "args": { + "External id": 3878,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3877 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913438, "dur": 5, + "args": { + "External id": 3879,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3878 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790913447, "dur": 22, + "args": { + "External id": 3880,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3879 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913449, "dur": 4, + "args": { + "External id": 3881,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3880 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790913454, "dur": 15, + "args": { + "External id": 3882,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3881 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790913456, "dur": 12, + "args": { + "External id": 3883,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3882 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913487, "dur": 5, + "args": { + "External id": 3884,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3883 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790913496, "dur": 22, + "args": { + "External id": 3885,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3884 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913497, "dur": 5, + "args": { + "External id": 3886,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3885 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790913503, "dur": 14, + "args": { + "External id": 3887,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3886 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790913505, "dur": 11, + "args": { + "External id": 3888,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3887 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913536, "dur": 5, + "args": { + "External id": 3889,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3888 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790913545, "dur": 21, + "args": { + "External id": 3890,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3889 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913546, "dur": 4, + "args": { + "External id": 3891,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3890 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790913551, "dur": 15, + "args": { + "External id": 3892,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3891 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790913553, "dur": 11, + "args": { + "External id": 3893,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3892 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913584, "dur": 5, + "args": { + "External id": 3894,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3893 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790913593, "dur": 21, + "args": { + "External id": 3895,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3894 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913594, "dur": 5, + "args": { + "External id": 3896,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3895 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790913600, "dur": 14, + "args": { + "External id": 3897,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3896 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790913601, "dur": 12, + "args": { + "External id": 3898,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3897 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913632, "dur": 5, + "args": { + "External id": 3899,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3898 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790913641, "dur": 21, + "args": { + "External id": 3900,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3899 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913642, "dur": 4, + "args": { + "External id": 3901,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3900 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790913648, "dur": 14, + "args": { + "External id": 3902,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3901 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790913649, "dur": 12, + "args": { + "External id": 3903,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3902 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913730, "dur": 7, + "args": { + "External id": 3904,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3903 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790913741, "dur": 25, + "args": { + "External id": 3905,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3904 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913742, "dur": 5, + "args": { + "External id": 3906,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3905 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790913748, "dur": 17, + "args": { + "External id": 3907,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3906 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790913750, "dur": 14, + "args": { + "External id": 3908,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3907 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913786, "dur": 6, + "args": { + "External id": 3909,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3908 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790913796, "dur": 21, + "args": { + "External id": 3910,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3909 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913797, "dur": 4, + "args": { + "External id": 3911,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3910 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790913803, "dur": 14, + "args": { + "External id": 3912,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3911 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790913804, "dur": 12, + "args": { + "External id": 3913,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3912 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913836, "dur": 5, + "args": { + "External id": 3914,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3913 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790913845, "dur": 21, + "args": { + "External id": 3915,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3914 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913846, "dur": 4, + "args": { + "External id": 3916,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3915 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790913852, "dur": 14, + "args": { + "External id": 3917,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3916 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790913853, "dur": 12, + "args": { + "External id": 3918,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3917 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913884, "dur": 5, + "args": { + "External id": 3919,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3918 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790913893, "dur": 22, + "args": { + "External id": 3920,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3919 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913894, "dur": 5, + "args": { + "External id": 3921,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3920 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790913900, "dur": 14, + "args": { + "External id": 3922,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3921 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790913902, "dur": 11, + "args": { + "External id": 3923,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3922 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913933, "dur": 5, + "args": { + "External id": 3924,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3923 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790913942, "dur": 22, + "args": { + "External id": 3925,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3924 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913943, "dur": 5, + "args": { + "External id": 3926,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3925 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790913949, "dur": 14, + "args": { + "External id": 3927,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3926 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790913951, "dur": 11, + "args": { + "External id": 3928,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3927 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913982, "dur": 5, + "args": { + "External id": 3929,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3928 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790913991, "dur": 21, + "args": { + "External id": 3930,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3929 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790913992, "dur": 5, + "args": { + "External id": 3931,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3930 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790913998, "dur": 14, + "args": { + "External id": 3932,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3931 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790914000, "dur": 11, + "args": { + "External id": 3933,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3932 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914030, "dur": 6, + "args": { + "External id": 3934,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3933 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790914040, "dur": 21, + "args": { + "External id": 3935,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3934 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914041, "dur": 5, + "args": { + "External id": 3936,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3935 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790914047, "dur": 14, + "args": { + "External id": 3937,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3936 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790914048, "dur": 12, + "args": { + "External id": 3938,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3937 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914080, "dur": 5, + "args": { + "External id": 3939,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3938 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790914089, "dur": 22, + "args": { + "External id": 3940,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3939 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914090, "dur": 5, + "args": { + "External id": 3941,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3940 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790914096, "dur": 14, + "args": { + "External id": 3942,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3941 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790914098, "dur": 11, + "args": { + "External id": 3943,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3942 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914129, "dur": 5, + "args": { + "External id": 3944,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3943 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790914138, "dur": 21, + "args": { + "External id": 3945,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3944 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914139, "dur": 5, + "args": { + "External id": 3946,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3945 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790914145, "dur": 14, + "args": { + "External id": 3947,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3946 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790914147, "dur": 10, + "args": { + "External id": 3948,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3947 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914177, "dur": 5, + "args": { + "External id": 3949,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3948 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790914186, "dur": 22, + "args": { + "External id": 3950,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3949 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914187, "dur": 5, + "args": { + "External id": 3951,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3950 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790914193, "dur": 14, + "args": { + "External id": 3952,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3951 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790914195, "dur": 11, + "args": { + "External id": 3953,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3952 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914225, "dur": 6, + "args": { + "External id": 3954,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3953 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790914235, "dur": 21, + "args": { + "External id": 3955,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3954 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914236, "dur": 4, + "args": { + "External id": 3956,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3955 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790914241, "dur": 15, + "args": { + "External id": 3957,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3956 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790914243, "dur": 12, + "args": { + "External id": 3958,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3957 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914274, "dur": 6, + "args": { + "External id": 3959,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3958 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790914284, "dur": 21, + "args": { + "External id": 3960,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3959 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914285, "dur": 4, + "args": { + "External id": 3961,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3960 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790914291, "dur": 14, + "args": { + "External id": 3962,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3961 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790914292, "dur": 11, + "args": { + "External id": 3963,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3962 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914323, "dur": 5, + "args": { + "External id": 3964,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3963 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790914332, "dur": 22, + "args": { + "External id": 3965,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3964 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914333, "dur": 5, + "args": { + "External id": 3966,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3965 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790914339, "dur": 14, + "args": { + "External id": 3967,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3966 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790914341, "dur": 11, + "args": { + "External id": 3968,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3967 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914371, "dur": 6, + "args": { + "External id": 3969,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3968 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790914381, "dur": 21, + "args": { + "External id": 3970,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3969 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914382, "dur": 4, + "args": { + "External id": 3971,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3970 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790914387, "dur": 15, + "args": { + "External id": 3972,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3971 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790914389, "dur": 11, + "args": { + "External id": 3973,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3972 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914420, "dur": 5, + "args": { + "External id": 3974,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3973 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790914430, "dur": 21, + "args": { + "External id": 3975,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3974 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914431, "dur": 4, + "args": { + "External id": 3976,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3975 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790914437, "dur": 14, + "args": { + "External id": 3977,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3976 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790914438, "dur": 11, + "args": { + "External id": 3978,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3977 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914469, "dur": 5, + "args": { + "External id": 3979,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3978 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790914478, "dur": 21, + "args": { + "External id": 3980,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3979 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914479, "dur": 5, + "args": { + "External id": 3981,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3980 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790914485, "dur": 14, + "args": { + "External id": 3982,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3981 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790914486, "dur": 11, + "args": { + "External id": 3983,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3982 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914517, "dur": 5, + "args": { + "External id": 3984,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3983 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790914526, "dur": 21, + "args": { + "External id": 3985,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3984 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914527, "dur": 4, + "args": { + "External id": 3986,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3985 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790914533, "dur": 14, + "args": { + "External id": 3987,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3986 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790914534, "dur": 12, + "args": { + "External id": 3988,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3987 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914565, "dur": 5, + "args": { + "External id": 3989,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3988 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790914574, "dur": 22, + "args": { + "External id": 3990,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3989 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914575, "dur": 5, + "args": { + "External id": 3991,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3990 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790914581, "dur": 14, + "args": { + "External id": 3992,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3991 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790914583, "dur": 11, + "args": { + "External id": 3993,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3992 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914614, "dur": 5, + "args": { + "External id": 3994,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3993 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790914623, "dur": 21, + "args": { + "External id": 3995,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3994 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914624, "dur": 4, + "args": { + "External id": 3996,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3995 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790914629, "dur": 15, + "args": { + "External id": 3997,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 3996 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790914631, "dur": 11, + "args": { + "External id": 3998,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 3997 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914662, "dur": 5, + "args": { + "External id": 3999,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3998 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790914671, "dur": 21, + "args": { + "External id": 4000,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 3999 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914672, "dur": 5, + "args": { + "External id": 4001,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4000 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790914678, "dur": 14, + "args": { + "External id": 4002,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4001 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790914680, "dur": 11, + "args": { + "External id": 4003,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4002 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914710, "dur": 6, + "args": { + "External id": 4004,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4003 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790914720, "dur": 21, + "args": { + "External id": 4005,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4004 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914721, "dur": 4, + "args": { + "External id": 4006,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4005 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790914726, "dur": 14, + "args": { + "External id": 4007,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4006 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790914728, "dur": 11, + "args": { + "External id": 4008,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4007 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914758, "dur": 6, + "args": { + "External id": 4009,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4008 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790914768, "dur": 21, + "args": { + "External id": 4010,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4009 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914769, "dur": 4, + "args": { + "External id": 4011,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4010 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790914774, "dur": 14, + "args": { + "External id": 4012,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4011 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790914776, "dur": 11, + "args": { + "External id": 4013,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4012 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914807, "dur": 5, + "args": { + "External id": 4014,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4013 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790914816, "dur": 21, + "args": { + "External id": 4015,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4014 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914817, "dur": 4, + "args": { + "External id": 4016,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4015 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790914823, "dur": 14, + "args": { + "External id": 4017,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4016 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790914824, "dur": 11, + "args": { + "External id": 4018,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4017 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914855, "dur": 5, + "args": { + "External id": 4019,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4018 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790914864, "dur": 21, + "args": { + "External id": 4020,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4019 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914865, "dur": 5, + "args": { + "External id": 4021,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4020 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790914871, "dur": 14, + "args": { + "External id": 4022,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4021 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790914872, "dur": 11, + "args": { + "External id": 4023,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4022 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914903, "dur": 5, + "args": { + "External id": 4024,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4023 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790914912, "dur": 21, + "args": { + "External id": 4025,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4024 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914913, "dur": 5, + "args": { + "External id": 4026,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4025 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790914919, "dur": 14, + "args": { + "External id": 4027,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4026 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790914921, "dur": 11, + "args": { + "External id": 4028,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4027 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914951, "dur": 5, + "args": { + "External id": 4029,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4028 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790914960, "dur": 21, + "args": { + "External id": 4030,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4029 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914961, "dur": 5, + "args": { + "External id": 4031,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4030 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790914967, "dur": 14, + "args": { + "External id": 4032,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4031 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790914969, "dur": 11, + "args": { + "External id": 4033,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4032 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790914999, "dur": 5, + "args": { + "External id": 4034,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4033 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790915008, "dur": 21, + "args": { + "External id": 4035,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4034 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915010, "dur": 4, + "args": { + "External id": 4036,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4035 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790915015, "dur": 14, + "args": { + "External id": 4037,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4036 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790915017, "dur": 11, + "args": { + "External id": 4038,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4037 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915047, "dur": 5, + "args": { + "External id": 4039,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4038 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790915056, "dur": 21, + "args": { + "External id": 4040,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4039 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915057, "dur": 5, + "args": { + "External id": 4041,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4040 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790915063, "dur": 14, + "args": { + "External id": 4042,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4041 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790915065, "dur": 11, + "args": { + "External id": 4043,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4042 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915098, "dur": 7, + "args": { + "External id": 4044,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4043 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790915109, "dur": 22, + "args": { + "External id": 4045,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4044 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915111, "dur": 4, + "args": { + "External id": 4046,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4045 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790915116, "dur": 15, + "args": { + "External id": 4047,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4046 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790915118, "dur": 11, + "args": { + "External id": 4048,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4047 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915149, "dur": 5, + "args": { + "External id": 4049,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4048 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790915158, "dur": 21, + "args": { + "External id": 4050,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4049 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915159, "dur": 5, + "args": { + "External id": 4051,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4050 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790915165, "dur": 14, + "args": { + "External id": 4052,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4051 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790915167, "dur": 11, + "args": { + "External id": 4053,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4052 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915197, "dur": 5, + "args": { + "External id": 4054,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4053 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790915206, "dur": 22, + "args": { + "External id": 4055,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4054 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915208, "dur": 4, + "args": { + "External id": 4056,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4055 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790915213, "dur": 14, + "args": { + "External id": 4057,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4056 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790915215, "dur": 11, + "args": { + "External id": 4058,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4057 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915246, "dur": 5, + "args": { + "External id": 4059,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4058 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790915255, "dur": 21, + "args": { + "External id": 4060,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4059 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915256, "dur": 4, + "args": { + "External id": 4061,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4060 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790915262, "dur": 14, + "args": { + "External id": 4062,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4061 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790915263, "dur": 11, + "args": { + "External id": 4063,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4062 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915294, "dur": 6, + "args": { + "External id": 4064,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4063 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790915304, "dur": 21, + "args": { + "External id": 4065,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4064 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915305, "dur": 4, + "args": { + "External id": 4066,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4065 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790915310, "dur": 15, + "args": { + "External id": 4067,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4066 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790915312, "dur": 11, + "args": { + "External id": 4068,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4067 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915343, "dur": 5, + "args": { + "External id": 4069,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4068 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790915352, "dur": 21, + "args": { + "External id": 4070,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4069 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915353, "dur": 5, + "args": { + "External id": 4071,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4070 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790915359, "dur": 14, + "args": { + "External id": 4072,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4071 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790915361, "dur": 11, + "args": { + "External id": 4073,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4072 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915392, "dur": 5, + "args": { + "External id": 4074,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4073 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790915401, "dur": 21, + "args": { + "External id": 4075,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4074 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915402, "dur": 5, + "args": { + "External id": 4076,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4075 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790915408, "dur": 14, + "args": { + "External id": 4077,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4076 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790915409, "dur": 12, + "args": { + "External id": 4078,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4077 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915440, "dur": 5, + "args": { + "External id": 4079,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4078 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790915449, "dur": 21, + "args": { + "External id": 4080,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4079 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915450, "dur": 5, + "args": { + "External id": 4081,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4080 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790915456, "dur": 14, + "args": { + "External id": 4082,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4081 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790915457, "dur": 11, + "args": { + "External id": 4083,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4082 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915487, "dur": 5, + "args": { + "External id": 4084,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4083 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790915496, "dur": 21, + "args": { + "External id": 4085,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4084 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915497, "dur": 5, + "args": { + "External id": 4086,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4085 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790915503, "dur": 14, + "args": { + "External id": 4087,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4086 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790915505, "dur": 11, + "args": { + "External id": 4088,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4087 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915535, "dur": 5, + "args": { + "External id": 4089,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4088 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790915544, "dur": 21, + "args": { + "External id": 4090,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4089 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915545, "dur": 5, + "args": { + "External id": 4091,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4090 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790915551, "dur": 14, + "args": { + "External id": 4092,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4091 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790915553, "dur": 11, + "args": { + "External id": 4093,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4092 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915583, "dur": 6, + "args": { + "External id": 4094,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4093 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790915593, "dur": 31, + "args": { + "External id": 4095,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4094 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915594, "dur": 4, + "args": { + "External id": 4096,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4095 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790915609, "dur": 15, + "args": { + "External id": 4097,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4096 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790915611, "dur": 11, + "args": { + "External id": 4098,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4097 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915642, "dur": 5, + "args": { + "External id": 4099,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4098 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790915652, "dur": 21, + "args": { + "External id": 4100,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4099 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915653, "dur": 4, + "args": { + "External id": 4101,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4100 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790915658, "dur": 14, + "args": { + "External id": 4102,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4101 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790915660, "dur": 11, + "args": { + "External id": 4103,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4102 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915691, "dur": 5, + "args": { + "External id": 4104,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4103 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790915700, "dur": 21, + "args": { + "External id": 4105,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4104 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915701, "dur": 5, + "args": { + "External id": 4106,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4105 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790915707, "dur": 14, + "args": { + "External id": 4107,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4106 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790915708, "dur": 12, + "args": { + "External id": 4108,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4107 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915740, "dur": 6, + "args": { + "External id": 4109,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4108 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790915750, "dur": 21, + "args": { + "External id": 4110,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4109 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915751, "dur": 4, + "args": { + "External id": 4111,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4110 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790915756, "dur": 15, + "args": { + "External id": 4112,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4111 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790915758, "dur": 11, + "args": { + "External id": 4113,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4112 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915789, "dur": 5, + "args": { + "External id": 4114,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4113 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790915798, "dur": 21, + "args": { + "External id": 4115,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4114 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915799, "dur": 5, + "args": { + "External id": 4116,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4115 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790915805, "dur": 14, + "args": { + "External id": 4117,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4116 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790915807, "dur": 11, + "args": { + "External id": 4118,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4117 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915837, "dur": 6, + "args": { + "External id": 4119,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4118 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790915847, "dur": 21, + "args": { + "External id": 4120,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4119 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915848, "dur": 4, + "args": { + "External id": 4121,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4120 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790915853, "dur": 14, + "args": { + "External id": 4122,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4121 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790915855, "dur": 11, + "args": { + "External id": 4123,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4122 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915886, "dur": 5, + "args": { + "External id": 4124,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4123 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790915895, "dur": 21, + "args": { + "External id": 4125,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4124 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915896, "dur": 4, + "args": { + "External id": 4126,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4125 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790915902, "dur": 14, + "args": { + "External id": 4127,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4126 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790915903, "dur": 12, + "args": { + "External id": 4128,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4127 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915934, "dur": 6, + "args": { + "External id": 4129,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4128 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790915943, "dur": 22, + "args": { + "External id": 4130,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4129 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915945, "dur": 4, + "args": { + "External id": 4131,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4130 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790915950, "dur": 14, + "args": { + "External id": 4132,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4131 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790915952, "dur": 11, + "args": { + "External id": 4133,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4132 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915983, "dur": 5, + "args": { + "External id": 4134,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4133 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790915992, "dur": 22, + "args": { + "External id": 4135,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4134 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790915993, "dur": 5, + "args": { + "External id": 4136,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4135 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790915999, "dur": 14, + "args": { + "External id": 4137,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4136 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790916001, "dur": 11, + "args": { + "External id": 4138,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4137 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916032, "dur": 5, + "args": { + "External id": 4139,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4138 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790916041, "dur": 22, + "args": { + "External id": 4140,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4139 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916042, "dur": 5, + "args": { + "External id": 4141,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4140 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790916048, "dur": 14, + "args": { + "External id": 4142,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4141 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790916049, "dur": 12, + "args": { + "External id": 4143,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4142 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916081, "dur": 5, + "args": { + "External id": 4144,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4143 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790916090, "dur": 21, + "args": { + "External id": 4145,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4144 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916091, "dur": 5, + "args": { + "External id": 4146,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4145 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790916097, "dur": 14, + "args": { + "External id": 4147,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4146 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790916098, "dur": 12, + "args": { + "External id": 4148,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4147 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916129, "dur": 5, + "args": { + "External id": 4149,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4148 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790916138, "dur": 21, + "args": { + "External id": 4150,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4149 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916139, "dur": 5, + "args": { + "External id": 4151,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4150 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790916145, "dur": 14, + "args": { + "External id": 4152,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4151 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790916146, "dur": 11, + "args": { + "External id": 4153,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4152 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916256, "dur": 7, + "args": { + "External id": 4154,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4153 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790916269, "dur": 25, + "args": { + "External id": 4155,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4154 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916270, "dur": 4, + "args": { + "External id": 4156,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4155 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790916276, "dur": 18, + "args": { + "External id": 4157,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4156 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790916278, "dur": 15, + "args": { + "External id": 4158,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4157 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916316, "dur": 5, + "args": { + "External id": 4159,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4158 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790916325, "dur": 22, + "args": { + "External id": 4160,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4159 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916326, "dur": 5, + "args": { + "External id": 4161,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4160 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790916332, "dur": 14, + "args": { + "External id": 4162,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4161 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790916334, "dur": 11, + "args": { + "External id": 4163,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4162 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916365, "dur": 5, + "args": { + "External id": 4164,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4163 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790916374, "dur": 22, + "args": { + "External id": 4165,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4164 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916375, "dur": 5, + "args": { + "External id": 4166,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4165 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790916381, "dur": 15, + "args": { + "External id": 4167,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4166 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790916383, "dur": 11, + "args": { + "External id": 4168,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4167 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916414, "dur": 5, + "args": { + "External id": 4169,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4168 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790916423, "dur": 21, + "args": { + "External id": 4170,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4169 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916424, "dur": 5, + "args": { + "External id": 4171,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4170 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790916430, "dur": 14, + "args": { + "External id": 4172,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4171 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790916431, "dur": 12, + "args": { + "External id": 4173,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4172 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916462, "dur": 5, + "args": { + "External id": 4174,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4173 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790916471, "dur": 22, + "args": { + "External id": 4175,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4174 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916472, "dur": 5, + "args": { + "External id": 4176,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4175 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790916478, "dur": 14, + "args": { + "External id": 4177,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4176 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790916480, "dur": 11, + "args": { + "External id": 4178,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4177 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916511, "dur": 5, + "args": { + "External id": 4179,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4178 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790916520, "dur": 21, + "args": { + "External id": 4180,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4179 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916521, "dur": 5, + "args": { + "External id": 4181,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4180 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790916527, "dur": 14, + "args": { + "External id": 4182,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4181 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790916529, "dur": 11, + "args": { + "External id": 4183,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4182 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916559, "dur": 6, + "args": { + "External id": 4184,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4183 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790916568, "dur": 21, + "args": { + "External id": 4185,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4184 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916570, "dur": 4, + "args": { + "External id": 4186,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4185 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790916575, "dur": 14, + "args": { + "External id": 4187,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4186 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790916577, "dur": 11, + "args": { + "External id": 4188,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4187 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916607, "dur": 5, + "args": { + "External id": 4189,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4188 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790916616, "dur": 22, + "args": { + "External id": 4190,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4189 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916618, "dur": 4, + "args": { + "External id": 4191,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4190 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790916623, "dur": 14, + "args": { + "External id": 4192,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4191 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790916625, "dur": 11, + "args": { + "External id": 4193,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4192 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916656, "dur": 5, + "args": { + "External id": 4194,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4193 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790916665, "dur": 21, + "args": { + "External id": 4195,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4194 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916666, "dur": 4, + "args": { + "External id": 4196,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4195 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790916671, "dur": 14, + "args": { + "External id": 4197,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4196 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790916673, "dur": 11, + "args": { + "External id": 4198,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4197 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916703, "dur": 6, + "args": { + "External id": 4199,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4198 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790916713, "dur": 21, + "args": { + "External id": 4200,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4199 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916714, "dur": 4, + "args": { + "External id": 4201,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4200 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790916719, "dur": 14, + "args": { + "External id": 4202,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4201 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790916721, "dur": 11, + "args": { + "External id": 4203,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4202 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916802, "dur": 6, + "args": { + "External id": 4204,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4203 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790916813, "dur": 24, + "args": { + "External id": 4205,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4204 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916814, "dur": 5, + "args": { + "External id": 4206,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4205 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790916820, "dur": 17, + "args": { + "External id": 4207,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4206 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790916822, "dur": 14, + "args": { + "External id": 4208,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4207 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916858, "dur": 5, + "args": { + "External id": 4209,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4208 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790916867, "dur": 22, + "args": { + "External id": 4210,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4209 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916868, "dur": 5, + "args": { + "External id": 4211,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4210 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790916874, "dur": 15, + "args": { + "External id": 4212,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4211 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790916876, "dur": 11, + "args": { + "External id": 4213,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4212 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916908, "dur": 5, + "args": { + "External id": 4214,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4213 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790916917, "dur": 21, + "args": { + "External id": 4215,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4214 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916918, "dur": 4, + "args": { + "External id": 4216,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4215 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790916923, "dur": 15, + "args": { + "External id": 4217,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4216 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790916925, "dur": 12, + "args": { + "External id": 4218,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4217 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916956, "dur": 5, + "args": { + "External id": 4219,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4218 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790916965, "dur": 22, + "args": { + "External id": 4220,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4219 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790916967, "dur": 4, + "args": { + "External id": 4221,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4220 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790916972, "dur": 14, + "args": { + "External id": 4222,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4221 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790916974, "dur": 11, + "args": { + "External id": 4223,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4222 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917004, "dur": 6, + "args": { + "External id": 4224,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4223 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790917014, "dur": 21, + "args": { + "External id": 4225,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4224 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917015, "dur": 4, + "args": { + "External id": 4226,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4225 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790917020, "dur": 15, + "args": { + "External id": 4227,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4226 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790917022, "dur": 11, + "args": { + "External id": 4228,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4227 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917053, "dur": 5, + "args": { + "External id": 4229,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4228 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790917062, "dur": 21, + "args": { + "External id": 4230,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4229 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917063, "dur": 5, + "args": { + "External id": 4231,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4230 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790917069, "dur": 14, + "args": { + "External id": 4232,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4231 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790917070, "dur": 12, + "args": { + "External id": 4233,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4232 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917101, "dur": 5, + "args": { + "External id": 4234,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4233 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790917110, "dur": 22, + "args": { + "External id": 4235,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4234 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917112, "dur": 4, + "args": { + "External id": 4236,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4235 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790917118, "dur": 14, + "args": { + "External id": 4237,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4236 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790917119, "dur": 11, + "args": { + "External id": 4238,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4237 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917150, "dur": 5, + "args": { + "External id": 4239,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4238 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790917159, "dur": 21, + "args": { + "External id": 4240,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4239 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917160, "dur": 4, + "args": { + "External id": 4241,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4240 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790917166, "dur": 14, + "args": { + "External id": 4242,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4241 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790917167, "dur": 11, + "args": { + "External id": 4243,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4242 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917197, "dur": 6, + "args": { + "External id": 4244,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4243 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790917207, "dur": 21, + "args": { + "External id": 4245,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4244 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917208, "dur": 4, + "args": { + "External id": 4246,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4245 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790917214, "dur": 14, + "args": { + "External id": 4247,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4246 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790917215, "dur": 11, + "args": { + "External id": 4248,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4247 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917246, "dur": 5, + "args": { + "External id": 4249,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4248 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790917255, "dur": 21, + "args": { + "External id": 4250,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4249 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917256, "dur": 5, + "args": { + "External id": 4251,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4250 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790917262, "dur": 14, + "args": { + "External id": 4252,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4251 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790917264, "dur": 11, + "args": { + "External id": 4253,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4252 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917294, "dur": 6, + "args": { + "External id": 4254,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4253 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790917304, "dur": 21, + "args": { + "External id": 4255,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4254 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917305, "dur": 4, + "args": { + "External id": 4256,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4255 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790917310, "dur": 15, + "args": { + "External id": 4257,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4256 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790917312, "dur": 11, + "args": { + "External id": 4258,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4257 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917343, "dur": 5, + "args": { + "External id": 4259,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4258 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790917352, "dur": 22, + "args": { + "External id": 4260,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4259 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917353, "dur": 5, + "args": { + "External id": 4261,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4260 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790917359, "dur": 14, + "args": { + "External id": 4262,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4261 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790917361, "dur": 11, + "args": { + "External id": 4263,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4262 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917391, "dur": 6, + "args": { + "External id": 4264,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4263 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790917401, "dur": 21, + "args": { + "External id": 4265,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4264 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917402, "dur": 4, + "args": { + "External id": 4266,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4265 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790917407, "dur": 15, + "args": { + "External id": 4267,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4266 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790917409, "dur": 11, + "args": { + "External id": 4268,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4267 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917440, "dur": 5, + "args": { + "External id": 4269,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4268 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790917449, "dur": 21, + "args": { + "External id": 4270,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4269 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917450, "dur": 5, + "args": { + "External id": 4271,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4270 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790917456, "dur": 14, + "args": { + "External id": 4272,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4271 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790917458, "dur": 11, + "args": { + "External id": 4273,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4272 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917488, "dur": 5, + "args": { + "External id": 4274,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4273 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790917497, "dur": 21, + "args": { + "External id": 4275,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4274 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917498, "dur": 5, + "args": { + "External id": 4276,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4275 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790917504, "dur": 14, + "args": { + "External id": 4277,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4276 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790917505, "dur": 11, + "args": { + "External id": 4278,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4277 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917536, "dur": 5, + "args": { + "External id": 4279,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4278 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790917545, "dur": 24, + "args": { + "External id": 4280,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4279 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917548, "dur": 5, + "args": { + "External id": 4281,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4280 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790917554, "dur": 14, + "args": { + "External id": 4282,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4281 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790917555, "dur": 12, + "args": { + "External id": 4283,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4282 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917587, "dur": 5, + "args": { + "External id": 4284,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4283 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790917596, "dur": 21, + "args": { + "External id": 4285,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4284 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917597, "dur": 5, + "args": { + "External id": 4286,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4285 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790917603, "dur": 14, + "args": { + "External id": 4287,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4286 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790917604, "dur": 12, + "args": { + "External id": 4288,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4287 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917635, "dur": 5, + "args": { + "External id": 4289,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4288 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790917644, "dur": 21, + "args": { + "External id": 4290,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4289 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917645, "dur": 5, + "args": { + "External id": 4291,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4290 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790917651, "dur": 14, + "args": { + "External id": 4292,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4291 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790917652, "dur": 12, + "args": { + "External id": 4293,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4292 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917683, "dur": 6, + "args": { + "External id": 4294,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4293 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790917692, "dur": 22, + "args": { + "External id": 4295,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4294 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917694, "dur": 4, + "args": { + "External id": 4296,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4295 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790917699, "dur": 14, + "args": { + "External id": 4297,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4296 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790917701, "dur": 11, + "args": { + "External id": 4298,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4297 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917732, "dur": 5, + "args": { + "External id": 4299,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4298 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790917741, "dur": 22, + "args": { + "External id": 4300,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4299 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917742, "dur": 5, + "args": { + "External id": 4301,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4300 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790917748, "dur": 14, + "args": { + "External id": 4302,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4301 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790917750, "dur": 11, + "args": { + "External id": 4303,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4302 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917781, "dur": 5, + "args": { + "External id": 4304,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4303 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790917790, "dur": 21, + "args": { + "External id": 4305,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4304 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917791, "dur": 5, + "args": { + "External id": 4306,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4305 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790917797, "dur": 14, + "args": { + "External id": 4307,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4306 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790917799, "dur": 11, + "args": { + "External id": 4308,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4307 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917830, "dur": 5, + "args": { + "External id": 4309,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4308 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790917839, "dur": 21, + "args": { + "External id": 4310,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4309 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917840, "dur": 5, + "args": { + "External id": 4311,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4310 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790917846, "dur": 14, + "args": { + "External id": 4312,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4311 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790917847, "dur": 12, + "args": { + "External id": 4313,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4312 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917878, "dur": 5, + "args": { + "External id": 4314,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4313 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790917887, "dur": 23, + "args": { + "External id": 4315,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4314 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917889, "dur": 4, + "args": { + "External id": 4316,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4315 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790917894, "dur": 16, + "args": { + "External id": 4317,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4316 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790917896, "dur": 13, + "args": { + "External id": 4318,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4317 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917928, "dur": 5, + "args": { + "External id": 4319,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4318 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790917937, "dur": 22, + "args": { + "External id": 4320,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4319 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917939, "dur": 4, + "args": { + "External id": 4321,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4320 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790917944, "dur": 14, + "args": { + "External id": 4322,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4321 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790917946, "dur": 11, + "args": { + "External id": 4323,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4322 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917977, "dur": 5, + "args": { + "External id": 4324,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4323 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790917986, "dur": 21, + "args": { + "External id": 4325,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4324 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790917987, "dur": 5, + "args": { + "External id": 4326,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4325 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790917993, "dur": 14, + "args": { + "External id": 4327,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4326 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790917995, "dur": 11, + "args": { + "External id": 4328,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4327 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918026, "dur": 5, + "args": { + "External id": 4329,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4328 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790918035, "dur": 22, + "args": { + "External id": 4330,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4329 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918036, "dur": 5, + "args": { + "External id": 4331,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4330 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790918042, "dur": 14, + "args": { + "External id": 4332,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4331 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790918044, "dur": 11, + "args": { + "External id": 4333,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4332 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918075, "dur": 6, + "args": { + "External id": 4334,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4333 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790918086, "dur": 24, + "args": { + "External id": 4335,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4334 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918087, "dur": 5, + "args": { + "External id": 4336,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4335 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790918093, "dur": 16, + "args": { + "External id": 4337,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4336 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790918095, "dur": 13, + "args": { + "External id": 4338,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4337 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918128, "dur": 5, + "args": { + "External id": 4339,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4338 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790918137, "dur": 21, + "args": { + "External id": 4340,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4339 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918138, "dur": 4, + "args": { + "External id": 4341,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4340 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790918143, "dur": 15, + "args": { + "External id": 4342,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4341 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790918145, "dur": 11, + "args": { + "External id": 4343,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4342 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918176, "dur": 5, + "args": { + "External id": 4344,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4343 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790918185, "dur": 22, + "args": { + "External id": 4345,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4344 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918186, "dur": 5, + "args": { + "External id": 4346,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4345 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790918192, "dur": 14, + "args": { + "External id": 4347,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4346 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790918194, "dur": 11, + "args": { + "External id": 4348,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4347 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918225, "dur": 5, + "args": { + "External id": 4349,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4348 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790918234, "dur": 21, + "args": { + "External id": 4350,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4349 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918235, "dur": 4, + "args": { + "External id": 4351,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4350 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790918240, "dur": 14, + "args": { + "External id": 4352,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4351 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790918242, "dur": 11, + "args": { + "External id": 4353,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4352 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918273, "dur": 5, + "args": { + "External id": 4354,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4353 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790918282, "dur": 21, + "args": { + "External id": 4355,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4354 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918283, "dur": 4, + "args": { + "External id": 4356,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4355 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790918288, "dur": 14, + "args": { + "External id": 4357,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4356 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790918290, "dur": 11, + "args": { + "External id": 4358,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4357 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918321, "dur": 5, + "args": { + "External id": 4359,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4358 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790918330, "dur": 21, + "args": { + "External id": 4360,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4359 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918331, "dur": 4, + "args": { + "External id": 4361,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4360 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790918337, "dur": 14, + "args": { + "External id": 4362,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4361 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790918338, "dur": 11, + "args": { + "External id": 4363,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4362 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918369, "dur": 5, + "args": { + "External id": 4364,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4363 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790918378, "dur": 21, + "args": { + "External id": 4365,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4364 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918379, "dur": 4, + "args": { + "External id": 4366,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4365 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790918384, "dur": 15, + "args": { + "External id": 4367,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4366 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790918386, "dur": 11, + "args": { + "External id": 4368,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4367 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918417, "dur": 5, + "args": { + "External id": 4369,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4368 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790918426, "dur": 21, + "args": { + "External id": 4370,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4369 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918427, "dur": 4, + "args": { + "External id": 4371,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4370 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790918433, "dur": 14, + "args": { + "External id": 4372,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4371 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790918434, "dur": 11, + "args": { + "External id": 4373,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4372 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918465, "dur": 5, + "args": { + "External id": 4374,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4373 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790918474, "dur": 21, + "args": { + "External id": 4375,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4374 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918475, "dur": 5, + "args": { + "External id": 4376,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4375 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790918481, "dur": 14, + "args": { + "External id": 4377,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4376 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790918483, "dur": 11, + "args": { + "External id": 4378,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4377 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918513, "dur": 5, + "args": { + "External id": 4379,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4378 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790918522, "dur": 21, + "args": { + "External id": 4380,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4379 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918523, "dur": 5, + "args": { + "External id": 4381,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4380 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790918529, "dur": 14, + "args": { + "External id": 4382,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4381 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790918531, "dur": 11, + "args": { + "External id": 4383,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4382 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918562, "dur": 5, + "args": { + "External id": 4384,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4383 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790918571, "dur": 21, + "args": { + "External id": 4385,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4384 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918572, "dur": 4, + "args": { + "External id": 4386,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4385 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790918578, "dur": 14, + "args": { + "External id": 4387,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4386 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790918579, "dur": 11, + "args": { + "External id": 4388,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4387 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918610, "dur": 5, + "args": { + "External id": 4389,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4388 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790918619, "dur": 21, + "args": { + "External id": 4390,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4389 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918620, "dur": 4, + "args": { + "External id": 4391,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4390 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790918626, "dur": 14, + "args": { + "External id": 4392,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4391 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790918627, "dur": 11, + "args": { + "External id": 4393,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4392 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918658, "dur": 5, + "args": { + "External id": 4394,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4393 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790918666, "dur": 22, + "args": { + "External id": 4395,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4394 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918668, "dur": 4, + "args": { + "External id": 4396,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4395 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790918673, "dur": 14, + "args": { + "External id": 4397,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4396 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790918675, "dur": 11, + "args": { + "External id": 4398,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4397 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918705, "dur": 5, + "args": { + "External id": 4399,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4398 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790918714, "dur": 22, + "args": { + "External id": 4400,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4399 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918716, "dur": 4, + "args": { + "External id": 4401,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4400 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790918721, "dur": 14, + "args": { + "External id": 4402,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4401 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790918723, "dur": 11, + "args": { + "External id": 4403,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4402 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918753, "dur": 6, + "args": { + "External id": 4404,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4403 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790918762, "dur": 22, + "args": { + "External id": 4405,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4404 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918764, "dur": 4, + "args": { + "External id": 4406,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4405 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790918769, "dur": 14, + "args": { + "External id": 4407,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4406 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790918771, "dur": 11, + "args": { + "External id": 4408,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4407 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918801, "dur": 6, + "args": { + "External id": 4409,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4408 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790918811, "dur": 21, + "args": { + "External id": 4410,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4409 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918812, "dur": 4, + "args": { + "External id": 4411,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4410 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790918817, "dur": 15, + "args": { + "External id": 4412,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4411 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790918819, "dur": 11, + "args": { + "External id": 4413,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4412 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918850, "dur": 5, + "args": { + "External id": 4414,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4413 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790918859, "dur": 21, + "args": { + "External id": 4415,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4414 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918860, "dur": 4, + "args": { + "External id": 4416,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4415 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790918866, "dur": 14, + "args": { + "External id": 4417,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4416 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790918867, "dur": 12, + "args": { + "External id": 4418,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4417 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918898, "dur": 5, + "args": { + "External id": 4419,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4418 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790918907, "dur": 22, + "args": { + "External id": 4420,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4419 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918908, "dur": 5, + "args": { + "External id": 4421,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4420 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790918914, "dur": 14, + "args": { + "External id": 4422,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4421 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790918916, "dur": 11, + "args": { + "External id": 4423,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4422 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918946, "dur": 5, + "args": { + "External id": 4424,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4423 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790918955, "dur": 22, + "args": { + "External id": 4425,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4424 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918957, "dur": 4, + "args": { + "External id": 4426,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4425 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790918962, "dur": 14, + "args": { + "External id": 4427,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4426 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790918964, "dur": 11, + "args": { + "External id": 4428,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4427 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790918995, "dur": 5, + "args": { + "External id": 4429,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4428 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790919004, "dur": 21, + "args": { + "External id": 4430,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4429 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919005, "dur": 5, + "args": { + "External id": 4431,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4430 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790919011, "dur": 14, + "args": { + "External id": 4432,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4431 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790919012, "dur": 12, + "args": { + "External id": 4433,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4432 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919043, "dur": 6, + "args": { + "External id": 4434,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4433 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790919052, "dur": 22, + "args": { + "External id": 4435,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4434 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919054, "dur": 4, + "args": { + "External id": 4436,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4435 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790919059, "dur": 14, + "args": { + "External id": 4437,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4436 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790919061, "dur": 11, + "args": { + "External id": 4438,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4437 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919091, "dur": 6, + "args": { + "External id": 4439,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4438 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790919101, "dur": 21, + "args": { + "External id": 4440,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4439 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919102, "dur": 4, + "args": { + "External id": 4441,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4440 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790919107, "dur": 14, + "args": { + "External id": 4442,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4441 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790919109, "dur": 11, + "args": { + "External id": 4443,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4442 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919140, "dur": 5, + "args": { + "External id": 4444,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4443 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790919148, "dur": 22, + "args": { + "External id": 4445,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4444 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919150, "dur": 4, + "args": { + "External id": 4446,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4445 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790919155, "dur": 14, + "args": { + "External id": 4447,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4446 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790919157, "dur": 11, + "args": { + "External id": 4448,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4447 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919187, "dur": 6, + "args": { + "External id": 4449,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4448 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790919197, "dur": 21, + "args": { + "External id": 4450,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4449 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919198, "dur": 4, + "args": { + "External id": 4451,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4450 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790919203, "dur": 14, + "args": { + "External id": 4452,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4451 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790919205, "dur": 11, + "args": { + "External id": 4453,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4452 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919294, "dur": 12, + "args": { + "External id": 4454,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4453 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790919312, "dur": 26, + "args": { + "External id": 4455,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4454 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919314, "dur": 4, + "args": { + "External id": 4456,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4455 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790919320, "dur": 18, + "args": { + "External id": 4457,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4456 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790919322, "dur": 15, + "args": { + "External id": 4458,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4457 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919360, "dur": 5, + "args": { + "External id": 4459,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4458 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790919369, "dur": 22, + "args": { + "External id": 4460,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4459 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919370, "dur": 5, + "args": { + "External id": 4461,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4460 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790919376, "dur": 14, + "args": { + "External id": 4462,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4461 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790919378, "dur": 11, + "args": { + "External id": 4463,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4462 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919410, "dur": 5, + "args": { + "External id": 4464,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4463 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790919419, "dur": 22, + "args": { + "External id": 4465,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4464 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919420, "dur": 5, + "args": { + "External id": 4466,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4465 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790919426, "dur": 14, + "args": { + "External id": 4467,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4466 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790919428, "dur": 11, + "args": { + "External id": 4468,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4467 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919459, "dur": 6, + "args": { + "External id": 4469,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4468 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790919469, "dur": 21, + "args": { + "External id": 4470,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4469 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919470, "dur": 4, + "args": { + "External id": 4471,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4470 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790919475, "dur": 14, + "args": { + "External id": 4472,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4471 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790919477, "dur": 11, + "args": { + "External id": 4473,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4472 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919508, "dur": 5, + "args": { + "External id": 4474,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4473 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790919517, "dur": 21, + "args": { + "External id": 4475,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4474 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919518, "dur": 5, + "args": { + "External id": 4476,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4475 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790919524, "dur": 14, + "args": { + "External id": 4477,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4476 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790919526, "dur": 11, + "args": { + "External id": 4478,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4477 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919557, "dur": 5, + "args": { + "External id": 4479,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4478 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790919566, "dur": 21, + "args": { + "External id": 4480,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4479 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919567, "dur": 5, + "args": { + "External id": 4481,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4480 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790919573, "dur": 14, + "args": { + "External id": 4482,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4481 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790919574, "dur": 12, + "args": { + "External id": 4483,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4482 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919605, "dur": 6, + "args": { + "External id": 4484,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4483 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790919614, "dur": 22, + "args": { + "External id": 4485,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4484 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919616, "dur": 4, + "args": { + "External id": 4486,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4485 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790919621, "dur": 14, + "args": { + "External id": 4487,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4486 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790919623, "dur": 11, + "args": { + "External id": 4488,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4487 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919653, "dur": 5, + "args": { + "External id": 4489,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4488 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790919662, "dur": 21, + "args": { + "External id": 4490,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4489 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919664, "dur": 4, + "args": { + "External id": 4491,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4490 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790919669, "dur": 14, + "args": { + "External id": 4492,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4491 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790919671, "dur": 11, + "args": { + "External id": 4493,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4492 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919701, "dur": 6, + "args": { + "External id": 4494,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4493 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790919710, "dur": 22, + "args": { + "External id": 4495,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4494 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919712, "dur": 4, + "args": { + "External id": 4496,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4495 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790919717, "dur": 14, + "args": { + "External id": 4497,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4496 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790919719, "dur": 11, + "args": { + "External id": 4498,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4497 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919750, "dur": 5, + "args": { + "External id": 4499,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4498 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790919759, "dur": 21, + "args": { + "External id": 4500,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4499 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919760, "dur": 4, + "args": { + "External id": 4501,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4500 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790919765, "dur": 14, + "args": { + "External id": 4502,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4501 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790919767, "dur": 11, + "args": { + "External id": 4503,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4502 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919848, "dur": 6, + "args": { + "External id": 4504,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4503 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790919859, "dur": 25, + "args": { + "External id": 4505,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4504 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919860, "dur": 5, + "args": { + "External id": 4506,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4505 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790919866, "dur": 17, + "args": { + "External id": 4507,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4506 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790919868, "dur": 14, + "args": { + "External id": 4508,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4507 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919904, "dur": 5, + "args": { + "External id": 4509,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4508 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790919913, "dur": 22, + "args": { + "External id": 4510,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4509 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919915, "dur": 4, + "args": { + "External id": 4511,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4510 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790919920, "dur": 15, + "args": { + "External id": 4512,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4511 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790919922, "dur": 11, + "args": { + "External id": 4513,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4512 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919953, "dur": 6, + "args": { + "External id": 4514,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4513 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790919963, "dur": 21, + "args": { + "External id": 4515,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4514 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790919964, "dur": 4, + "args": { + "External id": 4516,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4515 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790919969, "dur": 15, + "args": { + "External id": 4517,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4516 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790919971, "dur": 11, + "args": { + "External id": 4518,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4517 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920002, "dur": 5, + "args": { + "External id": 4519,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4518 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790920011, "dur": 22, + "args": { + "External id": 4520,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4519 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920013, "dur": 4, + "args": { + "External id": 4521,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4520 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790920018, "dur": 14, + "args": { + "External id": 4522,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4521 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790920020, "dur": 11, + "args": { + "External id": 4523,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4522 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920051, "dur": 5, + "args": { + "External id": 4524,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4523 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790920060, "dur": 22, + "args": { + "External id": 4525,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4524 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920061, "dur": 5, + "args": { + "External id": 4526,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4525 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790920067, "dur": 14, + "args": { + "External id": 4527,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4526 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790920069, "dur": 11, + "args": { + "External id": 4528,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4527 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920100, "dur": 5, + "args": { + "External id": 4529,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4528 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790920109, "dur": 21, + "args": { + "External id": 4530,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4529 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920110, "dur": 4, + "args": { + "External id": 4531,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4530 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790920115, "dur": 14, + "args": { + "External id": 4532,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4531 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790920117, "dur": 11, + "args": { + "External id": 4533,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4532 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920148, "dur": 5, + "args": { + "External id": 4534,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4533 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790920157, "dur": 21, + "args": { + "External id": 4535,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4534 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920158, "dur": 4, + "args": { + "External id": 4536,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4535 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790920164, "dur": 14, + "args": { + "External id": 4537,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4536 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790920165, "dur": 12, + "args": { + "External id": 4538,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4537 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920196, "dur": 5, + "args": { + "External id": 4539,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4538 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790920205, "dur": 42, + "args": { + "External id": 4540,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4539 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920207, "dur": 4, + "args": { + "External id": 4541,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4540 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790920212, "dur": 35, + "args": { + "External id": 4542,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4541 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790920214, "dur": 31, + "args": { + "External id": 4543,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4542 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920269, "dur": 5, + "args": { + "External id": 4544,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4543 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790920278, "dur": 22, + "args": { + "External id": 4545,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4544 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920279, "dur": 5, + "args": { + "External id": 4546,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4545 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790920285, "dur": 15, + "args": { + "External id": 4547,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4546 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790920287, "dur": 11, + "args": { + "External id": 4548,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4547 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920318, "dur": 5, + "args": { + "External id": 4549,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4548 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790920327, "dur": 25, + "args": { + "External id": 4550,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4549 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920328, "dur": 5, + "args": { + "External id": 4551,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4550 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790920334, "dur": 18, + "args": { + "External id": 4552,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4551 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790920339, "dur": 11, + "args": { + "External id": 4553,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4552 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920370, "dur": 6, + "args": { + "External id": 4554,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4553 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790920379, "dur": 22, + "args": { + "External id": 4555,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4554 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920381, "dur": 4, + "args": { + "External id": 4556,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4555 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790920386, "dur": 14, + "args": { + "External id": 4557,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4556 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790920388, "dur": 11, + "args": { + "External id": 4558,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4557 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920419, "dur": 5, + "args": { + "External id": 4559,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4558 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790920428, "dur": 22, + "args": { + "External id": 4560,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4559 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920429, "dur": 5, + "args": { + "External id": 4561,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4560 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790920435, "dur": 14, + "args": { + "External id": 4562,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4561 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790920437, "dur": 11, + "args": { + "External id": 4563,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4562 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920467, "dur": 6, + "args": { + "External id": 4564,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4563 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790920477, "dur": 21, + "args": { + "External id": 4565,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4564 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920478, "dur": 4, + "args": { + "External id": 4566,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4565 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790920483, "dur": 15, + "args": { + "External id": 4567,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4566 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790920485, "dur": 11, + "args": { + "External id": 4568,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4567 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920516, "dur": 5, + "args": { + "External id": 4569,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4568 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790920525, "dur": 21, + "args": { + "External id": 4570,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4569 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920526, "dur": 5, + "args": { + "External id": 4571,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4570 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790920532, "dur": 14, + "args": { + "External id": 4572,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4571 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790920534, "dur": 11, + "args": { + "External id": 4573,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4572 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920565, "dur": 5, + "args": { + "External id": 4574,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4573 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790920574, "dur": 21, + "args": { + "External id": 4575,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4574 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920575, "dur": 4, + "args": { + "External id": 4576,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4575 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790920581, "dur": 13, + "args": { + "External id": 4577,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4576 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790920582, "dur": 11, + "args": { + "External id": 4578,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4577 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920613, "dur": 5, + "args": { + "External id": 4579,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4578 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790920622, "dur": 21, + "args": { + "External id": 4580,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4579 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920623, "dur": 5, + "args": { + "External id": 4581,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4580 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790920629, "dur": 14, + "args": { + "External id": 4582,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4581 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790920630, "dur": 12, + "args": { + "External id": 4583,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4582 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920661, "dur": 5, + "args": { + "External id": 4584,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4583 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790920670, "dur": 22, + "args": { + "External id": 4585,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4584 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920671, "dur": 5, + "args": { + "External id": 4586,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4585 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790920677, "dur": 14, + "args": { + "External id": 4587,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4586 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790920679, "dur": 11, + "args": { + "External id": 4588,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4587 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920710, "dur": 5, + "args": { + "External id": 4589,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4588 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790920719, "dur": 22, + "args": { + "External id": 4590,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4589 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920721, "dur": 4, + "args": { + "External id": 4591,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4590 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790920726, "dur": 14, + "args": { + "External id": 4592,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4591 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790920728, "dur": 11, + "args": { + "External id": 4593,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4592 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920759, "dur": 5, + "args": { + "External id": 4594,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4593 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790920768, "dur": 22, + "args": { + "External id": 4595,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4594 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920769, "dur": 4, + "args": { + "External id": 4596,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4595 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790920775, "dur": 15, + "args": { + "External id": 4597,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4596 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790920777, "dur": 11, + "args": { + "External id": 4598,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4597 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920808, "dur": 5, + "args": { + "External id": 4599,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4598 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790920817, "dur": 22, + "args": { + "External id": 4600,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4599 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920819, "dur": 4, + "args": { + "External id": 4601,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4600 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790920824, "dur": 15, + "args": { + "External id": 4602,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4601 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790920826, "dur": 11, + "args": { + "External id": 4603,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4602 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920857, "dur": 5, + "args": { + "External id": 4604,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4603 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790920866, "dur": 21, + "args": { + "External id": 4605,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4604 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920868, "dur": 4, + "args": { + "External id": 4606,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4605 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790920873, "dur": 14, + "args": { + "External id": 4607,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4606 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790920875, "dur": 11, + "args": { + "External id": 4608,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4607 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920914, "dur": 6, + "args": { + "External id": 4609,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4608 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790920924, "dur": 21, + "args": { + "External id": 4610,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4609 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920925, "dur": 4, + "args": { + "External id": 4611,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4610 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790920931, "dur": 14, + "args": { + "External id": 4612,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4611 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790920932, "dur": 12, + "args": { + "External id": 4613,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4612 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920964, "dur": 5, + "args": { + "External id": 4614,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4613 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790920973, "dur": 21, + "args": { + "External id": 4615,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4614 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790920974, "dur": 4, + "args": { + "External id": 4616,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4615 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790920979, "dur": 15, + "args": { + "External id": 4617,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4616 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790920981, "dur": 12, + "args": { + "External id": 4618,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4617 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921012, "dur": 5, + "args": { + "External id": 4619,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4618 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790921021, "dur": 22, + "args": { + "External id": 4620,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4619 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921023, "dur": 4, + "args": { + "External id": 4621,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4620 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790921028, "dur": 14, + "args": { + "External id": 4622,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4621 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790921030, "dur": 11, + "args": { + "External id": 4623,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4622 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921060, "dur": 6, + "args": { + "External id": 4624,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4623 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790921069, "dur": 22, + "args": { + "External id": 4625,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4624 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921071, "dur": 4, + "args": { + "External id": 4626,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4625 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790921076, "dur": 14, + "args": { + "External id": 4627,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4626 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790921078, "dur": 11, + "args": { + "External id": 4628,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4627 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921109, "dur": 5, + "args": { + "External id": 4629,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4628 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790921118, "dur": 21, + "args": { + "External id": 4630,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4629 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921119, "dur": 4, + "args": { + "External id": 4631,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4630 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790921125, "dur": 14, + "args": { + "External id": 4632,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4631 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790921126, "dur": 11, + "args": { + "External id": 4633,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4632 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921157, "dur": 5, + "args": { + "External id": 4634,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4633 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790921166, "dur": 21, + "args": { + "External id": 4635,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4634 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921167, "dur": 4, + "args": { + "External id": 4636,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4635 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790921172, "dur": 14, + "args": { + "External id": 4637,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4636 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790921174, "dur": 11, + "args": { + "External id": 4638,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4637 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921205, "dur": 5, + "args": { + "External id": 4639,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4638 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790921214, "dur": 21, + "args": { + "External id": 4640,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4639 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921215, "dur": 4, + "args": { + "External id": 4641,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4640 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790921220, "dur": 15, + "args": { + "External id": 4642,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4641 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790921222, "dur": 12, + "args": { + "External id": 4643,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4642 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921253, "dur": 5, + "args": { + "External id": 4644,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4643 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790921262, "dur": 22, + "args": { + "External id": 4645,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4644 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921264, "dur": 4, + "args": { + "External id": 4646,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4645 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790921269, "dur": 14, + "args": { + "External id": 4647,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4646 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790921271, "dur": 11, + "args": { + "External id": 4648,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4647 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921302, "dur": 5, + "args": { + "External id": 4649,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4648 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790921311, "dur": 21, + "args": { + "External id": 4650,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4649 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921312, "dur": 4, + "args": { + "External id": 4651,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4650 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790921317, "dur": 14, + "args": { + "External id": 4652,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4651 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790921319, "dur": 11, + "args": { + "External id": 4653,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4652 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921349, "dur": 6, + "args": { + "External id": 4654,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4653 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790921358, "dur": 22, + "args": { + "External id": 4655,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4654 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921360, "dur": 4, + "args": { + "External id": 4656,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4655 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790921365, "dur": 14, + "args": { + "External id": 4657,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4656 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790921367, "dur": 11, + "args": { + "External id": 4658,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4657 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921397, "dur": 5, + "args": { + "External id": 4659,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4658 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790921406, "dur": 22, + "args": { + "External id": 4660,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4659 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921408, "dur": 4, + "args": { + "External id": 4661,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4660 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790921413, "dur": 14, + "args": { + "External id": 4662,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4661 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790921415, "dur": 11, + "args": { + "External id": 4663,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4662 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921446, "dur": 5, + "args": { + "External id": 4664,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4663 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790921455, "dur": 21, + "args": { + "External id": 4665,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4664 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921456, "dur": 4, + "args": { + "External id": 4666,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4665 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790921462, "dur": 14, + "args": { + "External id": 4667,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4666 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790921463, "dur": 11, + "args": { + "External id": 4668,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4667 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921494, "dur": 5, + "args": { + "External id": 4669,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4668 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790921502, "dur": 21, + "args": { + "External id": 4670,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4669 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921504, "dur": 4, + "args": { + "External id": 4671,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4670 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790921509, "dur": 14, + "args": { + "External id": 4672,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4671 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790921511, "dur": 11, + "args": { + "External id": 4673,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4672 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921541, "dur": 6, + "args": { + "External id": 4674,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4673 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790921550, "dur": 21, + "args": { + "External id": 4675,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4674 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921552, "dur": 4, + "args": { + "External id": 4676,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4675 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790921557, "dur": 14, + "args": { + "External id": 4677,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4676 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790921559, "dur": 11, + "args": { + "External id": 4678,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4677 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921589, "dur": 6, + "args": { + "External id": 4679,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4678 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790921598, "dur": 21, + "args": { + "External id": 4680,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4679 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921600, "dur": 4, + "args": { + "External id": 4681,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4680 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790921605, "dur": 14, + "args": { + "External id": 4682,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4681 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790921607, "dur": 11, + "args": { + "External id": 4683,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4682 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921637, "dur": 5, + "args": { + "External id": 4684,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4683 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790921646, "dur": 22, + "args": { + "External id": 4685,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4684 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921647, "dur": 5, + "args": { + "External id": 4686,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4685 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790921653, "dur": 14, + "args": { + "External id": 4687,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4686 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790921655, "dur": 11, + "args": { + "External id": 4688,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4687 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921685, "dur": 6, + "args": { + "External id": 4689,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4688 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790921695, "dur": 21, + "args": { + "External id": 4690,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4689 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921696, "dur": 4, + "args": { + "External id": 4691,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4690 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790921701, "dur": 14, + "args": { + "External id": 4692,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4691 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790921703, "dur": 11, + "args": { + "External id": 4693,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4692 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921733, "dur": 5, + "args": { + "External id": 4694,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4693 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790921742, "dur": 21, + "args": { + "External id": 4695,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4694 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921743, "dur": 5, + "args": { + "External id": 4696,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4695 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790921749, "dur": 14, + "args": { + "External id": 4697,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4696 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790921751, "dur": 11, + "args": { + "External id": 4698,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4697 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921781, "dur": 5, + "args": { + "External id": 4699,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4698 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790921790, "dur": 22, + "args": { + "External id": 4700,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4699 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921791, "dur": 5, + "args": { + "External id": 4701,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4700 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790921797, "dur": 14, + "args": { + "External id": 4702,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4701 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790921799, "dur": 11, + "args": { + "External id": 4703,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4702 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921830, "dur": 5, + "args": { + "External id": 4704,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4703 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790921839, "dur": 21, + "args": { + "External id": 4705,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4704 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921840, "dur": 5, + "args": { + "External id": 4706,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4705 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790921846, "dur": 14, + "args": { + "External id": 4707,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4706 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790921848, "dur": 11, + "args": { + "External id": 4708,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4707 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921879, "dur": 5, + "args": { + "External id": 4709,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4708 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790921888, "dur": 21, + "args": { + "External id": 4710,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4709 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921889, "dur": 4, + "args": { + "External id": 4711,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4710 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790921895, "dur": 14, + "args": { + "External id": 4712,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4711 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790921896, "dur": 12, + "args": { + "External id": 4713,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4712 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921927, "dur": 5, + "args": { + "External id": 4714,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4713 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790921936, "dur": 22, + "args": { + "External id": 4715,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4714 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921938, "dur": 4, + "args": { + "External id": 4716,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4715 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790921943, "dur": 14, + "args": { + "External id": 4717,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4716 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790921945, "dur": 11, + "args": { + "External id": 4718,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4717 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921976, "dur": 5, + "args": { + "External id": 4719,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4718 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790921985, "dur": 21, + "args": { + "External id": 4720,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4719 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790921986, "dur": 5, + "args": { + "External id": 4721,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4720 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790921992, "dur": 14, + "args": { + "External id": 4722,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4721 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790921993, "dur": 11, + "args": { + "External id": 4723,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4722 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790922024, "dur": 5, + "args": { + "External id": 4724,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4723 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790922033, "dur": 21, + "args": { + "External id": 4725,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4724 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790922034, "dur": 5, + "args": { + "External id": 4726,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4725 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790922040, "dur": 14, + "args": { + "External id": 4727,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4726 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790922041, "dur": 11, + "args": { + "External id": 4728,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4727 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790922072, "dur": 5, + "args": { + "External id": 4729,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4728 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790922081, "dur": 21, + "args": { + "External id": 4730,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4729 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790922082, "dur": 4, + "args": { + "External id": 4731,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4730 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790922087, "dur": 15, + "args": { + "External id": 4732,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4731 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790922089, "dur": 12, + "args": { + "External id": 4733,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4732 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790922120, "dur": 5, + "args": { + "External id": 4734,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4733 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790922129, "dur": 21, + "args": { + "External id": 4735,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4734 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790922130, "dur": 5, + "args": { + "External id": 4736,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4735 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790922136, "dur": 14, + "args": { + "External id": 4737,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4736 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790922138, "dur": 11, + "args": { + "External id": 4738,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4737 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790922168, "dur": 5, + "args": { + "External id": 4739,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4738 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790922177, "dur": 22, + "args": { + "External id": 4740,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4739 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790922178, "dur": 5, + "args": { + "External id": 4741,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4740 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790922184, "dur": 15, + "args": { + "External id": 4742,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4741 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790922186, "dur": 11, + "args": { + "External id": 4743,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4742 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790922217, "dur": 5, + "args": { + "External id": 4744,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4743 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790922226, "dur": 21, + "args": { + "External id": 4745,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4744 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790922227, "dur": 4, + "args": { + "External id": 4746,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4745 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790922233, "dur": 14, + "args": { + "External id": 4747,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4746 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790922234, "dur": 11, + "args": { + "External id": 4748,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4747 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790922265, "dur": 5, + "args": { + "External id": 4749,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4748 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649790922274, "dur": 21, + "args": { + "External id": 4750,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4749 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790922276, "dur": 4, + "args": { + "External id": 4751,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4750 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649790922281, "dur": 14, + "args": { + "External id": 4752,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 4751 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649790922283, "dur": 11, + "args": { + "External id": 4753,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 4752 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::randn", "pid": 435, "tid": 435, + "ts": 1776649790922450, "dur": 54, + "args": { + "External id": 4754,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 4753 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790922455, "dur": 8, + "args": { + "External id": 4755,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4754 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::normal_", "pid": 435, "tid": 435, + "ts": 1776649790922467, "dur": 36, + "args": { + "External id": 4756,"Record function id": 0, "Concrete Inputs": ["", "0.", "1.", ""], "Input type": ["c10::Half", "Scalar", "Scalar", ""], "Input Dims": [[8, 4096], [], [], []], "Ev Idx": 4755 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::randn", "pid": 435, "tid": 435, + "ts": 1776649790922529, "dur": 28, + "args": { + "External id": 4757,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "", "5", "", "", "False"], "Input type": ["ScalarList", "", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4756 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790922532, "dur": 5, + "args": { + "External id": 4758,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4757 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::normal_", "pid": 435, "tid": 435, + "ts": 1776649790922539, "dur": 18, + "args": { + "External id": 4759,"Record function id": 0, "Concrete Inputs": ["", "0.", "1.", ""], "Input type": ["c10::Half", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 4758 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 435, "tid": 435, + "ts": 1776649790922586, "dur": 41, + "args": { + "External id": 4760,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "double"], "Input Dims": [[4096, 4096], []], "Ev Idx": 4759 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790922641, "dur": 13, + "args": { + "External id": 4761,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 4096], []], "Ev Idx": 4760 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790922644, "dur": 9, + "args": { + "External id": 4762,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 4096], []], "Ev Idx": 4761 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 435, "tid": 435, + "ts": 1776649790922658, "dur": 47, + "args": { + "External id": 4763,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[32, 128, 4096]], "Ev Idx": 4762 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790922663, "dur": 5, + "args": { + "External id": 4764,"Record function id": 0, "Concrete Inputs": ["[0]", "5", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4763 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 435, "tid": 435, + "ts": 1776649790922671, "dur": 33, + "args": { + "External id": 4765,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[32, 128, 4096], [0]], "Ev Idx": 4764 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 435, "tid": 435, + "ts": 1776649790922679, "dur": 9, + "args": { + "External id": 4766,"Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]", ""], "Input type": ["c10::Half", "ScalarList", ""], "Input Dims": [[0], [], []], "Ev Idx": 4765 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 435, "tid": 435, + "ts": 1776649790922714, "dur": 43, + "args": { + "External id": 4767,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1]", "False"], "Input type": ["c10::Half", "ScalarList", "Scalar"], "Input Dims": [[32, 128, 4096], [], []], "Ev Idx": 4766 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790922731, "dur": 2, + "args": { + "External id": 4768,"Record function id": 0, "Concrete Inputs": ["", "[32, 1, 4096]", "[4096, 0, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 4767 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 435, "tid": 435, + "ts": 1776649790922768, "dur": 26, + "args": { + "External id": 4769,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "double"], "Input Dims": [[32, 4096], []], "Ev Idx": 4768 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 435, "tid": 435, + "ts": 1776649790922804, "dur": 34, + "args": { + "External id": 4770,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "9.9999999999999995e-07", ""], "Input type": ["c10::Half", "Scalar", ""], "Input Dims": [[32, 4096], [], []], "Ev Idx": 4769 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790922810, "dur": 2, + "args": { + "External id": 4771,"Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 4096], [], [], [], []], "Ev Idx": 4770 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790922844, "dur": 0, + "args": { + "External id": 4772,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 4096], [], [], [], []], "Ev Idx": 4771 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649790922850, "dur": 8, + "args": { + "External id": 4773,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 4096], []], "Ev Idx": 4772 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790922856, "dur": 1, + "args": { + "External id": 4774,"Record function id": 0, "Concrete Inputs": ["", "[32, 1, 4096]", "[4096, 4096, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 4773 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 435, "tid": 435, + "ts": 1776649790922861, "dur": 24, + "args": { + "External id": 4775,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[32, 128, 4096], [32, 1, 4096]], "Ev Idx": 4774 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 435, "tid": 435, + "ts": 1776649790922891, "dur": 22, + "args": { + "External id": 4776,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[32, 128, 4096]], "Ev Idx": 4775 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 435, "tid": 435, + "ts": 1776649790922919, "dur": 17, + "args": { + "External id": 4777,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "-8", "7"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[32, 128, 4096], [], []], "Ev Idx": 4776 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790922921, "dur": 0, + "args": { + "External id": 4778,"Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 128, 4096], [], [], [], []], "Ev Idx": 4777 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790922941, "dur": 40, + "args": { + "External id": 4779,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 128, 4096], [], [], [], []], "Ev Idx": 4778 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790922943, "dur": 38, + "args": { + "External id": 4780,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["c10::Half", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[32, 128, 4096], [], [], [], [], [], []], "Ev Idx": 4779 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790922950, "dur": 7, + "args": { + "External id": 4781,"Record function id": 0, "Concrete Inputs": ["[32, 128, 4096]", "[524288, 4096, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4780 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790922960, "dur": 20, + "args": { + "External id": 4782,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "c10::Half", "Scalar"], "Input Dims": [[32, 128, 4096], [32, 128, 4096], []], "Ev Idx": 4781 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 435, "tid": 435, + "ts": 1776649790922993, "dur": 30, + "args": { + "External id": 4783,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["signed char", "long int", "Scalar"], "Input Dims": [[32, 128, 4096], [], []], "Ev Idx": 4782 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790923027, "dur": 26, + "args": { + "External id": 4784,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "False", "False", ""], "Input type": ["signed char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 128, 4096], [], [], [], []], "Ev Idx": 4783 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790923028, "dur": 25, + "args": { + "External id": 4785,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "", "", "", "False", ""], "Input type": ["signed char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[32, 128, 4096], [], [], [], [], [], []], "Ev Idx": 4784 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790923031, "dur": 5, + "args": { + "External id": 4786,"Record function id": 0, "Concrete Inputs": ["[32, 128, 4096]", "[524288, 4096, 1]", "0", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4785 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790923038, "dur": 14, + "args": { + "External id": 4787,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["unsigned char", "signed char", "Scalar"], "Input Dims": [[32, 128, 4096], [32, 128, 4096], []], "Ev Idx": 4786 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790923060, "dur": 5, + "args": { + "External id": 4788,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["unsigned char", "ScalarList"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 4787 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790923062, "dur": 2, + "args": { + "External id": 4789,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["unsigned char", "ScalarList"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 4788 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649790923076, "dur": 8, + "args": { + "External id": 4790,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 4789 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790923080, "dur": 2, + "args": { + "External id": 4791,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]", "[4096, 1]", "0"], "Input type": ["unsigned char", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 4790 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649790923086, "dur": 4, + "args": { + "External id": 4792,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "2"], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 4791 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790923087, "dur": 2, + "args": { + "External id": 4793,"Record function id": 0, "Concrete Inputs": ["", "[4096, 2048]", "[4096, 2]", "0"], "Input type": ["unsigned char", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 4792 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649790923098, "dur": 4, + "args": { + "External id": 4794,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 4793 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790923100, "dur": 1, + "args": { + "External id": 4795,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]", "[4096, 1]", "0"], "Input type": ["unsigned char", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 4794 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649790923103, "dur": 4, + "args": { + "External id": 4796,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "1", "9223372036854775807", "2"], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 4795 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790923105, "dur": 1, + "args": { + "External id": 4797,"Record function id": 0, "Concrete Inputs": ["", "[4096, 2048]", "[4096, 2]", "1"], "Input type": ["unsigned char", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 4796 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__lshift__", "pid": 435, "tid": 435, + "ts": 1776649790923115, "dur": 42, + "args": { + "External id": 4798,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 4797 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790923126, "dur": 6, + "args": { + "External id": 4799,"Record function id": 0, "Concrete Inputs": ["[4096, 2048]", "0", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4798 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__or__", "pid": 435, "tid": 435, + "ts": 1776649790923163, "dur": 29, + "args": { + "External id": 4800,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["unsigned char", "unsigned char"], "Input Dims": [[4096, 2048], [4096, 2048]], "Ev Idx": 4799 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_or", "pid": 435, "tid": 435, + "ts": 1776649790923166, "dur": 25, + "args": { + "External id": 4801,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["unsigned char", "unsigned char"], "Input Dims": [[4096, 2048], [4096, 2048]], "Ev Idx": 4800 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790923197, "dur": 0, + "args": { + "External id": 4802,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], []], "Ev Idx": 4801 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__and__", "pid": 435, "tid": 435, + "ts": 1776649790923214, "dur": 33, + "args": { + "External id": 4803,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "15"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 4802 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_and", "pid": 435, "tid": 435, + "ts": 1776649790923217, "dur": 30, + "args": { + "External id": 4804,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "15"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 4803 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_and", "pid": 435, "tid": 435, + "ts": 1776649790923221, "dur": 24, + "args": { + "External id": 4805,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["unsigned char", "long int"], "Input Dims": [[4096, 2048], []], "Ev Idx": 4804 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790923252, "dur": 26, + "args": { + "External id": 4806,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], []], "Ev Idx": 4805 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790923253, "dur": 25, + "args": { + "External id": 4807,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["unsigned char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], [], [], []], "Ev Idx": 4806 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790923256, "dur": 6, + "args": { + "External id": 4808,"Record function id": 0, "Concrete Inputs": ["[4096, 2048]", "[2048, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4807 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790923264, "dur": 13, + "args": { + "External id": 4809,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "unsigned char", "Scalar"], "Input Dims": [[4096, 2048], [4096, 2048], []], "Ev Idx": 4808 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__rshift__", "pid": 435, "tid": 435, + "ts": 1776649790923285, "dur": 28, + "args": { + "External id": 4810,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 4809 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790923291, "dur": 5, + "args": { + "External id": 4811,"Record function id": 0, "Concrete Inputs": ["[4096, 2048]", "0", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4810 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790923316, "dur": 23, + "args": { + "External id": 4812,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], []], "Ev Idx": 4811 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790923317, "dur": 21, + "args": { + "External id": 4813,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["unsigned char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], [], [], []], "Ev Idx": 4812 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790923319, "dur": 5, + "args": { + "External id": 4814,"Record function id": 0, "Concrete Inputs": ["[4096, 2048]", "[2048, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4813 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790923326, "dur": 12, + "args": { + "External id": 4815,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "unsigned char", "Scalar"], "Input Dims": [[4096, 2048], [4096, 2048], []], "Ev Idx": 4814 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::stack", "pid": 435, "tid": 435, + "ts": 1776649790923352, "dur": 55, + "args": { + "External id": 4816,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["TensorList", "Scalar"], "Input Dims": [[[4096, 2048], [4096, 2048]], []], "Ev Idx": 4815 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649790923357, "dur": 3, + "args": { + "External id": 4817,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["signed char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 4816 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790923359, "dur": 1, + "args": { + "External id": 4818,"Record function id": 0, "Concrete Inputs": ["", "[4096, 2048, 1]", "[2048, 1, 1]", ""], "Input type": ["signed char", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 2048], [], [], []], "Ev Idx": 4817 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649790923361, "dur": 2, + "args": { + "External id": 4819,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["signed char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 4818 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790923363, "dur": 0, + "args": { + "External id": 4820,"Record function id": 0, "Concrete Inputs": ["", "[4096, 2048, 1]", "[2048, 1, 1]", ""], "Input type": ["signed char", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 2048], [], [], []], "Ev Idx": 4819 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 435, "tid": 435, + "ts": 1776649790923366, "dur": 38, + "args": { + "External id": 4821,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["TensorList", "Scalar"], "Input Dims": [[[4096, 2048, 1], [4096, 2048, 1]], []], "Ev Idx": 4820 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790923413, "dur": 5, + "args": { + "External id": 4822,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["signed char", "ScalarList"], "Input Dims": [[4096, 2048, 2], []], "Ev Idx": 4821 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790923414, "dur": 3, + "args": { + "External id": 4823,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["signed char", "ScalarList"], "Input Dims": [[4096, 2048, 2], []], "Ev Idx": 4822 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::repeat_interleave", "pid": 435, "tid": 435, + "ts": 1776649790923425, "dur": 65, + "args": { + "External id": 4824,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "128", "0", ""], "Input type": ["c10::Half", "Scalar", "Scalar", ""], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 4823 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649790923429, "dur": 3, + "args": { + "External id": 4825,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 4096], []], "Ev Idx": 4824 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790923430, "dur": 1, + "args": { + "External id": 4826,"Record function id": 0, "Concrete Inputs": ["", "[32, 1, 4096]", "[4096, 4096, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 4825 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 435, "tid": 435, + "ts": 1776649790923435, "dur": 7, + "args": { + "External id": 4827,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]", "False"], "Input type": ["c10::Half", "ScalarList", "Scalar"], "Input Dims": [[32, 1, 4096], [], []], "Ev Idx": 4826 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790923440, "dur": 1, + "args": { + "External id": 4828,"Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]", "[4096, 0, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 1, 4096], [], [], []], "Ev Idx": 4827 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649790923445, "dur": 37, + "args": { + "External id": 4829,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 4828 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649790923450, "dur": 11, + "args": { + "External id": 4830,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[32, 128, 4096], [], [], [], [], []], "Ev Idx": 4829 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790923453, "dur": 7, + "args": { + "External id": 4831,"Record function id": 0, "Concrete Inputs": ["[32, 128, 4096]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4830 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790923462, "dur": 19, + "args": { + "External id": 4832,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[32, 128, 4096], [32, 128, 4096], []], "Ev Idx": 4831 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::flatten", "pid": 435, "tid": 435, + "ts": 1776649790923484, "dur": 5, + "args": { + "External id": 4833,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[32, 128, 4096], [], []], "Ev Idx": 4832 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790923487, "dur": 2, + "args": { + "External id": 4834,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 4833 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 435, "tid": 435, + "ts": 1776649790923499, "dur": 22, + "args": { + "External id": 4835,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["signed char", "long int", "Scalar"], "Input Dims": [[4096, 4096], [], []], "Ev Idx": 4834 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790923527, "dur": 25, + "args": { + "External id": 4836,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["signed char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 4835 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790923528, "dur": 24, + "args": { + "External id": 4837,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "", "", "", "False", ""], "Input type": ["signed char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], [], []], "Ev Idx": 4836 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790923531, "dur": 5, + "args": { + "External id": 4838,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "[4096, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4837 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790923537, "dur": 14, + "args": { + "External id": 4839,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "signed char", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 4838 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 435, "tid": 435, + "ts": 1776649790923558, "dur": 19, + "args": { + "External id": 4840,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4096, 4096], [4096, 4096]], "Ev Idx": 4839 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790923642, "dur": 155, + "args": { + "External id": 4841,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4840 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790923648, "dur": 148, + "args": { + "External id": 4842,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4841 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790923805, "dur": 29, + "args": { + "External id": 4843,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4842 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790923807, "dur": 27, + "args": { + "External id": 4844,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4843 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790923839, "dur": 27, + "args": { + "External id": 4845,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4844 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790923842, "dur": 23, + "args": { + "External id": 4846,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4845 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790923871, "dur": 26, + "args": { + "External id": 4847,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4846 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790923873, "dur": 24, + "args": { + "External id": 4848,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4847 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790923902, "dur": 26, + "args": { + "External id": 4849,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4848 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790923904, "dur": 23, + "args": { + "External id": 4850,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4849 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790923932, "dur": 26, + "args": { + "External id": 4851,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4850 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790923934, "dur": 24, + "args": { + "External id": 4852,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4851 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790923963, "dur": 26, + "args": { + "External id": 4853,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4852 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790923965, "dur": 23, + "args": { + "External id": 4854,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4853 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790923993, "dur": 26, + "args": { + "External id": 4855,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4854 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790923996, "dur": 23, + "args": { + "External id": 4856,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4855 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924024, "dur": 26, + "args": { + "External id": 4857,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4856 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924026, "dur": 23, + "args": { + "External id": 4858,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4857 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924054, "dur": 26, + "args": { + "External id": 4859,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4858 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924057, "dur": 23, + "args": { + "External id": 4860,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4859 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924144, "dur": 36, + "args": { + "External id": 4861,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4860 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924147, "dur": 32, + "args": { + "External id": 4862,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4861 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924186, "dur": 27, + "args": { + "External id": 4863,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4862 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924188, "dur": 24, + "args": { + "External id": 4864,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4863 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924237, "dur": 30, + "args": { + "External id": 4865,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4864 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924240, "dur": 26, + "args": { + "External id": 4866,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4865 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924272, "dur": 26, + "args": { + "External id": 4867,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4866 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924274, "dur": 24, + "args": { + "External id": 4868,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4867 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924303, "dur": 26, + "args": { + "External id": 4869,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4868 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924305, "dur": 23, + "args": { + "External id": 4870,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4869 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924333, "dur": 26, + "args": { + "External id": 4871,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4870 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924335, "dur": 23, + "args": { + "External id": 4872,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4871 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924363, "dur": 26, + "args": { + "External id": 4873,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4872 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924365, "dur": 23, + "args": { + "External id": 4874,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4873 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924394, "dur": 25, + "args": { + "External id": 4875,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4874 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924396, "dur": 23, + "args": { + "External id": 4876,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4875 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924424, "dur": 26, + "args": { + "External id": 4877,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4876 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924426, "dur": 23, + "args": { + "External id": 4878,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4877 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924454, "dur": 26, + "args": { + "External id": 4879,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4878 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924456, "dur": 24, + "args": { + "External id": 4880,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4879 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924485, "dur": 26, + "args": { + "External id": 4881,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4880 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924487, "dur": 23, + "args": { + "External id": 4882,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4881 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924516, "dur": 26, + "args": { + "External id": 4883,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4882 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924518, "dur": 23, + "args": { + "External id": 4884,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4883 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924547, "dur": 25, + "args": { + "External id": 4885,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4884 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924549, "dur": 22, + "args": { + "External id": 4886,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4885 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924577, "dur": 25, + "args": { + "External id": 4887,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4886 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924579, "dur": 23, + "args": { + "External id": 4888,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4887 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924607, "dur": 26, + "args": { + "External id": 4889,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4888 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924609, "dur": 23, + "args": { + "External id": 4890,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4889 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924637, "dur": 26, + "args": { + "External id": 4891,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4890 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924639, "dur": 23, + "args": { + "External id": 4892,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4891 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924667, "dur": 26, + "args": { + "External id": 4893,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4892 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924669, "dur": 23, + "args": { + "External id": 4894,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4893 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924698, "dur": 25, + "args": { + "External id": 4895,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4894 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924700, "dur": 22, + "args": { + "External id": 4896,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4895 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924727, "dur": 26, + "args": { + "External id": 4897,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4896 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924729, "dur": 23, + "args": { + "External id": 4898,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4897 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924758, "dur": 25, + "args": { + "External id": 4899,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4898 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924760, "dur": 22, + "args": { + "External id": 4900,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4899 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924787, "dur": 27, + "args": { + "External id": 4901,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4900 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924790, "dur": 23, + "args": { + "External id": 4902,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4901 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924818, "dur": 26, + "args": { + "External id": 4903,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4902 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924821, "dur": 22, + "args": { + "External id": 4904,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4903 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924848, "dur": 25, + "args": { + "External id": 4905,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4904 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924850, "dur": 23, + "args": { + "External id": 4906,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4905 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924878, "dur": 25, + "args": { + "External id": 4907,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4906 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924880, "dur": 22, + "args": { + "External id": 4908,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4907 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924907, "dur": 26, + "args": { + "External id": 4909,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4908 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924909, "dur": 23, + "args": { + "External id": 4910,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4909 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924937, "dur": 25, + "args": { + "External id": 4911,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4910 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924939, "dur": 23, + "args": { + "External id": 4912,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4911 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924967, "dur": 25, + "args": { + "External id": 4913,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4912 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924969, "dur": 23, + "args": { + "External id": 4914,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4913 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790924997, "dur": 25, + "args": { + "External id": 4915,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4914 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790924999, "dur": 22, + "args": { + "External id": 4916,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4915 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925026, "dur": 26, + "args": { + "External id": 4917,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4916 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925028, "dur": 23, + "args": { + "External id": 4918,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4917 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925056, "dur": 26, + "args": { + "External id": 4919,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4918 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925058, "dur": 23, + "args": { + "External id": 4920,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4919 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925086, "dur": 26, + "args": { + "External id": 4921,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4920 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925089, "dur": 22, + "args": { + "External id": 4922,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4921 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925116, "dur": 25, + "args": { + "External id": 4923,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4922 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925118, "dur": 23, + "args": { + "External id": 4924,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4923 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925146, "dur": 25, + "args": { + "External id": 4925,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4924 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925148, "dur": 22, + "args": { + "External id": 4926,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4925 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925175, "dur": 26, + "args": { + "External id": 4927,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4926 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925178, "dur": 22, + "args": { + "External id": 4928,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4927 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925205, "dur": 25, + "args": { + "External id": 4929,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4928 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925207, "dur": 23, + "args": { + "External id": 4930,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4929 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925235, "dur": 25, + "args": { + "External id": 4931,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4930 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925237, "dur": 22, + "args": { + "External id": 4932,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4931 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925264, "dur": 26, + "args": { + "External id": 4933,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4932 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925266, "dur": 23, + "args": { + "External id": 4934,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4933 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925294, "dur": 25, + "args": { + "External id": 4935,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4934 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925296, "dur": 22, + "args": { + "External id": 4936,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4935 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925323, "dur": 26, + "args": { + "External id": 4937,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4936 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925325, "dur": 23, + "args": { + "External id": 4938,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4937 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925353, "dur": 26, + "args": { + "External id": 4939,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4938 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925356, "dur": 22, + "args": { + "External id": 4940,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4939 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925383, "dur": 26, + "args": { + "External id": 4941,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4940 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925385, "dur": 23, + "args": { + "External id": 4942,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4941 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925413, "dur": 26, + "args": { + "External id": 4943,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4942 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925415, "dur": 23, + "args": { + "External id": 4944,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4943 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925443, "dur": 25, + "args": { + "External id": 4945,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4944 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925445, "dur": 23, + "args": { + "External id": 4946,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4945 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925472, "dur": 26, + "args": { + "External id": 4947,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4946 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925475, "dur": 22, + "args": { + "External id": 4948,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4947 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925502, "dur": 26, + "args": { + "External id": 4949,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4948 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925504, "dur": 23, + "args": { + "External id": 4950,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4949 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925532, "dur": 25, + "args": { + "External id": 4951,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4950 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925534, "dur": 23, + "args": { + "External id": 4952,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4951 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925562, "dur": 25, + "args": { + "External id": 4953,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4952 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925564, "dur": 22, + "args": { + "External id": 4954,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4953 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925591, "dur": 26, + "args": { + "External id": 4955,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4954 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925593, "dur": 23, + "args": { + "External id": 4956,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4955 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925621, "dur": 25, + "args": { + "External id": 4957,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4956 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925623, "dur": 23, + "args": { + "External id": 4958,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4957 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925651, "dur": 25, + "args": { + "External id": 4959,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4958 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925653, "dur": 22, + "args": { + "External id": 4960,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4959 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925742, "dur": 35, + "args": { + "External id": 4961,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4960 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925746, "dur": 30, + "args": { + "External id": 4962,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4961 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925783, "dur": 27, + "args": { + "External id": 4963,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4962 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925786, "dur": 23, + "args": { + "External id": 4964,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4963 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925815, "dur": 26, + "args": { + "External id": 4965,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4964 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925817, "dur": 23, + "args": { + "External id": 4966,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4965 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925846, "dur": 25, + "args": { + "External id": 4967,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4966 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925848, "dur": 23, + "args": { + "External id": 4968,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4967 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925876, "dur": 26, + "args": { + "External id": 4969,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4968 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925878, "dur": 23, + "args": { + "External id": 4970,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4969 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925906, "dur": 26, + "args": { + "External id": 4971,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4970 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925908, "dur": 23, + "args": { + "External id": 4972,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4971 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925937, "dur": 26, + "args": { + "External id": 4973,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4972 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925939, "dur": 23, + "args": { + "External id": 4974,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4973 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925967, "dur": 26, + "args": { + "External id": 4975,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4974 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925969, "dur": 23, + "args": { + "External id": 4976,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4975 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790925997, "dur": 26, + "args": { + "External id": 4977,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4976 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790925999, "dur": 24, + "args": { + "External id": 4978,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4977 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926028, "dur": 26, + "args": { + "External id": 4979,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4978 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926030, "dur": 23, + "args": { + "External id": 4980,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4979 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926111, "dur": 34, + "args": { + "External id": 4981,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4980 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926115, "dur": 29, + "args": { + "External id": 4982,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4981 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926150, "dur": 27, + "args": { + "External id": 4983,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4982 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926153, "dur": 23, + "args": { + "External id": 4984,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4983 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926181, "dur": 26, + "args": { + "External id": 4985,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4984 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926184, "dur": 23, + "args": { + "External id": 4986,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4985 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926212, "dur": 26, + "args": { + "External id": 4987,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4986 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926214, "dur": 23, + "args": { + "External id": 4988,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4987 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926242, "dur": 26, + "args": { + "External id": 4989,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4988 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926244, "dur": 24, + "args": { + "External id": 4990,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4989 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926273, "dur": 26, + "args": { + "External id": 4991,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4990 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926275, "dur": 23, + "args": { + "External id": 4992,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4991 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926303, "dur": 26, + "args": { + "External id": 4993,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4992 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926305, "dur": 23, + "args": { + "External id": 4994,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4993 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926333, "dur": 26, + "args": { + "External id": 4995,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4994 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926336, "dur": 22, + "args": { + "External id": 4996,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4995 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926364, "dur": 25, + "args": { + "External id": 4997,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4996 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926366, "dur": 22, + "args": { + "External id": 4998,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4997 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926394, "dur": 25, + "args": { + "External id": 4999,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4998 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926396, "dur": 22, + "args": { + "External id": 5000,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 4999 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926423, "dur": 26, + "args": { + "External id": 5001,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5000 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926426, "dur": 22, + "args": { + "External id": 5002,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5001 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926453, "dur": 26, + "args": { + "External id": 5003,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5002 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926456, "dur": 22, + "args": { + "External id": 5004,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5003 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926483, "dur": 26, + "args": { + "External id": 5005,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5004 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926485, "dur": 23, + "args": { + "External id": 5006,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5005 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926513, "dur": 26, + "args": { + "External id": 5007,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5006 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926515, "dur": 23, + "args": { + "External id": 5008,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5007 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926543, "dur": 25, + "args": { + "External id": 5009,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5008 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926545, "dur": 23, + "args": { + "External id": 5010,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5009 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926573, "dur": 25, + "args": { + "External id": 5011,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5010 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926575, "dur": 23, + "args": { + "External id": 5012,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5011 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926603, "dur": 25, + "args": { + "External id": 5013,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5012 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926605, "dur": 23, + "args": { + "External id": 5014,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5013 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926633, "dur": 25, + "args": { + "External id": 5015,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5014 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926635, "dur": 23, + "args": { + "External id": 5016,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5015 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926663, "dur": 25, + "args": { + "External id": 5017,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5016 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926665, "dur": 23, + "args": { + "External id": 5018,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5017 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926693, "dur": 25, + "args": { + "External id": 5019,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5018 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926695, "dur": 22, + "args": { + "External id": 5020,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5019 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926722, "dur": 26, + "args": { + "External id": 5021,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5020 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926725, "dur": 22, + "args": { + "External id": 5022,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5021 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926752, "dur": 26, + "args": { + "External id": 5023,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5022 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926755, "dur": 22, + "args": { + "External id": 5024,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5023 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926782, "dur": 26, + "args": { + "External id": 5025,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5024 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926785, "dur": 22, + "args": { + "External id": 5026,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5025 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926812, "dur": 26, + "args": { + "External id": 5027,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5026 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926814, "dur": 23, + "args": { + "External id": 5028,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5027 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926842, "dur": 26, + "args": { + "External id": 5029,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5028 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926845, "dur": 23, + "args": { + "External id": 5030,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5029 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926873, "dur": 25, + "args": { + "External id": 5031,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5030 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926875, "dur": 23, + "args": { + "External id": 5032,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5031 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926903, "dur": 25, + "args": { + "External id": 5033,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5032 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926905, "dur": 23, + "args": { + "External id": 5034,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5033 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926933, "dur": 25, + "args": { + "External id": 5035,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5034 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926935, "dur": 22, + "args": { + "External id": 5036,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5035 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926962, "dur": 26, + "args": { + "External id": 5037,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5036 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926965, "dur": 22, + "args": { + "External id": 5038,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5037 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790926993, "dur": 25, + "args": { + "External id": 5039,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5038 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790926995, "dur": 22, + "args": { + "External id": 5040,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5039 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927023, "dur": 25, + "args": { + "External id": 5041,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5040 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927025, "dur": 22, + "args": { + "External id": 5042,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5041 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927053, "dur": 25, + "args": { + "External id": 5043,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5042 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927055, "dur": 23, + "args": { + "External id": 5044,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5043 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927083, "dur": 25, + "args": { + "External id": 5045,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5044 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927085, "dur": 23, + "args": { + "External id": 5046,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5045 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927113, "dur": 25, + "args": { + "External id": 5047,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5046 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927115, "dur": 23, + "args": { + "External id": 5048,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5047 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927143, "dur": 25, + "args": { + "External id": 5049,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5048 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927145, "dur": 23, + "args": { + "External id": 5050,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5049 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927173, "dur": 25, + "args": { + "External id": 5051,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5050 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927175, "dur": 23, + "args": { + "External id": 5052,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5051 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927203, "dur": 25, + "args": { + "External id": 5053,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5052 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927205, "dur": 22, + "args": { + "External id": 5054,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5053 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927233, "dur": 26, + "args": { + "External id": 5055,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5054 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927235, "dur": 23, + "args": { + "External id": 5056,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5055 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927263, "dur": 25, + "args": { + "External id": 5057,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5056 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927265, "dur": 23, + "args": { + "External id": 5058,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5057 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927293, "dur": 25, + "args": { + "External id": 5059,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5058 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927295, "dur": 22, + "args": { + "External id": 5060,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5059 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927322, "dur": 26, + "args": { + "External id": 5061,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5060 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927325, "dur": 23, + "args": { + "External id": 5062,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5061 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927353, "dur": 25, + "args": { + "External id": 5063,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5062 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927355, "dur": 22, + "args": { + "External id": 5064,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5063 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927382, "dur": 26, + "args": { + "External id": 5065,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5064 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927385, "dur": 22, + "args": { + "External id": 5066,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5065 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927412, "dur": 25, + "args": { + "External id": 5067,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5066 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927414, "dur": 23, + "args": { + "External id": 5068,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5067 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927442, "dur": 25, + "args": { + "External id": 5069,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5068 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927444, "dur": 22, + "args": { + "External id": 5070,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5069 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927472, "dur": 26, + "args": { + "External id": 5071,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5070 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927475, "dur": 22, + "args": { + "External id": 5072,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5071 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927502, "dur": 26, + "args": { + "External id": 5073,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5072 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927505, "dur": 23, + "args": { + "External id": 5074,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5073 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927533, "dur": 25, + "args": { + "External id": 5075,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5074 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927535, "dur": 22, + "args": { + "External id": 5076,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5075 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927562, "dur": 26, + "args": { + "External id": 5077,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5076 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927565, "dur": 22, + "args": { + "External id": 5078,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5077 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927592, "dur": 26, + "args": { + "External id": 5079,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5078 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927594, "dur": 23, + "args": { + "External id": 5080,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5079 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927681, "dur": 34, + "args": { + "External id": 5081,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5080 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927684, "dur": 30, + "args": { + "External id": 5082,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5081 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927720, "dur": 27, + "args": { + "External id": 5083,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5082 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927723, "dur": 24, + "args": { + "External id": 5084,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5083 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927752, "dur": 26, + "args": { + "External id": 5085,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5084 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927754, "dur": 23, + "args": { + "External id": 5086,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5085 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927783, "dur": 25, + "args": { + "External id": 5087,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5086 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927785, "dur": 22, + "args": { + "External id": 5088,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5087 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927812, "dur": 26, + "args": { + "External id": 5089,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5088 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927815, "dur": 22, + "args": { + "External id": 5090,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5089 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927843, "dur": 26, + "args": { + "External id": 5091,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5090 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927845, "dur": 23, + "args": { + "External id": 5092,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5091 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927873, "dur": 26, + "args": { + "External id": 5093,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5092 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927876, "dur": 22, + "args": { + "External id": 5094,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5093 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927904, "dur": 26, + "args": { + "External id": 5095,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5094 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927906, "dur": 23, + "args": { + "External id": 5096,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5095 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927934, "dur": 25, + "args": { + "External id": 5097,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5096 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927936, "dur": 22, + "args": { + "External id": 5098,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5097 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790927963, "dur": 26, + "args": { + "External id": 5099,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5098 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790927966, "dur": 23, + "args": { + "External id": 5100,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5099 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928047, "dur": 32, + "args": { + "External id": 5101,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5100 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928050, "dur": 28, + "args": { + "External id": 5102,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5101 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928085, "dur": 26, + "args": { + "External id": 5103,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5102 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928087, "dur": 23, + "args": { + "External id": 5104,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5103 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928116, "dur": 25, + "args": { + "External id": 5105,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5104 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928118, "dur": 23, + "args": { + "External id": 5106,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5105 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928146, "dur": 26, + "args": { + "External id": 5107,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5106 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928148, "dur": 23, + "args": { + "External id": 5108,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5107 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928177, "dur": 26, + "args": { + "External id": 5109,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5108 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928179, "dur": 23, + "args": { + "External id": 5110,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5109 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928208, "dur": 67, + "args": { + "External id": 5111,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5110 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928210, "dur": 64, + "args": { + "External id": 5112,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5111 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928281, "dur": 27, + "args": { + "External id": 5113,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5112 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928284, "dur": 24, + "args": { + "External id": 5114,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5113 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928313, "dur": 25, + "args": { + "External id": 5115,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5114 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928315, "dur": 23, + "args": { + "External id": 5116,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5115 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928343, "dur": 25, + "args": { + "External id": 5117,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5116 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928345, "dur": 22, + "args": { + "External id": 5118,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5117 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928373, "dur": 25, + "args": { + "External id": 5119,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5118 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928375, "dur": 23, + "args": { + "External id": 5120,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5119 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928412, "dur": 26, + "args": { + "External id": 5121,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5120 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928415, "dur": 23, + "args": { + "External id": 5122,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5121 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928443, "dur": 25, + "args": { + "External id": 5123,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5122 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928445, "dur": 23, + "args": { + "External id": 5124,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5123 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928473, "dur": 25, + "args": { + "External id": 5125,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5124 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928475, "dur": 23, + "args": { + "External id": 5126,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5125 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928502, "dur": 26, + "args": { + "External id": 5127,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5126 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928505, "dur": 22, + "args": { + "External id": 5128,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5127 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928532, "dur": 25, + "args": { + "External id": 5129,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5128 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928534, "dur": 23, + "args": { + "External id": 5130,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5129 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928562, "dur": 25, + "args": { + "External id": 5131,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5130 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928564, "dur": 23, + "args": { + "External id": 5132,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5131 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928592, "dur": 25, + "args": { + "External id": 5133,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5132 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928594, "dur": 22, + "args": { + "External id": 5134,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5133 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928621, "dur": 26, + "args": { + "External id": 5135,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5134 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928623, "dur": 23, + "args": { + "External id": 5136,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5135 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928651, "dur": 25, + "args": { + "External id": 5137,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5136 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928653, "dur": 23, + "args": { + "External id": 5138,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5137 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928681, "dur": 25, + "args": { + "External id": 5139,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5138 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928683, "dur": 22, + "args": { + "External id": 5140,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5139 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928710, "dur": 26, + "args": { + "External id": 5141,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5140 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928712, "dur": 23, + "args": { + "External id": 5142,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5141 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928740, "dur": 25, + "args": { + "External id": 5143,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5142 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928742, "dur": 22, + "args": { + "External id": 5144,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5143 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928769, "dur": 26, + "args": { + "External id": 5145,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5144 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928772, "dur": 22, + "args": { + "External id": 5146,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5145 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928799, "dur": 26, + "args": { + "External id": 5147,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5146 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928802, "dur": 22, + "args": { + "External id": 5148,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5147 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928829, "dur": 26, + "args": { + "External id": 5149,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5148 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928831, "dur": 24, + "args": { + "External id": 5150,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5149 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928860, "dur": 25, + "args": { + "External id": 5151,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5150 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928862, "dur": 22, + "args": { + "External id": 5152,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5151 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928889, "dur": 26, + "args": { + "External id": 5153,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5152 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928892, "dur": 22, + "args": { + "External id": 5154,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5153 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928919, "dur": 25, + "args": { + "External id": 5155,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5154 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928921, "dur": 23, + "args": { + "External id": 5156,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5155 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928948, "dur": 27, + "args": { + "External id": 5157,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5156 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928952, "dur": 22, + "args": { + "External id": 5158,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5157 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790928979, "dur": 25, + "args": { + "External id": 5159,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5158 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790928981, "dur": 23, + "args": { + "External id": 5160,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5159 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929008, "dur": 26, + "args": { + "External id": 5161,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5160 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929011, "dur": 22, + "args": { + "External id": 5162,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5161 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929038, "dur": 25, + "args": { + "External id": 5163,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5162 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929040, "dur": 23, + "args": { + "External id": 5164,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5163 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929068, "dur": 25, + "args": { + "External id": 5165,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5164 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929070, "dur": 22, + "args": { + "External id": 5166,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5165 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929097, "dur": 25, + "args": { + "External id": 5167,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5166 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929099, "dur": 23, + "args": { + "External id": 5168,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5167 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929127, "dur": 25, + "args": { + "External id": 5169,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5168 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929129, "dur": 23, + "args": { + "External id": 5170,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5169 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929157, "dur": 25, + "args": { + "External id": 5171,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5170 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929159, "dur": 22, + "args": { + "External id": 5172,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5171 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929186, "dur": 26, + "args": { + "External id": 5173,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5172 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929189, "dur": 22, + "args": { + "External id": 5174,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5173 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929216, "dur": 25, + "args": { + "External id": 5175,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5174 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929218, "dur": 23, + "args": { + "External id": 5176,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5175 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929246, "dur": 25, + "args": { + "External id": 5177,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5176 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929248, "dur": 22, + "args": { + "External id": 5178,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5177 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929275, "dur": 31, + "args": { + "External id": 5179,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5178 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929277, "dur": 28, + "args": { + "External id": 5180,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5179 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929311, "dur": 26, + "args": { + "External id": 5181,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5180 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929313, "dur": 24, + "args": { + "External id": 5182,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5181 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929342, "dur": 25, + "args": { + "External id": 5183,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5182 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929344, "dur": 22, + "args": { + "External id": 5184,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5183 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929371, "dur": 26, + "args": { + "External id": 5185,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5184 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929373, "dur": 23, + "args": { + "External id": 5186,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5185 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929401, "dur": 25, + "args": { + "External id": 5187,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5186 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929403, "dur": 22, + "args": { + "External id": 5188,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5187 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929430, "dur": 25, + "args": { + "External id": 5189,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5188 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929432, "dur": 23, + "args": { + "External id": 5190,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5189 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929460, "dur": 25, + "args": { + "External id": 5191,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5190 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929462, "dur": 23, + "args": { + "External id": 5192,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5191 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929490, "dur": 25, + "args": { + "External id": 5193,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5192 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929492, "dur": 22, + "args": { + "External id": 5194,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5193 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929519, "dur": 26, + "args": { + "External id": 5195,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5194 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929522, "dur": 22, + "args": { + "External id": 5196,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5195 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929549, "dur": 25, + "args": { + "External id": 5197,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5196 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929551, "dur": 23, + "args": { + "External id": 5198,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5197 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929579, "dur": 25, + "args": { + "External id": 5199,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5198 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929581, "dur": 22, + "args": { + "External id": 5200,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5199 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929669, "dur": 35, + "args": { + "External id": 5201,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5200 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929673, "dur": 30, + "args": { + "External id": 5202,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5201 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929709, "dur": 27, + "args": { + "External id": 5203,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5202 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929712, "dur": 23, + "args": { + "External id": 5204,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5203 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929741, "dur": 26, + "args": { + "External id": 5205,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5204 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929743, "dur": 23, + "args": { + "External id": 5206,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5205 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929771, "dur": 26, + "args": { + "External id": 5207,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5206 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929774, "dur": 22, + "args": { + "External id": 5208,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5207 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929801, "dur": 26, + "args": { + "External id": 5209,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5208 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929804, "dur": 22, + "args": { + "External id": 5210,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5209 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929832, "dur": 25, + "args": { + "External id": 5211,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5210 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929834, "dur": 22, + "args": { + "External id": 5212,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5211 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929861, "dur": 26, + "args": { + "External id": 5213,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5212 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929864, "dur": 22, + "args": { + "External id": 5214,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5213 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929892, "dur": 25, + "args": { + "External id": 5215,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5214 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929894, "dur": 22, + "args": { + "External id": 5216,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5215 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929922, "dur": 25, + "args": { + "External id": 5217,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5216 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929924, "dur": 22, + "args": { + "External id": 5218,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5217 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790929952, "dur": 25, + "args": { + "External id": 5219,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5218 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790929954, "dur": 22, + "args": { + "External id": 5220,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5219 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930035, "dur": 33, + "args": { + "External id": 5221,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5220 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930038, "dur": 29, + "args": { + "External id": 5222,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5221 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930074, "dur": 26, + "args": { + "External id": 5223,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5222 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930076, "dur": 23, + "args": { + "External id": 5224,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5223 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930105, "dur": 26, + "args": { + "External id": 5225,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5224 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930107, "dur": 23, + "args": { + "External id": 5226,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5225 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930135, "dur": 26, + "args": { + "External id": 5227,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5226 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930137, "dur": 23, + "args": { + "External id": 5228,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5227 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930165, "dur": 26, + "args": { + "External id": 5229,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5228 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930168, "dur": 22, + "args": { + "External id": 5230,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5229 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930195, "dur": 26, + "args": { + "External id": 5231,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5230 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930198, "dur": 22, + "args": { + "External id": 5232,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5231 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930225, "dur": 26, + "args": { + "External id": 5233,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5232 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930227, "dur": 23, + "args": { + "External id": 5234,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5233 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930255, "dur": 26, + "args": { + "External id": 5235,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5234 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930257, "dur": 23, + "args": { + "External id": 5236,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5235 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930285, "dur": 26, + "args": { + "External id": 5237,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5236 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930287, "dur": 23, + "args": { + "External id": 5238,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5237 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930315, "dur": 26, + "args": { + "External id": 5239,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5238 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930317, "dur": 23, + "args": { + "External id": 5240,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5239 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930345, "dur": 26, + "args": { + "External id": 5241,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5240 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930347, "dur": 23, + "args": { + "External id": 5242,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5241 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930375, "dur": 25, + "args": { + "External id": 5243,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5242 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930377, "dur": 22, + "args": { + "External id": 5244,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5243 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930404, "dur": 26, + "args": { + "External id": 5245,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5244 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930406, "dur": 23, + "args": { + "External id": 5246,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5245 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930434, "dur": 25, + "args": { + "External id": 5247,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5246 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930436, "dur": 23, + "args": { + "External id": 5248,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5247 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930464, "dur": 25, + "args": { + "External id": 5249,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5248 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930466, "dur": 23, + "args": { + "External id": 5250,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5249 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930494, "dur": 25, + "args": { + "External id": 5251,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5250 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930496, "dur": 22, + "args": { + "External id": 5252,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5251 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930523, "dur": 26, + "args": { + "External id": 5253,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5252 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930525, "dur": 23, + "args": { + "External id": 5254,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5253 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930553, "dur": 25, + "args": { + "External id": 5255,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5254 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930555, "dur": 23, + "args": { + "External id": 5256,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5255 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930583, "dur": 25, + "args": { + "External id": 5257,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5256 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930585, "dur": 23, + "args": { + "External id": 5258,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5257 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930612, "dur": 26, + "args": { + "External id": 5259,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5258 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930615, "dur": 22, + "args": { + "External id": 5260,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5259 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930642, "dur": 26, + "args": { + "External id": 5261,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5260 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930644, "dur": 23, + "args": { + "External id": 5262,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5261 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930672, "dur": 26, + "args": { + "External id": 5263,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5262 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930674, "dur": 23, + "args": { + "External id": 5264,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5263 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930702, "dur": 29, + "args": { + "External id": 5265,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5264 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930707, "dur": 23, + "args": { + "External id": 5266,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5265 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930735, "dur": 26, + "args": { + "External id": 5267,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5266 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930738, "dur": 22, + "args": { + "External id": 5268,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5267 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930765, "dur": 26, + "args": { + "External id": 5269,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5268 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930767, "dur": 23, + "args": { + "External id": 5270,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5269 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930795, "dur": 26, + "args": { + "External id": 5271,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5270 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930797, "dur": 24, + "args": { + "External id": 5272,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5271 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930826, "dur": 25, + "args": { + "External id": 5273,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5272 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930828, "dur": 23, + "args": { + "External id": 5274,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5273 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930856, "dur": 25, + "args": { + "External id": 5275,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5274 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930858, "dur": 23, + "args": { + "External id": 5276,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5275 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930886, "dur": 26, + "args": { + "External id": 5277,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5276 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930888, "dur": 23, + "args": { + "External id": 5278,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5277 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930916, "dur": 25, + "args": { + "External id": 5279,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5278 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930918, "dur": 23, + "args": { + "External id": 5280,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5279 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930946, "dur": 25, + "args": { + "External id": 5281,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5280 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930948, "dur": 22, + "args": { + "External id": 5282,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5281 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790930975, "dur": 26, + "args": { + "External id": 5283,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5282 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790930978, "dur": 22, + "args": { + "External id": 5284,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5283 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931005, "dur": 27, + "args": { + "External id": 5285,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5284 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931008, "dur": 23, + "args": { + "External id": 5286,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5285 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931036, "dur": 25, + "args": { + "External id": 5287,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5286 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931038, "dur": 23, + "args": { + "External id": 5288,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5287 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931066, "dur": 25, + "args": { + "External id": 5289,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5288 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931068, "dur": 23, + "args": { + "External id": 5290,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5289 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931096, "dur": 25, + "args": { + "External id": 5291,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5290 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931098, "dur": 23, + "args": { + "External id": 5292,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5291 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931126, "dur": 26, + "args": { + "External id": 5293,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5292 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931128, "dur": 24, + "args": { + "External id": 5294,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5293 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931157, "dur": 25, + "args": { + "External id": 5295,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5294 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931159, "dur": 22, + "args": { + "External id": 5296,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5295 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931186, "dur": 26, + "args": { + "External id": 5297,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5296 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931188, "dur": 23, + "args": { + "External id": 5298,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5297 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931216, "dur": 26, + "args": { + "External id": 5299,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5298 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931219, "dur": 22, + "args": { + "External id": 5300,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5299 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931246, "dur": 26, + "args": { + "External id": 5301,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5300 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931248, "dur": 23, + "args": { + "External id": 5302,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5301 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931276, "dur": 26, + "args": { + "External id": 5303,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5302 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931278, "dur": 23, + "args": { + "External id": 5304,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5303 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931306, "dur": 25, + "args": { + "External id": 5305,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5304 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931308, "dur": 23, + "args": { + "External id": 5306,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5305 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931336, "dur": 25, + "args": { + "External id": 5307,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5306 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931338, "dur": 23, + "args": { + "External id": 5308,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5307 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931366, "dur": 26, + "args": { + "External id": 5309,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5308 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931368, "dur": 23, + "args": { + "External id": 5310,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5309 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931396, "dur": 26, + "args": { + "External id": 5311,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5310 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931398, "dur": 23, + "args": { + "External id": 5312,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5311 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931426, "dur": 26, + "args": { + "External id": 5313,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5312 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931428, "dur": 23, + "args": { + "External id": 5314,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5313 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931456, "dur": 26, + "args": { + "External id": 5315,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5314 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931458, "dur": 23, + "args": { + "External id": 5316,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5315 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931486, "dur": 25, + "args": { + "External id": 5317,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5316 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931488, "dur": 23, + "args": { + "External id": 5318,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5317 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931516, "dur": 25, + "args": { + "External id": 5319,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5318 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931518, "dur": 23, + "args": { + "External id": 5320,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5319 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931604, "dur": 34, + "args": { + "External id": 5321,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5320 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931607, "dur": 30, + "args": { + "External id": 5322,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5321 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931643, "dur": 27, + "args": { + "External id": 5323,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5322 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931646, "dur": 23, + "args": { + "External id": 5324,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5323 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931674, "dur": 27, + "args": { + "External id": 5325,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5324 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931677, "dur": 23, + "args": { + "External id": 5326,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5325 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931705, "dur": 26, + "args": { + "External id": 5327,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5326 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931707, "dur": 23, + "args": { + "External id": 5328,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5327 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931736, "dur": 26, + "args": { + "External id": 5329,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5328 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931738, "dur": 23, + "args": { + "External id": 5330,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5329 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931766, "dur": 26, + "args": { + "External id": 5331,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5330 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931769, "dur": 22, + "args": { + "External id": 5332,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5331 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931797, "dur": 25, + "args": { + "External id": 5333,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5332 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931799, "dur": 23, + "args": { + "External id": 5334,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5333 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931827, "dur": 25, + "args": { + "External id": 5335,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5334 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931829, "dur": 22, + "args": { + "External id": 5336,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5335 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931856, "dur": 26, + "args": { + "External id": 5337,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5336 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931858, "dur": 23, + "args": { + "External id": 5338,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5337 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931886, "dur": 26, + "args": { + "External id": 5339,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5338 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931888, "dur": 23, + "args": { + "External id": 5340,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5339 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790931969, "dur": 32, + "args": { + "External id": 5341,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5340 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790931972, "dur": 28, + "args": { + "External id": 5342,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5341 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932006, "dur": 26, + "args": { + "External id": 5343,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5342 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932009, "dur": 22, + "args": { + "External id": 5344,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5343 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932037, "dur": 25, + "args": { + "External id": 5345,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5344 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932039, "dur": 23, + "args": { + "External id": 5346,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5345 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932066, "dur": 26, + "args": { + "External id": 5347,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5346 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932069, "dur": 22, + "args": { + "External id": 5348,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5347 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932096, "dur": 25, + "args": { + "External id": 5349,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5348 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932098, "dur": 23, + "args": { + "External id": 5350,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5349 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932125, "dur": 26, + "args": { + "External id": 5351,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5350 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932128, "dur": 22, + "args": { + "External id": 5352,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5351 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932155, "dur": 25, + "args": { + "External id": 5353,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5352 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932157, "dur": 22, + "args": { + "External id": 5354,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5353 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932184, "dur": 25, + "args": { + "External id": 5355,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5354 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932186, "dur": 23, + "args": { + "External id": 5356,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5355 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932213, "dur": 51, + "args": { + "External id": 5357,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5356 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932236, "dur": 26, + "args": { + "External id": 5358,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5357 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932269, "dur": 26, + "args": { + "External id": 5359,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5358 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932271, "dur": 23, + "args": { + "External id": 5360,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5359 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932299, "dur": 25, + "args": { + "External id": 5361,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5360 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932301, "dur": 22, + "args": { + "External id": 5362,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5361 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932328, "dur": 25, + "args": { + "External id": 5363,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5362 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932330, "dur": 23, + "args": { + "External id": 5364,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5363 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932358, "dur": 25, + "args": { + "External id": 5365,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5364 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932360, "dur": 22, + "args": { + "External id": 5366,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5365 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932387, "dur": 26, + "args": { + "External id": 5367,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5366 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932389, "dur": 23, + "args": { + "External id": 5368,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5367 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932417, "dur": 25, + "args": { + "External id": 5369,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5368 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932419, "dur": 22, + "args": { + "External id": 5370,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5369 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932446, "dur": 25, + "args": { + "External id": 5371,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5370 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932448, "dur": 22, + "args": { + "External id": 5372,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5371 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932475, "dur": 25, + "args": { + "External id": 5373,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5372 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932477, "dur": 22, + "args": { + "External id": 5374,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5373 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932504, "dur": 25, + "args": { + "External id": 5375,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5374 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932506, "dur": 22, + "args": { + "External id": 5376,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5375 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932533, "dur": 25, + "args": { + "External id": 5377,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5376 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932535, "dur": 22, + "args": { + "External id": 5378,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5377 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932562, "dur": 25, + "args": { + "External id": 5379,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5378 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932564, "dur": 22, + "args": { + "External id": 5380,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5379 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932591, "dur": 25, + "args": { + "External id": 5381,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5380 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932593, "dur": 22, + "args": { + "External id": 5382,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5381 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932620, "dur": 25, + "args": { + "External id": 5383,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5382 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932622, "dur": 22, + "args": { + "External id": 5384,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5383 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932649, "dur": 25, + "args": { + "External id": 5385,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5384 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932651, "dur": 22, + "args": { + "External id": 5386,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5385 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932678, "dur": 24, + "args": { + "External id": 5387,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5386 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932680, "dur": 22, + "args": { + "External id": 5388,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5387 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932706, "dur": 25, + "args": { + "External id": 5389,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5388 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932709, "dur": 22, + "args": { + "External id": 5390,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5389 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932736, "dur": 25, + "args": { + "External id": 5391,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5390 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932738, "dur": 22, + "args": { + "External id": 5392,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5391 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932765, "dur": 25, + "args": { + "External id": 5393,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5392 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932767, "dur": 22, + "args": { + "External id": 5394,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5393 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932794, "dur": 25, + "args": { + "External id": 5395,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5394 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932796, "dur": 22, + "args": { + "External id": 5396,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5395 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932823, "dur": 24, + "args": { + "External id": 5397,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5396 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932825, "dur": 22, + "args": { + "External id": 5398,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5397 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932852, "dur": 25, + "args": { + "External id": 5399,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5398 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932854, "dur": 22, + "args": { + "External id": 5400,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5399 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932881, "dur": 24, + "args": { + "External id": 5401,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5400 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932883, "dur": 22, + "args": { + "External id": 5402,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5401 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932909, "dur": 25, + "args": { + "External id": 5403,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5402 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932912, "dur": 22, + "args": { + "External id": 5404,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5403 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932938, "dur": 25, + "args": { + "External id": 5405,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5404 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932940, "dur": 22, + "args": { + "External id": 5406,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5405 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932967, "dur": 25, + "args": { + "External id": 5407,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5406 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932969, "dur": 23, + "args": { + "External id": 5408,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5407 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790932996, "dur": 26, + "args": { + "External id": 5409,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5408 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790932999, "dur": 22, + "args": { + "External id": 5410,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5409 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790933026, "dur": 24, + "args": { + "External id": 5411,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5410 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790933028, "dur": 22, + "args": { + "External id": 5412,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5411 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790933055, "dur": 25, + "args": { + "External id": 5413,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5412 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790933057, "dur": 23, + "args": { + "External id": 5414,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5413 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790933084, "dur": 25, + "args": { + "External id": 5415,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5414 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790933087, "dur": 22, + "args": { + "External id": 5416,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5415 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790933113, "dur": 26, + "args": { + "External id": 5417,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5416 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790933116, "dur": 22, + "args": { + "External id": 5418,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5417 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790933143, "dur": 25, + "args": { + "External id": 5419,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5418 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790933145, "dur": 22, + "args": { + "External id": 5420,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5419 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790933172, "dur": 25, + "args": { + "External id": 5421,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5420 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790933174, "dur": 22, + "args": { + "External id": 5422,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5421 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790933201, "dur": 25, + "args": { + "External id": 5423,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5422 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790933203, "dur": 22, + "args": { + "External id": 5424,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5423 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790933230, "dur": 25, + "args": { + "External id": 5425,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5424 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790933232, "dur": 22, + "args": { + "External id": 5426,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5425 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790933259, "dur": 29, + "args": { + "External id": 5427,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5426 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790933261, "dur": 26, + "args": { + "External id": 5428,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5427 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790933293, "dur": 25, + "args": { + "External id": 5429,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5428 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790933295, "dur": 22, + "args": { + "External id": 5430,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5429 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790933322, "dur": 25, + "args": { + "External id": 5431,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5430 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790933324, "dur": 22, + "args": { + "External id": 5432,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5431 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790933351, "dur": 25, + "args": { + "External id": 5433,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5432 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790933353, "dur": 23, + "args": { + "External id": 5434,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5433 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790933380, "dur": 25, + "args": { + "External id": 5435,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5434 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790933382, "dur": 23, + "args": { + "External id": 5436,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5435 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790933409, "dur": 25, + "args": { + "External id": 5437,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5436 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790933412, "dur": 22, + "args": { + "External id": 5438,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5437 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649790933439, "dur": 25, + "args": { + "External id": 5439,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5438 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649790933441, "dur": 22, + "args": { + "External id": 5440,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 4096]], "Ev Idx": 5439 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__and__", "pid": 435, "tid": 435, + "ts": 1776649790933568, "dur": 42, + "args": { + "External id": 5441,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "15"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 5440 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_and", "pid": 435, "tid": 435, + "ts": 1776649790933570, "dur": 39, + "args": { + "External id": 5442,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "15"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 5441 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_and", "pid": 435, "tid": 435, + "ts": 1776649790933576, "dur": 31, + "args": { + "External id": 5443,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["unsigned char", "long int"], "Input Dims": [[4096, 2048], []], "Ev Idx": 5442 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790933616, "dur": 34, + "args": { + "External id": 5444,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], []], "Ev Idx": 5443 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790933617, "dur": 33, + "args": { + "External id": 5445,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["unsigned char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], [], [], []], "Ev Idx": 5444 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790933623, "dur": 7, + "args": { + "External id": 5446,"Record function id": 0, "Concrete Inputs": ["[4096, 2048]", "[2048, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5445 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790933632, "dur": 17, + "args": { + "External id": 5447,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "unsigned char", "Scalar"], "Input Dims": [[4096, 2048], [4096, 2048], []], "Ev Idx": 5446 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__rshift__", "pid": 435, "tid": 435, + "ts": 1776649790933657, "dur": 27, + "args": { + "External id": 5448,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 5447 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790933665, "dur": 5, + "args": { + "External id": 5449,"Record function id": 0, "Concrete Inputs": ["[4096, 2048]", "0", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5448 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790933688, "dur": 23, + "args": { + "External id": 5450,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], []], "Ev Idx": 5449 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790933689, "dur": 21, + "args": { + "External id": 5451,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["unsigned char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 2048], [], [], [], [], [], []], "Ev Idx": 5450 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790933691, "dur": 5, + "args": { + "External id": 5452,"Record function id": 0, "Concrete Inputs": ["[4096, 2048]", "[2048, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5451 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790933698, "dur": 12, + "args": { + "External id": 5453,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "unsigned char", "Scalar"], "Input Dims": [[4096, 2048], [4096, 2048], []], "Ev Idx": 5452 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::stack", "pid": 435, "tid": 435, + "ts": 1776649790933760, "dur": 45, + "args": { + "External id": 5454,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["TensorList", "Scalar"], "Input Dims": [[[4096, 2048], [4096, 2048]], []], "Ev Idx": 5453 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649790933764, "dur": 4, + "args": { + "External id": 5455,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["signed char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 5454 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790933766, "dur": 2, + "args": { + "External id": 5456,"Record function id": 0, "Concrete Inputs": ["", "[4096, 2048, 1]", "[2048, 1, 1]", ""], "Input type": ["signed char", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 2048], [], [], []], "Ev Idx": 5455 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649790933769, "dur": 2, + "args": { + "External id": 5457,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["signed char", "Scalar"], "Input Dims": [[4096, 2048], []], "Ev Idx": 5456 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790933771, "dur": 0, + "args": { + "External id": 5458,"Record function id": 0, "Concrete Inputs": ["", "[4096, 2048, 1]", "[2048, 1, 1]", ""], "Input type": ["signed char", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 2048], [], [], []], "Ev Idx": 5457 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 435, "tid": 435, + "ts": 1776649790933773, "dur": 31, + "args": { + "External id": 5459,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["TensorList", "Scalar"], "Input Dims": [[[4096, 2048, 1], [4096, 2048, 1]], []], "Ev Idx": 5458 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649790933813, "dur": 5, + "args": { + "External id": 5460,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["signed char", "ScalarList"], "Input Dims": [[4096, 2048, 2], []], "Ev Idx": 5459 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790933814, "dur": 4, + "args": { + "External id": 5461,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["signed char", "ScalarList"], "Input Dims": [[4096, 2048, 2], []], "Ev Idx": 5460 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::repeat_interleave", "pid": 435, "tid": 435, + "ts": 1776649790933824, "dur": 59, + "args": { + "External id": 5462,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "128", "0", ""], "Input type": ["c10::Half", "Scalar", "Scalar", ""], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 5461 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649790933827, "dur": 4, + "args": { + "External id": 5463,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 4096], []], "Ev Idx": 5462 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790933829, "dur": 1, + "args": { + "External id": 5464,"Record function id": 0, "Concrete Inputs": ["", "[32, 1, 4096]", "[4096, 4096, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 5463 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 435, "tid": 435, + "ts": 1776649790933833, "dur": 5, + "args": { + "External id": 5465,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]", "False"], "Input type": ["c10::Half", "ScalarList", "Scalar"], "Input Dims": [[32, 1, 4096], [], []], "Ev Idx": 5464 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790933836, "dur": 1, + "args": { + "External id": 5466,"Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]", "[4096, 0, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 1, 4096], [], [], []], "Ev Idx": 5465 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649790933840, "dur": 35, + "args": { + "External id": 5467,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 5466 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649790933844, "dur": 9, + "args": { + "External id": 5468,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[32, 128, 4096], [], [], [], [], []], "Ev Idx": 5467 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790933846, "dur": 7, + "args": { + "External id": 5469,"Record function id": 0, "Concrete Inputs": ["[32, 128, 4096]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5468 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790933855, "dur": 20, + "args": { + "External id": 5470,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[32, 128, 4096], [32, 128, 4096], []], "Ev Idx": 5469 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::flatten", "pid": 435, "tid": 435, + "ts": 1776649790933878, "dur": 4, + "args": { + "External id": 5471,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[32, 128, 4096], [], []], "Ev Idx": 5470 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649790933880, "dur": 2, + "args": { + "External id": 5472,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 5471 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 435, "tid": 435, + "ts": 1776649790933890, "dur": 23, + "args": { + "External id": 5473,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["signed char", "long int", "Scalar"], "Input Dims": [[4096, 4096], [], []], "Ev Idx": 5472 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649790933918, "dur": 26, + "args": { + "External id": 5474,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["signed char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 5473 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649790933919, "dur": 25, + "args": { + "External id": 5475,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "", "", "", "False", ""], "Input type": ["signed char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], [], []], "Ev Idx": 5474 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649790933923, "dur": 5, + "args": { + "External id": 5476,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "[4096, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5475 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790933930, "dur": 13, + "args": { + "External id": 5477,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "signed char", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 5476 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 435, "tid": 435, + "ts": 1776649790933950, "dur": 19, + "args": { + "External id": 5478,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4096, 4096], [4096, 4096]], "Ev Idx": 5477 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 435, "tid": 435, + "ts": 1776649790933981, "dur": 12, + "args": { + "External id": 5479,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[4096, 4096]], "Ev Idx": 5478 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 435, "tid": 435, + "ts": 1776649790933986, "dur": 6, + "args": { + "External id": 5480,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[4096, 4096], [], []], "Ev Idx": 5479 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649790933990, "dur": 2, + "args": { + "External id": 5481,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]", "[1, 4096]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 5480 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 435, "tid": 435, + "ts": 1776649790934000, "dur": 28, + "args": { + "External id": 5482,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 5481 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649790934001, "dur": 27, + "args": { + "External id": 5483,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 5482 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649790934003, "dur": 8, + "args": { + "External id": 5484,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[4096, 4096], [], [], [], [], []], "Ev Idx": 5483 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790934005, "dur": 5, + "args": { + "External id": 5485,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5484 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649790934012, "dur": 15, + "args": { + "External id": 5486,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 5485 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649790934082, "dur": 42, + "args": { + "External id": 5487,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "", "", "", "False", ""], "Input type": ["ScalarList", "", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5486 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 435, "tid": 435, + "ts": 1776649790934135, "dur": 7, + "args": { + "External id": 5488,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Dims": [[4096, 4096]], "Ev Idx": 5487 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 435, "tid": 435, + "ts": 1776649790934138, "dur": 4, + "args": { + "External id": 5489,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Dims": [[4096, 4096]], "Ev Idx": 5488 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::uniform_", "pid": 435, "tid": 435, + "ts": 1776649790934214, "dur": 106304, + "args": { + "External id": 5490,"Record function id": 0, "Concrete Inputs": ["", "-0.015624999999999998", "0.015624999999999998", ""], "Input type": ["float", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 5489 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791040686, "dur": 40220, + "args": { + "External id": 5491,"Record function id": 0, "Concrete Inputs": ["", "", "5", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], []], "Ev Idx": 5490 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791040693, "dur": 40211, + "args": { + "External id": 5492,"Record function id": 0, "Concrete Inputs": ["", "5", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], [], []], "Ev Idx": 5491 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791040719, "dur": 60, + "args": { + "External id": 5493,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "[4096, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5492 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791040782, "dur": 40113, + "args": { + "External id": 5494,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "float", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 5493 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791040802, "dur": 31928, + "args": { + "External id": 5495,"Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 5494 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791040804, "dur": 31925, + "args": { + "External id": 5496,"Record function id": 0, "Concrete Inputs": ["", "5", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], [], []], "Ev Idx": 5495 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791040806, "dur": 52, + "args": { + "External id": 5497,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "[4096, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5496 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791040859, "dur": 31868, + "args": { + "External id": 5498,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "float", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 5497 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::expand_as", "pid": 435, "tid": 435, + "ts": 1776649791072735, "dur": 18, + "args": { + "External id": 5499,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4096, 4096], [4096, 4096]], "Ev Idx": 5498 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 435, "tid": 435, + "ts": 1776649791072741, "dur": 11, + "args": { + "External id": 5500,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]", "False"], "Input type": ["c10::Half", "ScalarList", "Scalar"], "Input Dims": [[4096, 4096], [], []], "Ev Idx": 5499 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791072747, "dur": 5, + "args": { + "External id": 5501,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]", "[4096, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 5500 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791072756, "dur": 3805, + "args": { + "External id": 5502,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 5501 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_has_compatible_shallow_copy_type", "pid": 435, "tid": 435, + "ts": 1776649791080997, "dur": 3, + "args": { + "External id": 5503,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "c10::Half"], "Input Dims": [[4096, 4096], [4096, 4096]], "Ev Idx": 5502 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_has_compatible_shallow_copy_type", "pid": 435, "tid": 435, + "ts": 1776649791081024, "dur": 0, + "args": { + "External id": 5504,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "c10::Half"], "Input Dims": [[4096, 4096], [4096, 4096]], "Ev Idx": 5503 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791177404, "dur": 120, + "args": { + "External id": 5505,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 5504 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791177626, "dur": 49, + "args": { + "External id": 5506,"Record function id": 0, "Concrete Inputs": ["[256, 8192]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5505 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791177695, "dur": 2, + "args": { + "External id": 5507,"Record function id": 0, "Concrete Inputs": ["[32, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5506 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791177710, "dur": 10, + "args": { + "External id": 5508,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5507 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791177714, "dur": 1, + "args": { + "External id": 5509,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5508 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791177718, "dur": 2, + "args": { + "External id": 5510,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5509 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791177758, "dur": 2478, + "args": { + "External id": 5511,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 8192], [], [], [], [], [], [], []], "Ev Idx": 5510 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791177762, "dur": 2474, + "args": { + "External id": 5512,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[256, 8192], [], [], [], [], [], []], "Ev Idx": 5511 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791177774, "dur": 23, + "args": { + "External id": 5513,"Record function id": 0, "Concrete Inputs": ["[256, 8192]", "[8192, 1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5512 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791177799, "dur": 2416, + "args": { + "External id": 5514,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 8192], [256, 8192], []], "Ev Idx": 5513 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791180249, "dur": 76, + "args": { + "External id": 5515,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[32, 4096], [], [], [], [], [], [], []], "Ev Idx": 5514 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791180250, "dur": 74, + "args": { + "External id": 5516,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[32, 4096], [], [], [], [], [], []], "Ev Idx": 5515 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791180254, "dur": 8, + "args": { + "External id": 5517,"Record function id": 0, "Concrete Inputs": ["[32, 4096]", "[4096, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5516 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791180263, "dur": 60, + "args": { + "External id": 5518,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[32, 4096], [32, 4096], []], "Ev Idx": 5517 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791180331, "dur": 32, + "args": { + "External id": 5519,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[512], [], [], [], [], [], [], []], "Ev Idx": 5518 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791180332, "dur": 30, + "args": { + "External id": 5520,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[512], [], [], [], [], [], []], "Ev Idx": 5519 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791180334, "dur": 6, + "args": { + "External id": 5521,"Record function id": 0, "Concrete Inputs": ["[512]", "[1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5520 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791180341, "dur": 21, + "args": { + "External id": 5522,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[512], [512], []], "Ev Idx": 5521 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 435, "tid": 435, + "ts": 1776649791180377, "dur": 24, + "args": { + "External id": 5523,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[32, 4096]], "Ev Idx": 5522 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 435, "tid": 435, + "ts": 1776649791180383, "dur": 14, + "args": { + "External id": 5524,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[32, 4096], [], []], "Ev Idx": 5523 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791180391, "dur": 6, + "args": { + "External id": 5525,"Record function id": 0, "Concrete Inputs": ["", "[4096, 32]", "[1, 4096]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 5524 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 435, "tid": 435, + "ts": 1776649791180406, "dur": 4, + "args": { + "External id": 5526,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[4096, 4096]], "Ev Idx": 5525 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 435, "tid": 435, + "ts": 1776649791180406, "dur": 3, + "args": { + "External id": 5527,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[4096, 4096], [], []], "Ev Idx": 5526 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791180408, "dur": 0, + "args": { + "External id": 5528,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]", "[1, 4096]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 5527 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791180419, "dur": 24, + "args": { + "External id": 5529,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 128, 4096]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 4096], []], "Ev Idx": 5528 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 435, "tid": 435, + "ts": 1776649791180436, "dur": 7, + "args": { + "External id": 5530,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]", "[128, 1, 4096]"], "Input type": ["c10::Half", "ScalarList", "ScalarList"], "Input Dims": [[4096, 4096], [], []], "Ev Idx": 5529 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 435, "tid": 435, + "ts": 1776649791180451, "dur": 8, + "args": { + "External id": 5531,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1, 0, 2]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 5530 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791180457, "dur": 1, + "args": { + "External id": 5532,"Record function id": 0, "Concrete Inputs": ["", "[128, 32, 4096]", "[1, 128, 4096]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 128, 4096], [], [], []], "Ev Idx": 5531 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791180462, "dur": 74, + "args": { + "External id": 5533,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[128, -1]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[128, 32, 4096], []], "Ev Idx": 5532 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649791180465, "dur": 61, + "args": { + "External id": 5534,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[128, 32, 4096], []], "Ev Idx": 5533 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649791180470, "dur": 15, + "args": { + "External id": 5535,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[128, 32, 4096], [], [], [], [], []], "Ev Idx": 5534 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791180474, "dur": 10, + "args": { + "External id": 5536,"Record function id": 0, "Concrete Inputs": ["[128, 32, 4096]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5535 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791180486, "dur": 39, + "args": { + "External id": 5537,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[128, 32, 4096], [128, 32, 4096], []], "Ev Idx": 5536 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 435, "tid": 435, + "ts": 1776649791180530, "dur": 5, + "args": { + "External id": 5538,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[128, 131072]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[128, 32, 4096], []], "Ev Idx": 5537 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791180543, "dur": 32, + "args": { + "External id": 5539,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1, -1]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 32], []], "Ev Idx": 5538 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649791180545, "dur": 26, + "args": { + "External id": 5540,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[4096, 32], []], "Ev Idx": 5539 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649791180547, "dur": 9, + "args": { + "External id": 5541,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[4096, 32], [], [], [], [], []], "Ev Idx": 5540 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791180548, "dur": 7, + "args": { + "External id": 5542,"Record function id": 0, "Concrete Inputs": ["[4096, 32]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5541 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791180557, "dur": 14, + "args": { + "External id": 5543,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[4096, 32], [4096, 32], []], "Ev Idx": 5542 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 435, "tid": 435, + "ts": 1776649791180573, "dur": 1, + "args": { + "External id": 5544,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1, 131072]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 32], []], "Ev Idx": 5543 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 435, "tid": 435, + "ts": 1776649791180584, "dur": 36, + "args": { + "External id": 5545,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[128, 131072], [1, 131072]], "Ev Idx": 5544 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 435, "tid": 435, + "ts": 1776649791180626, "dur": 23, + "args": { + "External id": 5546,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[128, 131072]], "Ev Idx": 5545 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791180656, "dur": 29, + "args": { + "External id": 5547,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[128, 131072], [], [], [], []], "Ev Idx": 5546 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791180658, "dur": 27, + "args": { + "External id": 5548,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "", "", "", "False", ""], "Input type": ["c10::Half", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[128, 131072], [], [], [], [], [], []], "Ev Idx": 5547 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791180661, "dur": 5, + "args": { + "External id": 5549,"Record function id": 0, "Concrete Inputs": ["[128, 131072]", "[131072, 1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5548 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791180668, "dur": 16, + "args": { + "External id": 5550,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "c10::Half", "Scalar"], "Input Dims": [[128, 131072], [128, 131072], []], "Ev Idx": 5549 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 435, "tid": 435, + "ts": 1776649791180700, "dur": 30, + "args": { + "External id": 5551,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["int", "long int", "Scalar"], "Input Dims": [[128, 131072], [], []], "Ev Idx": 5550 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 435, "tid": 435, + "ts": 1776649791180739, "dur": 31, + "args": { + "External id": 5552,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "15"], "Input type": ["int", "Scalar", "Scalar"], "Input Dims": [[128, 131072], [], []], "Ev Idx": 5551 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791180744, "dur": 1, + "args": { + "External id": 5553,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[128, 131072], [], [], [], []], "Ev Idx": 5552 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791180778, "dur": 8, + "args": { + "External id": 5554,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[128, -1, 4096]"], "Input type": ["int", "ScalarList"], "Input Dims": [[128, 131072], []], "Ev Idx": 5553 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791180780, "dur": 6, + "args": { + "External id": 5555,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[128, -1, 4096]"], "Input type": ["int", "ScalarList"], "Input Dims": [[128, 131072], []], "Ev Idx": 5554 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 435, "tid": 435, + "ts": 1776649791180789, "dur": 5, + "args": { + "External id": 5556,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1, 0, 2]"], "Input type": ["int", "ScalarList"], "Input Dims": [[128, 32, 4096], []], "Ev Idx": 5555 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791180792, "dur": 1, + "args": { + "External id": 5557,"Record function id": 0, "Concrete Inputs": ["", "[32, 128, 4096]", "[4096, 131072, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Dims": [[128, 32, 4096], [], [], []], "Ev Idx": 5556 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791180797, "dur": 35, + "args": { + "External id": 5558,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["int", "ScalarList"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 5557 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649791180800, "dur": 29, + "args": { + "External id": 5559,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 5558 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649791180802, "dur": 8, + "args": { + "External id": 5560,"Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "0"], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[32, 128, 4096], [], [], [], [], []], "Ev Idx": 5559 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791180803, "dur": 6, + "args": { + "External id": 5561,"Record function id": 0, "Concrete Inputs": ["[32, 128, 4096]", "3", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5560 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791180811, "dur": 17, + "args": { + "External id": 5562,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[32, 128, 4096], [32, 128, 4096], []], "Ev Idx": 5561 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 435, "tid": 435, + "ts": 1776649791180831, "dur": 1, + "args": { + "External id": 5563,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]"], "Input type": ["int", "ScalarList"], "Input Dims": [[32, 128, 4096], []], "Ev Idx": 5562 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791180841, "dur": 2, + "args": { + "External id": 5564,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 64]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[1, 131072], []], "Ev Idx": 5563 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791180842, "dur": 1, + "args": { + "External id": 5565,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 64]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[1, 131072], []], "Ev Idx": 5564 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649791180857, "dur": 11, + "args": { + "External id": 5566,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[2048, 64], [], [], [], []], "Ev Idx": 5565 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791180864, "dur": 1, + "args": { + "External id": 5567,"Record function id": 0, "Concrete Inputs": ["", "[2048, 64]", "[64, 1]", "0"], "Input type": ["c10::Half", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[2048, 64], [], [], []], "Ev Idx": 5566 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791180884, "dur": 2, + "args": { + "External id": 5568,"Record function id": 0, "Concrete Inputs": ["[64]", "4", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5567 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791180894, "dur": 407, + "args": { + "External id": 5569,"Record function id": 0, "Concrete Inputs": ["", "", "4", "False", "False", ""], "Input type": ["long int", "", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[64], [], [], [], [], []], "Ev Idx": 5568 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791180896, "dur": 404, + "args": { + "External id": 5570,"Record function id": 0, "Concrete Inputs": ["", "4", "", "", "", "False", ""], "Input type": ["long int", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[64], [], [], [], [], [], []], "Ev Idx": 5569 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791180898, "dur": 6, + "args": { + "External id": 5571,"Record function id": 0, "Concrete Inputs": ["[64]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5570 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791180906, "dur": 394, + "args": { + "External id": 5572,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Dims": [[64], [64], []], "Ev Idx": 5571 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 435, "tid": 435, + "ts": 1776649791181305, "dur": 1, + "args": { + "External id": 5573,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Dims": [[64]], "Ev Idx": 5572 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::index", "pid": 435, "tid": 435, + "ts": 1776649791181313, "dur": 50, + "args": { + "External id": 5574,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", ""], "Input Dims": [[2048, 64], []], "Ev Idx": 5573 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791181325, "dur": 2, + "args": { + "External id": 5575,"Record function id": 0, "Concrete Inputs": ["", "[2048, 64]", "[64, 0]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[2048, 64], [], [], []], "Ev Idx": 5574 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791181328, "dur": 2, + "args": { + "External id": 5576,"Record function id": 0, "Concrete Inputs": ["", "[1, 64]"], "Input type": ["long int", "ScalarList"], "Input Dims": [[64], []], "Ev Idx": 5575 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791181329, "dur": 1, + "args": { + "External id": 5577,"Record function id": 0, "Concrete Inputs": ["", "[1, 64]"], "Input type": ["long int", "ScalarList"], "Input Dims": [[64], []], "Ev Idx": 5576 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791181376, "dur": 4, + "args": { + "External id": 5578,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 4096]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[2048, 64], []], "Ev Idx": 5577 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791181377, "dur": 2, + "args": { + "External id": 5579,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 4096]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[2048, 64], []], "Ev Idx": 5578 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791181383, "dur": 3, + "args": { + "External id": 5580,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 16, 256, 16]"], "Input type": ["int", "ScalarList"], "Input Dims": [[4096, 4096], []], "Ev Idx": 5579 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791181384, "dur": 2, + "args": { + "External id": 5581,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 16, 256, 16]"], "Input type": ["int", "ScalarList"], "Input Dims": [[4096, 4096], []], "Ev Idx": 5580 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 435, "tid": 435, + "ts": 1776649791181389, "dur": 5, + "args": { + "External id": 5582,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[0, 2, 1, 3]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 16, 256, 16], []], "Ev Idx": 5581 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791181392, "dur": 1, + "args": { + "External id": 5583,"Record function id": 0, "Concrete Inputs": ["", "[256, 256, 16, 16]", "[65536, 16, 4096, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Dims": [[256, 16, 256, 16], [], [], []], "Ev Idx": 5582 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791181397, "dur": 36, + "args": { + "External id": 5584,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 65536]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 256, 16, 16], []], "Ev Idx": 5583 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649791181399, "dur": 30, + "args": { + "External id": 5585,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[256, 256, 16, 16], []], "Ev Idx": 5584 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649791181401, "dur": 9, + "args": { + "External id": 5586,"Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "0"], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[256, 256, 16, 16], [], [], [], [], []], "Ev Idx": 5585 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791181403, "dur": 6, + "args": { + "External id": 5587,"Record function id": 0, "Concrete Inputs": ["[256, 256, 16, 16]", "3", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5586 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791181411, "dur": 17, + "args": { + "External id": 5588,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 256, 16, 16], [256, 256, 16, 16], []], "Ev Idx": 5587 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 435, "tid": 435, + "ts": 1776649791181431, "dur": 1, + "args": { + "External id": 5589,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 65536]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 256, 16, 16], []], "Ev Idx": 5588 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791181441, "dur": 3, + "args": { + "External id": 5590,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 65536], []], "Ev Idx": 5589 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791181442, "dur": 1, + "args": { + "External id": 5591,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 65536], []], "Ev Idx": 5590 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649791181449, "dur": 5, + "args": { + "External id": 5592,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[16384, 1024], [], [], [], []], "Ev Idx": 5591 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791181452, "dur": 0, + "args": { + "External id": 5593,"Record function id": 0, "Concrete Inputs": ["", "[16384, 1024]", "[1024, 1]", "0"], "Input type": ["int", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[16384, 1024], [], [], []], "Ev Idx": 5592 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::index", "pid": 435, "tid": 435, + "ts": 1776649791181456, "dur": 128, + "args": { + "External id": 5594,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["int", ""], "Input Dims": [[16384, 1024], []], "Ev Idx": 5593 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791181459, "dur": 93, + "args": { + "External id": 5595,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "False", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Dims": [[1024], [], [], [], [], [], [], []], "Ev Idx": 5594 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791181462, "dur": 90, + "args": { + "External id": 5596,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Dims": [[1024], [], [], [], [], [], []], "Ev Idx": 5595 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791181480, "dur": 7, + "args": { + "External id": 5597,"Record function id": 0, "Concrete Inputs": ["[1024]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5596 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791181488, "dur": 63, + "args": { + "External id": 5598,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Dims": [[1024], [1024], []], "Ev Idx": 5597 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791181555, "dur": 1, + "args": { + "External id": 5599,"Record function id": 0, "Concrete Inputs": ["", "[16384, 1024]", "[1024, 0]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Dims": [[16384, 1024], [], [], []], "Ev Idx": 5598 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791181557, "dur": 2, + "args": { + "External id": 5600,"Record function id": 0, "Concrete Inputs": ["", "[1, 1024]"], "Input type": ["long int", "ScalarList"], "Input Dims": [[1024], []], "Ev Idx": 5599 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791181558, "dur": 1, + "args": { + "External id": 5601,"Record function id": 0, "Concrete Inputs": ["", "[1, 1024]"], "Input type": ["long int", "ScalarList"], "Input Dims": [[1024], []], "Ev Idx": 5600 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791181595, "dur": 4, + "args": { + "External id": 5602,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 65536]"], "Input type": ["int", "ScalarList"], "Input Dims": [[16384, 1024], []], "Ev Idx": 5601 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791181597, "dur": 2, + "args": { + "External id": 5603,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 65536]"], "Input type": ["int", "ScalarList"], "Input Dims": [[16384, 1024], []], "Ev Idx": 5602 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791195648, "dur": 49080, + "args": { + "External id": 5604,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 65536], [], [], [], [], [], [], []], "Ev Idx": 5603 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791195650, "dur": 49077, + "args": { + "External id": 5605,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[256, 65536], [], [], [], [], [], []], "Ev Idx": 5604 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791195656, "dur": 33, + "args": { + "External id": 5606,"Record function id": 0, "Concrete Inputs": ["[256, 65536]", "[65536, 1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5605 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791195691, "dur": 49031, + "args": { + "External id": 5607,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 65536], [256, 65536], []], "Ev Idx": 5606 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 435, "tid": 435, + "ts": 1776649791244756, "dur": 16, + "args": { + "External id": 5608,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 65536]], "Ev Idx": 5607 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 435, "tid": 435, + "ts": 1776649791244760, "dur": 11, + "args": { + "External id": 5609,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 65536]], "Ev Idx": 5608 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791244774, "dur": 0, + "args": { + "External id": 5610,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 65536], [], [], [], [], [], [], []], "Ev Idx": 5609 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::resolve_conj", "pid": 435, "tid": 435, + "ts": 1776649791244778, "dur": 1, + "args": { + "External id": 5611,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 65536]], "Ev Idx": 5610 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::resolve_neg", "pid": 435, "tid": 435, + "ts": 1776649791244781, "dur": 0, + "args": { + "External id": 5612,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 65536]], "Ev Idx": 5611 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 435, "tid": 435, + "ts": 1776649791304813, "dur": 8, + "args": { + "External id": 5613,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 8192]], "Ev Idx": 5612 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791304886, "dur": 798, + "args": { + "External id": 5614,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 8192], [], [], [], [], [], [], []], "Ev Idx": 5613 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791304891, "dur": 792, + "args": { + "External id": 5615,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[256, 8192], [], [], [], [], [], []], "Ev Idx": 5614 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791304913, "dur": 57, + "args": { + "External id": 5616,"Record function id": 0, "Concrete Inputs": ["[256, 8192]", "[8192, 1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5615 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791304974, "dur": 707, + "args": { + "External id": 5617,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 8192], [256, 8192], []], "Ev Idx": 5616 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791305755, "dur": 1, + "args": { + "External id": 5618,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 8192], [], [], [], [], [], [], []], "Ev Idx": 5617 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649791305785, "dur": 27, + "args": { + "External id": 5619,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[256, 8192], [], [], [], []], "Ev Idx": 5618 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791305799, "dur": 5, + "args": { + "External id": 5620,"Record function id": 0, "Concrete Inputs": ["", "[256, 8192]", "[8192, 1]", "0"], "Input type": ["int", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[256, 8192], [], [], []], "Ev Idx": 5619 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649791305814, "dur": 7, + "args": { + "External id": 5621,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[256, 8192], [], [], [], []], "Ev Idx": 5620 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791305818, "dur": 1, + "args": { + "External id": 5622,"Record function id": 0, "Concrete Inputs": ["", "[256, 8192]", "[8192, 1]", "0"], "Input type": ["int", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[256, 8192], [], [], []], "Ev Idx": 5621 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791305825, "dur": 54, + "args": { + "External id": 5623,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 8192], [256, 8192], []], "Ev Idx": 5622 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791305890, "dur": 1, + "args": { + "External id": 5624,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[32, 4096], [], [], [], [], [], [], []], "Ev Idx": 5623 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649791305896, "dur": 5, + "args": { + "External id": 5625,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[32, 4096], [], [], [], []], "Ev Idx": 5624 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791305899, "dur": 1, + "args": { + "External id": 5626,"Record function id": 0, "Concrete Inputs": ["", "[32, 4096]", "[4096, 1]", "0"], "Input type": ["c10::Half", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 5625 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649791305903, "dur": 4, + "args": { + "External id": 5627,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[32, 4096], [], [], [], []], "Ev Idx": 5626 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791305905, "dur": 1, + "args": { + "External id": 5628,"Record function id": 0, "Concrete Inputs": ["", "[32, 4096]", "[4096, 1]", "0"], "Input type": ["c10::Half", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[32, 4096], [], [], []], "Ev Idx": 5627 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791305908, "dur": 20, + "args": { + "External id": 5629,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[32, 4096], [32, 4096], []], "Ev Idx": 5628 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791306560, "dur": 22, + "args": { + "External id": 5630,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5629 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791306615, "dur": 74, + "args": { + "External id": 5631,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5630 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791306618, "dur": 6, + "args": { + "External id": 5632,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5631 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791306637, "dur": 51, + "args": { + "External id": 5633,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5632 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791306643, "dur": 44, + "args": { + "External id": 5634,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5633 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791306732, "dur": 6, + "args": { + "External id": 5635,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5634 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791306743, "dur": 23, + "args": { + "External id": 5636,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5635 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791306744, "dur": 5, + "args": { + "External id": 5637,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5636 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791306750, "dur": 16, + "args": { + "External id": 5638,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5637 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791306752, "dur": 12, + "args": { + "External id": 5639,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5638 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791306785, "dur": 6, + "args": { + "External id": 5640,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5639 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791306795, "dur": 22, + "args": { + "External id": 5641,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5640 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791306796, "dur": 4, + "args": { + "External id": 5642,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5641 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791306802, "dur": 14, + "args": { + "External id": 5643,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5642 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791306803, "dur": 12, + "args": { + "External id": 5644,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5643 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791306835, "dur": 5, + "args": { + "External id": 5645,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5644 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791306844, "dur": 22, + "args": { + "External id": 5646,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5645 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791306845, "dur": 5, + "args": { + "External id": 5647,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5646 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791306851, "dur": 14, + "args": { + "External id": 5648,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5647 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791306853, "dur": 11, + "args": { + "External id": 5649,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5648 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791306884, "dur": 6, + "args": { + "External id": 5650,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5649 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791306894, "dur": 21, + "args": { + "External id": 5651,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5650 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791306895, "dur": 4, + "args": { + "External id": 5652,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5651 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791306900, "dur": 15, + "args": { + "External id": 5653,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5652 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791306902, "dur": 11, + "args": { + "External id": 5654,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5653 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791306934, "dur": 5, + "args": { + "External id": 5655,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5654 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791306943, "dur": 22, + "args": { + "External id": 5656,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5655 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791306945, "dur": 4, + "args": { + "External id": 5657,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5656 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791306950, "dur": 15, + "args": { + "External id": 5658,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5657 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791306952, "dur": 11, + "args": { + "External id": 5659,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5658 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791306983, "dur": 5, + "args": { + "External id": 5660,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5659 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791306992, "dur": 21, + "args": { + "External id": 5661,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5660 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791306993, "dur": 4, + "args": { + "External id": 5662,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5661 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791306999, "dur": 13, + "args": { + "External id": 5663,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5662 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791307000, "dur": 11, + "args": { + "External id": 5664,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5663 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307030, "dur": 6, + "args": { + "External id": 5665,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5664 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791307040, "dur": 21, + "args": { + "External id": 5666,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5665 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307041, "dur": 4, + "args": { + "External id": 5667,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5666 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791307046, "dur": 15, + "args": { + "External id": 5668,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5667 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791307048, "dur": 12, + "args": { + "External id": 5669,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5668 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307079, "dur": 5, + "args": { + "External id": 5670,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5669 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791307088, "dur": 21, + "args": { + "External id": 5671,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5670 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307089, "dur": 4, + "args": { + "External id": 5672,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5671 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791307095, "dur": 14, + "args": { + "External id": 5673,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5672 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791307096, "dur": 11, + "args": { + "External id": 5674,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5673 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307127, "dur": 5, + "args": { + "External id": 5675,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5674 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791307136, "dur": 21, + "args": { + "External id": 5676,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5675 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307137, "dur": 5, + "args": { + "External id": 5677,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5676 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791307143, "dur": 14, + "args": { + "External id": 5678,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5677 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791307145, "dur": 10, + "args": { + "External id": 5679,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5678 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307321, "dur": 7, + "args": { + "External id": 5680,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5679 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791307333, "dur": 27, + "args": { + "External id": 5681,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5680 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307335, "dur": 4, + "args": { + "External id": 5682,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5681 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791307341, "dur": 18, + "args": { + "External id": 5683,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5682 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791307343, "dur": 15, + "args": { + "External id": 5684,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5683 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307382, "dur": 5, + "args": { + "External id": 5685,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5684 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791307391, "dur": 22, + "args": { + "External id": 5686,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5685 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307392, "dur": 5, + "args": { + "External id": 5687,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5686 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791307398, "dur": 15, + "args": { + "External id": 5688,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5687 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791307400, "dur": 11, + "args": { + "External id": 5689,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5688 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307432, "dur": 5, + "args": { + "External id": 5690,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5689 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791307441, "dur": 22, + "args": { + "External id": 5691,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5690 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307442, "dur": 5, + "args": { + "External id": 5692,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5691 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791307448, "dur": 14, + "args": { + "External id": 5693,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5692 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791307449, "dur": 12, + "args": { + "External id": 5694,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5693 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307481, "dur": 5, + "args": { + "External id": 5695,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5694 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791307490, "dur": 21, + "args": { + "External id": 5696,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5695 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307491, "dur": 4, + "args": { + "External id": 5697,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5696 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791307497, "dur": 14, + "args": { + "External id": 5698,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5697 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791307498, "dur": 11, + "args": { + "External id": 5699,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5698 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307530, "dur": 5, + "args": { + "External id": 5700,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5699 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791307539, "dur": 21, + "args": { + "External id": 5701,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5700 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307540, "dur": 4, + "args": { + "External id": 5702,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5701 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791307546, "dur": 14, + "args": { + "External id": 5703,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5702 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791307547, "dur": 11, + "args": { + "External id": 5704,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5703 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307578, "dur": 5, + "args": { + "External id": 5705,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5704 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791307588, "dur": 21, + "args": { + "External id": 5706,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5705 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307589, "dur": 4, + "args": { + "External id": 5707,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5706 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791307595, "dur": 13, + "args": { + "External id": 5708,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5707 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791307596, "dur": 11, + "args": { + "External id": 5709,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5708 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307626, "dur": 6, + "args": { + "External id": 5710,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5709 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791307635, "dur": 22, + "args": { + "External id": 5711,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5710 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307637, "dur": 4, + "args": { + "External id": 5712,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5711 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791307642, "dur": 14, + "args": { + "External id": 5713,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5712 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791307644, "dur": 11, + "args": { + "External id": 5714,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5713 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307675, "dur": 5, + "args": { + "External id": 5715,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5714 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791307684, "dur": 21, + "args": { + "External id": 5716,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5715 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307685, "dur": 4, + "args": { + "External id": 5717,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5716 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791307690, "dur": 14, + "args": { + "External id": 5718,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5717 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791307692, "dur": 11, + "args": { + "External id": 5719,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5718 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307723, "dur": 5, + "args": { + "External id": 5720,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5719 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791307732, "dur": 21, + "args": { + "External id": 5721,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5720 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307733, "dur": 5, + "args": { + "External id": 5722,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5721 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791307739, "dur": 14, + "args": { + "External id": 5723,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5722 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791307741, "dur": 11, + "args": { + "External id": 5724,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5723 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307771, "dur": 5, + "args": { + "External id": 5725,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5724 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791307780, "dur": 21, + "args": { + "External id": 5726,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5725 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307781, "dur": 5, + "args": { + "External id": 5727,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5726 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791307787, "dur": 14, + "args": { + "External id": 5728,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5727 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791307789, "dur": 10, + "args": { + "External id": 5729,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5728 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307818, "dur": 6, + "args": { + "External id": 5730,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5729 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791307828, "dur": 22, + "args": { + "External id": 5731,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5730 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307829, "dur": 4, + "args": { + "External id": 5732,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5731 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791307835, "dur": 14, + "args": { + "External id": 5733,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5732 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791307836, "dur": 12, + "args": { + "External id": 5734,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5733 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307868, "dur": 5, + "args": { + "External id": 5735,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5734 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791307877, "dur": 21, + "args": { + "External id": 5736,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5735 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307878, "dur": 5, + "args": { + "External id": 5737,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5736 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791307884, "dur": 14, + "args": { + "External id": 5738,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5737 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791307885, "dur": 12, + "args": { + "External id": 5739,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5738 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307916, "dur": 5, + "args": { + "External id": 5740,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5739 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791307925, "dur": 21, + "args": { + "External id": 5741,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5740 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307926, "dur": 4, + "args": { + "External id": 5742,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5741 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791307932, "dur": 13, + "args": { + "External id": 5743,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5742 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791307933, "dur": 11, + "args": { + "External id": 5744,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5743 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307963, "dur": 5, + "args": { + "External id": 5745,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5744 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791307972, "dur": 21, + "args": { + "External id": 5746,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5745 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791307974, "dur": 4, + "args": { + "External id": 5747,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5746 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791307979, "dur": 14, + "args": { + "External id": 5748,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5747 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791307981, "dur": 11, + "args": { + "External id": 5749,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5748 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308011, "dur": 5, + "args": { + "External id": 5750,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5749 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791308020, "dur": 21, + "args": { + "External id": 5751,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5750 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308021, "dur": 5, + "args": { + "External id": 5752,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5751 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791308027, "dur": 14, + "args": { + "External id": 5753,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5752 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791308028, "dur": 12, + "args": { + "External id": 5754,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5753 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308059, "dur": 5, + "args": { + "External id": 5755,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5754 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791308068, "dur": 21, + "args": { + "External id": 5756,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5755 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308069, "dur": 5, + "args": { + "External id": 5757,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5756 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791308075, "dur": 14, + "args": { + "External id": 5758,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5757 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791308077, "dur": 11, + "args": { + "External id": 5759,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5758 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308107, "dur": 5, + "args": { + "External id": 5760,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5759 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791308116, "dur": 21, + "args": { + "External id": 5761,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5760 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308117, "dur": 5, + "args": { + "External id": 5762,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5761 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791308123, "dur": 14, + "args": { + "External id": 5763,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5762 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791308124, "dur": 12, + "args": { + "External id": 5764,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5763 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308155, "dur": 5, + "args": { + "External id": 5765,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5764 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791308164, "dur": 21, + "args": { + "External id": 5766,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5765 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308165, "dur": 5, + "args": { + "External id": 5767,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5766 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791308171, "dur": 14, + "args": { + "External id": 5768,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5767 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791308173, "dur": 11, + "args": { + "External id": 5769,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5768 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308203, "dur": 5, + "args": { + "External id": 5770,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5769 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791308212, "dur": 63, + "args": { + "External id": 5771,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5770 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308213, "dur": 43, + "args": { + "External id": 5772,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5771 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791308258, "dur": 17, + "args": { + "External id": 5773,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5772 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791308260, "dur": 13, + "args": { + "External id": 5774,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5773 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308297, "dur": 6, + "args": { + "External id": 5775,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5774 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791308307, "dur": 22, + "args": { + "External id": 5776,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5775 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308308, "dur": 5, + "args": { + "External id": 5777,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5776 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791308314, "dur": 15, + "args": { + "External id": 5778,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5777 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791308316, "dur": 11, + "args": { + "External id": 5779,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5778 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308347, "dur": 5, + "args": { + "External id": 5780,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5779 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791308356, "dur": 22, + "args": { + "External id": 5781,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5780 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308357, "dur": 5, + "args": { + "External id": 5782,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5781 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791308363, "dur": 14, + "args": { + "External id": 5783,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5782 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791308365, "dur": 11, + "args": { + "External id": 5784,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5783 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308396, "dur": 5, + "args": { + "External id": 5785,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5784 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791308405, "dur": 21, + "args": { + "External id": 5786,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5785 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308406, "dur": 4, + "args": { + "External id": 5787,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5786 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791308412, "dur": 14, + "args": { + "External id": 5788,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5787 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791308413, "dur": 11, + "args": { + "External id": 5789,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5788 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308444, "dur": 5, + "args": { + "External id": 5790,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5789 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791308453, "dur": 21, + "args": { + "External id": 5791,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5790 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308454, "dur": 4, + "args": { + "External id": 5792,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5791 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791308460, "dur": 14, + "args": { + "External id": 5793,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5792 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791308461, "dur": 11, + "args": { + "External id": 5794,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5793 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308492, "dur": 5, + "args": { + "External id": 5795,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5794 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791308501, "dur": 21, + "args": { + "External id": 5796,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5795 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308502, "dur": 4, + "args": { + "External id": 5797,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5796 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791308508, "dur": 14, + "args": { + "External id": 5798,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5797 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791308509, "dur": 11, + "args": { + "External id": 5799,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5798 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308540, "dur": 5, + "args": { + "External id": 5800,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5799 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791308549, "dur": 21, + "args": { + "External id": 5801,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5800 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308550, "dur": 4, + "args": { + "External id": 5802,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5801 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791308556, "dur": 14, + "args": { + "External id": 5803,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5802 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791308557, "dur": 11, + "args": { + "External id": 5804,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5803 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308588, "dur": 5, + "args": { + "External id": 5805,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5804 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791308597, "dur": 21, + "args": { + "External id": 5806,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5805 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308598, "dur": 5, + "args": { + "External id": 5807,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5806 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791308604, "dur": 14, + "args": { + "External id": 5808,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5807 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791308606, "dur": 11, + "args": { + "External id": 5809,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5808 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308636, "dur": 5, + "args": { + "External id": 5810,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5809 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791308645, "dur": 21, + "args": { + "External id": 5811,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5810 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308646, "dur": 5, + "args": { + "External id": 5812,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5811 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791308652, "dur": 14, + "args": { + "External id": 5813,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5812 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791308653, "dur": 11, + "args": { + "External id": 5814,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5813 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308684, "dur": 5, + "args": { + "External id": 5815,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5814 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791308693, "dur": 21, + "args": { + "External id": 5816,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5815 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308694, "dur": 4, + "args": { + "External id": 5817,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5816 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791308700, "dur": 14, + "args": { + "External id": 5818,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5817 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791308701, "dur": 11, + "args": { + "External id": 5819,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5818 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308732, "dur": 5, + "args": { + "External id": 5820,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5819 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791308741, "dur": 21, + "args": { + "External id": 5821,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5820 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308742, "dur": 5, + "args": { + "External id": 5822,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5821 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791308748, "dur": 14, + "args": { + "External id": 5823,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5822 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791308750, "dur": 11, + "args": { + "External id": 5824,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5823 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308780, "dur": 5, + "args": { + "External id": 5825,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5824 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791308789, "dur": 21, + "args": { + "External id": 5826,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5825 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308790, "dur": 5, + "args": { + "External id": 5827,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5826 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791308796, "dur": 14, + "args": { + "External id": 5828,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5827 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791308798, "dur": 11, + "args": { + "External id": 5829,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5828 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308828, "dur": 5, + "args": { + "External id": 5830,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5829 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791308837, "dur": 21, + "args": { + "External id": 5831,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5830 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308838, "dur": 5, + "args": { + "External id": 5832,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5831 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791308844, "dur": 14, + "args": { + "External id": 5833,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5832 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791308846, "dur": 11, + "args": { + "External id": 5834,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5833 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308876, "dur": 5, + "args": { + "External id": 5835,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5834 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791308885, "dur": 21, + "args": { + "External id": 5836,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5835 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308886, "dur": 5, + "args": { + "External id": 5837,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5836 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791308892, "dur": 14, + "args": { + "External id": 5838,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5837 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791308894, "dur": 11, + "args": { + "External id": 5839,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5838 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308924, "dur": 5, + "args": { + "External id": 5840,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5839 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791308933, "dur": 21, + "args": { + "External id": 5841,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5840 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308934, "dur": 5, + "args": { + "External id": 5842,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5841 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791308940, "dur": 14, + "args": { + "External id": 5843,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5842 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791308942, "dur": 11, + "args": { + "External id": 5844,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5843 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308972, "dur": 5, + "args": { + "External id": 5845,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5844 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791308981, "dur": 21, + "args": { + "External id": 5846,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5845 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791308982, "dur": 4, + "args": { + "External id": 5847,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5846 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791308987, "dur": 15, + "args": { + "External id": 5848,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5847 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791308989, "dur": 11, + "args": { + "External id": 5849,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5848 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309020, "dur": 5, + "args": { + "External id": 5850,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5849 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791309029, "dur": 21, + "args": { + "External id": 5851,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5850 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309030, "dur": 4, + "args": { + "External id": 5852,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5851 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791309036, "dur": 14, + "args": { + "External id": 5853,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5852 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791309037, "dur": 11, + "args": { + "External id": 5854,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5853 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309068, "dur": 5, + "args": { + "External id": 5855,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5854 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791309077, "dur": 21, + "args": { + "External id": 5856,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5855 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309078, "dur": 4, + "args": { + "External id": 5857,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5856 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791309084, "dur": 14, + "args": { + "External id": 5858,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5857 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791309085, "dur": 11, + "args": { + "External id": 5859,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5858 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309117, "dur": 5, + "args": { + "External id": 5860,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5859 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791309126, "dur": 21, + "args": { + "External id": 5861,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5860 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309127, "dur": 5, + "args": { + "External id": 5862,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5861 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791309133, "dur": 14, + "args": { + "External id": 5863,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5862 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791309134, "dur": 12, + "args": { + "External id": 5864,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5863 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309165, "dur": 5, + "args": { + "External id": 5865,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5864 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791309174, "dur": 21, + "args": { + "External id": 5866,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5865 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309175, "dur": 4, + "args": { + "External id": 5867,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5866 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791309181, "dur": 14, + "args": { + "External id": 5868,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5867 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791309182, "dur": 11, + "args": { + "External id": 5869,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5868 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309213, "dur": 5, + "args": { + "External id": 5870,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5869 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791309222, "dur": 22, + "args": { + "External id": 5871,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5870 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309223, "dur": 5, + "args": { + "External id": 5872,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5871 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791309229, "dur": 14, + "args": { + "External id": 5873,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5872 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791309231, "dur": 11, + "args": { + "External id": 5874,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5873 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309262, "dur": 5, + "args": { + "External id": 5875,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5874 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791309271, "dur": 21, + "args": { + "External id": 5876,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5875 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309272, "dur": 4, + "args": { + "External id": 5877,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5876 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791309278, "dur": 14, + "args": { + "External id": 5878,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5877 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791309279, "dur": 11, + "args": { + "External id": 5879,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5878 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309310, "dur": 5, + "args": { + "External id": 5880,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5879 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791309319, "dur": 21, + "args": { + "External id": 5881,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5880 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309320, "dur": 4, + "args": { + "External id": 5882,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5881 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791309326, "dur": 14, + "args": { + "External id": 5883,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5882 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791309327, "dur": 11, + "args": { + "External id": 5884,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5883 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309358, "dur": 5, + "args": { + "External id": 5885,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5884 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791309367, "dur": 21, + "args": { + "External id": 5886,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5885 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309368, "dur": 4, + "args": { + "External id": 5887,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5886 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791309374, "dur": 13, + "args": { + "External id": 5888,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5887 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791309375, "dur": 11, + "args": { + "External id": 5889,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5888 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309405, "dur": 5, + "args": { + "External id": 5890,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5889 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791309414, "dur": 22, + "args": { + "External id": 5891,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5890 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309416, "dur": 4, + "args": { + "External id": 5892,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5891 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791309421, "dur": 14, + "args": { + "External id": 5893,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5892 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791309423, "dur": 11, + "args": { + "External id": 5894,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5893 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309453, "dur": 5, + "args": { + "External id": 5895,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5894 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791309462, "dur": 21, + "args": { + "External id": 5896,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5895 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309463, "dur": 4, + "args": { + "External id": 5897,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5896 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791309469, "dur": 14, + "args": { + "External id": 5898,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5897 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791309470, "dur": 11, + "args": { + "External id": 5899,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5898 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309501, "dur": 5, + "args": { + "External id": 5900,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5899 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791309510, "dur": 21, + "args": { + "External id": 5901,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5900 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309511, "dur": 4, + "args": { + "External id": 5902,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5901 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791309517, "dur": 14, + "args": { + "External id": 5903,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5902 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791309518, "dur": 11, + "args": { + "External id": 5904,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5903 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309549, "dur": 5, + "args": { + "External id": 5905,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5904 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791309558, "dur": 21, + "args": { + "External id": 5906,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5905 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309559, "dur": 4, + "args": { + "External id": 5907,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5906 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791309565, "dur": 14, + "args": { + "External id": 5908,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5907 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791309566, "dur": 11, + "args": { + "External id": 5909,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5908 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309598, "dur": 6, + "args": { + "External id": 5910,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5909 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791309608, "dur": 21, + "args": { + "External id": 5911,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5910 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309609, "dur": 4, + "args": { + "External id": 5912,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5911 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791309615, "dur": 14, + "args": { + "External id": 5913,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5912 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791309616, "dur": 11, + "args": { + "External id": 5914,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5913 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309647, "dur": 5, + "args": { + "External id": 5915,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5914 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791309656, "dur": 21, + "args": { + "External id": 5916,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5915 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309657, "dur": 4, + "args": { + "External id": 5917,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5916 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791309663, "dur": 13, + "args": { + "External id": 5918,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5917 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791309664, "dur": 11, + "args": { + "External id": 5919,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5918 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309695, "dur": 5, + "args": { + "External id": 5920,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5919 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791309704, "dur": 21, + "args": { + "External id": 5921,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5920 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309705, "dur": 4, + "args": { + "External id": 5922,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5921 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791309711, "dur": 14, + "args": { + "External id": 5923,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5922 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791309712, "dur": 11, + "args": { + "External id": 5924,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5923 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309743, "dur": 5, + "args": { + "External id": 5925,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5924 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791309752, "dur": 21, + "args": { + "External id": 5926,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5925 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309753, "dur": 4, + "args": { + "External id": 5927,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5926 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791309758, "dur": 15, + "args": { + "External id": 5928,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5927 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791309760, "dur": 11, + "args": { + "External id": 5929,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5928 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309862, "dur": 6, + "args": { + "External id": 5930,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5929 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791309873, "dur": 25, + "args": { + "External id": 5931,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5930 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309874, "dur": 5, + "args": { + "External id": 5932,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5931 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791309880, "dur": 18, + "args": { + "External id": 5933,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5932 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791309882, "dur": 14, + "args": { + "External id": 5934,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5933 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309919, "dur": 5, + "args": { + "External id": 5935,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5934 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791309929, "dur": 22, + "args": { + "External id": 5936,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5935 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309930, "dur": 4, + "args": { + "External id": 5937,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5936 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791309936, "dur": 14, + "args": { + "External id": 5938,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5937 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791309937, "dur": 12, + "args": { + "External id": 5939,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5938 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309969, "dur": 5, + "args": { + "External id": 5940,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5939 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791309978, "dur": 22, + "args": { + "External id": 5941,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5940 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791309979, "dur": 5, + "args": { + "External id": 5942,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5941 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791309985, "dur": 14, + "args": { + "External id": 5943,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5942 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791309987, "dur": 11, + "args": { + "External id": 5944,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5943 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310018, "dur": 5, + "args": { + "External id": 5945,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5944 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791310027, "dur": 21, + "args": { + "External id": 5946,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5945 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310028, "dur": 4, + "args": { + "External id": 5947,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5946 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791310034, "dur": 14, + "args": { + "External id": 5948,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5947 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791310035, "dur": 12, + "args": { + "External id": 5949,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5948 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310066, "dur": 5, + "args": { + "External id": 5950,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5949 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791310075, "dur": 21, + "args": { + "External id": 5951,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5950 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310076, "dur": 5, + "args": { + "External id": 5952,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5951 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791310082, "dur": 14, + "args": { + "External id": 5953,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5952 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791310084, "dur": 11, + "args": { + "External id": 5954,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5953 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310115, "dur": 5, + "args": { + "External id": 5955,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5954 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791310123, "dur": 22, + "args": { + "External id": 5956,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5955 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310125, "dur": 4, + "args": { + "External id": 5957,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5956 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791310130, "dur": 15, + "args": { + "External id": 5958,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5957 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791310132, "dur": 11, + "args": { + "External id": 5959,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5958 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310163, "dur": 5, + "args": { + "External id": 5960,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5959 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791310172, "dur": 21, + "args": { + "External id": 5961,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5960 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310173, "dur": 5, + "args": { + "External id": 5962,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5961 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791310179, "dur": 14, + "args": { + "External id": 5963,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5962 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791310181, "dur": 11, + "args": { + "External id": 5964,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5963 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310211, "dur": 6, + "args": { + "External id": 5965,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5964 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791310221, "dur": 21, + "args": { + "External id": 5966,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5965 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310222, "dur": 4, + "args": { + "External id": 5967,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5966 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791310227, "dur": 14, + "args": { + "External id": 5968,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5967 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791310229, "dur": 11, + "args": { + "External id": 5969,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5968 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310259, "dur": 6, + "args": { + "External id": 5970,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5969 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791310268, "dur": 21, + "args": { + "External id": 5971,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5970 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310270, "dur": 4, + "args": { + "External id": 5972,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5971 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791310275, "dur": 14, + "args": { + "External id": 5973,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5972 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791310277, "dur": 11, + "args": { + "External id": 5974,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5973 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310308, "dur": 5, + "args": { + "External id": 5975,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5974 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791310317, "dur": 21, + "args": { + "External id": 5976,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5975 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310318, "dur": 4, + "args": { + "External id": 5977,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5976 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791310324, "dur": 14, + "args": { + "External id": 5978,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5977 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791310325, "dur": 11, + "args": { + "External id": 5979,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5978 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310406, "dur": 6, + "args": { + "External id": 5980,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5979 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791310417, "dur": 24, + "args": { + "External id": 5981,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5980 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310418, "dur": 5, + "args": { + "External id": 5982,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5981 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791310424, "dur": 17, + "args": { + "External id": 5983,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5982 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791310426, "dur": 14, + "args": { + "External id": 5984,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5983 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310462, "dur": 6, + "args": { + "External id": 5985,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5984 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791310472, "dur": 21, + "args": { + "External id": 5986,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5985 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310473, "dur": 4, + "args": { + "External id": 5987,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5986 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791310479, "dur": 14, + "args": { + "External id": 5988,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5987 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791310480, "dur": 12, + "args": { + "External id": 5989,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5988 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310512, "dur": 5, + "args": { + "External id": 5990,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5989 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791310521, "dur": 21, + "args": { + "External id": 5991,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5990 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310522, "dur": 5, + "args": { + "External id": 5992,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5991 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791310528, "dur": 14, + "args": { + "External id": 5993,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5992 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791310530, "dur": 11, + "args": { + "External id": 5994,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5993 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310560, "dur": 6, + "args": { + "External id": 5995,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5994 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791310570, "dur": 21, + "args": { + "External id": 5996,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 5995 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310571, "dur": 4, + "args": { + "External id": 5997,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5996 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791310577, "dur": 14, + "args": { + "External id": 5998,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 5997 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791310578, "dur": 11, + "args": { + "External id": 5999,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 5998 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310609, "dur": 5, + "args": { + "External id": 6000,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5999 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791310618, "dur": 22, + "args": { + "External id": 6001,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6000 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310620, "dur": 4, + "args": { + "External id": 6002,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6001 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791310625, "dur": 14, + "args": { + "External id": 6003,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6002 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791310627, "dur": 11, + "args": { + "External id": 6004,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6003 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310657, "dur": 6, + "args": { + "External id": 6005,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6004 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791310667, "dur": 21, + "args": { + "External id": 6006,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6005 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310668, "dur": 4, + "args": { + "External id": 6007,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6006 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791310673, "dur": 14, + "args": { + "External id": 6008,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6007 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791310675, "dur": 11, + "args": { + "External id": 6009,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6008 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310705, "dur": 6, + "args": { + "External id": 6010,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6009 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791310715, "dur": 21, + "args": { + "External id": 6011,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6010 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310716, "dur": 4, + "args": { + "External id": 6012,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6011 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791310721, "dur": 15, + "args": { + "External id": 6013,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6012 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791310723, "dur": 11, + "args": { + "External id": 6014,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6013 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310754, "dur": 5, + "args": { + "External id": 6015,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6014 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791310763, "dur": 21, + "args": { + "External id": 6016,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6015 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310764, "dur": 5, + "args": { + "External id": 6017,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6016 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791310770, "dur": 14, + "args": { + "External id": 6018,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6017 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791310772, "dur": 11, + "args": { + "External id": 6019,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6018 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310802, "dur": 5, + "args": { + "External id": 6020,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6019 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791310811, "dur": 21, + "args": { + "External id": 6021,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6020 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310812, "dur": 5, + "args": { + "External id": 6022,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6021 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791310818, "dur": 14, + "args": { + "External id": 6023,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6022 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791310820, "dur": 11, + "args": { + "External id": 6024,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6023 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310850, "dur": 5, + "args": { + "External id": 6025,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6024 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791310859, "dur": 22, + "args": { + "External id": 6026,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6025 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310860, "dur": 5, + "args": { + "External id": 6027,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6026 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791310866, "dur": 14, + "args": { + "External id": 6028,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6027 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791310868, "dur": 11, + "args": { + "External id": 6029,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6028 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310898, "dur": 5, + "args": { + "External id": 6030,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6029 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791310907, "dur": 21, + "args": { + "External id": 6031,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6030 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310908, "dur": 5, + "args": { + "External id": 6032,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6031 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791310914, "dur": 14, + "args": { + "External id": 6033,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6032 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791310916, "dur": 11, + "args": { + "External id": 6034,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6033 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310946, "dur": 5, + "args": { + "External id": 6035,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6034 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791310955, "dur": 21, + "args": { + "External id": 6036,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6035 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310956, "dur": 5, + "args": { + "External id": 6037,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6036 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791310962, "dur": 14, + "args": { + "External id": 6038,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6037 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791310964, "dur": 11, + "args": { + "External id": 6039,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6038 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791310994, "dur": 5, + "args": { + "External id": 6040,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6039 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791311003, "dur": 21, + "args": { + "External id": 6041,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6040 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311004, "dur": 5, + "args": { + "External id": 6042,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6041 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791311010, "dur": 14, + "args": { + "External id": 6043,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6042 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791311011, "dur": 11, + "args": { + "External id": 6044,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6043 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311042, "dur": 5, + "args": { + "External id": 6045,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6044 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791311051, "dur": 21, + "args": { + "External id": 6046,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6045 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311052, "dur": 5, + "args": { + "External id": 6047,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6046 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791311058, "dur": 14, + "args": { + "External id": 6048,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6047 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791311060, "dur": 10, + "args": { + "External id": 6049,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6048 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311090, "dur": 5, + "args": { + "External id": 6050,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6049 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791311099, "dur": 21, + "args": { + "External id": 6051,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6050 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311100, "dur": 5, + "args": { + "External id": 6052,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6051 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791311106, "dur": 14, + "args": { + "External id": 6053,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6052 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791311108, "dur": 10, + "args": { + "External id": 6054,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6053 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311138, "dur": 5, + "args": { + "External id": 6055,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6054 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791311147, "dur": 21, + "args": { + "External id": 6056,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6055 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311148, "dur": 5, + "args": { + "External id": 6057,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6056 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791311154, "dur": 14, + "args": { + "External id": 6058,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6057 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791311156, "dur": 11, + "args": { + "External id": 6059,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6058 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311186, "dur": 5, + "args": { + "External id": 6060,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6059 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791311195, "dur": 21, + "args": { + "External id": 6061,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6060 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311196, "dur": 5, + "args": { + "External id": 6062,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6061 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791311202, "dur": 14, + "args": { + "External id": 6063,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6062 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791311204, "dur": 11, + "args": { + "External id": 6064,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6063 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311235, "dur": 5, + "args": { + "External id": 6065,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6064 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791311244, "dur": 22, + "args": { + "External id": 6066,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6065 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311245, "dur": 5, + "args": { + "External id": 6067,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6066 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791311251, "dur": 14, + "args": { + "External id": 6068,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6067 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791311253, "dur": 11, + "args": { + "External id": 6069,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6068 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311284, "dur": 5, + "args": { + "External id": 6070,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6069 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791311293, "dur": 21, + "args": { + "External id": 6071,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6070 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311294, "dur": 5, + "args": { + "External id": 6072,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6071 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791311300, "dur": 14, + "args": { + "External id": 6073,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6072 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791311301, "dur": 11, + "args": { + "External id": 6074,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6073 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311332, "dur": 5, + "args": { + "External id": 6075,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6074 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791311341, "dur": 22, + "args": { + "External id": 6076,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6075 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311342, "dur": 5, + "args": { + "External id": 6077,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6076 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791311348, "dur": 14, + "args": { + "External id": 6078,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6077 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791311350, "dur": 11, + "args": { + "External id": 6079,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6078 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311381, "dur": 5, + "args": { + "External id": 6080,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6079 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791311390, "dur": 21, + "args": { + "External id": 6081,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6080 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311391, "dur": 4, + "args": { + "External id": 6082,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6081 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791311397, "dur": 14, + "args": { + "External id": 6083,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6082 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791311398, "dur": 11, + "args": { + "External id": 6084,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6083 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311429, "dur": 5, + "args": { + "External id": 6085,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6084 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791311438, "dur": 21, + "args": { + "External id": 6086,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6085 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311439, "dur": 4, + "args": { + "External id": 6087,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6086 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791311444, "dur": 14, + "args": { + "External id": 6088,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6087 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791311446, "dur": 11, + "args": { + "External id": 6089,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6088 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311477, "dur": 5, + "args": { + "External id": 6090,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6089 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791311486, "dur": 21, + "args": { + "External id": 6091,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6090 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311487, "dur": 5, + "args": { + "External id": 6092,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6091 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791311493, "dur": 14, + "args": { + "External id": 6093,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6092 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791311495, "dur": 11, + "args": { + "External id": 6094,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6093 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311525, "dur": 5, + "args": { + "External id": 6095,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6094 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791311534, "dur": 21, + "args": { + "External id": 6096,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6095 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311535, "dur": 5, + "args": { + "External id": 6097,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6096 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791311541, "dur": 14, + "args": { + "External id": 6098,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6097 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791311543, "dur": 11, + "args": { + "External id": 6099,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6098 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311573, "dur": 5, + "args": { + "External id": 6100,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6099 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791311582, "dur": 22, + "args": { + "External id": 6101,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6100 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311584, "dur": 4, + "args": { + "External id": 6102,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6101 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791311589, "dur": 14, + "args": { + "External id": 6103,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6102 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791311591, "dur": 11, + "args": { + "External id": 6104,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6103 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311622, "dur": 5, + "args": { + "External id": 6105,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6104 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791311631, "dur": 21, + "args": { + "External id": 6106,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6105 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311632, "dur": 4, + "args": { + "External id": 6107,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6106 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791311638, "dur": 13, + "args": { + "External id": 6108,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6107 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791311639, "dur": 11, + "args": { + "External id": 6109,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6108 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311670, "dur": 5, + "args": { + "External id": 6110,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6109 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791311679, "dur": 21, + "args": { + "External id": 6111,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6110 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311680, "dur": 5, + "args": { + "External id": 6112,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6111 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791311686, "dur": 14, + "args": { + "External id": 6113,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6112 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791311687, "dur": 11, + "args": { + "External id": 6114,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6113 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311718, "dur": 5, + "args": { + "External id": 6115,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6114 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791311727, "dur": 21, + "args": { + "External id": 6116,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6115 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311728, "dur": 4, + "args": { + "External id": 6117,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6116 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791311734, "dur": 14, + "args": { + "External id": 6118,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6117 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791311735, "dur": 11, + "args": { + "External id": 6119,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6118 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311766, "dur": 5, + "args": { + "External id": 6120,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6119 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791311775, "dur": 21, + "args": { + "External id": 6121,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6120 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311776, "dur": 5, + "args": { + "External id": 6122,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6121 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791311782, "dur": 14, + "args": { + "External id": 6123,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6122 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791311784, "dur": 11, + "args": { + "External id": 6124,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6123 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311814, "dur": 6, + "args": { + "External id": 6125,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6124 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791311824, "dur": 21, + "args": { + "External id": 6126,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6125 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311825, "dur": 4, + "args": { + "External id": 6127,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6126 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791311830, "dur": 14, + "args": { + "External id": 6128,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6127 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791311832, "dur": 11, + "args": { + "External id": 6129,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6128 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311862, "dur": 5, + "args": { + "External id": 6130,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6129 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791311871, "dur": 21, + "args": { + "External id": 6131,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6130 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311872, "dur": 5, + "args": { + "External id": 6132,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6131 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791311878, "dur": 14, + "args": { + "External id": 6133,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6132 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791311880, "dur": 11, + "args": { + "External id": 6134,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6133 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311910, "dur": 5, + "args": { + "External id": 6135,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6134 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791311919, "dur": 21, + "args": { + "External id": 6136,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6135 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311920, "dur": 5, + "args": { + "External id": 6137,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6136 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791311926, "dur": 14, + "args": { + "External id": 6138,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6137 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791311928, "dur": 11, + "args": { + "External id": 6139,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6138 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311958, "dur": 5, + "args": { + "External id": 6140,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6139 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791311967, "dur": 22, + "args": { + "External id": 6141,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6140 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791311968, "dur": 5, + "args": { + "External id": 6142,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6141 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791311974, "dur": 14, + "args": { + "External id": 6143,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6142 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791311976, "dur": 11, + "args": { + "External id": 6144,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6143 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312016, "dur": 5, + "args": { + "External id": 6145,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6144 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791312026, "dur": 21, + "args": { + "External id": 6146,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6145 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312027, "dur": 4, + "args": { + "External id": 6147,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6146 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791312033, "dur": 14, + "args": { + "External id": 6148,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6147 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791312034, "dur": 11, + "args": { + "External id": 6149,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6148 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312065, "dur": 5, + "args": { + "External id": 6150,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6149 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791312074, "dur": 21, + "args": { + "External id": 6151,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6150 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312075, "dur": 5, + "args": { + "External id": 6152,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6151 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791312081, "dur": 14, + "args": { + "External id": 6153,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6152 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791312083, "dur": 11, + "args": { + "External id": 6154,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6153 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312113, "dur": 5, + "args": { + "External id": 6155,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6154 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791312122, "dur": 21, + "args": { + "External id": 6156,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6155 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312124, "dur": 4, + "args": { + "External id": 6157,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6156 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791312129, "dur": 14, + "args": { + "External id": 6158,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6157 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791312131, "dur": 11, + "args": { + "External id": 6159,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6158 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312161, "dur": 5, + "args": { + "External id": 6160,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6159 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791312170, "dur": 21, + "args": { + "External id": 6161,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6160 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312171, "dur": 5, + "args": { + "External id": 6162,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6161 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791312177, "dur": 14, + "args": { + "External id": 6163,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6162 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791312179, "dur": 11, + "args": { + "External id": 6164,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6163 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312209, "dur": 5, + "args": { + "External id": 6165,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6164 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791312237, "dur": 25, + "args": { + "External id": 6166,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6165 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312238, "dur": 6, + "args": { + "External id": 6167,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6166 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791312245, "dur": 16, + "args": { + "External id": 6168,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6167 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791312247, "dur": 13, + "args": { + "External id": 6169,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6168 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312282, "dur": 5, + "args": { + "External id": 6170,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6169 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791312291, "dur": 22, + "args": { + "External id": 6171,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6170 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312292, "dur": 5, + "args": { + "External id": 6172,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6171 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791312298, "dur": 14, + "args": { + "External id": 6173,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6172 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791312300, "dur": 11, + "args": { + "External id": 6174,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6173 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312331, "dur": 5, + "args": { + "External id": 6175,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6174 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791312340, "dur": 22, + "args": { + "External id": 6176,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6175 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312341, "dur": 5, + "args": { + "External id": 6177,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6176 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791312347, "dur": 14, + "args": { + "External id": 6178,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6177 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791312349, "dur": 11, + "args": { + "External id": 6179,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6178 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312380, "dur": 5, + "args": { + "External id": 6180,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6179 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791312389, "dur": 21, + "args": { + "External id": 6181,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6180 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312390, "dur": 4, + "args": { + "External id": 6182,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6181 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791312395, "dur": 15, + "args": { + "External id": 6183,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6182 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791312397, "dur": 11, + "args": { + "External id": 6184,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6183 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312428, "dur": 5, + "args": { + "External id": 6185,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6184 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791312437, "dur": 21, + "args": { + "External id": 6186,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6185 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312438, "dur": 5, + "args": { + "External id": 6187,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6186 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791312444, "dur": 14, + "args": { + "External id": 6188,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6187 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791312445, "dur": 11, + "args": { + "External id": 6189,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6188 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312476, "dur": 5, + "args": { + "External id": 6190,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6189 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791312485, "dur": 22, + "args": { + "External id": 6191,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6190 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312486, "dur": 5, + "args": { + "External id": 6192,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6191 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791312492, "dur": 14, + "args": { + "External id": 6193,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6192 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791312494, "dur": 11, + "args": { + "External id": 6194,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6193 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312525, "dur": 5, + "args": { + "External id": 6195,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6194 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791312534, "dur": 21, + "args": { + "External id": 6196,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6195 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312535, "dur": 4, + "args": { + "External id": 6197,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6196 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791312541, "dur": 14, + "args": { + "External id": 6198,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6197 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791312542, "dur": 11, + "args": { + "External id": 6199,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6198 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312573, "dur": 5, + "args": { + "External id": 6200,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6199 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791312582, "dur": 22, + "args": { + "External id": 6201,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6200 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312583, "dur": 5, + "args": { + "External id": 6202,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6201 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791312589, "dur": 14, + "args": { + "External id": 6203,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6202 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791312591, "dur": 11, + "args": { + "External id": 6204,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6203 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312621, "dur": 6, + "args": { + "External id": 6205,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6204 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791312631, "dur": 21, + "args": { + "External id": 6206,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6205 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312632, "dur": 4, + "args": { + "External id": 6207,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6206 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791312637, "dur": 15, + "args": { + "External id": 6208,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6207 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791312639, "dur": 11, + "args": { + "External id": 6209,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6208 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312669, "dur": 6, + "args": { + "External id": 6210,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6209 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791312679, "dur": 21, + "args": { + "External id": 6211,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6210 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312680, "dur": 4, + "args": { + "External id": 6212,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6211 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791312685, "dur": 14, + "args": { + "External id": 6213,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6212 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791312687, "dur": 11, + "args": { + "External id": 6214,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6213 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312717, "dur": 6, + "args": { + "External id": 6215,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6214 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791312727, "dur": 21, + "args": { + "External id": 6216,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6215 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312728, "dur": 4, + "args": { + "External id": 6217,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6216 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791312733, "dur": 14, + "args": { + "External id": 6218,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6217 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791312735, "dur": 11, + "args": { + "External id": 6219,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6218 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312766, "dur": 5, + "args": { + "External id": 6220,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6219 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791312775, "dur": 21, + "args": { + "External id": 6221,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6220 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312776, "dur": 4, + "args": { + "External id": 6222,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6221 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791312782, "dur": 14, + "args": { + "External id": 6223,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6222 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791312783, "dur": 11, + "args": { + "External id": 6224,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6223 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312814, "dur": 5, + "args": { + "External id": 6225,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6224 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791312823, "dur": 21, + "args": { + "External id": 6226,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6225 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312824, "dur": 5, + "args": { + "External id": 6227,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6226 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791312830, "dur": 14, + "args": { + "External id": 6228,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6227 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791312832, "dur": 11, + "args": { + "External id": 6229,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6228 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312925, "dur": 7, + "args": { + "External id": 6230,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6229 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791312936, "dur": 26, + "args": { + "External id": 6231,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6230 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312938, "dur": 4, + "args": { + "External id": 6232,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6231 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791312944, "dur": 17, + "args": { + "External id": 6233,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6232 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791312945, "dur": 15, + "args": { + "External id": 6234,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6233 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312982, "dur": 5, + "args": { + "External id": 6235,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6234 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791312991, "dur": 22, + "args": { + "External id": 6236,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6235 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791312993, "dur": 4, + "args": { + "External id": 6237,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6236 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791312998, "dur": 15, + "args": { + "External id": 6238,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6237 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791313000, "dur": 12, + "args": { + "External id": 6239,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6238 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313032, "dur": 5, + "args": { + "External id": 6240,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6239 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791313041, "dur": 21, + "args": { + "External id": 6241,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6240 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313042, "dur": 5, + "args": { + "External id": 6242,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6241 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791313048, "dur": 14, + "args": { + "External id": 6243,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6242 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791313049, "dur": 12, + "args": { + "External id": 6244,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6243 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313080, "dur": 6, + "args": { + "External id": 6245,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6244 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791313090, "dur": 21, + "args": { + "External id": 6246,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6245 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313091, "dur": 4, + "args": { + "External id": 6247,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6246 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791313097, "dur": 14, + "args": { + "External id": 6248,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6247 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791313098, "dur": 11, + "args": { + "External id": 6249,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6248 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313129, "dur": 5, + "args": { + "External id": 6250,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6249 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791313138, "dur": 21, + "args": { + "External id": 6251,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6250 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313139, "dur": 5, + "args": { + "External id": 6252,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6251 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791313145, "dur": 14, + "args": { + "External id": 6253,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6252 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791313147, "dur": 11, + "args": { + "External id": 6254,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6253 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313177, "dur": 5, + "args": { + "External id": 6255,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6254 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791313186, "dur": 21, + "args": { + "External id": 6256,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6255 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313187, "dur": 4, + "args": { + "External id": 6257,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6256 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791313193, "dur": 14, + "args": { + "External id": 6258,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6257 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791313194, "dur": 11, + "args": { + "External id": 6259,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6258 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313225, "dur": 5, + "args": { + "External id": 6260,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6259 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791313234, "dur": 22, + "args": { + "External id": 6261,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6260 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313236, "dur": 4, + "args": { + "External id": 6262,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6261 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791313241, "dur": 14, + "args": { + "External id": 6263,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6262 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791313243, "dur": 11, + "args": { + "External id": 6264,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6263 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313274, "dur": 5, + "args": { + "External id": 6265,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6264 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791313283, "dur": 22, + "args": { + "External id": 6266,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6265 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313284, "dur": 4, + "args": { + "External id": 6267,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6266 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791313290, "dur": 14, + "args": { + "External id": 6268,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6267 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791313292, "dur": 11, + "args": { + "External id": 6269,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6268 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313323, "dur": 5, + "args": { + "External id": 6270,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6269 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791313332, "dur": 21, + "args": { + "External id": 6271,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6270 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313333, "dur": 4, + "args": { + "External id": 6272,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6271 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791313339, "dur": 14, + "args": { + "External id": 6273,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6272 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791313340, "dur": 12, + "args": { + "External id": 6274,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6273 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313371, "dur": 5, + "args": { + "External id": 6275,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6274 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791313380, "dur": 21, + "args": { + "External id": 6276,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6275 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313381, "dur": 5, + "args": { + "External id": 6277,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6276 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791313387, "dur": 14, + "args": { + "External id": 6278,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6277 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791313388, "dur": 12, + "args": { + "External id": 6279,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6278 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313469, "dur": 6, + "args": { + "External id": 6280,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6279 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791313480, "dur": 24, + "args": { + "External id": 6281,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6280 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313481, "dur": 4, + "args": { + "External id": 6282,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6281 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791313487, "dur": 17, + "args": { + "External id": 6283,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6282 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791313489, "dur": 14, + "args": { + "External id": 6284,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6283 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313525, "dur": 5, + "args": { + "External id": 6285,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6284 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791313534, "dur": 22, + "args": { + "External id": 6286,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6285 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313535, "dur": 5, + "args": { + "External id": 6287,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6286 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791313541, "dur": 15, + "args": { + "External id": 6288,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6287 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791313543, "dur": 12, + "args": { + "External id": 6289,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6288 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313575, "dur": 5, + "args": { + "External id": 6290,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6289 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791313584, "dur": 22, + "args": { + "External id": 6291,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6290 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313585, "dur": 5, + "args": { + "External id": 6292,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6291 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791313591, "dur": 14, + "args": { + "External id": 6293,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6292 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791313593, "dur": 11, + "args": { + "External id": 6294,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6293 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313624, "dur": 5, + "args": { + "External id": 6295,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6294 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791313633, "dur": 21, + "args": { + "External id": 6296,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6295 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313634, "dur": 5, + "args": { + "External id": 6297,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6296 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791313640, "dur": 14, + "args": { + "External id": 6298,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6297 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791313641, "dur": 12, + "args": { + "External id": 6299,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6298 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313672, "dur": 5, + "args": { + "External id": 6300,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6299 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791313681, "dur": 22, + "args": { + "External id": 6301,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6300 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313682, "dur": 5, + "args": { + "External id": 6302,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6301 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791313688, "dur": 15, + "args": { + "External id": 6303,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6302 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791313690, "dur": 11, + "args": { + "External id": 6304,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6303 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313721, "dur": 5, + "args": { + "External id": 6305,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6304 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791313730, "dur": 21, + "args": { + "External id": 6306,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6305 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313731, "dur": 5, + "args": { + "External id": 6307,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6306 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791313737, "dur": 14, + "args": { + "External id": 6308,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6307 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791313739, "dur": 11, + "args": { + "External id": 6309,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6308 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313770, "dur": 5, + "args": { + "External id": 6310,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6309 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791313779, "dur": 21, + "args": { + "External id": 6311,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6310 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313780, "dur": 5, + "args": { + "External id": 6312,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6311 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791313786, "dur": 14, + "args": { + "External id": 6313,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6312 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791313787, "dur": 12, + "args": { + "External id": 6314,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6313 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313818, "dur": 5, + "args": { + "External id": 6315,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6314 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791313827, "dur": 21, + "args": { + "External id": 6316,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6315 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313828, "dur": 5, + "args": { + "External id": 6317,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6316 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791313834, "dur": 14, + "args": { + "External id": 6318,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6317 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791313835, "dur": 12, + "args": { + "External id": 6319,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6318 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313866, "dur": 6, + "args": { + "External id": 6320,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6319 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791313876, "dur": 21, + "args": { + "External id": 6321,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6320 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313877, "dur": 4, + "args": { + "External id": 6322,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6321 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791313882, "dur": 15, + "args": { + "External id": 6323,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6322 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791313884, "dur": 11, + "args": { + "External id": 6324,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6323 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313915, "dur": 5, + "args": { + "External id": 6325,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6324 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791313924, "dur": 21, + "args": { + "External id": 6326,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6325 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313925, "dur": 5, + "args": { + "External id": 6327,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6326 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791313931, "dur": 14, + "args": { + "External id": 6328,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6327 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791313933, "dur": 11, + "args": { + "External id": 6329,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6328 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313963, "dur": 6, + "args": { + "External id": 6330,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6329 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791313973, "dur": 21, + "args": { + "External id": 6331,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6330 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791313974, "dur": 4, + "args": { + "External id": 6332,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6331 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791313979, "dur": 15, + "args": { + "External id": 6333,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6332 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791313981, "dur": 11, + "args": { + "External id": 6334,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6333 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314012, "dur": 5, + "args": { + "External id": 6335,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6334 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791314021, "dur": 21, + "args": { + "External id": 6336,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6335 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314022, "dur": 4, + "args": { + "External id": 6337,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6336 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791314028, "dur": 14, + "args": { + "External id": 6338,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6337 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791314029, "dur": 11, + "args": { + "External id": 6339,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6338 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314060, "dur": 5, + "args": { + "External id": 6340,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6339 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791314069, "dur": 21, + "args": { + "External id": 6341,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6340 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314070, "dur": 5, + "args": { + "External id": 6342,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6341 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791314076, "dur": 14, + "args": { + "External id": 6343,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6342 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791314078, "dur": 10, + "args": { + "External id": 6344,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6343 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314108, "dur": 5, + "args": { + "External id": 6345,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6344 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791314117, "dur": 21, + "args": { + "External id": 6346,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6345 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314118, "dur": 5, + "args": { + "External id": 6347,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6346 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791314124, "dur": 14, + "args": { + "External id": 6348,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6347 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791314126, "dur": 11, + "args": { + "External id": 6349,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6348 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314156, "dur": 6, + "args": { + "External id": 6350,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6349 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791314166, "dur": 21, + "args": { + "External id": 6351,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6350 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314167, "dur": 4, + "args": { + "External id": 6352,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6351 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791314173, "dur": 14, + "args": { + "External id": 6353,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6352 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791314174, "dur": 11, + "args": { + "External id": 6354,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6353 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314205, "dur": 5, + "args": { + "External id": 6355,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6354 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791314214, "dur": 21, + "args": { + "External id": 6356,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6355 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314215, "dur": 4, + "args": { + "External id": 6357,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6356 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791314221, "dur": 14, + "args": { + "External id": 6358,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6357 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791314222, "dur": 12, + "args": { + "External id": 6359,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6358 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314253, "dur": 6, + "args": { + "External id": 6360,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6359 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791314262, "dur": 22, + "args": { + "External id": 6361,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6360 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314263, "dur": 5, + "args": { + "External id": 6362,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6361 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791314269, "dur": 15, + "args": { + "External id": 6363,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6362 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791314271, "dur": 11, + "args": { + "External id": 6364,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6363 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314302, "dur": 5, + "args": { + "External id": 6365,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6364 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791314311, "dur": 21, + "args": { + "External id": 6366,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6365 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314312, "dur": 4, + "args": { + "External id": 6367,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6366 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791314317, "dur": 15, + "args": { + "External id": 6368,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6367 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791314319, "dur": 11, + "args": { + "External id": 6369,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6368 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314350, "dur": 5, + "args": { + "External id": 6370,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6369 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791314359, "dur": 22, + "args": { + "External id": 6371,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6370 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314361, "dur": 4, + "args": { + "External id": 6372,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6371 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791314366, "dur": 14, + "args": { + "External id": 6373,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6372 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791314368, "dur": 11, + "args": { + "External id": 6374,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6373 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314398, "dur": 5, + "args": { + "External id": 6375,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6374 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791314407, "dur": 21, + "args": { + "External id": 6376,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6375 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314408, "dur": 5, + "args": { + "External id": 6377,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6376 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791314414, "dur": 14, + "args": { + "External id": 6378,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6377 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791314416, "dur": 11, + "args": { + "External id": 6379,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6378 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314446, "dur": 5, + "args": { + "External id": 6380,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6379 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791314455, "dur": 22, + "args": { + "External id": 6381,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6380 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314457, "dur": 4, + "args": { + "External id": 6382,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6381 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791314462, "dur": 15, + "args": { + "External id": 6383,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6382 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791314464, "dur": 12, + "args": { + "External id": 6384,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6383 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314495, "dur": 5, + "args": { + "External id": 6385,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6384 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791314504, "dur": 21, + "args": { + "External id": 6386,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6385 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314505, "dur": 5, + "args": { + "External id": 6387,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6386 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791314511, "dur": 14, + "args": { + "External id": 6388,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6387 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791314513, "dur": 11, + "args": { + "External id": 6389,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6388 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314543, "dur": 5, + "args": { + "External id": 6390,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6389 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791314552, "dur": 21, + "args": { + "External id": 6391,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6390 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314553, "dur": 5, + "args": { + "External id": 6392,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6391 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791314559, "dur": 14, + "args": { + "External id": 6393,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6392 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791314561, "dur": 10, + "args": { + "External id": 6394,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6393 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314591, "dur": 5, + "args": { + "External id": 6395,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6394 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791314600, "dur": 21, + "args": { + "External id": 6396,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6395 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314601, "dur": 5, + "args": { + "External id": 6397,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6396 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791314607, "dur": 14, + "args": { + "External id": 6398,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6397 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791314608, "dur": 12, + "args": { + "External id": 6399,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6398 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314639, "dur": 5, + "args": { + "External id": 6400,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6399 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791314648, "dur": 22, + "args": { + "External id": 6401,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6400 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314650, "dur": 4, + "args": { + "External id": 6402,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6401 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791314655, "dur": 14, + "args": { + "External id": 6403,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6402 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791314657, "dur": 11, + "args": { + "External id": 6404,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6403 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314688, "dur": 5, + "args": { + "External id": 6405,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6404 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791314697, "dur": 20, + "args": { + "External id": 6406,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6405 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314698, "dur": 4, + "args": { + "External id": 6407,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6406 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791314703, "dur": 14, + "args": { + "External id": 6408,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6407 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791314705, "dur": 11, + "args": { + "External id": 6409,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6408 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314736, "dur": 5, + "args": { + "External id": 6410,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6409 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791314745, "dur": 21, + "args": { + "External id": 6411,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6410 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314746, "dur": 4, + "args": { + "External id": 6412,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6411 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791314751, "dur": 15, + "args": { + "External id": 6413,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6412 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791314753, "dur": 11, + "args": { + "External id": 6414,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6413 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314784, "dur": 5, + "args": { + "External id": 6415,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6414 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791314793, "dur": 21, + "args": { + "External id": 6416,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6415 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314794, "dur": 4, + "args": { + "External id": 6417,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6416 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791314799, "dur": 15, + "args": { + "External id": 6418,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6417 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791314801, "dur": 11, + "args": { + "External id": 6419,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6418 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314832, "dur": 5, + "args": { + "External id": 6420,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6419 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791314843, "dur": 22, + "args": { + "External id": 6421,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6420 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314845, "dur": 4, + "args": { + "External id": 6422,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6421 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791314850, "dur": 15, + "args": { + "External id": 6423,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6422 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791314852, "dur": 11, + "args": { + "External id": 6424,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6423 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314891, "dur": 5, + "args": { + "External id": 6425,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6424 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791314901, "dur": 21, + "args": { + "External id": 6426,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6425 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314902, "dur": 4, + "args": { + "External id": 6427,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6426 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791314907, "dur": 15, + "args": { + "External id": 6428,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6427 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791314909, "dur": 12, + "args": { + "External id": 6429,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6428 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314940, "dur": 5, + "args": { + "External id": 6430,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6429 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791314950, "dur": 21, + "args": { + "External id": 6431,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6430 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314951, "dur": 4, + "args": { + "External id": 6432,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6431 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791314956, "dur": 14, + "args": { + "External id": 6433,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6432 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791314958, "dur": 11, + "args": { + "External id": 6434,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6433 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314989, "dur": 5, + "args": { + "External id": 6435,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6434 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791314998, "dur": 21, + "args": { + "External id": 6436,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6435 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791314999, "dur": 4, + "args": { + "External id": 6437,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6436 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791315005, "dur": 14, + "args": { + "External id": 6438,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6437 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791315006, "dur": 12, + "args": { + "External id": 6439,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6438 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315037, "dur": 5, + "args": { + "External id": 6440,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6439 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791315046, "dur": 22, + "args": { + "External id": 6441,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6440 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315047, "dur": 5, + "args": { + "External id": 6442,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6441 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791315053, "dur": 14, + "args": { + "External id": 6443,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6442 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791315055, "dur": 11, + "args": { + "External id": 6444,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6443 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315085, "dur": 5, + "args": { + "External id": 6445,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6444 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791315094, "dur": 21, + "args": { + "External id": 6446,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6445 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315095, "dur": 5, + "args": { + "External id": 6447,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6446 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791315101, "dur": 14, + "args": { + "External id": 6448,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6447 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791315103, "dur": 11, + "args": { + "External id": 6449,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6448 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315133, "dur": 5, + "args": { + "External id": 6450,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6449 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791315142, "dur": 21, + "args": { + "External id": 6451,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6450 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315143, "dur": 5, + "args": { + "External id": 6452,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6451 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791315149, "dur": 14, + "args": { + "External id": 6453,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6452 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791315151, "dur": 11, + "args": { + "External id": 6454,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6453 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315181, "dur": 5, + "args": { + "External id": 6455,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6454 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791315190, "dur": 21, + "args": { + "External id": 6456,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6455 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315191, "dur": 5, + "args": { + "External id": 6457,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6456 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791315197, "dur": 14, + "args": { + "External id": 6458,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6457 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791315199, "dur": 11, + "args": { + "External id": 6459,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6458 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315229, "dur": 5, + "args": { + "External id": 6460,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6459 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791315238, "dur": 21, + "args": { + "External id": 6461,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6460 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315239, "dur": 4, + "args": { + "External id": 6462,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6461 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791315245, "dur": 14, + "args": { + "External id": 6463,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6462 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791315246, "dur": 11, + "args": { + "External id": 6464,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6463 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315277, "dur": 5, + "args": { + "External id": 6465,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6464 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791315286, "dur": 21, + "args": { + "External id": 6466,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6465 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315287, "dur": 5, + "args": { + "External id": 6467,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6466 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791315293, "dur": 14, + "args": { + "External id": 6468,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6467 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791315295, "dur": 11, + "args": { + "External id": 6469,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6468 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315325, "dur": 5, + "args": { + "External id": 6470,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6469 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791315334, "dur": 22, + "args": { + "External id": 6471,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6470 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315335, "dur": 5, + "args": { + "External id": 6472,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6471 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791315341, "dur": 15, + "args": { + "External id": 6473,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6472 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791315343, "dur": 12, + "args": { + "External id": 6474,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6473 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315374, "dur": 6, + "args": { + "External id": 6475,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6474 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791315384, "dur": 21, + "args": { + "External id": 6476,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6475 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315385, "dur": 4, + "args": { + "External id": 6477,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6476 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791315390, "dur": 14, + "args": { + "External id": 6478,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6477 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791315392, "dur": 11, + "args": { + "External id": 6479,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6478 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315423, "dur": 5, + "args": { + "External id": 6480,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6479 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791315431, "dur": 22, + "args": { + "External id": 6481,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6480 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315433, "dur": 4, + "args": { + "External id": 6482,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6481 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791315438, "dur": 14, + "args": { + "External id": 6483,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6482 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791315440, "dur": 11, + "args": { + "External id": 6484,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6483 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315470, "dur": 6, + "args": { + "External id": 6485,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6484 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791315479, "dur": 22, + "args": { + "External id": 6486,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6485 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315481, "dur": 4, + "args": { + "External id": 6487,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6486 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791315486, "dur": 14, + "args": { + "External id": 6488,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6487 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791315488, "dur": 11, + "args": { + "External id": 6489,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6488 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315518, "dur": 5, + "args": { + "External id": 6490,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6489 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791315527, "dur": 22, + "args": { + "External id": 6491,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6490 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315528, "dur": 5, + "args": { + "External id": 6492,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6491 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791315534, "dur": 14, + "args": { + "External id": 6493,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6492 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791315536, "dur": 11, + "args": { + "External id": 6494,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6493 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315566, "dur": 6, + "args": { + "External id": 6495,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6494 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791315576, "dur": 21, + "args": { + "External id": 6496,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6495 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315577, "dur": 4, + "args": { + "External id": 6497,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6496 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791315582, "dur": 14, + "args": { + "External id": 6498,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6497 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791315584, "dur": 11, + "args": { + "External id": 6499,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6498 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315614, "dur": 5, + "args": { + "External id": 6500,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6499 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791315623, "dur": 22, + "args": { + "External id": 6501,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6500 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315624, "dur": 5, + "args": { + "External id": 6502,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6501 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791315630, "dur": 14, + "args": { + "External id": 6503,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6502 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791315632, "dur": 11, + "args": { + "External id": 6504,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6503 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315663, "dur": 5, + "args": { + "External id": 6505,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6504 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791315672, "dur": 21, + "args": { + "External id": 6506,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6505 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315673, "dur": 5, + "args": { + "External id": 6507,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6506 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791315679, "dur": 14, + "args": { + "External id": 6508,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6507 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791315681, "dur": 10, + "args": { + "External id": 6509,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6508 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315711, "dur": 5, + "args": { + "External id": 6510,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6509 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791315720, "dur": 21, + "args": { + "External id": 6511,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6510 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315721, "dur": 4, + "args": { + "External id": 6512,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6511 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791315726, "dur": 14, + "args": { + "External id": 6513,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6512 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791315728, "dur": 11, + "args": { + "External id": 6514,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6513 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315758, "dur": 5, + "args": { + "External id": 6515,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6514 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791315767, "dur": 21, + "args": { + "External id": 6516,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6515 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315768, "dur": 5, + "args": { + "External id": 6517,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6516 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791315774, "dur": 14, + "args": { + "External id": 6518,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6517 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791315776, "dur": 11, + "args": { + "External id": 6519,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6518 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315806, "dur": 5, + "args": { + "External id": 6520,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6519 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791315815, "dur": 21, + "args": { + "External id": 6521,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6520 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315816, "dur": 5, + "args": { + "External id": 6522,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6521 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791315822, "dur": 14, + "args": { + "External id": 6523,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6522 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791315824, "dur": 11, + "args": { + "External id": 6524,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6523 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315854, "dur": 5, + "args": { + "External id": 6525,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6524 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791315863, "dur": 21, + "args": { + "External id": 6526,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6525 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315864, "dur": 5, + "args": { + "External id": 6527,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6526 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791315870, "dur": 13, + "args": { + "External id": 6528,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6527 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791315871, "dur": 11, + "args": { + "External id": 6529,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6528 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315962, "dur": 7, + "args": { + "External id": 6530,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6529 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791315974, "dur": 24, + "args": { + "External id": 6531,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6530 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791315975, "dur": 4, + "args": { + "External id": 6532,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6531 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791315981, "dur": 17, + "args": { + "External id": 6533,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6532 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791315983, "dur": 14, + "args": { + "External id": 6534,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6533 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316019, "dur": 5, + "args": { + "External id": 6535,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6534 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791316028, "dur": 22, + "args": { + "External id": 6536,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6535 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316030, "dur": 4, + "args": { + "External id": 6537,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6536 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791316035, "dur": 15, + "args": { + "External id": 6538,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6537 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791316037, "dur": 11, + "args": { + "External id": 6539,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6538 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316068, "dur": 5, + "args": { + "External id": 6540,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6539 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791316077, "dur": 22, + "args": { + "External id": 6541,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6540 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316079, "dur": 4, + "args": { + "External id": 6542,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6541 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791316084, "dur": 15, + "args": { + "External id": 6543,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6542 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791316086, "dur": 11, + "args": { + "External id": 6544,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6543 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316117, "dur": 5, + "args": { + "External id": 6545,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6544 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791316126, "dur": 22, + "args": { + "External id": 6546,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6545 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316127, "dur": 5, + "args": { + "External id": 6547,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6546 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791316133, "dur": 15, + "args": { + "External id": 6548,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6547 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791316135, "dur": 11, + "args": { + "External id": 6549,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6548 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316167, "dur": 6, + "args": { + "External id": 6550,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6549 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791316177, "dur": 21, + "args": { + "External id": 6551,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6550 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316178, "dur": 4, + "args": { + "External id": 6552,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6551 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791316183, "dur": 15, + "args": { + "External id": 6553,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6552 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791316185, "dur": 11, + "args": { + "External id": 6554,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6553 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316236, "dur": 6, + "args": { + "External id": 6555,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6554 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791316248, "dur": 23, + "args": { + "External id": 6556,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6555 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316250, "dur": 4, + "args": { + "External id": 6557,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6556 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791316255, "dur": 16, + "args": { + "External id": 6558,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6557 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791316257, "dur": 13, + "args": { + "External id": 6559,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6558 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316291, "dur": 5, + "args": { + "External id": 6560,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6559 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791316301, "dur": 21, + "args": { + "External id": 6561,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6560 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316302, "dur": 4, + "args": { + "External id": 6562,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6561 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791316308, "dur": 14, + "args": { + "External id": 6563,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6562 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791316309, "dur": 12, + "args": { + "External id": 6564,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6563 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316340, "dur": 6, + "args": { + "External id": 6565,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6564 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791316349, "dur": 22, + "args": { + "External id": 6566,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6565 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316351, "dur": 4, + "args": { + "External id": 6567,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6566 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791316356, "dur": 15, + "args": { + "External id": 6568,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6567 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791316358, "dur": 11, + "args": { + "External id": 6569,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6568 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316389, "dur": 5, + "args": { + "External id": 6570,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6569 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791316398, "dur": 22, + "args": { + "External id": 6571,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6570 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316400, "dur": 4, + "args": { + "External id": 6572,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6571 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791316405, "dur": 14, + "args": { + "External id": 6573,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6572 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791316407, "dur": 11, + "args": { + "External id": 6574,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6573 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316437, "dur": 6, + "args": { + "External id": 6575,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6574 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791316447, "dur": 21, + "args": { + "External id": 6576,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6575 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316448, "dur": 4, + "args": { + "External id": 6577,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6576 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791316453, "dur": 15, + "args": { + "External id": 6578,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6577 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791316455, "dur": 11, + "args": { + "External id": 6579,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6578 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316537, "dur": 6, + "args": { + "External id": 6580,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6579 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791316548, "dur": 25, + "args": { + "External id": 6581,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6580 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316549, "dur": 5, + "args": { + "External id": 6582,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6581 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791316555, "dur": 17, + "args": { + "External id": 6583,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6582 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791316557, "dur": 14, + "args": { + "External id": 6584,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6583 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316593, "dur": 5, + "args": { + "External id": 6585,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6584 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791316602, "dur": 22, + "args": { + "External id": 6586,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6585 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316603, "dur": 5, + "args": { + "External id": 6587,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6586 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791316609, "dur": 14, + "args": { + "External id": 6588,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6587 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791316611, "dur": 11, + "args": { + "External id": 6589,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6588 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316642, "dur": 5, + "args": { + "External id": 6590,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6589 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791316651, "dur": 21, + "args": { + "External id": 6591,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6590 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316652, "dur": 5, + "args": { + "External id": 6592,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6591 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791316658, "dur": 14, + "args": { + "External id": 6593,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6592 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791316660, "dur": 11, + "args": { + "External id": 6594,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6593 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316691, "dur": 5, + "args": { + "External id": 6595,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6594 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791316700, "dur": 21, + "args": { + "External id": 6596,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6595 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316701, "dur": 5, + "args": { + "External id": 6597,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6596 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791316707, "dur": 14, + "args": { + "External id": 6598,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6597 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791316708, "dur": 12, + "args": { + "External id": 6599,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6598 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316739, "dur": 5, + "args": { + "External id": 6600,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6599 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791316748, "dur": 21, + "args": { + "External id": 6601,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6600 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316749, "dur": 5, + "args": { + "External id": 6602,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6601 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791316755, "dur": 14, + "args": { + "External id": 6603,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6602 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791316757, "dur": 11, + "args": { + "External id": 6604,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6603 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316788, "dur": 5, + "args": { + "External id": 6605,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6604 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791316797, "dur": 21, + "args": { + "External id": 6606,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6605 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316798, "dur": 4, + "args": { + "External id": 6607,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6606 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791316804, "dur": 13, + "args": { + "External id": 6608,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6607 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791316805, "dur": 11, + "args": { + "External id": 6609,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6608 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316838, "dur": 5, + "args": { + "External id": 6610,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6609 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791316848, "dur": 23, + "args": { + "External id": 6611,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6610 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316849, "dur": 6, + "args": { + "External id": 6612,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6611 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791316856, "dur": 14, + "args": { + "External id": 6613,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6612 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791316858, "dur": 11, + "args": { + "External id": 6614,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6613 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316889, "dur": 5, + "args": { + "External id": 6615,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6614 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791316898, "dur": 21, + "args": { + "External id": 6616,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6615 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316899, "dur": 4, + "args": { + "External id": 6617,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6616 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791316905, "dur": 14, + "args": { + "External id": 6618,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6617 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791316906, "dur": 11, + "args": { + "External id": 6619,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6618 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316937, "dur": 5, + "args": { + "External id": 6620,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6619 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791316946, "dur": 21, + "args": { + "External id": 6621,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6620 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316947, "dur": 5, + "args": { + "External id": 6622,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6621 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791316953, "dur": 14, + "args": { + "External id": 6623,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6622 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791316955, "dur": 11, + "args": { + "External id": 6624,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6623 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316985, "dur": 5, + "args": { + "External id": 6625,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6624 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791316994, "dur": 22, + "args": { + "External id": 6626,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6625 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791316995, "dur": 5, + "args": { + "External id": 6627,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6626 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791317001, "dur": 14, + "args": { + "External id": 6628,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6627 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791317003, "dur": 11, + "args": { + "External id": 6629,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6628 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317033, "dur": 6, + "args": { + "External id": 6630,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6629 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791317043, "dur": 21, + "args": { + "External id": 6631,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6630 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317044, "dur": 4, + "args": { + "External id": 6632,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6631 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791317050, "dur": 14, + "args": { + "External id": 6633,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6632 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791317051, "dur": 11, + "args": { + "External id": 6634,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6633 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317082, "dur": 5, + "args": { + "External id": 6635,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6634 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791317091, "dur": 21, + "args": { + "External id": 6636,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6635 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317092, "dur": 5, + "args": { + "External id": 6637,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6636 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791317098, "dur": 14, + "args": { + "External id": 6638,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6637 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791317099, "dur": 12, + "args": { + "External id": 6639,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6638 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317130, "dur": 5, + "args": { + "External id": 6640,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6639 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791317139, "dur": 21, + "args": { + "External id": 6641,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6640 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317140, "dur": 5, + "args": { + "External id": 6642,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6641 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791317146, "dur": 14, + "args": { + "External id": 6643,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6642 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791317148, "dur": 11, + "args": { + "External id": 6644,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6643 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317178, "dur": 6, + "args": { + "External id": 6645,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6644 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791317188, "dur": 21, + "args": { + "External id": 6646,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6645 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317189, "dur": 4, + "args": { + "External id": 6647,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6646 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791317194, "dur": 14, + "args": { + "External id": 6648,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6647 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791317196, "dur": 11, + "args": { + "External id": 6649,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6648 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317226, "dur": 6, + "args": { + "External id": 6650,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6649 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791317235, "dur": 22, + "args": { + "External id": 6651,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6650 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317237, "dur": 4, + "args": { + "External id": 6652,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6651 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791317242, "dur": 14, + "args": { + "External id": 6653,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6652 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791317244, "dur": 11, + "args": { + "External id": 6654,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6653 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317274, "dur": 6, + "args": { + "External id": 6655,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6654 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791317284, "dur": 30, + "args": { + "External id": 6656,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6655 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317294, "dur": 4, + "args": { + "External id": 6657,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6656 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791317300, "dur": 14, + "args": { + "External id": 6658,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6657 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791317301, "dur": 11, + "args": { + "External id": 6659,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6658 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317332, "dur": 6, + "args": { + "External id": 6660,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6659 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791317341, "dur": 21, + "args": { + "External id": 6661,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6660 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317342, "dur": 5, + "args": { + "External id": 6662,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6661 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791317348, "dur": 14, + "args": { + "External id": 6663,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6662 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791317350, "dur": 11, + "args": { + "External id": 6664,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6663 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317380, "dur": 5, + "args": { + "External id": 6665,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6664 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791317389, "dur": 22, + "args": { + "External id": 6666,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6665 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317391, "dur": 4, + "args": { + "External id": 6667,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6666 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791317396, "dur": 14, + "args": { + "External id": 6668,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6667 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791317398, "dur": 11, + "args": { + "External id": 6669,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6668 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317429, "dur": 5, + "args": { + "External id": 6670,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6669 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791317438, "dur": 22, + "args": { + "External id": 6671,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6670 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317439, "dur": 5, + "args": { + "External id": 6672,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6671 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791317445, "dur": 14, + "args": { + "External id": 6673,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6672 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791317447, "dur": 11, + "args": { + "External id": 6674,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6673 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317478, "dur": 5, + "args": { + "External id": 6675,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6674 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791317487, "dur": 22, + "args": { + "External id": 6676,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6675 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317488, "dur": 5, + "args": { + "External id": 6677,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6676 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791317494, "dur": 15, + "args": { + "External id": 6678,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6677 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791317496, "dur": 11, + "args": { + "External id": 6679,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6678 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317527, "dur": 5, + "args": { + "External id": 6680,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6679 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791317536, "dur": 21, + "args": { + "External id": 6681,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6680 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317537, "dur": 5, + "args": { + "External id": 6682,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6681 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791317543, "dur": 14, + "args": { + "External id": 6683,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6682 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791317545, "dur": 10, + "args": { + "External id": 6684,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6683 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317575, "dur": 5, + "args": { + "External id": 6685,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6684 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791317584, "dur": 21, + "args": { + "External id": 6686,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6685 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317585, "dur": 5, + "args": { + "External id": 6687,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6686 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791317591, "dur": 14, + "args": { + "External id": 6688,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6687 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791317593, "dur": 11, + "args": { + "External id": 6689,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6688 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317623, "dur": 5, + "args": { + "External id": 6690,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6689 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791317632, "dur": 21, + "args": { + "External id": 6691,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6690 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317633, "dur": 5, + "args": { + "External id": 6692,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6691 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791317639, "dur": 14, + "args": { + "External id": 6693,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6692 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791317641, "dur": 11, + "args": { + "External id": 6694,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6693 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317672, "dur": 5, + "args": { + "External id": 6695,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6694 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791317681, "dur": 21, + "args": { + "External id": 6696,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6695 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317682, "dur": 4, + "args": { + "External id": 6697,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6696 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791317688, "dur": 13, + "args": { + "External id": 6698,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6697 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791317689, "dur": 11, + "args": { + "External id": 6699,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6698 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317720, "dur": 5, + "args": { + "External id": 6700,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6699 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791317729, "dur": 21, + "args": { + "External id": 6701,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6700 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317730, "dur": 4, + "args": { + "External id": 6702,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6701 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791317736, "dur": 14, + "args": { + "External id": 6703,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6702 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791317737, "dur": 12, + "args": { + "External id": 6704,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6703 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317768, "dur": 5, + "args": { + "External id": 6705,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6704 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791317777, "dur": 21, + "args": { + "External id": 6706,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6705 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317778, "dur": 5, + "args": { + "External id": 6707,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6706 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791317784, "dur": 14, + "args": { + "External id": 6708,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6707 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791317786, "dur": 11, + "args": { + "External id": 6709,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6708 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317816, "dur": 5, + "args": { + "External id": 6710,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6709 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791317825, "dur": 21, + "args": { + "External id": 6711,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6710 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317826, "dur": 5, + "args": { + "External id": 6712,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6711 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791317832, "dur": 14, + "args": { + "External id": 6713,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6712 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791317834, "dur": 11, + "args": { + "External id": 6714,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6713 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317865, "dur": 5, + "args": { + "External id": 6715,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6714 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791317874, "dur": 21, + "args": { + "External id": 6716,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6715 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317875, "dur": 4, + "args": { + "External id": 6717,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6716 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791317881, "dur": 14, + "args": { + "External id": 6718,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6717 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791317882, "dur": 12, + "args": { + "External id": 6719,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6718 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317913, "dur": 5, + "args": { + "External id": 6720,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6719 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791317922, "dur": 22, + "args": { + "External id": 6721,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6720 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317923, "dur": 5, + "args": { + "External id": 6722,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6721 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791317929, "dur": 14, + "args": { + "External id": 6723,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6722 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791317931, "dur": 11, + "args": { + "External id": 6724,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6723 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317962, "dur": 5, + "args": { + "External id": 6725,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6724 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791317971, "dur": 21, + "args": { + "External id": 6726,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6725 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791317972, "dur": 4, + "args": { + "External id": 6727,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6726 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791317978, "dur": 13, + "args": { + "External id": 6728,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6727 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791317979, "dur": 11, + "args": { + "External id": 6729,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6728 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318010, "dur": 5, + "args": { + "External id": 6730,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6729 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791318019, "dur": 21, + "args": { + "External id": 6731,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6730 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318020, "dur": 4, + "args": { + "External id": 6732,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6731 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791318026, "dur": 14, + "args": { + "External id": 6733,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6732 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791318027, "dur": 11, + "args": { + "External id": 6734,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6733 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318058, "dur": 5, + "args": { + "External id": 6735,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6734 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791318067, "dur": 21, + "args": { + "External id": 6736,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6735 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318068, "dur": 5, + "args": { + "External id": 6737,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6736 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791318074, "dur": 14, + "args": { + "External id": 6738,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6737 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791318076, "dur": 11, + "args": { + "External id": 6739,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6738 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318106, "dur": 5, + "args": { + "External id": 6740,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6739 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791318115, "dur": 21, + "args": { + "External id": 6741,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6740 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318116, "dur": 5, + "args": { + "External id": 6742,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6741 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791318122, "dur": 14, + "args": { + "External id": 6743,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6742 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791318124, "dur": 10, + "args": { + "External id": 6744,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6743 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318154, "dur": 5, + "args": { + "External id": 6745,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6744 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791318163, "dur": 21, + "args": { + "External id": 6746,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6745 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318164, "dur": 4, + "args": { + "External id": 6747,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6746 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791318170, "dur": 13, + "args": { + "External id": 6748,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6747 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791318171, "dur": 11, + "args": { + "External id": 6749,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6748 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318202, "dur": 5, + "args": { + "External id": 6750,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6749 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791318211, "dur": 21, + "args": { + "External id": 6751,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6750 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318212, "dur": 5, + "args": { + "External id": 6752,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6751 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791318218, "dur": 14, + "args": { + "External id": 6753,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6752 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791318220, "dur": 11, + "args": { + "External id": 6754,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6753 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318250, "dur": 6, + "args": { + "External id": 6755,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6754 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791318260, "dur": 21, + "args": { + "External id": 6756,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6755 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318261, "dur": 4, + "args": { + "External id": 6757,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6756 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791318266, "dur": 14, + "args": { + "External id": 6758,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6757 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791318268, "dur": 11, + "args": { + "External id": 6759,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6758 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318299, "dur": 5, + "args": { + "External id": 6760,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6759 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791318308, "dur": 21, + "args": { + "External id": 6761,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6760 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318309, "dur": 5, + "args": { + "External id": 6762,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6761 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791318315, "dur": 14, + "args": { + "External id": 6763,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6762 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791318317, "dur": 11, + "args": { + "External id": 6764,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6763 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318347, "dur": 5, + "args": { + "External id": 6765,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6764 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791318356, "dur": 22, + "args": { + "External id": 6766,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6765 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318357, "dur": 5, + "args": { + "External id": 6767,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6766 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791318363, "dur": 14, + "args": { + "External id": 6768,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6767 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791318364, "dur": 12, + "args": { + "External id": 6769,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6768 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318396, "dur": 5, + "args": { + "External id": 6770,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6769 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791318405, "dur": 21, + "args": { + "External id": 6771,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6770 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318406, "dur": 4, + "args": { + "External id": 6772,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6771 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791318412, "dur": 14, + "args": { + "External id": 6773,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6772 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791318413, "dur": 11, + "args": { + "External id": 6774,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6773 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318444, "dur": 5, + "args": { + "External id": 6775,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6774 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791318453, "dur": 21, + "args": { + "External id": 6776,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6775 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318454, "dur": 5, + "args": { + "External id": 6777,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6776 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791318460, "dur": 14, + "args": { + "External id": 6778,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6777 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791318462, "dur": 11, + "args": { + "External id": 6779,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6778 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318492, "dur": 5, + "args": { + "External id": 6780,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6779 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791318501, "dur": 22, + "args": { + "External id": 6781,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6780 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318502, "dur": 5, + "args": { + "External id": 6782,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6781 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791318508, "dur": 14, + "args": { + "External id": 6783,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6782 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791318510, "dur": 11, + "args": { + "External id": 6784,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6783 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318541, "dur": 5, + "args": { + "External id": 6785,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6784 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791318550, "dur": 21, + "args": { + "External id": 6786,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6785 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318551, "dur": 4, + "args": { + "External id": 6787,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6786 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791318557, "dur": 14, + "args": { + "External id": 6788,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6787 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791318558, "dur": 11, + "args": { + "External id": 6789,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6788 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318589, "dur": 5, + "args": { + "External id": 6790,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6789 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791318598, "dur": 21, + "args": { + "External id": 6791,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6790 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318599, "dur": 5, + "args": { + "External id": 6792,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6791 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791318605, "dur": 14, + "args": { + "External id": 6793,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6792 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791318607, "dur": 11, + "args": { + "External id": 6794,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6793 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318637, "dur": 6, + "args": { + "External id": 6795,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6794 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791318646, "dur": 22, + "args": { + "External id": 6796,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6795 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318648, "dur": 4, + "args": { + "External id": 6797,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6796 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791318653, "dur": 14, + "args": { + "External id": 6798,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6797 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791318655, "dur": 11, + "args": { + "External id": 6799,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6798 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318686, "dur": 5, + "args": { + "External id": 6800,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6799 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791318695, "dur": 21, + "args": { + "External id": 6801,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6800 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318696, "dur": 4, + "args": { + "External id": 6802,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6801 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791318702, "dur": 14, + "args": { + "External id": 6803,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6802 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791318703, "dur": 11, + "args": { + "External id": 6804,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6803 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318734, "dur": 5, + "args": { + "External id": 6805,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6804 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791318743, "dur": 21, + "args": { + "External id": 6806,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6805 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318744, "dur": 5, + "args": { + "External id": 6807,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6806 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791318750, "dur": 14, + "args": { + "External id": 6808,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6807 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791318752, "dur": 11, + "args": { + "External id": 6809,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6808 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318782, "dur": 5, + "args": { + "External id": 6810,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6809 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791318791, "dur": 22, + "args": { + "External id": 6811,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6810 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318792, "dur": 5, + "args": { + "External id": 6812,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6811 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791318798, "dur": 14, + "args": { + "External id": 6813,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6812 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791318800, "dur": 11, + "args": { + "External id": 6814,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6813 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318831, "dur": 5, + "args": { + "External id": 6815,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6814 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791318840, "dur": 24, + "args": { + "External id": 6816,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6815 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318841, "dur": 5, + "args": { + "External id": 6817,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6816 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791318847, "dur": 16, + "args": { + "External id": 6818,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6817 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791318849, "dur": 13, + "args": { + "External id": 6819,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6818 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318882, "dur": 5, + "args": { + "External id": 6820,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6819 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791318891, "dur": 23, + "args": { + "External id": 6821,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6820 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318893, "dur": 4, + "args": { + "External id": 6822,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6821 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791318898, "dur": 15, + "args": { + "External id": 6823,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6822 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791318900, "dur": 12, + "args": { + "External id": 6824,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6823 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318932, "dur": 5, + "args": { + "External id": 6825,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6824 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791318941, "dur": 24, + "args": { + "External id": 6826,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6825 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791318942, "dur": 4, + "args": { + "External id": 6827,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6826 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791318947, "dur": 18, + "args": { + "External id": 6828,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6827 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791318952, "dur": 11, + "args": { + "External id": 6829,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6828 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319044, "dur": 6, + "args": { + "External id": 6830,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6829 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791319055, "dur": 25, + "args": { + "External id": 6831,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6830 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319056, "dur": 5, + "args": { + "External id": 6832,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6831 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791319062, "dur": 17, + "args": { + "External id": 6833,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6832 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791319064, "dur": 14, + "args": { + "External id": 6834,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6833 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319100, "dur": 6, + "args": { + "External id": 6835,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6834 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791319110, "dur": 22, + "args": { + "External id": 6836,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6835 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319111, "dur": 4, + "args": { + "External id": 6837,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6836 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791319117, "dur": 14, + "args": { + "External id": 6838,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6837 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791319118, "dur": 12, + "args": { + "External id": 6839,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6838 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319150, "dur": 5, + "args": { + "External id": 6840,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6839 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791319159, "dur": 22, + "args": { + "External id": 6841,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6840 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319161, "dur": 4, + "args": { + "External id": 6842,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6841 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791319166, "dur": 14, + "args": { + "External id": 6843,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6842 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791319168, "dur": 11, + "args": { + "External id": 6844,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6843 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319199, "dur": 5, + "args": { + "External id": 6845,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6844 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791319209, "dur": 21, + "args": { + "External id": 6846,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6845 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319210, "dur": 4, + "args": { + "External id": 6847,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6846 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791319215, "dur": 15, + "args": { + "External id": 6848,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6847 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791319217, "dur": 12, + "args": { + "External id": 6849,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6848 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319249, "dur": 5, + "args": { + "External id": 6850,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6849 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791319258, "dur": 21, + "args": { + "External id": 6851,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6850 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319259, "dur": 4, + "args": { + "External id": 6852,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6851 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791319264, "dur": 14, + "args": { + "External id": 6853,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6852 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791319266, "dur": 11, + "args": { + "External id": 6854,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6853 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319297, "dur": 5, + "args": { + "External id": 6855,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6854 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791319306, "dur": 21, + "args": { + "External id": 6856,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6855 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319307, "dur": 5, + "args": { + "External id": 6857,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6856 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791319313, "dur": 14, + "args": { + "External id": 6858,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6857 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791319314, "dur": 11, + "args": { + "External id": 6859,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6858 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319345, "dur": 5, + "args": { + "External id": 6860,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6859 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791319354, "dur": 22, + "args": { + "External id": 6861,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6860 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319355, "dur": 5, + "args": { + "External id": 6862,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6861 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791319361, "dur": 14, + "args": { + "External id": 6863,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6862 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791319362, "dur": 12, + "args": { + "External id": 6864,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6863 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319393, "dur": 6, + "args": { + "External id": 6865,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6864 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791319402, "dur": 22, + "args": { + "External id": 6866,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6865 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319404, "dur": 4, + "args": { + "External id": 6867,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6866 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791319409, "dur": 14, + "args": { + "External id": 6868,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6867 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791319411, "dur": 11, + "args": { + "External id": 6869,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6868 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319441, "dur": 5, + "args": { + "External id": 6870,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6869 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791319450, "dur": 21, + "args": { + "External id": 6871,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6870 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319452, "dur": 4, + "args": { + "External id": 6872,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6871 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791319457, "dur": 14, + "args": { + "External id": 6873,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6872 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791319459, "dur": 11, + "args": { + "External id": 6874,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6873 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319489, "dur": 5, + "args": { + "External id": 6875,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6874 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791319498, "dur": 22, + "args": { + "External id": 6876,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6875 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319500, "dur": 4, + "args": { + "External id": 6877,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6876 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791319505, "dur": 14, + "args": { + "External id": 6878,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6877 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791319507, "dur": 11, + "args": { + "External id": 6879,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6878 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319587, "dur": 6, + "args": { + "External id": 6880,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6879 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791319598, "dur": 25, + "args": { + "External id": 6881,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6880 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319599, "dur": 5, + "args": { + "External id": 6882,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6881 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791319606, "dur": 17, + "args": { + "External id": 6883,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6882 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791319608, "dur": 13, + "args": { + "External id": 6884,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6883 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319643, "dur": 5, + "args": { + "External id": 6885,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6884 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791319652, "dur": 22, + "args": { + "External id": 6886,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6885 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319654, "dur": 4, + "args": { + "External id": 6887,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6886 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791319660, "dur": 14, + "args": { + "External id": 6888,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6887 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791319661, "dur": 12, + "args": { + "External id": 6889,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6888 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319692, "dur": 5, + "args": { + "External id": 6890,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6889 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791319701, "dur": 22, + "args": { + "External id": 6891,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6890 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319703, "dur": 4, + "args": { + "External id": 6892,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6891 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791319708, "dur": 14, + "args": { + "External id": 6893,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6892 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791319710, "dur": 11, + "args": { + "External id": 6894,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6893 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319740, "dur": 6, + "args": { + "External id": 6895,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6894 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791319749, "dur": 23, + "args": { + "External id": 6896,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6895 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319751, "dur": 4, + "args": { + "External id": 6897,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6896 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791319757, "dur": 14, + "args": { + "External id": 6898,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6897 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791319759, "dur": 11, + "args": { + "External id": 6899,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6898 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319790, "dur": 5, + "args": { + "External id": 6900,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6899 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791319799, "dur": 21, + "args": { + "External id": 6901,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6900 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319800, "dur": 4, + "args": { + "External id": 6902,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6901 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791319806, "dur": 14, + "args": { + "External id": 6903,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6902 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791319807, "dur": 11, + "args": { + "External id": 6904,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6903 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319840, "dur": 6, + "args": { + "External id": 6905,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6904 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791319850, "dur": 24, + "args": { + "External id": 6906,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6905 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319851, "dur": 6, + "args": { + "External id": 6907,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6906 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791319858, "dur": 15, + "args": { + "External id": 6908,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6907 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791319860, "dur": 12, + "args": { + "External id": 6909,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6908 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319891, "dur": 5, + "args": { + "External id": 6910,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6909 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791319900, "dur": 22, + "args": { + "External id": 6911,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6910 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319901, "dur": 5, + "args": { + "External id": 6912,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6911 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791319907, "dur": 14, + "args": { + "External id": 6913,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6912 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791319909, "dur": 11, + "args": { + "External id": 6914,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6913 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319940, "dur": 5, + "args": { + "External id": 6915,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6914 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791319949, "dur": 21, + "args": { + "External id": 6916,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6915 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319950, "dur": 5, + "args": { + "External id": 6917,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6916 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791319956, "dur": 14, + "args": { + "External id": 6918,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6917 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791319957, "dur": 12, + "args": { + "External id": 6919,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6918 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319988, "dur": 5, + "args": { + "External id": 6920,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6919 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791319997, "dur": 22, + "args": { + "External id": 6921,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6920 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791319998, "dur": 5, + "args": { + "External id": 6922,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6921 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791320004, "dur": 14, + "args": { + "External id": 6923,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6922 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791320006, "dur": 11, + "args": { + "External id": 6924,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6923 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320036, "dur": 5, + "args": { + "External id": 6925,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6924 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791320045, "dur": 21, + "args": { + "External id": 6926,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6925 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320046, "dur": 5, + "args": { + "External id": 6927,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6926 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791320052, "dur": 14, + "args": { + "External id": 6928,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6927 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791320054, "dur": 11, + "args": { + "External id": 6929,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6928 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320085, "dur": 5, + "args": { + "External id": 6930,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6929 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791320094, "dur": 23, + "args": { + "External id": 6931,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6930 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320096, "dur": 5, + "args": { + "External id": 6932,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6931 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791320102, "dur": 14, + "args": { + "External id": 6933,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6932 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791320104, "dur": 11, + "args": { + "External id": 6934,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6933 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320135, "dur": 5, + "args": { + "External id": 6935,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6934 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791320144, "dur": 21, + "args": { + "External id": 6936,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6935 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320145, "dur": 5, + "args": { + "External id": 6937,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6936 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791320151, "dur": 14, + "args": { + "External id": 6938,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6937 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791320153, "dur": 11, + "args": { + "External id": 6939,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6938 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320183, "dur": 5, + "args": { + "External id": 6940,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6939 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791320192, "dur": 22, + "args": { + "External id": 6941,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6940 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320193, "dur": 5, + "args": { + "External id": 6942,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6941 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791320199, "dur": 15, + "args": { + "External id": 6943,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6942 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791320201, "dur": 11, + "args": { + "External id": 6944,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6943 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320256, "dur": 5, + "args": { + "External id": 6945,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6944 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791320266, "dur": 22, + "args": { + "External id": 6946,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6945 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320267, "dur": 5, + "args": { + "External id": 6947,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6946 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791320273, "dur": 15, + "args": { + "External id": 6948,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6947 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791320275, "dur": 11, + "args": { + "External id": 6949,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6948 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320307, "dur": 5, + "args": { + "External id": 6950,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6949 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791320316, "dur": 21, + "args": { + "External id": 6951,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6950 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320317, "dur": 5, + "args": { + "External id": 6952,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6951 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791320323, "dur": 14, + "args": { + "External id": 6953,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6952 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791320325, "dur": 11, + "args": { + "External id": 6954,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6953 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320355, "dur": 5, + "args": { + "External id": 6955,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6954 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791320364, "dur": 21, + "args": { + "External id": 6956,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6955 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320365, "dur": 5, + "args": { + "External id": 6957,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6956 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791320371, "dur": 14, + "args": { + "External id": 6958,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6957 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791320373, "dur": 11, + "args": { + "External id": 6959,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6958 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320403, "dur": 5, + "args": { + "External id": 6960,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6959 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791320412, "dur": 21, + "args": { + "External id": 6961,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6960 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320413, "dur": 5, + "args": { + "External id": 6962,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6961 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791320419, "dur": 14, + "args": { + "External id": 6963,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6962 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791320420, "dur": 11, + "args": { + "External id": 6964,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6963 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320451, "dur": 5, + "args": { + "External id": 6965,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6964 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791320460, "dur": 21, + "args": { + "External id": 6966,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6965 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320461, "dur": 5, + "args": { + "External id": 6967,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6966 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791320467, "dur": 14, + "args": { + "External id": 6968,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6967 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791320469, "dur": 11, + "args": { + "External id": 6969,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6968 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320499, "dur": 5, + "args": { + "External id": 6970,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6969 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791320508, "dur": 21, + "args": { + "External id": 6971,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6970 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320509, "dur": 5, + "args": { + "External id": 6972,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6971 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791320515, "dur": 14, + "args": { + "External id": 6973,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6972 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791320516, "dur": 12, + "args": { + "External id": 6974,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6973 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320547, "dur": 5, + "args": { + "External id": 6975,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6974 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791320556, "dur": 21, + "args": { + "External id": 6976,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6975 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320557, "dur": 4, + "args": { + "External id": 6977,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6976 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791320563, "dur": 13, + "args": { + "External id": 6978,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6977 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791320564, "dur": 11, + "args": { + "External id": 6979,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6978 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320594, "dur": 6, + "args": { + "External id": 6980,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6979 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791320603, "dur": 22, + "args": { + "External id": 6981,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6980 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320605, "dur": 4, + "args": { + "External id": 6982,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6981 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791320610, "dur": 14, + "args": { + "External id": 6983,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6982 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791320612, "dur": 11, + "args": { + "External id": 6984,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6983 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320643, "dur": 5, + "args": { + "External id": 6985,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6984 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791320652, "dur": 21, + "args": { + "External id": 6986,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6985 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320653, "dur": 4, + "args": { + "External id": 6987,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6986 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791320658, "dur": 14, + "args": { + "External id": 6988,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6987 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791320660, "dur": 11, + "args": { + "External id": 6989,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6988 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320690, "dur": 6, + "args": { + "External id": 6990,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6989 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791320699, "dur": 22, + "args": { + "External id": 6991,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6990 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320701, "dur": 4, + "args": { + "External id": 6992,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6991 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791320706, "dur": 14, + "args": { + "External id": 6993,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6992 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791320708, "dur": 11, + "args": { + "External id": 6994,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6993 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320738, "dur": 5, + "args": { + "External id": 6995,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6994 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791320747, "dur": 22, + "args": { + "External id": 6996,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 6995 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320749, "dur": 4, + "args": { + "External id": 6997,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6996 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791320754, "dur": 14, + "args": { + "External id": 6998,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 6997 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791320756, "dur": 11, + "args": { + "External id": 6999,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 6998 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320786, "dur": 5, + "args": { + "External id": 7000,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6999 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791320795, "dur": 21, + "args": { + "External id": 7001,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7000 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320796, "dur": 5, + "args": { + "External id": 7002,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7001 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791320802, "dur": 14, + "args": { + "External id": 7003,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 7002 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791320804, "dur": 11, + "args": { + "External id": 7004,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 7003 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320834, "dur": 5, + "args": { + "External id": 7005,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7004 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791320843, "dur": 21, + "args": { + "External id": 7006,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7005 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320844, "dur": 5, + "args": { + "External id": 7007,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7006 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791320850, "dur": 14, + "args": { + "External id": 7008,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 7007 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791320852, "dur": 11, + "args": { + "External id": 7009,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 7008 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320882, "dur": 6, + "args": { + "External id": 7010,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7009 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791320891, "dur": 22, + "args": { + "External id": 7011,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7010 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320892, "dur": 5, + "args": { + "External id": 7012,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7011 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791320898, "dur": 14, + "args": { + "External id": 7013,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 7012 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791320900, "dur": 11, + "args": { + "External id": 7014,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 7013 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320931, "dur": 5, + "args": { + "External id": 7015,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7014 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791320940, "dur": 21, + "args": { + "External id": 7016,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7015 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320941, "dur": 4, + "args": { + "External id": 7017,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7016 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791320946, "dur": 14, + "args": { + "External id": 7018,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 7017 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791320948, "dur": 11, + "args": { + "External id": 7019,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 7018 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320979, "dur": 5, + "args": { + "External id": 7020,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7019 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791320988, "dur": 21, + "args": { + "External id": 7021,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7020 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791320989, "dur": 4, + "args": { + "External id": 7022,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7021 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791320994, "dur": 15, + "args": { + "External id": 7023,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 7022 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791320996, "dur": 11, + "args": { + "External id": 7024,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 7023 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321027, "dur": 5, + "args": { + "External id": 7025,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7024 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791321036, "dur": 21, + "args": { + "External id": 7026,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7025 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321037, "dur": 4, + "args": { + "External id": 7027,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7026 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791321043, "dur": 14, + "args": { + "External id": 7028,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 7027 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791321044, "dur": 12, + "args": { + "External id": 7029,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 7028 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321075, "dur": 5, + "args": { + "External id": 7030,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7029 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791321084, "dur": 21, + "args": { + "External id": 7031,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7030 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321085, "dur": 5, + "args": { + "External id": 7032,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7031 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791321091, "dur": 14, + "args": { + "External id": 7033,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 7032 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791321093, "dur": 11, + "args": { + "External id": 7034,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 7033 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321123, "dur": 5, + "args": { + "External id": 7035,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7034 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791321132, "dur": 21, + "args": { + "External id": 7036,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7035 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321133, "dur": 5, + "args": { + "External id": 7037,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7036 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791321139, "dur": 14, + "args": { + "External id": 7038,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 7037 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791321141, "dur": 11, + "args": { + "External id": 7039,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 7038 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321171, "dur": 5, + "args": { + "External id": 7040,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7039 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791321180, "dur": 21, + "args": { + "External id": 7041,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7040 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321181, "dur": 4, + "args": { + "External id": 7042,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7041 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791321187, "dur": 14, + "args": { + "External id": 7043,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 7042 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791321188, "dur": 11, + "args": { + "External id": 7044,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 7043 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321219, "dur": 5, + "args": { + "External id": 7045,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7044 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791321228, "dur": 21, + "args": { + "External id": 7046,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7045 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321229, "dur": 4, + "args": { + "External id": 7047,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7046 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791321235, "dur": 14, + "args": { + "External id": 7048,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 7047 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791321236, "dur": 11, + "args": { + "External id": 7049,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 7048 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321267, "dur": 5, + "args": { + "External id": 7050,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7049 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791321276, "dur": 22, + "args": { + "External id": 7051,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7050 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321278, "dur": 4, + "args": { + "External id": 7052,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7051 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791321283, "dur": 14, + "args": { + "External id": 7053,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 7052 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791321285, "dur": 11, + "args": { + "External id": 7054,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 7053 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321316, "dur": 5, + "args": { + "External id": 7055,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7054 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791321325, "dur": 21, + "args": { + "External id": 7056,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7055 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321326, "dur": 4, + "args": { + "External id": 7057,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7056 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791321331, "dur": 14, + "args": { + "External id": 7058,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 7057 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791321333, "dur": 11, + "args": { + "External id": 7059,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 7058 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321364, "dur": 5, + "args": { + "External id": 7060,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7059 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791321372, "dur": 22, + "args": { + "External id": 7061,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7060 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321374, "dur": 4, + "args": { + "External id": 7062,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7061 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791321379, "dur": 14, + "args": { + "External id": 7063,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 7062 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791321381, "dur": 11, + "args": { + "External id": 7064,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 7063 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321411, "dur": 5, + "args": { + "External id": 7065,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7064 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791321420, "dur": 21, + "args": { + "External id": 7066,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7065 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321421, "dur": 5, + "args": { + "External id": 7067,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7066 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791321427, "dur": 14, + "args": { + "External id": 7068,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 7067 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791321428, "dur": 11, + "args": { + "External id": 7069,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 7068 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321459, "dur": 5, + "args": { + "External id": 7070,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7069 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791321468, "dur": 22, + "args": { + "External id": 7071,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7070 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321469, "dur": 5, + "args": { + "External id": 7072,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7071 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791321475, "dur": 14, + "args": { + "External id": 7073,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 7072 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791321477, "dur": 11, + "args": { + "External id": 7074,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 7073 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321507, "dur": 6, + "args": { + "External id": 7075,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7074 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791321516, "dur": 22, + "args": { + "External id": 7076,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7075 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321518, "dur": 4, + "args": { + "External id": 7077,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7076 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791321523, "dur": 14, + "args": { + "External id": 7078,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 7077 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791321525, "dur": 11, + "args": { + "External id": 7079,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 7078 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321556, "dur": 5, + "args": { + "External id": 7080,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7079 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791321565, "dur": 21, + "args": { + "External id": 7081,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7080 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321566, "dur": 4, + "args": { + "External id": 7082,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7081 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791321572, "dur": 14, + "args": { + "External id": 7083,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 7082 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791321573, "dur": 11, + "args": { + "External id": 7084,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 7083 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321604, "dur": 5, + "args": { + "External id": 7085,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7084 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791321613, "dur": 22, + "args": { + "External id": 7086,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7085 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321614, "dur": 5, + "args": { + "External id": 7087,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7086 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791321621, "dur": 14, + "args": { + "External id": 7088,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 7087 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791321622, "dur": 12, + "args": { + "External id": 7089,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 7088 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321653, "dur": 5, + "args": { + "External id": 7090,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7089 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791321662, "dur": 21, + "args": { + "External id": 7091,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7090 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321663, "dur": 5, + "args": { + "External id": 7092,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7091 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791321669, "dur": 14, + "args": { + "External id": 7093,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 7092 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791321671, "dur": 11, + "args": { + "External id": 7094,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 7093 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321701, "dur": 5, + "args": { + "External id": 7095,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7094 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791321710, "dur": 22, + "args": { + "External id": 7096,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7095 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321712, "dur": 4, + "args": { + "External id": 7097,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7096 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791321717, "dur": 15, + "args": { + "External id": 7098,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 7097 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791321719, "dur": 11, + "args": { + "External id": 7099,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 7098 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321750, "dur": 5, + "args": { + "External id": 7100,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7099 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791321759, "dur": 21, + "args": { + "External id": 7101,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7100 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321760, "dur": 5, + "args": { + "External id": 7102,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7101 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791321766, "dur": 14, + "args": { + "External id": 7103,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 7102 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791321768, "dur": 11, + "args": { + "External id": 7104,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 7103 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321798, "dur": 5, + "args": { + "External id": 7105,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7104 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791321807, "dur": 21, + "args": { + "External id": 7106,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7105 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321808, "dur": 5, + "args": { + "External id": 7107,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7106 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791321814, "dur": 14, + "args": { + "External id": 7108,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 7107 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791321815, "dur": 11, + "args": { + "External id": 7109,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 7108 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321846, "dur": 5, + "args": { + "External id": 7110,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7109 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791321855, "dur": 21, + "args": { + "External id": 7111,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7110 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321856, "dur": 4, + "args": { + "External id": 7112,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7111 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791321862, "dur": 13, + "args": { + "External id": 7113,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 7112 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791321863, "dur": 11, + "args": { + "External id": 7114,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 7113 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321893, "dur": 5, + "args": { + "External id": 7115,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7114 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791321902, "dur": 21, + "args": { + "External id": 7116,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7115 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321903, "dur": 5, + "args": { + "External id": 7117,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7116 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791321909, "dur": 14, + "args": { + "External id": 7118,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 7117 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791321911, "dur": 11, + "args": { + "External id": 7119,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 7118 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321941, "dur": 5, + "args": { + "External id": 7120,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7119 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791321950, "dur": 21, + "args": { + "External id": 7121,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7120 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321951, "dur": 5, + "args": { + "External id": 7122,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7121 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791321957, "dur": 14, + "args": { + "External id": 7123,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 7122 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791321959, "dur": 11, + "args": { + "External id": 7124,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 7123 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321989, "dur": 5, + "args": { + "External id": 7125,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7124 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791321998, "dur": 22, + "args": { + "External id": 7126,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7125 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791321999, "dur": 5, + "args": { + "External id": 7127,"Record function id": 0, "Concrete Inputs": ["[512]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7126 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791322005, "dur": 14, + "args": { + "External id": 7128,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[512]], "Ev Idx": 7127 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791322007, "dur": 11, + "args": { + "External id": 7129,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[512], []], "Ev Idx": 7128 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::randn", "pid": 435, "tid": 435, + "ts": 1776649791322217, "dur": 61, + "args": { + "External id": 7130,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7129 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791322223, "dur": 10, + "args": { + "External id": 7131,"Record function id": 0, "Concrete Inputs": ["[1, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7130 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::normal_", "pid": 435, "tid": 435, + "ts": 1776649791322237, "dur": 40, + "args": { + "External id": 7132,"Record function id": 0, "Concrete Inputs": ["", "0.", "1.", ""], "Input type": ["c10::Half", "Scalar", "Scalar", ""], "Input Dims": [[1, 4096], [], [], []], "Ev Idx": 7131 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::randn", "pid": 435, "tid": 435, + "ts": 1776649791322303, "dur": 31, + "args": { + "External id": 7133,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "", "5", "", "", "False"], "Input type": ["ScalarList", "", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7132 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791322307, "dur": 6, + "args": { + "External id": 7134,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7133 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::normal_", "pid": 435, "tid": 435, + "ts": 1776649791322314, "dur": 19, + "args": { + "External id": 7135,"Record function id": 0, "Concrete Inputs": ["", "0.", "1.", ""], "Input type": ["c10::Half", "Scalar", "Scalar", ""], "Input Dims": [[4096, 8192], [], [], []], "Ev Idx": 7134 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 435, "tid": 435, + "ts": 1776649791322363, "dur": 42, + "args": { + "External id": 7136,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "double"], "Input Dims": [[4096, 8192], []], "Ev Idx": 7135 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791322435, "dur": 16, + "args": { + "External id": 7137,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 8192], []], "Ev Idx": 7136 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791322437, "dur": 13, + "args": { + "External id": 7138,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 8192], []], "Ev Idx": 7137 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 435, "tid": 435, + "ts": 1776649791322455, "dur": 53, + "args": { + "External id": 7139,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[32, 128, 8192]], "Ev Idx": 7138 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791322462, "dur": 6, + "args": { + "External id": 7140,"Record function id": 0, "Concrete Inputs": ["[0]", "5", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7139 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 435, "tid": 435, + "ts": 1776649791322471, "dur": 36, + "args": { + "External id": 7141,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[32, 128, 8192], [0]], "Ev Idx": 7140 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 435, "tid": 435, + "ts": 1776649791322480, "dur": 9, + "args": { + "External id": 7142,"Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]", ""], "Input type": ["c10::Half", "ScalarList", ""], "Input Dims": [[0], [], []], "Ev Idx": 7141 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 435, "tid": 435, + "ts": 1776649791322526, "dur": 47, + "args": { + "External id": 7143,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1]", "False"], "Input type": ["c10::Half", "ScalarList", "Scalar"], "Input Dims": [[32, 128, 8192], [], []], "Ev Idx": 7142 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791322544, "dur": 3, + "args": { + "External id": 7144,"Record function id": 0, "Concrete Inputs": ["", "[32, 1, 8192]", "[8192, 0, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 7143 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 435, "tid": 435, + "ts": 1776649791322584, "dur": 26, + "args": { + "External id": 7145,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "double"], "Input Dims": [[32, 8192], []], "Ev Idx": 7144 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 435, "tid": 435, + "ts": 1776649791322621, "dur": 32, + "args": { + "External id": 7146,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "9.9999999999999995e-07", ""], "Input type": ["c10::Half", "Scalar", ""], "Input Dims": [[32, 8192], [], []], "Ev Idx": 7145 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791322626, "dur": 2, + "args": { + "External id": 7147,"Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 8192], [], [], [], []], "Ev Idx": 7146 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791322660, "dur": 0, + "args": { + "External id": 7148,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 8192], [], [], [], []], "Ev Idx": 7147 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649791322681, "dur": 10, + "args": { + "External id": 7149,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 8192], []], "Ev Idx": 7148 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791322688, "dur": 1, + "args": { + "External id": 7150,"Record function id": 0, "Concrete Inputs": ["", "[32, 1, 8192]", "[8192, 8192, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 7149 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 435, "tid": 435, + "ts": 1776649791322705, "dur": 28, + "args": { + "External id": 7151,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[32, 128, 8192], [32, 1, 8192]], "Ev Idx": 7150 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 435, "tid": 435, + "ts": 1776649791322739, "dur": 553, + "args": { + "External id": 7152,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[32, 128, 8192]], "Ev Idx": 7151 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 435, "tid": 435, + "ts": 1776649791323309, "dur": 24, + "args": { + "External id": 7153,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "-8", "7"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[32, 128, 8192], [], []], "Ev Idx": 7152 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791323312, "dur": 0, + "args": { + "External id": 7154,"Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 128, 8192], [], [], [], []], "Ev Idx": 7153 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791323338, "dur": 50, + "args": { + "External id": 7155,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 128, 8192], [], [], [], []], "Ev Idx": 7154 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791323341, "dur": 46, + "args": { + "External id": 7156,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["c10::Half", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[32, 128, 8192], [], [], [], [], [], []], "Ev Idx": 7155 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791323351, "dur": 8, + "args": { + "External id": 7157,"Record function id": 0, "Concrete Inputs": ["[32, 128, 8192]", "[1048576, 8192, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7156 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791323361, "dur": 25, + "args": { + "External id": 7158,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "c10::Half", "Scalar"], "Input Dims": [[32, 128, 8192], [32, 128, 8192], []], "Ev Idx": 7157 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 435, "tid": 435, + "ts": 1776649791323401, "dur": 31, + "args": { + "External id": 7159,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["signed char", "long int", "Scalar"], "Input Dims": [[32, 128, 8192], [], []], "Ev Idx": 7158 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791323437, "dur": 26, + "args": { + "External id": 7160,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "False", "False", ""], "Input type": ["signed char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 128, 8192], [], [], [], []], "Ev Idx": 7159 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791323438, "dur": 25, + "args": { + "External id": 7161,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "", "", "", "False", ""], "Input type": ["signed char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[32, 128, 8192], [], [], [], [], [], []], "Ev Idx": 7160 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791323441, "dur": 5, + "args": { + "External id": 7162,"Record function id": 0, "Concrete Inputs": ["[32, 128, 8192]", "[1048576, 8192, 1]", "0", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7161 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791323448, "dur": 14, + "args": { + "External id": 7163,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["unsigned char", "signed char", "Scalar"], "Input Dims": [[32, 128, 8192], [32, 128, 8192], []], "Ev Idx": 7162 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791323471, "dur": 6, + "args": { + "External id": 7164,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["unsigned char", "ScalarList"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 7163 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791323472, "dur": 4, + "args": { + "External id": 7165,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["unsigned char", "ScalarList"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 7164 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649791323492, "dur": 12, + "args": { + "External id": 7166,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 7165 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791323500, "dur": 1, + "args": { + "External id": 7167,"Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]", "[8192, 1]", "0"], "Input type": ["unsigned char", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 8192], [], [], []], "Ev Idx": 7166 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649791323506, "dur": 15, + "args": { + "External id": 7168,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "2"], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 7167 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791323518, "dur": 2, + "args": { + "External id": 7169,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]", "[8192, 2]", "0"], "Input type": ["unsigned char", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 8192], [], [], []], "Ev Idx": 7168 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649791323528, "dur": 3, + "args": { + "External id": 7170,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 7169 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791323530, "dur": 0, + "args": { + "External id": 7171,"Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]", "[8192, 1]", "0"], "Input type": ["unsigned char", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 8192], [], [], []], "Ev Idx": 7170 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649791323533, "dur": 3, + "args": { + "External id": 7172,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "1", "9223372036854775807", "2"], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 7171 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791323534, "dur": 1, + "args": { + "External id": 7173,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]", "[8192, 2]", "1"], "Input type": ["unsigned char", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 8192], [], [], []], "Ev Idx": 7172 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__lshift__", "pid": 435, "tid": 435, + "ts": 1776649791323544, "dur": 41, + "args": { + "External id": 7174,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 7173 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791323555, "dur": 6, + "args": { + "External id": 7175,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "0", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7174 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__or__", "pid": 435, "tid": 435, + "ts": 1776649791323591, "dur": 30, + "args": { + "External id": 7176,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["unsigned char", "unsigned char"], "Input Dims": [[4096, 4096], [4096, 4096]], "Ev Idx": 7175 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_or", "pid": 435, "tid": 435, + "ts": 1776649791323595, "dur": 26, + "args": { + "External id": 7177,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["unsigned char", "unsigned char"], "Input Dims": [[4096, 4096], [4096, 4096]], "Ev Idx": 7176 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791323627, "dur": 0, + "args": { + "External id": 7178,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 7177 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__and__", "pid": 435, "tid": 435, + "ts": 1776649791323647, "dur": 33, + "args": { + "External id": 7179,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "15"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 7178 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_and", "pid": 435, "tid": 435, + "ts": 1776649791323649, "dur": 31, + "args": { + "External id": 7180,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "15"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 7179 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_and", "pid": 435, "tid": 435, + "ts": 1776649791323654, "dur": 24, + "args": { + "External id": 7181,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["unsigned char", "long int"], "Input Dims": [[4096, 4096], []], "Ev Idx": 7180 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791323684, "dur": 27, + "args": { + "External id": 7182,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 7181 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791323685, "dur": 26, + "args": { + "External id": 7183,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["unsigned char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], [], []], "Ev Idx": 7182 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791323689, "dur": 6, + "args": { + "External id": 7184,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "[4096, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7183 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791323696, "dur": 14, + "args": { + "External id": 7185,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "unsigned char", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 7184 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__rshift__", "pid": 435, "tid": 435, + "ts": 1776649791323717, "dur": 29, + "args": { + "External id": 7186,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 7185 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791323724, "dur": 5, + "args": { + "External id": 7187,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "0", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7186 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791323750, "dur": 22, + "args": { + "External id": 7188,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 7187 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791323751, "dur": 20, + "args": { + "External id": 7189,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["unsigned char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], [], []], "Ev Idx": 7188 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791323753, "dur": 4, + "args": { + "External id": 7190,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "[4096, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7189 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791323759, "dur": 12, + "args": { + "External id": 7191,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "unsigned char", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 7190 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::stack", "pid": 435, "tid": 435, + "ts": 1776649791323805, "dur": 55, + "args": { + "External id": 7192,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["TensorList", "Scalar"], "Input Dims": [[[4096, 4096], [4096, 4096]], []], "Ev Idx": 7191 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649791323810, "dur": 4, + "args": { + "External id": 7193,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["signed char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 7192 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791323812, "dur": 2, + "args": { + "External id": 7194,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096, 1]", "[4096, 1, 1]", ""], "Input type": ["signed char", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 7193 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649791323815, "dur": 2, + "args": { + "External id": 7195,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["signed char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 7194 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791323816, "dur": 1, + "args": { + "External id": 7196,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096, 1]", "[4096, 1, 1]", ""], "Input type": ["signed char", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 7195 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 435, "tid": 435, + "ts": 1776649791323819, "dur": 39, + "args": { + "External id": 7197,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["TensorList", "Scalar"], "Input Dims": [[[4096, 4096, 1], [4096, 4096, 1]], []], "Ev Idx": 7196 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791323868, "dur": 4, + "args": { + "External id": 7198,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["signed char", "ScalarList"], "Input Dims": [[4096, 4096, 2], []], "Ev Idx": 7197 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791323869, "dur": 3, + "args": { + "External id": 7199,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["signed char", "ScalarList"], "Input Dims": [[4096, 4096, 2], []], "Ev Idx": 7198 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::repeat_interleave", "pid": 435, "tid": 435, + "ts": 1776649791323880, "dur": 66, + "args": { + "External id": 7200,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "128", "0", ""], "Input type": ["c10::Half", "Scalar", "Scalar", ""], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 7199 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649791323883, "dur": 3, + "args": { + "External id": 7201,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 8192], []], "Ev Idx": 7200 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791323885, "dur": 1, + "args": { + "External id": 7202,"Record function id": 0, "Concrete Inputs": ["", "[32, 1, 8192]", "[8192, 8192, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 7201 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 435, "tid": 435, + "ts": 1776649791323890, "dur": 6, + "args": { + "External id": 7203,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]", "False"], "Input type": ["c10::Half", "ScalarList", "Scalar"], "Input Dims": [[32, 1, 8192], [], []], "Ev Idx": 7202 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791323895, "dur": 0, + "args": { + "External id": 7204,"Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]", "[8192, 0, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 1, 8192], [], [], []], "Ev Idx": 7203 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649791323899, "dur": 38, + "args": { + "External id": 7205,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 7204 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649791323904, "dur": 11, + "args": { + "External id": 7206,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[32, 128, 8192], [], [], [], [], []], "Ev Idx": 7205 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791323907, "dur": 8, + "args": { + "External id": 7207,"Record function id": 0, "Concrete Inputs": ["[32, 128, 8192]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7206 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791323917, "dur": 20, + "args": { + "External id": 7208,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[32, 128, 8192], [32, 128, 8192], []], "Ev Idx": 7207 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::flatten", "pid": 435, "tid": 435, + "ts": 1776649791323940, "dur": 5, + "args": { + "External id": 7209,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[32, 128, 8192], [], []], "Ev Idx": 7208 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791323943, "dur": 2, + "args": { + "External id": 7210,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 7209 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 435, "tid": 435, + "ts": 1776649791323955, "dur": 22, + "args": { + "External id": 7211,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["signed char", "long int", "Scalar"], "Input Dims": [[4096, 8192], [], []], "Ev Idx": 7210 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791323982, "dur": 26, + "args": { + "External id": 7212,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["signed char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 7211 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791323984, "dur": 24, + "args": { + "External id": 7213,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "", "", "", "False", ""], "Input type": ["signed char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 8192], [], [], [], [], [], []], "Ev Idx": 7212 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791323987, "dur": 5, + "args": { + "External id": 7214,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7213 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791323993, "dur": 14, + "args": { + "External id": 7215,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "signed char", "Scalar"], "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 7214 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 435, "tid": 435, + "ts": 1776649791324015, "dur": 19, + "args": { + "External id": 7216,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4096, 8192], [4096, 8192]], "Ev Idx": 7215 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791324113, "dur": 396, + "args": { + "External id": 7217,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7216 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791324120, "dur": 388, + "args": { + "External id": 7218,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7217 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791324520, "dur": 35, + "args": { + "External id": 7219,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7218 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791324524, "dur": 30, + "args": { + "External id": 7220,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7219 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791324560, "dur": 28, + "args": { + "External id": 7221,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7220 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791324563, "dur": 24, + "args": { + "External id": 7222,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7221 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791324592, "dur": 26, + "args": { + "External id": 7223,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7222 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791324594, "dur": 23, + "args": { + "External id": 7224,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7223 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791324623, "dur": 25, + "args": { + "External id": 7225,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7224 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791324625, "dur": 22, + "args": { + "External id": 7226,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7225 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791324653, "dur": 26, + "args": { + "External id": 7227,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7226 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791324655, "dur": 23, + "args": { + "External id": 7228,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7227 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791324683, "dur": 26, + "args": { + "External id": 7229,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7228 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791324685, "dur": 23, + "args": { + "External id": 7230,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7229 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791324713, "dur": 26, + "args": { + "External id": 7231,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7230 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791324716, "dur": 22, + "args": { + "External id": 7232,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7231 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791324743, "dur": 25, + "args": { + "External id": 7233,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7232 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791324745, "dur": 22, + "args": { + "External id": 7234,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7233 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791324773, "dur": 25, + "args": { + "External id": 7235,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7234 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791324775, "dur": 22, + "args": { + "External id": 7236,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7235 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791325349, "dur": 37, + "args": { + "External id": 7237,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7236 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791325353, "dur": 32, + "args": { + "External id": 7238,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7237 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791325392, "dur": 27, + "args": { + "External id": 7239,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7238 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791325394, "dur": 24, + "args": { + "External id": 7240,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7239 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791325423, "dur": 27, + "args": { + "External id": 7241,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7240 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791325426, "dur": 23, + "args": { + "External id": 7242,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7241 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791325454, "dur": 26, + "args": { + "External id": 7243,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7242 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791325456, "dur": 23, + "args": { + "External id": 7244,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7243 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791325485, "dur": 26, + "args": { + "External id": 7245,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7244 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791325487, "dur": 23, + "args": { + "External id": 7246,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7245 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791325515, "dur": 26, + "args": { + "External id": 7247,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7246 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791325517, "dur": 23, + "args": { + "External id": 7248,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7247 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791325545, "dur": 26, + "args": { + "External id": 7249,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7248 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791325547, "dur": 23, + "args": { + "External id": 7250,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7249 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791325575, "dur": 26, + "args": { + "External id": 7251,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7250 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791325577, "dur": 23, + "args": { + "External id": 7252,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7251 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791325605, "dur": 26, + "args": { + "External id": 7253,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7252 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791325607, "dur": 23, + "args": { + "External id": 7254,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7253 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791325635, "dur": 26, + "args": { + "External id": 7255,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7254 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791325637, "dur": 23, + "args": { + "External id": 7256,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7255 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791325665, "dur": 26, + "args": { + "External id": 7257,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7256 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791325667, "dur": 23, + "args": { + "External id": 7258,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7257 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791325695, "dur": 26, + "args": { + "External id": 7259,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7258 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791325697, "dur": 23, + "args": { + "External id": 7260,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7259 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791325725, "dur": 25, + "args": { + "External id": 7261,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7260 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791325727, "dur": 23, + "args": { + "External id": 7262,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7261 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791325755, "dur": 26, + "args": { + "External id": 7263,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7262 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791325758, "dur": 22, + "args": { + "External id": 7264,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7263 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791325785, "dur": 26, + "args": { + "External id": 7265,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7264 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791325787, "dur": 23, + "args": { + "External id": 7266,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7265 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791325815, "dur": 25, + "args": { + "External id": 7267,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7266 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791325817, "dur": 23, + "args": { + "External id": 7268,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7267 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791325845, "dur": 27, + "args": { + "External id": 7269,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7268 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791325848, "dur": 23, + "args": { + "External id": 7270,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7269 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791325876, "dur": 26, + "args": { + "External id": 7271,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7270 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791325878, "dur": 23, + "args": { + "External id": 7272,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7271 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791325906, "dur": 26, + "args": { + "External id": 7273,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7272 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791325908, "dur": 23, + "args": { + "External id": 7274,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7273 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791325936, "dur": 25, + "args": { + "External id": 7275,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7274 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791325938, "dur": 23, + "args": { + "External id": 7276,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7275 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791325966, "dur": 26, + "args": { + "External id": 7277,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7276 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791325968, "dur": 23, + "args": { + "External id": 7278,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7277 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791325996, "dur": 25, + "args": { + "External id": 7279,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7278 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791325998, "dur": 23, + "args": { + "External id": 7280,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7279 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326026, "dur": 26, + "args": { + "External id": 7281,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7280 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326028, "dur": 23, + "args": { + "External id": 7282,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7281 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326056, "dur": 26, + "args": { + "External id": 7283,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7282 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326058, "dur": 23, + "args": { + "External id": 7284,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7283 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326086, "dur": 26, + "args": { + "External id": 7285,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7284 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326089, "dur": 22, + "args": { + "External id": 7286,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7285 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326116, "dur": 26, + "args": { + "External id": 7287,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7286 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326119, "dur": 22, + "args": { + "External id": 7288,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7287 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326147, "dur": 25, + "args": { + "External id": 7289,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7288 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326149, "dur": 23, + "args": { + "External id": 7290,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7289 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326177, "dur": 26, + "args": { + "External id": 7291,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7290 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326179, "dur": 23, + "args": { + "External id": 7292,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7291 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326207, "dur": 26, + "args": { + "External id": 7293,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7292 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326209, "dur": 23, + "args": { + "External id": 7294,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7293 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326237, "dur": 26, + "args": { + "External id": 7295,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7294 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326240, "dur": 23, + "args": { + "External id": 7296,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7295 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326268, "dur": 26, + "args": { + "External id": 7297,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7296 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326270, "dur": 23, + "args": { + "External id": 7298,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7297 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326298, "dur": 26, + "args": { + "External id": 7299,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7298 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326300, "dur": 23, + "args": { + "External id": 7300,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7299 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326328, "dur": 26, + "args": { + "External id": 7301,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7300 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326330, "dur": 23, + "args": { + "External id": 7302,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7301 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326358, "dur": 26, + "args": { + "External id": 7303,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7302 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326360, "dur": 23, + "args": { + "External id": 7304,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7303 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326388, "dur": 26, + "args": { + "External id": 7305,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7304 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326390, "dur": 23, + "args": { + "External id": 7306,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7305 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326418, "dur": 26, + "args": { + "External id": 7307,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7306 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326420, "dur": 23, + "args": { + "External id": 7308,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7307 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326448, "dur": 26, + "args": { + "External id": 7309,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7308 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326450, "dur": 23, + "args": { + "External id": 7310,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7309 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326478, "dur": 26, + "args": { + "External id": 7311,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7310 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326480, "dur": 23, + "args": { + "External id": 7312,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7311 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326508, "dur": 26, + "args": { + "External id": 7313,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7312 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326510, "dur": 23, + "args": { + "External id": 7314,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7313 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326538, "dur": 26, + "args": { + "External id": 7315,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7314 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326540, "dur": 23, + "args": { + "External id": 7316,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7315 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326568, "dur": 25, + "args": { + "External id": 7317,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7316 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326570, "dur": 23, + "args": { + "External id": 7318,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7317 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326598, "dur": 25, + "args": { + "External id": 7319,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7318 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326600, "dur": 22, + "args": { + "External id": 7320,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7319 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326627, "dur": 26, + "args": { + "External id": 7321,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7320 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326629, "dur": 23, + "args": { + "External id": 7322,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7321 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326657, "dur": 26, + "args": { + "External id": 7323,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7322 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326659, "dur": 23, + "args": { + "External id": 7324,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7323 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326687, "dur": 26, + "args": { + "External id": 7325,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7324 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326689, "dur": 23, + "args": { + "External id": 7326,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7325 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326717, "dur": 26, + "args": { + "External id": 7327,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7326 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326719, "dur": 23, + "args": { + "External id": 7328,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7327 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326747, "dur": 25, + "args": { + "External id": 7329,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7328 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326749, "dur": 23, + "args": { + "External id": 7330,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7329 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326777, "dur": 26, + "args": { + "External id": 7331,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7330 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326779, "dur": 23, + "args": { + "External id": 7332,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7331 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326807, "dur": 25, + "args": { + "External id": 7333,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7332 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326809, "dur": 23, + "args": { + "External id": 7334,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7333 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791326837, "dur": 25, + "args": { + "External id": 7335,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7334 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791326839, "dur": 22, + "args": { + "External id": 7336,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7335 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791327825, "dur": 35, + "args": { + "External id": 7337,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7336 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791327829, "dur": 30, + "args": { + "External id": 7338,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7337 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791327866, "dur": 27, + "args": { + "External id": 7339,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7338 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791327869, "dur": 24, + "args": { + "External id": 7340,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7339 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791327898, "dur": 26, + "args": { + "External id": 7341,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7340 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791327900, "dur": 23, + "args": { + "External id": 7342,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7341 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791327928, "dur": 26, + "args": { + "External id": 7343,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7342 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791327931, "dur": 23, + "args": { + "External id": 7344,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7343 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791327959, "dur": 25, + "args": { + "External id": 7345,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7344 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791327961, "dur": 23, + "args": { + "External id": 7346,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7345 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791327989, "dur": 26, + "args": { + "External id": 7347,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7346 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791327992, "dur": 23, + "args": { + "External id": 7348,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7347 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791328020, "dur": 26, + "args": { + "External id": 7349,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7348 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791328022, "dur": 23, + "args": { + "External id": 7350,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7349 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791328050, "dur": 26, + "args": { + "External id": 7351,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7350 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791328052, "dur": 23, + "args": { + "External id": 7352,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7351 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791328080, "dur": 26, + "args": { + "External id": 7353,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7352 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791328082, "dur": 23, + "args": { + "External id": 7354,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7353 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791328110, "dur": 26, + "args": { + "External id": 7355,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7354 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791328112, "dur": 23, + "args": { + "External id": 7356,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7355 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791328377, "dur": 39, + "args": { + "External id": 7357,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7356 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791328381, "dur": 34, + "args": { + "External id": 7358,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7357 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791328422, "dur": 28, + "args": { + "External id": 7359,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7358 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791328425, "dur": 24, + "args": { + "External id": 7360,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7359 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791328454, "dur": 26, + "args": { + "External id": 7361,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7360 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791328457, "dur": 23, + "args": { + "External id": 7362,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7361 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791328485, "dur": 26, + "args": { + "External id": 7363,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7362 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791328487, "dur": 23, + "args": { + "External id": 7364,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7363 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791328515, "dur": 26, + "args": { + "External id": 7365,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7364 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791328517, "dur": 23, + "args": { + "External id": 7366,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7365 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791328545, "dur": 26, + "args": { + "External id": 7367,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7366 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791328547, "dur": 23, + "args": { + "External id": 7368,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7367 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791328575, "dur": 26, + "args": { + "External id": 7369,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7368 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791328577, "dur": 23, + "args": { + "External id": 7370,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7369 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791328605, "dur": 26, + "args": { + "External id": 7371,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7370 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791328608, "dur": 23, + "args": { + "External id": 7372,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7371 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791328636, "dur": 25, + "args": { + "External id": 7373,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7372 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791328638, "dur": 23, + "args": { + "External id": 7374,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7373 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791328666, "dur": 26, + "args": { + "External id": 7375,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7374 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791328668, "dur": 23, + "args": { + "External id": 7376,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7375 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791328697, "dur": 25, + "args": { + "External id": 7377,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7376 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791328699, "dur": 23, + "args": { + "External id": 7378,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7377 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791328727, "dur": 25, + "args": { + "External id": 7379,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7378 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791328729, "dur": 23, + "args": { + "External id": 7380,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7379 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791328757, "dur": 26, + "args": { + "External id": 7381,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7380 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791328759, "dur": 23, + "args": { + "External id": 7382,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7381 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791328787, "dur": 25, + "args": { + "External id": 7383,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7382 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791328789, "dur": 23, + "args": { + "External id": 7384,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7383 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791328817, "dur": 26, + "args": { + "External id": 7385,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7384 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791328819, "dur": 24, + "args": { + "External id": 7386,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7385 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791328848, "dur": 25, + "args": { + "External id": 7387,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7386 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791328850, "dur": 23, + "args": { + "External id": 7388,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7387 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791328878, "dur": 26, + "args": { + "External id": 7389,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7388 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791328880, "dur": 23, + "args": { + "External id": 7390,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7389 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791328908, "dur": 26, + "args": { + "External id": 7391,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7390 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791328910, "dur": 23, + "args": { + "External id": 7392,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7391 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791328938, "dur": 25, + "args": { + "External id": 7393,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7392 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791328940, "dur": 23, + "args": { + "External id": 7394,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7393 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791328968, "dur": 25, + "args": { + "External id": 7395,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7394 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791328970, "dur": 23, + "args": { + "External id": 7396,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7395 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791328998, "dur": 26, + "args": { + "External id": 7397,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7396 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329001, "dur": 22, + "args": { + "External id": 7398,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7397 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329028, "dur": 26, + "args": { + "External id": 7399,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7398 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329030, "dur": 23, + "args": { + "External id": 7400,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7399 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329058, "dur": 26, + "args": { + "External id": 7401,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7400 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329061, "dur": 22, + "args": { + "External id": 7402,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7401 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329088, "dur": 26, + "args": { + "External id": 7403,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7402 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329090, "dur": 23, + "args": { + "External id": 7404,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7403 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329118, "dur": 26, + "args": { + "External id": 7405,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7404 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329120, "dur": 23, + "args": { + "External id": 7406,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7405 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329148, "dur": 26, + "args": { + "External id": 7407,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7406 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329150, "dur": 23, + "args": { + "External id": 7408,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7407 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329178, "dur": 26, + "args": { + "External id": 7409,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7408 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329180, "dur": 23, + "args": { + "External id": 7410,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7409 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329208, "dur": 26, + "args": { + "External id": 7411,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7410 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329210, "dur": 23, + "args": { + "External id": 7412,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7411 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329238, "dur": 25, + "args": { + "External id": 7413,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7412 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329240, "dur": 23, + "args": { + "External id": 7414,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7413 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329268, "dur": 26, + "args": { + "External id": 7415,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7414 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329270, "dur": 23, + "args": { + "External id": 7416,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7415 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329298, "dur": 26, + "args": { + "External id": 7417,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7416 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329300, "dur": 23, + "args": { + "External id": 7418,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7417 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329328, "dur": 26, + "args": { + "External id": 7419,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7418 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329330, "dur": 23, + "args": { + "External id": 7420,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7419 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329358, "dur": 26, + "args": { + "External id": 7421,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7420 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329360, "dur": 23, + "args": { + "External id": 7422,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7421 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329388, "dur": 26, + "args": { + "External id": 7423,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7422 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329390, "dur": 23, + "args": { + "External id": 7424,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7423 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329418, "dur": 25, + "args": { + "External id": 7425,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7424 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329420, "dur": 23, + "args": { + "External id": 7426,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7425 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329448, "dur": 26, + "args": { + "External id": 7427,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7426 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329450, "dur": 23, + "args": { + "External id": 7428,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7427 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329478, "dur": 26, + "args": { + "External id": 7429,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7428 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329480, "dur": 23, + "args": { + "External id": 7430,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7429 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329508, "dur": 26, + "args": { + "External id": 7431,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7430 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329510, "dur": 23, + "args": { + "External id": 7432,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7431 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329538, "dur": 26, + "args": { + "External id": 7433,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7432 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329540, "dur": 23, + "args": { + "External id": 7434,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7433 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329568, "dur": 25, + "args": { + "External id": 7435,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7434 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329570, "dur": 23, + "args": { + "External id": 7436,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7435 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329598, "dur": 25, + "args": { + "External id": 7437,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7436 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329600, "dur": 22, + "args": { + "External id": 7438,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7437 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329627, "dur": 26, + "args": { + "External id": 7439,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7438 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329629, "dur": 23, + "args": { + "External id": 7440,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7439 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329657, "dur": 26, + "args": { + "External id": 7441,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7440 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329659, "dur": 23, + "args": { + "External id": 7442,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7441 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329687, "dur": 26, + "args": { + "External id": 7443,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7442 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329689, "dur": 23, + "args": { + "External id": 7444,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7443 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329717, "dur": 26, + "args": { + "External id": 7445,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7444 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329719, "dur": 23, + "args": { + "External id": 7446,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7445 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329747, "dur": 26, + "args": { + "External id": 7447,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7446 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329749, "dur": 23, + "args": { + "External id": 7448,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7447 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329777, "dur": 26, + "args": { + "External id": 7449,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7448 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329779, "dur": 23, + "args": { + "External id": 7450,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7449 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329807, "dur": 26, + "args": { + "External id": 7451,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7450 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329809, "dur": 23, + "args": { + "External id": 7452,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7451 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329837, "dur": 28, + "args": { + "External id": 7453,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7452 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329840, "dur": 25, + "args": { + "External id": 7454,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7453 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791329870, "dur": 28, + "args": { + "External id": 7455,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7454 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791329873, "dur": 24, + "args": { + "External id": 7456,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7455 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791330861, "dur": 35, + "args": { + "External id": 7457,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7456 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791330864, "dur": 31, + "args": { + "External id": 7458,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7457 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791330902, "dur": 27, + "args": { + "External id": 7459,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7458 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791330904, "dur": 24, + "args": { + "External id": 7460,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7459 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791330934, "dur": 26, + "args": { + "External id": 7461,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7460 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791330936, "dur": 23, + "args": { + "External id": 7462,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7461 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791330965, "dur": 26, + "args": { + "External id": 7463,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7462 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791330967, "dur": 23, + "args": { + "External id": 7464,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7463 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791330995, "dur": 26, + "args": { + "External id": 7465,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7464 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791330997, "dur": 23, + "args": { + "External id": 7466,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7465 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791331025, "dur": 26, + "args": { + "External id": 7467,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7466 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791331027, "dur": 23, + "args": { + "External id": 7468,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7467 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791331055, "dur": 25, + "args": { + "External id": 7469,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7468 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791331057, "dur": 23, + "args": { + "External id": 7470,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7469 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791331085, "dur": 25, + "args": { + "External id": 7471,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7470 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791331087, "dur": 23, + "args": { + "External id": 7472,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7471 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791331114, "dur": 26, + "args": { + "External id": 7473,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7472 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791331117, "dur": 22, + "args": { + "External id": 7474,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7473 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791331144, "dur": 25, + "args": { + "External id": 7475,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7474 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791331146, "dur": 23, + "args": { + "External id": 7476,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7475 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791331408, "dur": 33, + "args": { + "External id": 7477,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7476 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791331412, "dur": 29, + "args": { + "External id": 7478,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7477 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791331447, "dur": 293, + "args": { + "External id": 7479,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7478 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791331449, "dur": 290, + "args": { + "External id": 7480,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7479 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791331747, "dur": 147, + "args": { + "External id": 7481,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7480 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791331750, "dur": 143, + "args": { + "External id": 7482,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7481 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791331903, "dur": 32, + "args": { + "External id": 7483,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7482 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791331906, "dur": 28, + "args": { + "External id": 7484,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7483 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791331940, "dur": 27, + "args": { + "External id": 7485,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7484 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791331942, "dur": 24, + "args": { + "External id": 7486,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7485 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791331971, "dur": 28, + "args": { + "External id": 7487,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7486 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791331974, "dur": 24, + "args": { + "External id": 7488,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7487 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791332004, "dur": 62, + "args": { + "External id": 7489,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7488 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791332006, "dur": 59, + "args": { + "External id": 7490,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7489 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791332072, "dur": 30, + "args": { + "External id": 7491,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7490 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791332075, "dur": 26, + "args": { + "External id": 7492,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7491 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791332106, "dur": 29, + "args": { + "External id": 7493,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7492 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791332111, "dur": 23, + "args": { + "External id": 7494,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7493 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791332139, "dur": 27, + "args": { + "External id": 7495,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7494 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791332142, "dur": 23, + "args": { + "External id": 7496,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7495 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791332170, "dur": 28, + "args": { + "External id": 7497,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7496 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791332173, "dur": 24, + "args": { + "External id": 7498,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7497 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791332202, "dur": 49, + "args": { + "External id": 7499,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7498 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791332205, "dur": 45, + "args": { + "External id": 7500,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7499 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791332256, "dur": 29, + "args": { + "External id": 7501,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7500 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791332260, "dur": 24, + "args": { + "External id": 7502,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7501 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791332290, "dur": 27, + "args": { + "External id": 7503,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7502 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791332294, "dur": 22, + "args": { + "External id": 7504,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7503 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791332322, "dur": 26, + "args": { + "External id": 7505,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7504 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791332324, "dur": 23, + "args": { + "External id": 7506,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7505 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791332352, "dur": 26, + "args": { + "External id": 7507,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7506 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791332355, "dur": 23, + "args": { + "External id": 7508,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7507 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791332383, "dur": 27, + "args": { + "External id": 7509,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7508 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791332387, "dur": 23, + "args": { + "External id": 7510,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7509 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791332415, "dur": 26, + "args": { + "External id": 7511,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7510 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791332417, "dur": 23, + "args": { + "External id": 7512,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7511 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791332445, "dur": 28, + "args": { + "External id": 7513,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7512 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791332448, "dur": 24, + "args": { + "External id": 7514,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7513 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791332477, "dur": 27, + "args": { + "External id": 7515,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7514 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791332480, "dur": 23, + "args": { + "External id": 7516,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7515 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791332508, "dur": 27, + "args": { + "External id": 7517,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7516 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791332511, "dur": 23, + "args": { + "External id": 7518,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7517 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791332539, "dur": 28, + "args": { + "External id": 7519,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7518 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791332543, "dur": 23, + "args": { + "External id": 7520,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7519 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791332571, "dur": 26, + "args": { + "External id": 7521,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7520 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791332573, "dur": 23, + "args": { + "External id": 7522,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7521 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791332601, "dur": 26, + "args": { + "External id": 7523,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7522 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791332603, "dur": 23, + "args": { + "External id": 7524,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7523 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791332631, "dur": 175, + "args": { + "External id": 7525,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7524 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791332780, "dur": 25, + "args": { + "External id": 7526,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7525 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791332811, "dur": 26, + "args": { + "External id": 7527,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7526 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791332813, "dur": 24, + "args": { + "External id": 7528,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7527 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791332842, "dur": 32, + "args": { + "External id": 7529,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7528 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791332844, "dur": 29, + "args": { + "External id": 7530,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7529 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791332878, "dur": 27, + "args": { + "External id": 7531,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7530 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791332881, "dur": 23, + "args": { + "External id": 7532,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7531 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791332909, "dur": 26, + "args": { + "External id": 7533,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7532 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791332911, "dur": 23, + "args": { + "External id": 7534,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7533 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791332939, "dur": 27, + "args": { + "External id": 7535,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7534 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791332942, "dur": 23, + "args": { + "External id": 7536,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7535 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791332970, "dur": 26, + "args": { + "External id": 7537,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7536 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791332972, "dur": 23, + "args": { + "External id": 7538,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7537 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791333000, "dur": 26, + "args": { + "External id": 7539,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7538 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791333002, "dur": 23, + "args": { + "External id": 7540,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7539 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791333030, "dur": 26, + "args": { + "External id": 7541,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7540 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791333033, "dur": 23, + "args": { + "External id": 7542,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7541 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791333061, "dur": 26, + "args": { + "External id": 7543,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7542 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791333063, "dur": 23, + "args": { + "External id": 7544,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7543 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791333091, "dur": 26, + "args": { + "External id": 7545,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7544 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791333093, "dur": 23, + "args": { + "External id": 7546,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7545 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791333121, "dur": 26, + "args": { + "External id": 7547,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7546 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791333123, "dur": 23, + "args": { + "External id": 7548,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7547 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791333151, "dur": 26, + "args": { + "External id": 7549,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7548 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791333154, "dur": 22, + "args": { + "External id": 7550,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7549 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791333181, "dur": 26, + "args": { + "External id": 7551,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7550 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791333184, "dur": 23, + "args": { + "External id": 7552,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7551 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791333212, "dur": 26, + "args": { + "External id": 7553,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7552 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791333214, "dur": 23, + "args": { + "External id": 7554,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7553 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791333242, "dur": 25, + "args": { + "External id": 7555,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7554 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791333244, "dur": 23, + "args": { + "External id": 7556,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7555 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791333271, "dur": 27, + "args": { + "External id": 7557,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7556 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791333274, "dur": 23, + "args": { + "External id": 7558,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7557 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791333302, "dur": 25, + "args": { + "External id": 7559,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7558 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791333304, "dur": 23, + "args": { + "External id": 7560,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7559 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791333332, "dur": 26, + "args": { + "External id": 7561,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7560 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791333334, "dur": 23, + "args": { + "External id": 7562,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7561 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791333362, "dur": 26, + "args": { + "External id": 7563,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7562 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791333365, "dur": 22, + "args": { + "External id": 7564,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7563 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791333392, "dur": 26, + "args": { + "External id": 7565,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7564 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791333395, "dur": 22, + "args": { + "External id": 7566,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7565 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791333422, "dur": 26, + "args": { + "External id": 7567,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7566 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791333425, "dur": 23, + "args": { + "External id": 7568,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7567 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791333453, "dur": 25, + "args": { + "External id": 7569,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7568 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791333455, "dur": 23, + "args": { + "External id": 7570,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7569 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791333483, "dur": 25, + "args": { + "External id": 7571,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7570 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791333485, "dur": 23, + "args": { + "External id": 7572,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7571 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791333513, "dur": 26, + "args": { + "External id": 7573,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7572 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791333516, "dur": 22, + "args": { + "External id": 7574,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7573 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791333543, "dur": 26, + "args": { + "External id": 7575,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7574 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791333545, "dur": 23, + "args": { + "External id": 7576,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7575 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791334232, "dur": 36, + "args": { + "External id": 7577,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7576 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791334236, "dur": 31, + "args": { + "External id": 7578,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7577 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791334274, "dur": 27, + "args": { + "External id": 7579,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7578 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791334277, "dur": 24, + "args": { + "External id": 7580,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7579 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791334306, "dur": 26, + "args": { + "External id": 7581,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7580 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791334309, "dur": 23, + "args": { + "External id": 7582,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7581 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791334337, "dur": 26, + "args": { + "External id": 7583,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7582 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791334339, "dur": 23, + "args": { + "External id": 7584,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7583 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791334367, "dur": 26, + "args": { + "External id": 7585,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7584 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791334370, "dur": 23, + "args": { + "External id": 7586,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7585 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791334398, "dur": 25, + "args": { + "External id": 7587,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7586 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791334400, "dur": 23, + "args": { + "External id": 7588,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7587 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791334428, "dur": 25, + "args": { + "External id": 7589,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7588 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791334430, "dur": 23, + "args": { + "External id": 7590,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7589 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791334458, "dur": 25, + "args": { + "External id": 7591,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7590 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791334460, "dur": 23, + "args": { + "External id": 7592,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7591 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791334488, "dur": 26, + "args": { + "External id": 7593,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7592 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791334490, "dur": 23, + "args": { + "External id": 7594,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7593 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791334519, "dur": 26, + "args": { + "External id": 7595,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7594 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791334521, "dur": 23, + "args": { + "External id": 7596,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7595 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791334772, "dur": 33, + "args": { + "External id": 7597,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7596 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791334776, "dur": 28, + "args": { + "External id": 7598,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7597 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791334811, "dur": 27, + "args": { + "External id": 7599,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7598 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791334814, "dur": 23, + "args": { + "External id": 7600,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7599 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791334842, "dur": 26, + "args": { + "External id": 7601,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7600 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791334845, "dur": 23, + "args": { + "External id": 7602,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7601 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791334873, "dur": 26, + "args": { + "External id": 7603,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7602 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791334875, "dur": 23, + "args": { + "External id": 7604,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7603 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791334903, "dur": 25, + "args": { + "External id": 7605,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7604 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791334906, "dur": 22, + "args": { + "External id": 7606,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7605 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791334933, "dur": 32, + "args": { + "External id": 7607,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7606 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791334935, "dur": 29, + "args": { + "External id": 7608,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7607 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791334970, "dur": 27, + "args": { + "External id": 7609,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7608 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791334972, "dur": 24, + "args": { + "External id": 7610,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7609 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335002, "dur": 25, + "args": { + "External id": 7611,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7610 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335004, "dur": 23, + "args": { + "External id": 7612,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7611 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335032, "dur": 25, + "args": { + "External id": 7613,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7612 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335034, "dur": 23, + "args": { + "External id": 7614,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7613 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335062, "dur": 25, + "args": { + "External id": 7615,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7614 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335064, "dur": 23, + "args": { + "External id": 7616,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7615 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335092, "dur": 25, + "args": { + "External id": 7617,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7616 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335094, "dur": 23, + "args": { + "External id": 7618,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7617 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335122, "dur": 26, + "args": { + "External id": 7619,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7618 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335124, "dur": 23, + "args": { + "External id": 7620,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7619 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335152, "dur": 78, + "args": { + "External id": 7621,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7620 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335205, "dur": 24, + "args": { + "External id": 7622,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7621 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335235, "dur": 26, + "args": { + "External id": 7623,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7622 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335237, "dur": 23, + "args": { + "External id": 7624,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7623 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335265, "dur": 26, + "args": { + "External id": 7625,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7624 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335267, "dur": 23, + "args": { + "External id": 7626,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7625 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335295, "dur": 26, + "args": { + "External id": 7627,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7626 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335297, "dur": 23, + "args": { + "External id": 7628,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7627 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335325, "dur": 25, + "args": { + "External id": 7629,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7628 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335327, "dur": 23, + "args": { + "External id": 7630,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7629 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335355, "dur": 25, + "args": { + "External id": 7631,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7630 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335357, "dur": 23, + "args": { + "External id": 7632,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7631 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335385, "dur": 25, + "args": { + "External id": 7633,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7632 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335387, "dur": 22, + "args": { + "External id": 7634,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7633 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335414, "dur": 26, + "args": { + "External id": 7635,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7634 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335417, "dur": 22, + "args": { + "External id": 7636,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7635 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335444, "dur": 26, + "args": { + "External id": 7637,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7636 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335447, "dur": 23, + "args": { + "External id": 7638,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7637 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335475, "dur": 25, + "args": { + "External id": 7639,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7638 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335477, "dur": 23, + "args": { + "External id": 7640,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7639 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335505, "dur": 26, + "args": { + "External id": 7641,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7640 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335507, "dur": 23, + "args": { + "External id": 7642,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7641 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335535, "dur": 26, + "args": { + "External id": 7643,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7642 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335537, "dur": 23, + "args": { + "External id": 7644,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7643 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335565, "dur": 25, + "args": { + "External id": 7645,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7644 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335567, "dur": 23, + "args": { + "External id": 7646,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7645 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335595, "dur": 25, + "args": { + "External id": 7647,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7646 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335597, "dur": 23, + "args": { + "External id": 7648,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7647 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335624, "dur": 26, + "args": { + "External id": 7649,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7648 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335627, "dur": 22, + "args": { + "External id": 7650,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7649 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335654, "dur": 26, + "args": { + "External id": 7651,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7650 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335656, "dur": 23, + "args": { + "External id": 7652,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7651 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335684, "dur": 27, + "args": { + "External id": 7653,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7652 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335687, "dur": 23, + "args": { + "External id": 7654,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7653 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335715, "dur": 26, + "args": { + "External id": 7655,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7654 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335717, "dur": 23, + "args": { + "External id": 7656,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7655 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335745, "dur": 26, + "args": { + "External id": 7657,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7656 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335747, "dur": 23, + "args": { + "External id": 7658,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7657 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335775, "dur": 25, + "args": { + "External id": 7659,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7658 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335777, "dur": 23, + "args": { + "External id": 7660,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7659 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335805, "dur": 25, + "args": { + "External id": 7661,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7660 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335807, "dur": 22, + "args": { + "External id": 7662,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7661 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335834, "dur": 26, + "args": { + "External id": 7663,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7662 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335837, "dur": 22, + "args": { + "External id": 7664,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7663 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335864, "dur": 26, + "args": { + "External id": 7665,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7664 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335866, "dur": 23, + "args": { + "External id": 7666,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7665 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335894, "dur": 25, + "args": { + "External id": 7667,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7666 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335896, "dur": 23, + "args": { + "External id": 7668,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7667 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335924, "dur": 25, + "args": { + "External id": 7669,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7668 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335926, "dur": 22, + "args": { + "External id": 7670,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7669 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335953, "dur": 26, + "args": { + "External id": 7671,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7670 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335956, "dur": 22, + "args": { + "External id": 7672,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7671 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791335983, "dur": 25, + "args": { + "External id": 7673,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7672 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791335985, "dur": 23, + "args": { + "External id": 7674,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7673 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791336012, "dur": 26, + "args": { + "External id": 7675,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7674 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791336015, "dur": 22, + "args": { + "External id": 7676,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7675 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791336042, "dur": 26, + "args": { + "External id": 7677,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7676 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791336045, "dur": 22, + "args": { + "External id": 7678,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7677 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791336072, "dur": 25, + "args": { + "External id": 7679,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7678 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791336074, "dur": 23, + "args": { + "External id": 7680,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7679 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791336112, "dur": 26, + "args": { + "External id": 7681,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7680 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791336114, "dur": 23, + "args": { + "External id": 7682,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7681 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791336142, "dur": 26, + "args": { + "External id": 7683,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7682 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791336145, "dur": 22, + "args": { + "External id": 7684,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7683 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791336172, "dur": 26, + "args": { + "External id": 7685,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7684 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791336175, "dur": 22, + "args": { + "External id": 7686,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7685 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791336202, "dur": 45, + "args": { + "External id": 7687,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7686 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791336204, "dur": 42, + "args": { + "External id": 7688,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7687 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791336253, "dur": 29, + "args": { + "External id": 7689,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7688 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791336257, "dur": 24, + "args": { + "External id": 7690,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7689 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791336287, "dur": 25, + "args": { + "External id": 7691,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7690 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791336289, "dur": 23, + "args": { + "External id": 7692,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7691 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791336317, "dur": 26, + "args": { + "External id": 7693,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7692 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791336319, "dur": 23, + "args": { + "External id": 7694,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7693 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791336347, "dur": 26, + "args": { + "External id": 7695,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7694 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791336349, "dur": 23, + "args": { + "External id": 7696,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7695 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791337261, "dur": 34, + "args": { + "External id": 7697,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7696 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791337264, "dur": 30, + "args": { + "External id": 7698,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7697 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791337300, "dur": 27, + "args": { + "External id": 7699,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7698 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791337303, "dur": 24, + "args": { + "External id": 7700,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7699 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791337332, "dur": 26, + "args": { + "External id": 7701,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7700 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791337334, "dur": 23, + "args": { + "External id": 7702,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7701 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791337362, "dur": 26, + "args": { + "External id": 7703,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7702 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791337365, "dur": 22, + "args": { + "External id": 7704,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7703 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791337392, "dur": 26, + "args": { + "External id": 7705,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7704 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791337395, "dur": 22, + "args": { + "External id": 7706,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7705 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791337422, "dur": 26, + "args": { + "External id": 7707,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7706 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791337425, "dur": 22, + "args": { + "External id": 7708,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7707 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791337452, "dur": 25, + "args": { + "External id": 7709,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7708 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791337454, "dur": 22, + "args": { + "External id": 7710,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7709 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791337481, "dur": 26, + "args": { + "External id": 7711,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7710 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791337484, "dur": 22, + "args": { + "External id": 7712,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7711 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791337511, "dur": 26, + "args": { + "External id": 7713,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7712 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791337514, "dur": 22, + "args": { + "External id": 7714,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7713 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791337541, "dur": 25, + "args": { + "External id": 7715,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7714 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791337543, "dur": 22, + "args": { + "External id": 7716,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7715 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791337805, "dur": 34, + "args": { + "External id": 7717,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7716 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791337809, "dur": 29, + "args": { + "External id": 7718,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7717 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791337845, "dur": 26, + "args": { + "External id": 7719,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7718 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791337847, "dur": 23, + "args": { + "External id": 7720,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7719 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791337876, "dur": 26, + "args": { + "External id": 7721,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7720 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791337878, "dur": 23, + "args": { + "External id": 7722,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7721 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791337906, "dur": 26, + "args": { + "External id": 7723,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7722 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791337909, "dur": 22, + "args": { + "External id": 7724,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7723 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791337937, "dur": 26, + "args": { + "External id": 7725,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7724 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791337939, "dur": 23, + "args": { + "External id": 7726,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7725 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791337967, "dur": 26, + "args": { + "External id": 7727,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7726 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791337970, "dur": 22, + "args": { + "External id": 7728,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7727 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791337998, "dur": 26, + "args": { + "External id": 7729,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7728 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338000, "dur": 23, + "args": { + "External id": 7730,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7729 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338028, "dur": 25, + "args": { + "External id": 7731,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7730 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338030, "dur": 22, + "args": { + "External id": 7732,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7731 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338058, "dur": 25, + "args": { + "External id": 7733,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7732 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338060, "dur": 23, + "args": { + "External id": 7734,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7733 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338088, "dur": 25, + "args": { + "External id": 7735,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7734 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338090, "dur": 22, + "args": { + "External id": 7736,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7735 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338118, "dur": 26, + "args": { + "External id": 7737,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7736 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338120, "dur": 23, + "args": { + "External id": 7738,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7737 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338148, "dur": 25, + "args": { + "External id": 7739,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7738 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338150, "dur": 23, + "args": { + "External id": 7740,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7739 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338178, "dur": 25, + "args": { + "External id": 7741,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7740 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338180, "dur": 22, + "args": { + "External id": 7742,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7741 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338207, "dur": 26, + "args": { + "External id": 7743,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7742 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338210, "dur": 22, + "args": { + "External id": 7744,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7743 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338238, "dur": 25, + "args": { + "External id": 7745,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7744 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338240, "dur": 22, + "args": { + "External id": 7746,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7745 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338267, "dur": 26, + "args": { + "External id": 7747,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7746 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338270, "dur": 22, + "args": { + "External id": 7748,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7747 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338297, "dur": 26, + "args": { + "External id": 7749,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7748 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338300, "dur": 22, + "args": { + "External id": 7750,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7749 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338327, "dur": 25, + "args": { + "External id": 7751,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7750 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338329, "dur": 23, + "args": { + "External id": 7752,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7751 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338357, "dur": 25, + "args": { + "External id": 7753,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7752 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338359, "dur": 22, + "args": { + "External id": 7754,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7753 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338386, "dur": 26, + "args": { + "External id": 7755,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7754 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338389, "dur": 22, + "args": { + "External id": 7756,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7755 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338416, "dur": 25, + "args": { + "External id": 7757,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7756 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338418, "dur": 23, + "args": { + "External id": 7758,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7757 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338446, "dur": 25, + "args": { + "External id": 7759,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7758 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338448, "dur": 22, + "args": { + "External id": 7760,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7759 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338475, "dur": 25, + "args": { + "External id": 7761,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7760 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338478, "dur": 22, + "args": { + "External id": 7762,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7761 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338505, "dur": 25, + "args": { + "External id": 7763,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7762 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338507, "dur": 23, + "args": { + "External id": 7764,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7763 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338535, "dur": 25, + "args": { + "External id": 7765,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7764 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338537, "dur": 23, + "args": { + "External id": 7766,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7765 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338565, "dur": 25, + "args": { + "External id": 7767,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7766 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338567, "dur": 22, + "args": { + "External id": 7768,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7767 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338594, "dur": 26, + "args": { + "External id": 7769,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7768 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338597, "dur": 22, + "args": { + "External id": 7770,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7769 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338624, "dur": 26, + "args": { + "External id": 7771,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7770 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338626, "dur": 23, + "args": { + "External id": 7772,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7771 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338654, "dur": 25, + "args": { + "External id": 7773,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7772 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338656, "dur": 23, + "args": { + "External id": 7774,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7773 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338684, "dur": 25, + "args": { + "External id": 7775,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7774 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338686, "dur": 23, + "args": { + "External id": 7776,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7775 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338714, "dur": 25, + "args": { + "External id": 7777,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7776 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338716, "dur": 22, + "args": { + "External id": 7778,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7777 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338743, "dur": 26, + "args": { + "External id": 7779,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7778 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338746, "dur": 22, + "args": { + "External id": 7780,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7779 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338773, "dur": 28, + "args": { + "External id": 7781,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7780 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338778, "dur": 22, + "args": { + "External id": 7782,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7781 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338805, "dur": 26, + "args": { + "External id": 7783,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7782 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338808, "dur": 22, + "args": { + "External id": 7784,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7783 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338836, "dur": 25, + "args": { + "External id": 7785,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7784 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338838, "dur": 23, + "args": { + "External id": 7786,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7785 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338866, "dur": 25, + "args": { + "External id": 7787,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7786 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338868, "dur": 22, + "args": { + "External id": 7788,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7787 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338895, "dur": 26, + "args": { + "External id": 7789,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7788 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338898, "dur": 22, + "args": { + "External id": 7790,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7789 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338925, "dur": 26, + "args": { + "External id": 7791,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7790 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338927, "dur": 23, + "args": { + "External id": 7792,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7791 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338955, "dur": 25, + "args": { + "External id": 7793,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7792 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338957, "dur": 23, + "args": { + "External id": 7794,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7793 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791338985, "dur": 25, + "args": { + "External id": 7795,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7794 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791338987, "dur": 22, + "args": { + "External id": 7796,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7795 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791339014, "dur": 26, + "args": { + "External id": 7797,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7796 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791339017, "dur": 22, + "args": { + "External id": 7798,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7797 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791339044, "dur": 25, + "args": { + "External id": 7799,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7798 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791339046, "dur": 23, + "args": { + "External id": 7800,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7799 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791339074, "dur": 26, + "args": { + "External id": 7801,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7800 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791339076, "dur": 23, + "args": { + "External id": 7802,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7801 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791339104, "dur": 25, + "args": { + "External id": 7803,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7802 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791339106, "dur": 23, + "args": { + "External id": 7804,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7803 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791339134, "dur": 25, + "args": { + "External id": 7805,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7804 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791339136, "dur": 23, + "args": { + "External id": 7806,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7805 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791339164, "dur": 26, + "args": { + "External id": 7807,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7806 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791339166, "dur": 23, + "args": { + "External id": 7808,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7807 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791339194, "dur": 25, + "args": { + "External id": 7809,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7808 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791339196, "dur": 23, + "args": { + "External id": 7810,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7809 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791339224, "dur": 26, + "args": { + "External id": 7811,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7810 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791339226, "dur": 23, + "args": { + "External id": 7812,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7811 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791339254, "dur": 25, + "args": { + "External id": 7813,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7812 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791339256, "dur": 23, + "args": { + "External id": 7814,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7813 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791339284, "dur": 26, + "args": { + "External id": 7815,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7814 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791339286, "dur": 23, + "args": { + "External id": 7816,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[1, 4096], [4096, 8192]], "Ev Idx": 7815 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__and__", "pid": 435, "tid": 435, + "ts": 1776649791340344, "dur": 59, + "args": { + "External id": 7817,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "15"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 7816 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_and", "pid": 435, "tid": 435, + "ts": 1776649791340347, "dur": 56, + "args": { + "External id": 7818,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "15"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 7817 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_and", "pid": 435, "tid": 435, + "ts": 1776649791340357, "dur": 43, + "args": { + "External id": 7819,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["unsigned char", "long int"], "Input Dims": [[4096, 4096], []], "Ev Idx": 7818 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791340411, "dur": 48, + "args": { + "External id": 7820,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 7819 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791340415, "dur": 44, + "args": { + "External id": 7821,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["unsigned char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], [], []], "Ev Idx": 7820 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791340425, "dur": 8, + "args": { + "External id": 7822,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "[4096, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7821 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791340436, "dur": 22, + "args": { + "External id": 7823,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "unsigned char", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 7822 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__rshift__", "pid": 435, "tid": 435, + "ts": 1776649791340467, "dur": 35, + "args": { + "External id": 7824,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 7823 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791340478, "dur": 6, + "args": { + "External id": 7825,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "0", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7824 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791340505, "dur": 23, + "args": { + "External id": 7826,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 7825 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791340506, "dur": 22, + "args": { + "External id": 7827,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["unsigned char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], [], []], "Ev Idx": 7826 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791340508, "dur": 5, + "args": { + "External id": 7828,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "[4096, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7827 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791340515, "dur": 12, + "args": { + "External id": 7829,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "unsigned char", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 7828 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::stack", "pid": 435, "tid": 435, + "ts": 1776649791340542, "dur": 53, + "args": { + "External id": 7830,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["TensorList", "Scalar"], "Input Dims": [[[4096, 4096], [4096, 4096]], []], "Ev Idx": 7829 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649791340548, "dur": 6, + "args": { + "External id": 7831,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["signed char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 7830 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791340551, "dur": 3, + "args": { + "External id": 7832,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096, 1]", "[4096, 1, 1]", ""], "Input type": ["signed char", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 7831 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649791340555, "dur": 2, + "args": { + "External id": 7833,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["signed char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 7832 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791340556, "dur": 1, + "args": { + "External id": 7834,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096, 1]", "[4096, 1, 1]", ""], "Input type": ["signed char", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 7833 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 435, "tid": 435, + "ts": 1776649791340560, "dur": 33, + "args": { + "External id": 7835,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["TensorList", "Scalar"], "Input Dims": [[[4096, 4096, 1], [4096, 4096, 1]], []], "Ev Idx": 7834 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791340604, "dur": 9, + "args": { + "External id": 7836,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["signed char", "ScalarList"], "Input Dims": [[4096, 4096, 2], []], "Ev Idx": 7835 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791340606, "dur": 6, + "args": { + "External id": 7837,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["signed char", "ScalarList"], "Input Dims": [[4096, 4096, 2], []], "Ev Idx": 7836 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::repeat_interleave", "pid": 435, "tid": 435, + "ts": 1776649791340620, "dur": 62, + "args": { + "External id": 7838,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "128", "0", ""], "Input type": ["c10::Half", "Scalar", "Scalar", ""], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 7837 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649791340623, "dur": 4, + "args": { + "External id": 7839,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 8192], []], "Ev Idx": 7838 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791340625, "dur": 1, + "args": { + "External id": 7840,"Record function id": 0, "Concrete Inputs": ["", "[32, 1, 8192]", "[8192, 8192, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 7839 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 435, "tid": 435, + "ts": 1776649791340629, "dur": 7, + "args": { + "External id": 7841,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]", "False"], "Input type": ["c10::Half", "ScalarList", "Scalar"], "Input Dims": [[32, 1, 8192], [], []], "Ev Idx": 7840 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791340634, "dur": 1, + "args": { + "External id": 7842,"Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]", "[8192, 0, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 1, 8192], [], [], []], "Ev Idx": 7841 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649791340639, "dur": 34, + "args": { + "External id": 7843,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 7842 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649791340644, "dur": 9, + "args": { + "External id": 7844,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[32, 128, 8192], [], [], [], [], []], "Ev Idx": 7843 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791340646, "dur": 7, + "args": { + "External id": 7845,"Record function id": 0, "Concrete Inputs": ["[32, 128, 8192]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7844 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791340655, "dur": 18, + "args": { + "External id": 7846,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[32, 128, 8192], [32, 128, 8192], []], "Ev Idx": 7845 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::flatten", "pid": 435, "tid": 435, + "ts": 1776649791340676, "dur": 5, + "args": { + "External id": 7847,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[32, 128, 8192], [], []], "Ev Idx": 7846 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791340679, "dur": 2, + "args": { + "External id": 7848,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 7847 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 435, "tid": 435, + "ts": 1776649791340691, "dur": 26, + "args": { + "External id": 7849,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["signed char", "long int", "Scalar"], "Input Dims": [[4096, 8192], [], []], "Ev Idx": 7848 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791340723, "dur": 25, + "args": { + "External id": 7850,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["signed char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 7849 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791340724, "dur": 24, + "args": { + "External id": 7851,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "", "", "", "False", ""], "Input type": ["signed char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 8192], [], [], [], [], [], []], "Ev Idx": 7850 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791340727, "dur": 5, + "args": { + "External id": 7852,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7851 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791340734, "dur": 13, + "args": { + "External id": 7853,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "signed char", "Scalar"], "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 7852 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 435, "tid": 435, + "ts": 1776649791340756, "dur": 466, + "args": { + "External id": 7854,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4096, 8192], [4096, 8192]], "Ev Idx": 7853 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 435, "tid": 435, + "ts": 1776649791341242, "dur": 13, + "args": { + "External id": 7855,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[4096, 8192]], "Ev Idx": 7854 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 435, "tid": 435, + "ts": 1776649791341247, "dur": 7, + "args": { + "External id": 7856,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[4096, 8192], [], []], "Ev Idx": 7855 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791341251, "dur": 2, + "args": { + "External id": 7857,"Record function id": 0, "Concrete Inputs": ["", "[8192, 4096]", "[1, 8192]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 8192], [], [], []], "Ev Idx": 7856 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 435, "tid": 435, + "ts": 1776649791341263, "dur": 32, + "args": { + "External id": 7858,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[8192, 4096], []], "Ev Idx": 7857 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649791341264, "dur": 31, + "args": { + "External id": 7859,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[8192, 4096], []], "Ev Idx": 7858 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649791341266, "dur": 9, + "args": { + "External id": 7860,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[8192, 4096], [], [], [], [], []], "Ev Idx": 7859 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791341268, "dur": 6, + "args": { + "External id": 7861,"Record function id": 0, "Concrete Inputs": ["[8192, 4096]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7860 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791341276, "dur": 18, + "args": { + "External id": 7862,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[8192, 4096], [8192, 4096], []], "Ev Idx": 7861 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791341351, "dur": 29, + "args": { + "External id": 7863,"Record function id": 0, "Concrete Inputs": ["[8192, 4096]", "", "", "", "False", ""], "Input type": ["ScalarList", "", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7862 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 435, "tid": 435, + "ts": 1776649791341391, "dur": 8, + "args": { + "External id": 7864,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Dims": [[8192, 4096]], "Ev Idx": 7863 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 435, "tid": 435, + "ts": 1776649791341395, "dur": 4, + "args": { + "External id": 7865,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Dims": [[8192, 4096]], "Ev Idx": 7864 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::uniform_", "pid": 435, "tid": 435, + "ts": 1776649791341494, "dur": 212488, + "args": { + "External id": 7866,"Record function id": 0, "Concrete Inputs": ["", "-0.015624999999999998", "0.015624999999999998", ""], "Input type": ["float", "Scalar", "Scalar", ""], "Input Dims": [[8192, 4096], [], [], []], "Ev Idx": 7865 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791554178, "dur": 125824, + "args": { + "External id": 7867,"Record function id": 0, "Concrete Inputs": ["", "", "5", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[8192, 4096], [], [], [], [], []], "Ev Idx": 7866 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791554184, "dur": 125815, + "args": { + "External id": 7868,"Record function id": 0, "Concrete Inputs": ["", "5", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[8192, 4096], [], [], [], [], [], []], "Ev Idx": 7867 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791554210, "dur": 59, + "args": { + "External id": 7869,"Record function id": 0, "Concrete Inputs": ["[8192, 4096]", "[4096, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7868 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791554274, "dur": 125711, + "args": { + "External id": 7870,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "float", "Scalar"], "Input Dims": [[8192, 4096], [8192, 4096], []], "Ev Idx": 7869 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791554295, "dur": 18864, + "args": { + "External id": 7871,"Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[8192, 4096], [], [], [], []], "Ev Idx": 7870 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791554297, "dur": 18862, + "args": { + "External id": 7872,"Record function id": 0, "Concrete Inputs": ["", "5", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[8192, 4096], [], [], [], [], [], []], "Ev Idx": 7871 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791554299, "dur": 42, + "args": { + "External id": 7873,"Record function id": 0, "Concrete Inputs": ["[8192, 4096]", "[4096, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7872 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791554343, "dur": 18810, + "args": { + "External id": 7874,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "float", "Scalar"], "Input Dims": [[8192, 4096], [8192, 4096], []], "Ev Idx": 7873 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::expand_as", "pid": 435, "tid": 435, + "ts": 1776649791573172, "dur": 24, + "args": { + "External id": 7875,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8192, 4096], [8192, 4096]], "Ev Idx": 7874 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 435, "tid": 435, + "ts": 1776649791573181, "dur": 14, + "args": { + "External id": 7876,"Record function id": 0, "Concrete Inputs": ["", "[8192, 4096]", "False"], "Input type": ["c10::Half", "ScalarList", "Scalar"], "Input Dims": [[8192, 4096], [], []], "Ev Idx": 7875 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791573188, "dur": 7, + "args": { + "External id": 7877,"Record function id": 0, "Concrete Inputs": ["", "[8192, 4096]", "[4096, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[8192, 4096], [], [], []], "Ev Idx": 7876 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791573198, "dur": 7586, + "args": { + "External id": 7878,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[8192, 4096], [8192, 4096], []], "Ev Idx": 7877 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_has_compatible_shallow_copy_type", "pid": 435, "tid": 435, + "ts": 1776649791680128, "dur": 3, + "args": { + "External id": 7879,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "c10::Half"], "Input Dims": [[8192, 4096], [8192, 4096]], "Ev Idx": 7878 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_has_compatible_shallow_copy_type", "pid": 435, "tid": 435, + "ts": 1776649791680154, "dur": 1, + "args": { + "External id": 7880,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "c10::Half"], "Input Dims": [[8192, 4096], [8192, 4096]], "Ev Idx": 7879 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791695671, "dur": 122, + "args": { + "External id": 7881,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[8192, 4096], [8192, 4096], []], "Ev Idx": 7880 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791695887, "dur": 41, + "args": { + "External id": 7882,"Record function id": 0, "Concrete Inputs": ["[256, 16384]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7881 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791695967, "dur": 1, + "args": { + "External id": 7883,"Record function id": 0, "Concrete Inputs": ["[32, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7882 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791695983, "dur": 10, + "args": { + "External id": 7884,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 7883 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791695987, "dur": 1, + "args": { + "External id": 7885,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7884 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791695991, "dur": 2, + "args": { + "External id": 7886,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 7885 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791696030, "dur": 4889, + "args": { + "External id": 7887,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 16384], [], [], [], [], [], [], []], "Ev Idx": 7886 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791696033, "dur": 4886, + "args": { + "External id": 7888,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[256, 16384], [], [], [], [], [], []], "Ev Idx": 7887 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791696044, "dur": 22, + "args": { + "External id": 7889,"Record function id": 0, "Concrete Inputs": ["[256, 16384]", "[16384, 1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7888 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791696069, "dur": 4847, + "args": { + "External id": 7890,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 16384], [256, 16384], []], "Ev Idx": 7889 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791701120, "dur": 108, + "args": { + "External id": 7891,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[32, 8192], [], [], [], [], [], [], []], "Ev Idx": 7890 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791701121, "dur": 106, + "args": { + "External id": 7892,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[32, 8192], [], [], [], [], [], []], "Ev Idx": 7891 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791701126, "dur": 10, + "args": { + "External id": 7893,"Record function id": 0, "Concrete Inputs": ["[32, 8192]", "[8192, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7892 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791701138, "dur": 89, + "args": { + "External id": 7894,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[32, 8192], [32, 8192], []], "Ev Idx": 7893 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791701235, "dur": 33, + "args": { + "External id": 7895,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[1024], [], [], [], [], [], [], []], "Ev Idx": 7894 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791701236, "dur": 31, + "args": { + "External id": 7896,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[1024], [], [], [], [], [], []], "Ev Idx": 7895 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791701239, "dur": 6, + "args": { + "External id": 7897,"Record function id": 0, "Concrete Inputs": ["[1024]", "[1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7896 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791701247, "dur": 20, + "args": { + "External id": 7898,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[1024], [1024], []], "Ev Idx": 7897 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 435, "tid": 435, + "ts": 1776649791701284, "dur": 26, + "args": { + "External id": 7899,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[32, 8192]], "Ev Idx": 7898 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 435, "tid": 435, + "ts": 1776649791701291, "dur": 15, + "args": { + "External id": 7900,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[32, 8192], [], []], "Ev Idx": 7899 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791701299, "dur": 6, + "args": { + "External id": 7901,"Record function id": 0, "Concrete Inputs": ["", "[8192, 32]", "[1, 8192]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 7900 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 435, "tid": 435, + "ts": 1776649791701314, "dur": 3, + "args": { + "External id": 7902,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[8192, 4096]], "Ev Idx": 7901 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 435, "tid": 435, + "ts": 1776649791701315, "dur": 2, + "args": { + "External id": 7903,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[8192, 4096], [], []], "Ev Idx": 7902 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791701316, "dur": 1, + "args": { + "External id": 7904,"Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]", "[1, 4096]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[8192, 4096], [], [], []], "Ev Idx": 7903 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791701327, "dur": 25, + "args": { + "External id": 7905,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 128, 8192]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 8192], []], "Ev Idx": 7904 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 435, "tid": 435, + "ts": 1776649791701343, "dur": 8, + "args": { + "External id": 7906,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]", "[128, 1, 4096]"], "Input type": ["c10::Half", "ScalarList", "ScalarList"], "Input Dims": [[4096, 8192], [], []], "Ev Idx": 7905 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 435, "tid": 435, + "ts": 1776649791701374, "dur": 9, + "args": { + "External id": 7907,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1, 0, 2]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 7906 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791701381, "dur": 1, + "args": { + "External id": 7908,"Record function id": 0, "Concrete Inputs": ["", "[128, 32, 8192]", "[1, 128, 4096]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 128, 8192], [], [], []], "Ev Idx": 7907 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791701386, "dur": 76, + "args": { + "External id": 7909,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[128, -1]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[128, 32, 8192], []], "Ev Idx": 7908 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649791701389, "dur": 63, + "args": { + "External id": 7910,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[128, 32, 8192], []], "Ev Idx": 7909 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649791701395, "dur": 15, + "args": { + "External id": 7911,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[128, 32, 8192], [], [], [], [], []], "Ev Idx": 7910 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791701398, "dur": 11, + "args": { + "External id": 7912,"Record function id": 0, "Concrete Inputs": ["[128, 32, 8192]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7911 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791701411, "dur": 41, + "args": { + "External id": 7913,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[128, 32, 8192], [128, 32, 8192], []], "Ev Idx": 7912 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 435, "tid": 435, + "ts": 1776649791701455, "dur": 7, + "args": { + "External id": 7914,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[128, 262144]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[128, 32, 8192], []], "Ev Idx": 7913 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791701470, "dur": 31, + "args": { + "External id": 7915,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1, -1]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[8192, 32], []], "Ev Idx": 7914 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649791701472, "dur": 25, + "args": { + "External id": 7916,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[8192, 32], []], "Ev Idx": 7915 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649791701473, "dur": 8, + "args": { + "External id": 7917,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[8192, 32], [], [], [], [], []], "Ev Idx": 7916 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791701474, "dur": 7, + "args": { + "External id": 7918,"Record function id": 0, "Concrete Inputs": ["[8192, 32]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7917 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791701482, "dur": 15, + "args": { + "External id": 7919,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[8192, 32], [8192, 32], []], "Ev Idx": 7918 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 435, "tid": 435, + "ts": 1776649791701499, "dur": 1, + "args": { + "External id": 7920,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1, 262144]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[8192, 32], []], "Ev Idx": 7919 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 435, "tid": 435, + "ts": 1776649791701510, "dur": 481, + "args": { + "External id": 7921,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[128, 262144], [1, 262144]], "Ev Idx": 7920 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 435, "tid": 435, + "ts": 1776649791702002, "dur": 312, + "args": { + "External id": 7922,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[128, 262144]], "Ev Idx": 7921 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791702328, "dur": 270, + "args": { + "External id": 7923,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[128, 262144], [], [], [], []], "Ev Idx": 7922 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791702330, "dur": 268, + "args": { + "External id": 7924,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "", "", "", "False", ""], "Input type": ["c10::Half", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[128, 262144], [], [], [], [], [], []], "Ev Idx": 7923 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791702335, "dur": 234, + "args": { + "External id": 7925,"Record function id": 0, "Concrete Inputs": ["[128, 262144]", "[262144, 1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7924 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791702572, "dur": 25, + "args": { + "External id": 7926,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "c10::Half", "Scalar"], "Input Dims": [[128, 262144], [128, 262144], []], "Ev Idx": 7925 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 435, "tid": 435, + "ts": 1776649791702617, "dur": 31, + "args": { + "External id": 7927,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["int", "long int", "Scalar"], "Input Dims": [[128, 262144], [], []], "Ev Idx": 7926 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 435, "tid": 435, + "ts": 1776649791702676, "dur": 274, + "args": { + "External id": 7928,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "15"], "Input type": ["int", "Scalar", "Scalar"], "Input Dims": [[128, 262144], [], []], "Ev Idx": 7927 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791702682, "dur": 0, + "args": { + "External id": 7929,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[128, 262144], [], [], [], []], "Ev Idx": 7928 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791702965, "dur": 9, + "args": { + "External id": 7930,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[128, -1, 8192]"], "Input type": ["int", "ScalarList"], "Input Dims": [[128, 262144], []], "Ev Idx": 7929 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791702967, "dur": 7, + "args": { + "External id": 7931,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[128, -1, 8192]"], "Input type": ["int", "ScalarList"], "Input Dims": [[128, 262144], []], "Ev Idx": 7930 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 435, "tid": 435, + "ts": 1776649791702978, "dur": 6, + "args": { + "External id": 7932,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1, 0, 2]"], "Input type": ["int", "ScalarList"], "Input Dims": [[128, 32, 8192], []], "Ev Idx": 7931 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791702982, "dur": 1, + "args": { + "External id": 7933,"Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]", "[8192, 262144, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Dims": [[128, 32, 8192], [], [], []], "Ev Idx": 7932 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791702988, "dur": 42, + "args": { + "External id": 7934,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["int", "ScalarList"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 7933 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649791702991, "dur": 34, + "args": { + "External id": 7935,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 7934 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649791702993, "dur": 10, + "args": { + "External id": 7936,"Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "0"], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[32, 128, 8192], [], [], [], [], []], "Ev Idx": 7935 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791702995, "dur": 8, + "args": { + "External id": 7937,"Record function id": 0, "Concrete Inputs": ["[32, 128, 8192]", "3", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7936 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791703005, "dur": 19, + "args": { + "External id": 7938,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[32, 128, 8192], [32, 128, 8192], []], "Ev Idx": 7937 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 435, "tid": 435, + "ts": 1776649791703028, "dur": 1, + "args": { + "External id": 7939,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["int", "ScalarList"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 7938 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791703038, "dur": 3, + "args": { + "External id": 7940,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 64]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[1, 262144], []], "Ev Idx": 7939 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791703039, "dur": 2, + "args": { + "External id": 7941,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 64]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[1, 262144], []], "Ev Idx": 7940 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649791703054, "dur": 11, + "args": { + "External id": 7942,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 64], [], [], [], []], "Ev Idx": 7941 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791703061, "dur": 1, + "args": { + "External id": 7943,"Record function id": 0, "Concrete Inputs": ["", "[4096, 64]", "[64, 1]", "0"], "Input type": ["c10::Half", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 64], [], [], []], "Ev Idx": 7942 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791703082, "dur": 2, + "args": { + "External id": 7944,"Record function id": 0, "Concrete Inputs": ["[64]", "4", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7943 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791703093, "dur": 443, + "args": { + "External id": 7945,"Record function id": 0, "Concrete Inputs": ["", "", "4", "False", "False", ""], "Input type": ["long int", "", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[64], [], [], [], [], []], "Ev Idx": 7944 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791703095, "dur": 440, + "args": { + "External id": 7946,"Record function id": 0, "Concrete Inputs": ["", "4", "", "", "", "False", ""], "Input type": ["long int", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[64], [], [], [], [], [], []], "Ev Idx": 7945 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791703097, "dur": 8, + "args": { + "External id": 7947,"Record function id": 0, "Concrete Inputs": ["[64]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7946 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791703106, "dur": 429, + "args": { + "External id": 7948,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Dims": [[64], [64], []], "Ev Idx": 7947 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 435, "tid": 435, + "ts": 1776649791703539, "dur": 1, + "args": { + "External id": 7949,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Dims": [[64]], "Ev Idx": 7948 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::index", "pid": 435, "tid": 435, + "ts": 1776649791703547, "dur": 280, + "args": { + "External id": 7950,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", ""], "Input Dims": [[4096, 64], []], "Ev Idx": 7949 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791703560, "dur": 1, + "args": { + "External id": 7951,"Record function id": 0, "Concrete Inputs": ["", "[4096, 64]", "[64, 0]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 64], [], [], []], "Ev Idx": 7950 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791703563, "dur": 2, + "args": { + "External id": 7952,"Record function id": 0, "Concrete Inputs": ["", "[1, 64]"], "Input type": ["long int", "ScalarList"], "Input Dims": [[64], []], "Ev Idx": 7951 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791703564, "dur": 1, + "args": { + "External id": 7953,"Record function id": 0, "Concrete Inputs": ["", "[1, 64]"], "Input type": ["long int", "ScalarList"], "Input Dims": [[64], []], "Ev Idx": 7952 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791703845, "dur": 6, + "args": { + "External id": 7954,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 8192]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 64], []], "Ev Idx": 7953 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791703847, "dur": 3, + "args": { + "External id": 7955,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 8192]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 64], []], "Ev Idx": 7954 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791703855, "dur": 3, + "args": { + "External id": 7956,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 16, 512, 16]"], "Input type": ["int", "ScalarList"], "Input Dims": [[4096, 8192], []], "Ev Idx": 7955 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791703856, "dur": 1, + "args": { + "External id": 7957,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 16, 512, 16]"], "Input type": ["int", "ScalarList"], "Input Dims": [[4096, 8192], []], "Ev Idx": 7956 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 435, "tid": 435, + "ts": 1776649791703861, "dur": 5, + "args": { + "External id": 7958,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[0, 2, 1, 3]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 16, 512, 16], []], "Ev Idx": 7957 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791703864, "dur": 1, + "args": { + "External id": 7959,"Record function id": 0, "Concrete Inputs": ["", "[256, 512, 16, 16]", "[131072, 16, 8192, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Dims": [[256, 16, 512, 16], [], [], []], "Ev Idx": 7958 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791703870, "dur": 41, + "args": { + "External id": 7960,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 131072]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 512, 16, 16], []], "Ev Idx": 7959 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649791703873, "dur": 34, + "args": { + "External id": 7961,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[256, 512, 16, 16], []], "Ev Idx": 7960 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649791703876, "dur": 10, + "args": { + "External id": 7962,"Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "0"], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[256, 512, 16, 16], [], [], [], [], []], "Ev Idx": 7961 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791703877, "dur": 8, + "args": { + "External id": 7963,"Record function id": 0, "Concrete Inputs": ["[256, 512, 16, 16]", "3", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7962 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791703887, "dur": 20, + "args": { + "External id": 7964,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 512, 16, 16], [256, 512, 16, 16], []], "Ev Idx": 7963 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 435, "tid": 435, + "ts": 1776649791703909, "dur": 2, + "args": { + "External id": 7965,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 131072]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 512, 16, 16], []], "Ev Idx": 7964 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791703920, "dur": 3, + "args": { + "External id": 7966,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 131072], []], "Ev Idx": 7965 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791703921, "dur": 2, + "args": { + "External id": 7967,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 131072], []], "Ev Idx": 7966 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649791703929, "dur": 5, + "args": { + "External id": 7968,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[32768, 1024], [], [], [], []], "Ev Idx": 7967 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791703932, "dur": 1, + "args": { + "External id": 7969,"Record function id": 0, "Concrete Inputs": ["", "[32768, 1024]", "[1024, 1]", "0"], "Input type": ["int", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[32768, 1024], [], [], []], "Ev Idx": 7968 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::index", "pid": 435, "tid": 435, + "ts": 1776649791703937, "dur": 229, + "args": { + "External id": 7970,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["int", ""], "Input Dims": [[32768, 1024], []], "Ev Idx": 7969 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791703941, "dur": 194, + "args": { + "External id": 7971,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "False", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Dims": [[1024], [], [], [], [], [], [], []], "Ev Idx": 7970 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791703944, "dur": 190, + "args": { + "External id": 7972,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Dims": [[1024], [], [], [], [], [], []], "Ev Idx": 7971 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791703947, "dur": 6, + "args": { + "External id": 7973,"Record function id": 0, "Concrete Inputs": ["[1024]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7972 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791703954, "dur": 180, + "args": { + "External id": 7974,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Dims": [[1024], [1024], []], "Ev Idx": 7973 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791704137, "dur": 1, + "args": { + "External id": 7975,"Record function id": 0, "Concrete Inputs": ["", "[32768, 1024]", "[1024, 0]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Dims": [[32768, 1024], [], [], []], "Ev Idx": 7974 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791704139, "dur": 2, + "args": { + "External id": 7976,"Record function id": 0, "Concrete Inputs": ["", "[1, 1024]"], "Input type": ["long int", "ScalarList"], "Input Dims": [[1024], []], "Ev Idx": 7975 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791704141, "dur": 0, + "args": { + "External id": 7977,"Record function id": 0, "Concrete Inputs": ["", "[1, 1024]"], "Input type": ["long int", "ScalarList"], "Input Dims": [[1024], []], "Ev Idx": 7976 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791704177, "dur": 4, + "args": { + "External id": 7978,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 131072]"], "Input type": ["int", "ScalarList"], "Input Dims": [[32768, 1024], []], "Ev Idx": 7977 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791704178, "dur": 2, + "args": { + "External id": 7979,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 131072]"], "Input type": ["int", "ScalarList"], "Input Dims": [[32768, 1024], []], "Ev Idx": 7978 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791704802, "dur": 92752, + "args": { + "External id": 7980,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 131072], [], [], [], [], [], [], []], "Ev Idx": 7979 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791704804, "dur": 92747, + "args": { + "External id": 7981,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[256, 131072], [], [], [], [], [], []], "Ev Idx": 7980 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791704807, "dur": 19, + "args": { + "External id": 7982,"Record function id": 0, "Concrete Inputs": ["[256, 131072]", "[131072, 1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7981 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791704828, "dur": 92714, + "args": { + "External id": 7983,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 131072], [256, 131072], []], "Ev Idx": 7982 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 435, "tid": 435, + "ts": 1776649791797601, "dur": 22, + "args": { + "External id": 7984,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 131072]], "Ev Idx": 7983 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 435, "tid": 435, + "ts": 1776649791797606, "dur": 17, + "args": { + "External id": 7985,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 131072]], "Ev Idx": 7984 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791797627, "dur": 0, + "args": { + "External id": 7986,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 131072], [], [], [], [], [], [], []], "Ev Idx": 7985 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::resolve_conj", "pid": 435, "tid": 435, + "ts": 1776649791797631, "dur": 1, + "args": { + "External id": 7987,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 131072]], "Ev Idx": 7986 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::resolve_neg", "pid": 435, "tid": 435, + "ts": 1776649791797634, "dur": 0, + "args": { + "External id": 7988,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 131072]], "Ev Idx": 7987 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 435, "tid": 435, + "ts": 1776649791928920, "dur": 17, + "args": { + "External id": 7989,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 16384]], "Ev Idx": 7988 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791929013, "dur": 1615, + "args": { + "External id": 7990,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 16384], [], [], [], [], [], [], []], "Ev Idx": 7989 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791929019, "dur": 1608, + "args": { + "External id": 7991,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[256, 16384], [], [], [], [], [], []], "Ev Idx": 7990 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791929049, "dur": 99, + "args": { + "External id": 7992,"Record function id": 0, "Concrete Inputs": ["[256, 16384]", "[16384, 1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7991 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791929158, "dur": 1462, + "args": { + "External id": 7993,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 16384], [256, 16384], []], "Ev Idx": 7992 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791932345, "dur": 2, + "args": { + "External id": 7994,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 16384], [], [], [], [], [], [], []], "Ev Idx": 7993 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649791932411, "dur": 46, + "args": { + "External id": 7995,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[256, 16384], [], [], [], []], "Ev Idx": 7994 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791932440, "dur": 9, + "args": { + "External id": 7996,"Record function id": 0, "Concrete Inputs": ["", "[256, 16384]", "[16384, 1]", "0"], "Input type": ["int", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[256, 16384], [], [], []], "Ev Idx": 7995 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649791932459, "dur": 7, + "args": { + "External id": 7997,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[256, 16384], [], [], [], []], "Ev Idx": 7996 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791932464, "dur": 0, + "args": { + "External id": 7998,"Record function id": 0, "Concrete Inputs": ["", "[256, 16384]", "[16384, 1]", "0"], "Input type": ["int", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[256, 16384], [], [], []], "Ev Idx": 7997 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791932472, "dur": 88, + "args": { + "External id": 7999,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 16384], [256, 16384], []], "Ev Idx": 7998 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791932583, "dur": 0, + "args": { + "External id": 8000,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[32, 8192], [], [], [], [], [], [], []], "Ev Idx": 7999 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649791932590, "dur": 6, + "args": { + "External id": 8001,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[32, 8192], [], [], [], []], "Ev Idx": 8000 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791932593, "dur": 1, + "args": { + "External id": 8002,"Record function id": 0, "Concrete Inputs": ["", "[32, 8192]", "[8192, 1]", "0"], "Input type": ["c10::Half", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 8001 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649791932597, "dur": 3, + "args": { + "External id": 8003,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[32, 8192], [], [], [], []], "Ev Idx": 8002 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791932599, "dur": 1, + "args": { + "External id": 8004,"Record function id": 0, "Concrete Inputs": ["", "[32, 8192]", "[8192, 1]", "0"], "Input type": ["c10::Half", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 8003 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791932602, "dur": 21, + "args": { + "External id": 8005,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[32, 8192], [32, 8192], []], "Ev Idx": 8004 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791934394, "dur": 26, + "args": { + "External id": 8006,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8005 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791934436, "dur": 68, + "args": { + "External id": 8007,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8006 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791934440, "dur": 5, + "args": { + "External id": 8008,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8007 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791934450, "dur": 54, + "args": { + "External id": 8009,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8008 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791934458, "dur": 44, + "args": { + "External id": 8010,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8009 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791934548, "dur": 6, + "args": { + "External id": 8011,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8010 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791934559, "dur": 23, + "args": { + "External id": 8012,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8011 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791934560, "dur": 4, + "args": { + "External id": 8013,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8012 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791934566, "dur": 15, + "args": { + "External id": 8014,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8013 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791934568, "dur": 12, + "args": { + "External id": 8015,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8014 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791934600, "dur": 6, + "args": { + "External id": 8016,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8015 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791934610, "dur": 21, + "args": { + "External id": 8017,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8016 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791934611, "dur": 4, + "args": { + "External id": 8018,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8017 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791934616, "dur": 15, + "args": { + "External id": 8019,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8018 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791934618, "dur": 11, + "args": { + "External id": 8020,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8019 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791934649, "dur": 5, + "args": { + "External id": 8021,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8020 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791934658, "dur": 22, + "args": { + "External id": 8022,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8021 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791934659, "dur": 5, + "args": { + "External id": 8023,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8022 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791934665, "dur": 15, + "args": { + "External id": 8024,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8023 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791934667, "dur": 12, + "args": { + "External id": 8025,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8024 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791934698, "dur": 6, + "args": { + "External id": 8026,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8025 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791934708, "dur": 21, + "args": { + "External id": 8027,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8026 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791934709, "dur": 4, + "args": { + "External id": 8028,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8027 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791934715, "dur": 14, + "args": { + "External id": 8029,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8028 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791934716, "dur": 11, + "args": { + "External id": 8030,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8029 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791934747, "dur": 5, + "args": { + "External id": 8031,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8030 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791934757, "dur": 21, + "args": { + "External id": 8032,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8031 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791934758, "dur": 4, + "args": { + "External id": 8033,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8032 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791934764, "dur": 14, + "args": { + "External id": 8034,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8033 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791934765, "dur": 12, + "args": { + "External id": 8035,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8034 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791934797, "dur": 5, + "args": { + "External id": 8036,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8035 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791934806, "dur": 21, + "args": { + "External id": 8037,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8036 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791934807, "dur": 4, + "args": { + "External id": 8038,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8037 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791934812, "dur": 14, + "args": { + "External id": 8039,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8038 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791934814, "dur": 11, + "args": { + "External id": 8040,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8039 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791934845, "dur": 5, + "args": { + "External id": 8041,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8040 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791934854, "dur": 21, + "args": { + "External id": 8042,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8041 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791934855, "dur": 5, + "args": { + "External id": 8043,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8042 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791934861, "dur": 14, + "args": { + "External id": 8044,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8043 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791934863, "dur": 10, + "args": { + "External id": 8045,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8044 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791934892, "dur": 6, + "args": { + "External id": 8046,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8045 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791934901, "dur": 25, + "args": { + "External id": 8047,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8046 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791934905, "dur": 5, + "args": { + "External id": 8048,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8047 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791934911, "dur": 14, + "args": { + "External id": 8049,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8048 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791934913, "dur": 11, + "args": { + "External id": 8050,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8049 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791934943, "dur": 6, + "args": { + "External id": 8051,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8050 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791934952, "dur": 22, + "args": { + "External id": 8052,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8051 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791934953, "dur": 5, + "args": { + "External id": 8053,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8052 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791934959, "dur": 14, + "args": { + "External id": 8054,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8053 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791934961, "dur": 11, + "args": { + "External id": 8055,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8054 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935253, "dur": 8, + "args": { + "External id": 8056,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8055 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791935267, "dur": 27, + "args": { + "External id": 8057,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8056 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935268, "dur": 5, + "args": { + "External id": 8058,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8057 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791935274, "dur": 20, + "args": { + "External id": 8059,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8058 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791935277, "dur": 16, + "args": { + "External id": 8060,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8059 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935317, "dur": 5, + "args": { + "External id": 8061,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8060 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791935326, "dur": 23, + "args": { + "External id": 8062,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8061 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935328, "dur": 4, + "args": { + "External id": 8063,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8062 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791935333, "dur": 15, + "args": { + "External id": 8064,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8063 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791935335, "dur": 12, + "args": { + "External id": 8065,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8064 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935367, "dur": 5, + "args": { + "External id": 8066,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8065 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791935376, "dur": 22, + "args": { + "External id": 8067,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8066 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935377, "dur": 5, + "args": { + "External id": 8068,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8067 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791935383, "dur": 14, + "args": { + "External id": 8069,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8068 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791935385, "dur": 11, + "args": { + "External id": 8070,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8069 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935416, "dur": 5, + "args": { + "External id": 8071,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8070 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791935425, "dur": 21, + "args": { + "External id": 8072,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8071 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935426, "dur": 5, + "args": { + "External id": 8073,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8072 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791935432, "dur": 14, + "args": { + "External id": 8074,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8073 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791935434, "dur": 11, + "args": { + "External id": 8075,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8074 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935464, "dur": 5, + "args": { + "External id": 8076,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8075 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791935473, "dur": 21, + "args": { + "External id": 8077,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8076 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935474, "dur": 5, + "args": { + "External id": 8078,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8077 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791935480, "dur": 14, + "args": { + "External id": 8079,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8078 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791935482, "dur": 11, + "args": { + "External id": 8080,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8079 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935512, "dur": 5, + "args": { + "External id": 8081,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8080 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791935521, "dur": 21, + "args": { + "External id": 8082,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8081 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935522, "dur": 4, + "args": { + "External id": 8083,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8082 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791935528, "dur": 13, + "args": { + "External id": 8084,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8083 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791935529, "dur": 11, + "args": { + "External id": 8085,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8084 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935559, "dur": 5, + "args": { + "External id": 8086,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8085 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791935568, "dur": 21, + "args": { + "External id": 8087,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8086 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935570, "dur": 4, + "args": { + "External id": 8088,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8087 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791935575, "dur": 14, + "args": { + "External id": 8089,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8088 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791935577, "dur": 11, + "args": { + "External id": 8090,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8089 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935607, "dur": 5, + "args": { + "External id": 8091,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8090 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791935616, "dur": 21, + "args": { + "External id": 8092,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8091 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935617, "dur": 4, + "args": { + "External id": 8093,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8092 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791935623, "dur": 13, + "args": { + "External id": 8094,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8093 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791935624, "dur": 11, + "args": { + "External id": 8095,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8094 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935656, "dur": 5, + "args": { + "External id": 8096,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8095 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791935665, "dur": 21, + "args": { + "External id": 8097,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8096 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935666, "dur": 5, + "args": { + "External id": 8098,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8097 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791935672, "dur": 14, + "args": { + "External id": 8099,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8098 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791935674, "dur": 11, + "args": { + "External id": 8100,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8099 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935704, "dur": 5, + "args": { + "External id": 8101,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8100 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791935713, "dur": 22, + "args": { + "External id": 8102,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8101 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935714, "dur": 5, + "args": { + "External id": 8103,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8102 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791935720, "dur": 14, + "args": { + "External id": 8104,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8103 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791935722, "dur": 11, + "args": { + "External id": 8105,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8104 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935752, "dur": 5, + "args": { + "External id": 8106,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8105 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791935761, "dur": 21, + "args": { + "External id": 8107,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8106 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935762, "dur": 5, + "args": { + "External id": 8108,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8107 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791935768, "dur": 14, + "args": { + "External id": 8109,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8108 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791935770, "dur": 11, + "args": { + "External id": 8110,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8109 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935800, "dur": 5, + "args": { + "External id": 8111,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8110 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791935810, "dur": 21, + "args": { + "External id": 8112,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8111 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935811, "dur": 4, + "args": { + "External id": 8113,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8112 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791935817, "dur": 13, + "args": { + "External id": 8114,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8113 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791935818, "dur": 11, + "args": { + "External id": 8115,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8114 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935848, "dur": 6, + "args": { + "External id": 8116,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8115 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791935857, "dur": 22, + "args": { + "External id": 8117,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8116 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935859, "dur": 4, + "args": { + "External id": 8118,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8117 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791935864, "dur": 14, + "args": { + "External id": 8119,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8118 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791935866, "dur": 11, + "args": { + "External id": 8120,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8119 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935897, "dur": 5, + "args": { + "External id": 8121,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8120 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791935906, "dur": 21, + "args": { + "External id": 8122,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8121 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935907, "dur": 4, + "args": { + "External id": 8123,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8122 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791935912, "dur": 14, + "args": { + "External id": 8124,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8123 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791935914, "dur": 11, + "args": { + "External id": 8125,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8124 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935944, "dur": 5, + "args": { + "External id": 8126,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8125 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791935953, "dur": 22, + "args": { + "External id": 8127,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8126 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935955, "dur": 4, + "args": { + "External id": 8128,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8127 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791935960, "dur": 14, + "args": { + "External id": 8129,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8128 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791935962, "dur": 11, + "args": { + "External id": 8130,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8129 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791935992, "dur": 5, + "args": { + "External id": 8131,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8130 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791936001, "dur": 21, + "args": { + "External id": 8132,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8131 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936002, "dur": 4, + "args": { + "External id": 8133,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8132 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791936008, "dur": 13, + "args": { + "External id": 8134,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8133 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791936009, "dur": 11, + "args": { + "External id": 8135,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8134 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936039, "dur": 5, + "args": { + "External id": 8136,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8135 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791936048, "dur": 21, + "args": { + "External id": 8137,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8136 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936049, "dur": 5, + "args": { + "External id": 8138,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8137 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791936055, "dur": 14, + "args": { + "External id": 8139,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8138 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791936057, "dur": 10, + "args": { + "External id": 8140,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8139 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936086, "dur": 6, + "args": { + "External id": 8141,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8140 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791936095, "dur": 22, + "args": { + "External id": 8142,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8141 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936097, "dur": 4, + "args": { + "External id": 8143,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8142 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791936102, "dur": 14, + "args": { + "External id": 8144,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8143 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791936104, "dur": 11, + "args": { + "External id": 8145,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8144 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936135, "dur": 5, + "args": { + "External id": 8146,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8145 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791936144, "dur": 20, + "args": { + "External id": 8147,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8146 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936145, "dur": 4, + "args": { + "External id": 8148,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8147 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791936150, "dur": 14, + "args": { + "External id": 8149,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8148 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791936152, "dur": 11, + "args": { + "External id": 8150,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8149 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936182, "dur": 5, + "args": { + "External id": 8151,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8150 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791936192, "dur": 21, + "args": { + "External id": 8152,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8151 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936193, "dur": 5, + "args": { + "External id": 8153,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8152 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791936199, "dur": 14, + "args": { + "External id": 8154,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8153 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791936201, "dur": 11, + "args": { + "External id": 8155,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8154 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936257, "dur": 5, + "args": { + "External id": 8156,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8155 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791936266, "dur": 22, + "args": { + "External id": 8157,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8156 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936268, "dur": 4, + "args": { + "External id": 8158,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8157 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791936273, "dur": 15, + "args": { + "External id": 8159,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8158 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791936275, "dur": 12, + "args": { + "External id": 8160,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8159 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936307, "dur": 5, + "args": { + "External id": 8161,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8160 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791936316, "dur": 21, + "args": { + "External id": 8162,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8161 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936317, "dur": 5, + "args": { + "External id": 8163,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8162 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791936323, "dur": 14, + "args": { + "External id": 8164,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8163 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791936325, "dur": 10, + "args": { + "External id": 8165,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8164 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936355, "dur": 5, + "args": { + "External id": 8166,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8165 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791936364, "dur": 21, + "args": { + "External id": 8167,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8166 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936365, "dur": 5, + "args": { + "External id": 8168,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8167 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791936371, "dur": 14, + "args": { + "External id": 8169,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8168 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791936373, "dur": 10, + "args": { + "External id": 8170,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8169 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936403, "dur": 5, + "args": { + "External id": 8171,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8170 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791936411, "dur": 22, + "args": { + "External id": 8172,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8171 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936413, "dur": 4, + "args": { + "External id": 8173,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8172 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791936419, "dur": 13, + "args": { + "External id": 8174,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8173 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791936420, "dur": 11, + "args": { + "External id": 8175,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8174 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936451, "dur": 5, + "args": { + "External id": 8176,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8175 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791936460, "dur": 21, + "args": { + "External id": 8177,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8176 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936461, "dur": 5, + "args": { + "External id": 8178,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8177 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791936467, "dur": 14, + "args": { + "External id": 8179,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8178 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791936469, "dur": 10, + "args": { + "External id": 8180,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8179 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936499, "dur": 5, + "args": { + "External id": 8181,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8180 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791936508, "dur": 21, + "args": { + "External id": 8182,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8181 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936509, "dur": 4, + "args": { + "External id": 8183,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8182 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791936514, "dur": 15, + "args": { + "External id": 8184,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8183 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791936516, "dur": 12, + "args": { + "External id": 8185,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8184 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936547, "dur": 5, + "args": { + "External id": 8186,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8185 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791936555, "dur": 21, + "args": { + "External id": 8187,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8186 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936557, "dur": 4, + "args": { + "External id": 8188,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8187 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791936562, "dur": 14, + "args": { + "External id": 8189,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8188 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791936564, "dur": 11, + "args": { + "External id": 8190,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8189 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936594, "dur": 5, + "args": { + "External id": 8191,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8190 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791936604, "dur": 31, + "args": { + "External id": 8192,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8191 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936614, "dur": 5, + "args": { + "External id": 8193,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8192 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791936620, "dur": 14, + "args": { + "External id": 8194,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8193 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791936622, "dur": 11, + "args": { + "External id": 8195,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8194 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936652, "dur": 5, + "args": { + "External id": 8196,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8195 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791936661, "dur": 21, + "args": { + "External id": 8197,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8196 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936662, "dur": 5, + "args": { + "External id": 8198,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8197 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791936668, "dur": 13, + "args": { + "External id": 8199,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8198 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791936670, "dur": 10, + "args": { + "External id": 8200,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8199 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936700, "dur": 5, + "args": { + "External id": 8201,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8200 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791936709, "dur": 21, + "args": { + "External id": 8202,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8201 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936710, "dur": 4, + "args": { + "External id": 8203,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8202 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791936716, "dur": 13, + "args": { + "External id": 8204,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8203 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791936717, "dur": 11, + "args": { + "External id": 8205,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8204 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936747, "dur": 5, + "args": { + "External id": 8206,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8205 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791936756, "dur": 21, + "args": { + "External id": 8207,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8206 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936758, "dur": 4, + "args": { + "External id": 8208,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8207 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791936763, "dur": 14, + "args": { + "External id": 8209,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8208 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791936765, "dur": 11, + "args": { + "External id": 8210,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8209 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936795, "dur": 5, + "args": { + "External id": 8211,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8210 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791936804, "dur": 20, + "args": { + "External id": 8212,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8211 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936805, "dur": 4, + "args": { + "External id": 8213,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8212 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791936810, "dur": 14, + "args": { + "External id": 8214,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8213 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791936812, "dur": 11, + "args": { + "External id": 8215,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8214 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936842, "dur": 5, + "args": { + "External id": 8216,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8215 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791936851, "dur": 21, + "args": { + "External id": 8217,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8216 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936852, "dur": 5, + "args": { + "External id": 8218,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8217 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791936858, "dur": 14, + "args": { + "External id": 8219,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8218 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791936860, "dur": 11, + "args": { + "External id": 8220,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8219 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936890, "dur": 5, + "args": { + "External id": 8221,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8220 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791936898, "dur": 23, + "args": { + "External id": 8222,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8221 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936900, "dur": 4, + "args": { + "External id": 8223,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8222 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791936905, "dur": 15, + "args": { + "External id": 8224,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8223 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791936907, "dur": 12, + "args": { + "External id": 8225,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8224 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936939, "dur": 5, + "args": { + "External id": 8226,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8225 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791936947, "dur": 22, + "args": { + "External id": 8227,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8226 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936949, "dur": 4, + "args": { + "External id": 8228,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8227 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791936954, "dur": 14, + "args": { + "External id": 8229,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8228 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791936956, "dur": 11, + "args": { + "External id": 8230,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8229 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936987, "dur": 5, + "args": { + "External id": 8231,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8230 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791936996, "dur": 21, + "args": { + "External id": 8232,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8231 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791936997, "dur": 5, + "args": { + "External id": 8233,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8232 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791937003, "dur": 14, + "args": { + "External id": 8234,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8233 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791937004, "dur": 12, + "args": { + "External id": 8235,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8234 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937036, "dur": 5, + "args": { + "External id": 8236,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8235 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791937045, "dur": 21, + "args": { + "External id": 8237,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8236 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937046, "dur": 4, + "args": { + "External id": 8238,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8237 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791937052, "dur": 14, + "args": { + "External id": 8239,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8238 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791937053, "dur": 12, + "args": { + "External id": 8240,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8239 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937084, "dur": 6, + "args": { + "External id": 8241,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8240 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791937093, "dur": 28, + "args": { + "External id": 8242,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8241 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937095, "dur": 10, + "args": { + "External id": 8243,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8242 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791937106, "dur": 15, + "args": { + "External id": 8244,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8243 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791937108, "dur": 12, + "args": { + "External id": 8245,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8244 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937141, "dur": 5, + "args": { + "External id": 8246,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8245 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791937150, "dur": 21, + "args": { + "External id": 8247,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8246 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937151, "dur": 4, + "args": { + "External id": 8248,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8247 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791937157, "dur": 14, + "args": { + "External id": 8249,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8248 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791937158, "dur": 11, + "args": { + "External id": 8250,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8249 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937189, "dur": 5, + "args": { + "External id": 8251,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8250 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791937198, "dur": 21, + "args": { + "External id": 8252,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8251 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937199, "dur": 4, + "args": { + "External id": 8253,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8252 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791937205, "dur": 13, + "args": { + "External id": 8254,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8253 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791937206, "dur": 11, + "args": { + "External id": 8255,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8254 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937237, "dur": 5, + "args": { + "External id": 8256,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8255 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791937246, "dur": 21, + "args": { + "External id": 8257,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8256 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937247, "dur": 5, + "args": { + "External id": 8258,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8257 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791937253, "dur": 14, + "args": { + "External id": 8259,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8258 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791937255, "dur": 10, + "args": { + "External id": 8260,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8259 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937285, "dur": 5, + "args": { + "External id": 8261,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8260 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791937294, "dur": 21, + "args": { + "External id": 8262,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8261 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937295, "dur": 4, + "args": { + "External id": 8263,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8262 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791937300, "dur": 15, + "args": { + "External id": 8264,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8263 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791937302, "dur": 12, + "args": { + "External id": 8265,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8264 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937334, "dur": 5, + "args": { + "External id": 8266,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8265 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791937342, "dur": 22, + "args": { + "External id": 8267,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8266 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937344, "dur": 4, + "args": { + "External id": 8268,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8267 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791937349, "dur": 14, + "args": { + "External id": 8269,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8268 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791937351, "dur": 11, + "args": { + "External id": 8270,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8269 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937381, "dur": 5, + "args": { + "External id": 8271,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8270 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791937391, "dur": 21, + "args": { + "External id": 8272,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8271 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937392, "dur": 4, + "args": { + "External id": 8273,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8272 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791937398, "dur": 14, + "args": { + "External id": 8274,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8273 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791937399, "dur": 11, + "args": { + "External id": 8275,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8274 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937430, "dur": 5, + "args": { + "External id": 8276,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8275 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791937439, "dur": 21, + "args": { + "External id": 8277,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8276 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937440, "dur": 5, + "args": { + "External id": 8278,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8277 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791937446, "dur": 14, + "args": { + "External id": 8279,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8278 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791937447, "dur": 12, + "args": { + "External id": 8280,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8279 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937479, "dur": 5, + "args": { + "External id": 8281,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8280 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791937487, "dur": 22, + "args": { + "External id": 8282,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8281 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937489, "dur": 4, + "args": { + "External id": 8283,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8282 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791937494, "dur": 14, + "args": { + "External id": 8284,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8283 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791937496, "dur": 11, + "args": { + "External id": 8285,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8284 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937526, "dur": 5, + "args": { + "External id": 8286,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8285 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791937535, "dur": 23, + "args": { + "External id": 8287,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8286 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937537, "dur": 5, + "args": { + "External id": 8288,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8287 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791937543, "dur": 14, + "args": { + "External id": 8289,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8288 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791937545, "dur": 11, + "args": { + "External id": 8290,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8289 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937575, "dur": 5, + "args": { + "External id": 8291,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8290 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791937584, "dur": 21, + "args": { + "External id": 8292,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8291 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937585, "dur": 5, + "args": { + "External id": 8293,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8292 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791937591, "dur": 14, + "args": { + "External id": 8294,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8293 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791937593, "dur": 11, + "args": { + "External id": 8295,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8294 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937624, "dur": 5, + "args": { + "External id": 8296,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8295 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791937633, "dur": 21, + "args": { + "External id": 8297,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8296 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937634, "dur": 5, + "args": { + "External id": 8298,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8297 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791937640, "dur": 14, + "args": { + "External id": 8299,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8298 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791937642, "dur": 11, + "args": { + "External id": 8300,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8299 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937672, "dur": 5, + "args": { + "External id": 8301,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8300 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791937681, "dur": 22, + "args": { + "External id": 8302,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8301 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937682, "dur": 5, + "args": { + "External id": 8303,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8302 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791937688, "dur": 15, + "args": { + "External id": 8304,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8303 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791937690, "dur": 11, + "args": { + "External id": 8305,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8304 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937790, "dur": 6, + "args": { + "External id": 8306,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8305 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791937801, "dur": 25, + "args": { + "External id": 8307,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8306 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937802, "dur": 5, + "args": { + "External id": 8308,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8307 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791937808, "dur": 18, + "args": { + "External id": 8309,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8308 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791937810, "dur": 14, + "args": { + "External id": 8310,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8309 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937847, "dur": 6, + "args": { + "External id": 8311,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8310 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791937857, "dur": 22, + "args": { + "External id": 8312,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8311 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937859, "dur": 4, + "args": { + "External id": 8313,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8312 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791937864, "dur": 15, + "args": { + "External id": 8314,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8313 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791937866, "dur": 12, + "args": { + "External id": 8315,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8314 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937898, "dur": 6, + "args": { + "External id": 8316,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8315 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791937908, "dur": 21, + "args": { + "External id": 8317,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8316 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937909, "dur": 5, + "args": { + "External id": 8318,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8317 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791937915, "dur": 14, + "args": { + "External id": 8319,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8318 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791937917, "dur": 11, + "args": { + "External id": 8320,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8319 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937949, "dur": 5, + "args": { + "External id": 8321,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8320 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791937957, "dur": 22, + "args": { + "External id": 8322,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8321 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937959, "dur": 4, + "args": { + "External id": 8323,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8322 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791937964, "dur": 15, + "args": { + "External id": 8324,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8323 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791937966, "dur": 11, + "args": { + "External id": 8325,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8324 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791937997, "dur": 5, + "args": { + "External id": 8326,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8325 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791938006, "dur": 22, + "args": { + "External id": 8327,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8326 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938008, "dur": 5, + "args": { + "External id": 8328,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8327 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791938014, "dur": 14, + "args": { + "External id": 8329,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8328 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791938015, "dur": 12, + "args": { + "External id": 8330,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8329 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938047, "dur": 5, + "args": { + "External id": 8331,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8330 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791938056, "dur": 21, + "args": { + "External id": 8332,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8331 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938057, "dur": 5, + "args": { + "External id": 8333,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8332 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791938063, "dur": 14, + "args": { + "External id": 8334,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8333 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791938064, "dur": 12, + "args": { + "External id": 8335,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8334 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938096, "dur": 5, + "args": { + "External id": 8336,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8335 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791938105, "dur": 21, + "args": { + "External id": 8337,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8336 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938106, "dur": 4, + "args": { + "External id": 8338,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8337 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791938112, "dur": 14, + "args": { + "External id": 8339,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8338 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791938113, "dur": 11, + "args": { + "External id": 8340,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8339 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938144, "dur": 5, + "args": { + "External id": 8341,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8340 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791938153, "dur": 21, + "args": { + "External id": 8342,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8341 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938154, "dur": 4, + "args": { + "External id": 8343,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8342 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791938160, "dur": 14, + "args": { + "External id": 8344,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8343 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791938161, "dur": 12, + "args": { + "External id": 8345,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8344 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938192, "dur": 5, + "args": { + "External id": 8346,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8345 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791938201, "dur": 21, + "args": { + "External id": 8347,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8346 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938202, "dur": 5, + "args": { + "External id": 8348,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8347 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791938208, "dur": 14, + "args": { + "External id": 8349,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8348 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791938209, "dur": 11, + "args": { + "External id": 8350,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8349 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938240, "dur": 5, + "args": { + "External id": 8351,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8350 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791938249, "dur": 21, + "args": { + "External id": 8352,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8351 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938250, "dur": 5, + "args": { + "External id": 8353,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8352 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791938256, "dur": 14, + "args": { + "External id": 8354,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8353 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791938257, "dur": 11, + "args": { + "External id": 8355,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8354 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938342, "dur": 6, + "args": { + "External id": 8356,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8355 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791938353, "dur": 25, + "args": { + "External id": 8357,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8356 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938355, "dur": 4, + "args": { + "External id": 8358,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8357 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791938360, "dur": 18, + "args": { + "External id": 8359,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8358 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791938362, "dur": 14, + "args": { + "External id": 8360,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8359 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938400, "dur": 5, + "args": { + "External id": 8361,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8360 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791938409, "dur": 21, + "args": { + "External id": 8362,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8361 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938410, "dur": 4, + "args": { + "External id": 8363,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8362 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791938415, "dur": 15, + "args": { + "External id": 8364,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8363 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791938417, "dur": 11, + "args": { + "External id": 8365,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8364 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938449, "dur": 5, + "args": { + "External id": 8366,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8365 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791938458, "dur": 22, + "args": { + "External id": 8367,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8366 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938460, "dur": 4, + "args": { + "External id": 8368,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8367 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791938466, "dur": 14, + "args": { + "External id": 8369,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8368 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791938467, "dur": 11, + "args": { + "External id": 8370,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8369 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938498, "dur": 5, + "args": { + "External id": 8371,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8370 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791938507, "dur": 21, + "args": { + "External id": 8372,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8371 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938508, "dur": 5, + "args": { + "External id": 8373,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8372 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791938514, "dur": 14, + "args": { + "External id": 8374,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8373 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791938515, "dur": 11, + "args": { + "External id": 8375,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8374 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938547, "dur": 5, + "args": { + "External id": 8376,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8375 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791938556, "dur": 21, + "args": { + "External id": 8377,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8376 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938557, "dur": 4, + "args": { + "External id": 8378,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8377 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791938563, "dur": 14, + "args": { + "External id": 8379,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8378 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791938564, "dur": 11, + "args": { + "External id": 8380,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8379 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938595, "dur": 5, + "args": { + "External id": 8381,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8380 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791938604, "dur": 23, + "args": { + "External id": 8382,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8381 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938606, "dur": 4, + "args": { + "External id": 8383,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8382 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791938611, "dur": 15, + "args": { + "External id": 8384,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8383 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791938613, "dur": 12, + "args": { + "External id": 8385,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8384 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938645, "dur": 5, + "args": { + "External id": 8386,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8385 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791938654, "dur": 21, + "args": { + "External id": 8387,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8386 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938655, "dur": 5, + "args": { + "External id": 8388,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8387 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791938661, "dur": 14, + "args": { + "External id": 8389,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8388 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791938662, "dur": 11, + "args": { + "External id": 8390,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8389 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938693, "dur": 5, + "args": { + "External id": 8391,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8390 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791938703, "dur": 21, + "args": { + "External id": 8392,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8391 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938704, "dur": 4, + "args": { + "External id": 8393,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8392 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791938710, "dur": 14, + "args": { + "External id": 8394,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8393 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791938711, "dur": 11, + "args": { + "External id": 8395,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8394 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938742, "dur": 5, + "args": { + "External id": 8396,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8395 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791938751, "dur": 21, + "args": { + "External id": 8397,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8396 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938753, "dur": 4, + "args": { + "External id": 8398,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8397 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791938758, "dur": 14, + "args": { + "External id": 8399,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8398 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791938760, "dur": 11, + "args": { + "External id": 8400,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8399 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938791, "dur": 5, + "args": { + "External id": 8401,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8400 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791938800, "dur": 21, + "args": { + "External id": 8402,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8401 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938801, "dur": 5, + "args": { + "External id": 8403,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8402 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791938807, "dur": 14, + "args": { + "External id": 8404,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8403 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791938808, "dur": 12, + "args": { + "External id": 8405,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8404 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938839, "dur": 5, + "args": { + "External id": 8406,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8405 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791938848, "dur": 21, + "args": { + "External id": 8407,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8406 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938849, "dur": 5, + "args": { + "External id": 8408,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8407 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791938855, "dur": 14, + "args": { + "External id": 8409,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8408 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791938857, "dur": 11, + "args": { + "External id": 8410,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8409 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938887, "dur": 5, + "args": { + "External id": 8411,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8410 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791938896, "dur": 21, + "args": { + "External id": 8412,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8411 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938897, "dur": 5, + "args": { + "External id": 8413,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8412 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791938903, "dur": 14, + "args": { + "External id": 8414,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8413 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791938904, "dur": 12, + "args": { + "External id": 8415,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8414 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938936, "dur": 5, + "args": { + "External id": 8416,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8415 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791938945, "dur": 21, + "args": { + "External id": 8417,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8416 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938946, "dur": 4, + "args": { + "External id": 8418,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8417 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791938951, "dur": 14, + "args": { + "External id": 8419,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8418 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791938953, "dur": 11, + "args": { + "External id": 8420,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8419 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938983, "dur": 6, + "args": { + "External id": 8421,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8420 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791938992, "dur": 22, + "args": { + "External id": 8422,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8421 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791938993, "dur": 5, + "args": { + "External id": 8423,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8422 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791938999, "dur": 15, + "args": { + "External id": 8424,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8423 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791939001, "dur": 12, + "args": { + "External id": 8425,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8424 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939032, "dur": 6, + "args": { + "External id": 8426,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8425 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791939041, "dur": 22, + "args": { + "External id": 8427,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8426 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939043, "dur": 4, + "args": { + "External id": 8428,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8427 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791939048, "dur": 14, + "args": { + "External id": 8429,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8428 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791939050, "dur": 11, + "args": { + "External id": 8430,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8429 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939081, "dur": 5, + "args": { + "External id": 8431,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8430 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791939090, "dur": 22, + "args": { + "External id": 8432,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8431 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939091, "dur": 5, + "args": { + "External id": 8433,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8432 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791939097, "dur": 14, + "args": { + "External id": 8434,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8433 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791939099, "dur": 11, + "args": { + "External id": 8435,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8434 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939129, "dur": 6, + "args": { + "External id": 8436,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8435 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791939139, "dur": 20, + "args": { + "External id": 8437,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8436 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939140, "dur": 4, + "args": { + "External id": 8438,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8437 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791939145, "dur": 14, + "args": { + "External id": 8439,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8438 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791939147, "dur": 11, + "args": { + "External id": 8440,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8439 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939178, "dur": 5, + "args": { + "External id": 8441,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8440 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791939187, "dur": 21, + "args": { + "External id": 8442,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8441 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939188, "dur": 4, + "args": { + "External id": 8443,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8442 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791939193, "dur": 15, + "args": { + "External id": 8444,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8443 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791939195, "dur": 11, + "args": { + "External id": 8445,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8444 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939226, "dur": 5, + "args": { + "External id": 8446,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8445 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791939235, "dur": 22, + "args": { + "External id": 8447,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8446 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939237, "dur": 4, + "args": { + "External id": 8448,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8447 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791939242, "dur": 14, + "args": { + "External id": 8449,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8448 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791939244, "dur": 11, + "args": { + "External id": 8450,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8449 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939274, "dur": 6, + "args": { + "External id": 8451,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8450 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791939283, "dur": 22, + "args": { + "External id": 8452,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8451 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939285, "dur": 4, + "args": { + "External id": 8453,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8452 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791939290, "dur": 14, + "args": { + "External id": 8454,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8453 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791939292, "dur": 11, + "args": { + "External id": 8455,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8454 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939323, "dur": 5, + "args": { + "External id": 8456,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8455 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791939332, "dur": 21, + "args": { + "External id": 8457,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8456 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939333, "dur": 5, + "args": { + "External id": 8458,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8457 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791939339, "dur": 14, + "args": { + "External id": 8459,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8458 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791939341, "dur": 11, + "args": { + "External id": 8460,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8459 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939370, "dur": 6, + "args": { + "External id": 8461,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8460 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791939379, "dur": 22, + "args": { + "External id": 8462,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8461 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939381, "dur": 4, + "args": { + "External id": 8463,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8462 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791939386, "dur": 14, + "args": { + "External id": 8464,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8463 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791939388, "dur": 11, + "args": { + "External id": 8465,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8464 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939418, "dur": 5, + "args": { + "External id": 8466,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8465 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791939427, "dur": 21, + "args": { + "External id": 8467,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8466 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939428, "dur": 5, + "args": { + "External id": 8468,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8467 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791939434, "dur": 14, + "args": { + "External id": 8469,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8468 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791939436, "dur": 11, + "args": { + "External id": 8470,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8469 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939466, "dur": 5, + "args": { + "External id": 8471,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8470 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791939476, "dur": 21, + "args": { + "External id": 8472,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8471 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939477, "dur": 4, + "args": { + "External id": 8473,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8472 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791939483, "dur": 13, + "args": { + "External id": 8474,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8473 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791939484, "dur": 11, + "args": { + "External id": 8475,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8474 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939514, "dur": 5, + "args": { + "External id": 8476,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8475 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791939523, "dur": 21, + "args": { + "External id": 8477,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8476 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939524, "dur": 5, + "args": { + "External id": 8478,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8477 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791939530, "dur": 14, + "args": { + "External id": 8479,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8478 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791939531, "dur": 11, + "args": { + "External id": 8480,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8479 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939563, "dur": 5, + "args": { + "External id": 8481,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8480 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791939572, "dur": 21, + "args": { + "External id": 8482,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8481 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939573, "dur": 5, + "args": { + "External id": 8483,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8482 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791939579, "dur": 14, + "args": { + "External id": 8484,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8483 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791939580, "dur": 11, + "args": { + "External id": 8485,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8484 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939611, "dur": 5, + "args": { + "External id": 8486,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8485 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791939620, "dur": 22, + "args": { + "External id": 8487,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8486 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939623, "dur": 4, + "args": { + "External id": 8488,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8487 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791939628, "dur": 14, + "args": { + "External id": 8489,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8488 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791939630, "dur": 11, + "args": { + "External id": 8490,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8489 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939660, "dur": 5, + "args": { + "External id": 8491,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8490 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791939669, "dur": 21, + "args": { + "External id": 8492,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8491 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939670, "dur": 5, + "args": { + "External id": 8493,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8492 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791939676, "dur": 14, + "args": { + "External id": 8494,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8493 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791939678, "dur": 11, + "args": { + "External id": 8495,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8494 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939709, "dur": 5, + "args": { + "External id": 8496,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8495 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791939718, "dur": 21, + "args": { + "External id": 8497,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8496 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939719, "dur": 5, + "args": { + "External id": 8498,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8497 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791939725, "dur": 14, + "args": { + "External id": 8499,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8498 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791939727, "dur": 11, + "args": { + "External id": 8500,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8499 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939757, "dur": 5, + "args": { + "External id": 8501,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8500 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791939766, "dur": 22, + "args": { + "External id": 8502,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8501 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939767, "dur": 4, + "args": { + "External id": 8503,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8502 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791939773, "dur": 14, + "args": { + "External id": 8504,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8503 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791939774, "dur": 12, + "args": { + "External id": 8505,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8504 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939806, "dur": 5, + "args": { + "External id": 8506,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8505 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791939814, "dur": 22, + "args": { + "External id": 8507,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8506 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939816, "dur": 4, + "args": { + "External id": 8508,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8507 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791939821, "dur": 14, + "args": { + "External id": 8509,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8508 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791939823, "dur": 11, + "args": { + "External id": 8510,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8509 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939853, "dur": 6, + "args": { + "External id": 8511,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8510 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791939864, "dur": 21, + "args": { + "External id": 8512,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8511 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939865, "dur": 4, + "args": { + "External id": 8513,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8512 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791939870, "dur": 14, + "args": { + "External id": 8514,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8513 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791939872, "dur": 11, + "args": { + "External id": 8515,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8514 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939902, "dur": 5, + "args": { + "External id": 8516,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8515 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791939911, "dur": 21, + "args": { + "External id": 8517,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8516 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939913, "dur": 4, + "args": { + "External id": 8518,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8517 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791939918, "dur": 14, + "args": { + "External id": 8519,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8518 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791939920, "dur": 11, + "args": { + "External id": 8520,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8519 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939951, "dur": 5, + "args": { + "External id": 8521,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8520 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791939960, "dur": 21, + "args": { + "External id": 8522,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8521 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939961, "dur": 4, + "args": { + "External id": 8523,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8522 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791939967, "dur": 14, + "args": { + "External id": 8524,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8523 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791939968, "dur": 11, + "args": { + "External id": 8525,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8524 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791939999, "dur": 5, + "args": { + "External id": 8526,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8525 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791940008, "dur": 22, + "args": { + "External id": 8527,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8526 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940010, "dur": 4, + "args": { + "External id": 8528,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8527 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791940015, "dur": 15, + "args": { + "External id": 8529,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8528 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791940017, "dur": 11, + "args": { + "External id": 8530,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8529 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940048, "dur": 5, + "args": { + "External id": 8531,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8530 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791940057, "dur": 21, + "args": { + "External id": 8532,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8531 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940058, "dur": 4, + "args": { + "External id": 8533,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8532 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791940063, "dur": 14, + "args": { + "External id": 8534,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8533 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791940065, "dur": 11, + "args": { + "External id": 8535,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8534 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940096, "dur": 5, + "args": { + "External id": 8536,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8535 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791940105, "dur": 21, + "args": { + "External id": 8537,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8536 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940106, "dur": 5, + "args": { + "External id": 8538,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8537 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791940112, "dur": 14, + "args": { + "External id": 8539,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8538 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791940114, "dur": 11, + "args": { + "External id": 8540,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8539 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940144, "dur": 5, + "args": { + "External id": 8541,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8540 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791940153, "dur": 22, + "args": { + "External id": 8542,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8541 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940154, "dur": 5, + "args": { + "External id": 8543,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8542 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791940160, "dur": 15, + "args": { + "External id": 8544,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8543 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791940162, "dur": 12, + "args": { + "External id": 8545,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8544 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940193, "dur": 5, + "args": { + "External id": 8546,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8545 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791940202, "dur": 43, + "args": { + "External id": 8547,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8546 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940203, "dur": 5, + "args": { + "External id": 8548,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8547 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791940209, "dur": 36, + "args": { + "External id": 8549,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8548 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791940210, "dur": 34, + "args": { + "External id": 8550,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8549 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940266, "dur": 6, + "args": { + "External id": 8551,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8550 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791940277, "dur": 22, + "args": { + "External id": 8552,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8551 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940278, "dur": 4, + "args": { + "External id": 8553,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8552 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791940284, "dur": 14, + "args": { + "External id": 8554,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8553 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791940285, "dur": 12, + "args": { + "External id": 8555,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8554 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940317, "dur": 5, + "args": { + "External id": 8556,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8555 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791940326, "dur": 21, + "args": { + "External id": 8557,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8556 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940327, "dur": 4, + "args": { + "External id": 8558,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8557 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791940333, "dur": 14, + "args": { + "External id": 8559,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8558 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791940334, "dur": 11, + "args": { + "External id": 8560,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8559 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940367, "dur": 5, + "args": { + "External id": 8561,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8560 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791940376, "dur": 21, + "args": { + "External id": 8562,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8561 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940378, "dur": 4, + "args": { + "External id": 8563,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8562 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791940383, "dur": 14, + "args": { + "External id": 8564,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8563 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791940385, "dur": 11, + "args": { + "External id": 8565,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8564 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940415, "dur": 5, + "args": { + "External id": 8566,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8565 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791940424, "dur": 21, + "args": { + "External id": 8567,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8566 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940425, "dur": 5, + "args": { + "External id": 8568,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8567 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791940431, "dur": 14, + "args": { + "External id": 8569,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8568 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791940432, "dur": 11, + "args": { + "External id": 8570,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8569 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940463, "dur": 5, + "args": { + "External id": 8571,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8570 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791940472, "dur": 21, + "args": { + "External id": 8572,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8571 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940473, "dur": 4, + "args": { + "External id": 8573,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8572 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791940478, "dur": 14, + "args": { + "External id": 8574,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8573 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791940480, "dur": 11, + "args": { + "External id": 8575,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8574 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940510, "dur": 5, + "args": { + "External id": 8576,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8575 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791940519, "dur": 21, + "args": { + "External id": 8577,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8576 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940520, "dur": 5, + "args": { + "External id": 8578,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8577 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791940526, "dur": 14, + "args": { + "External id": 8579,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8578 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791940527, "dur": 11, + "args": { + "External id": 8580,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8579 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940558, "dur": 5, + "args": { + "External id": 8581,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8580 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791940566, "dur": 22, + "args": { + "External id": 8582,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8581 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940568, "dur": 4, + "args": { + "External id": 8583,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8582 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791940574, "dur": 13, + "args": { + "External id": 8584,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8583 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791940575, "dur": 11, + "args": { + "External id": 8585,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8584 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940605, "dur": 5, + "args": { + "External id": 8586,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8585 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791940614, "dur": 21, + "args": { + "External id": 8587,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8586 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940615, "dur": 5, + "args": { + "External id": 8588,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8587 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791940621, "dur": 14, + "args": { + "External id": 8589,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8588 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791940622, "dur": 11, + "args": { + "External id": 8590,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8589 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940653, "dur": 5, + "args": { + "External id": 8591,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8590 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791940661, "dur": 21, + "args": { + "External id": 8592,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8591 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940663, "dur": 4, + "args": { + "External id": 8593,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8592 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791940668, "dur": 14, + "args": { + "External id": 8594,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8593 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791940670, "dur": 11, + "args": { + "External id": 8595,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8594 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940700, "dur": 5, + "args": { + "External id": 8596,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8595 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791940709, "dur": 21, + "args": { + "External id": 8597,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8596 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940710, "dur": 4, + "args": { + "External id": 8598,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8597 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791940716, "dur": 14, + "args": { + "External id": 8599,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8598 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791940718, "dur": 11, + "args": { + "External id": 8600,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8599 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940748, "dur": 5, + "args": { + "External id": 8601,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8600 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791940757, "dur": 21, + "args": { + "External id": 8602,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8601 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940758, "dur": 5, + "args": { + "External id": 8603,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8602 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791940764, "dur": 14, + "args": { + "External id": 8604,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8603 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791940766, "dur": 11, + "args": { + "External id": 8605,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8604 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940857, "dur": 7, + "args": { + "External id": 8606,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8605 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791940868, "dur": 25, + "args": { + "External id": 8607,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8606 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940870, "dur": 4, + "args": { + "External id": 8608,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8607 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791940876, "dur": 17, + "args": { + "External id": 8609,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8608 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791940878, "dur": 13, + "args": { + "External id": 8610,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8609 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940914, "dur": 5, + "args": { + "External id": 8611,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8610 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791940923, "dur": 22, + "args": { + "External id": 8612,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8611 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940924, "dur": 5, + "args": { + "External id": 8613,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8612 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791940930, "dur": 14, + "args": { + "External id": 8614,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8613 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791940932, "dur": 11, + "args": { + "External id": 8615,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8614 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940963, "dur": 5, + "args": { + "External id": 8616,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8615 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791940972, "dur": 21, + "args": { + "External id": 8617,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8616 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791940973, "dur": 5, + "args": { + "External id": 8618,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8617 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791940979, "dur": 14, + "args": { + "External id": 8619,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8618 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791940980, "dur": 11, + "args": { + "External id": 8620,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8619 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941012, "dur": 5, + "args": { + "External id": 8621,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8620 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791941021, "dur": 21, + "args": { + "External id": 8622,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8621 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941023, "dur": 4, + "args": { + "External id": 8623,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8622 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791941028, "dur": 14, + "args": { + "External id": 8624,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8623 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791941030, "dur": 11, + "args": { + "External id": 8625,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8624 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941061, "dur": 5, + "args": { + "External id": 8626,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8625 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791941070, "dur": 21, + "args": { + "External id": 8627,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8626 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941071, "dur": 4, + "args": { + "External id": 8628,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8627 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791941077, "dur": 14, + "args": { + "External id": 8629,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8628 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791941078, "dur": 11, + "args": { + "External id": 8630,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8629 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941109, "dur": 5, + "args": { + "External id": 8631,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8630 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791941118, "dur": 20, + "args": { + "External id": 8632,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8631 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941119, "dur": 4, + "args": { + "External id": 8633,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8632 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791941124, "dur": 14, + "args": { + "External id": 8634,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8633 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791941126, "dur": 11, + "args": { + "External id": 8635,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8634 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941156, "dur": 5, + "args": { + "External id": 8636,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8635 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791941165, "dur": 21, + "args": { + "External id": 8637,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8636 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941166, "dur": 4, + "args": { + "External id": 8638,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8637 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791941171, "dur": 14, + "args": { + "External id": 8639,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8638 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791941173, "dur": 11, + "args": { + "External id": 8640,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8639 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941205, "dur": 6, + "args": { + "External id": 8641,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8640 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791941216, "dur": 21, + "args": { + "External id": 8642,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8641 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941217, "dur": 5, + "args": { + "External id": 8643,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8642 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791941223, "dur": 14, + "args": { + "External id": 8644,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8643 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791941225, "dur": 11, + "args": { + "External id": 8645,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8644 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941254, "dur": 5, + "args": { + "External id": 8646,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8645 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791941263, "dur": 21, + "args": { + "External id": 8647,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8646 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941264, "dur": 5, + "args": { + "External id": 8648,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8647 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791941270, "dur": 14, + "args": { + "External id": 8649,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8648 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791941271, "dur": 11, + "args": { + "External id": 8650,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8649 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941302, "dur": 5, + "args": { + "External id": 8651,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8650 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791941310, "dur": 21, + "args": { + "External id": 8652,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8651 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941311, "dur": 5, + "args": { + "External id": 8653,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8652 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791941317, "dur": 14, + "args": { + "External id": 8654,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8653 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791941319, "dur": 11, + "args": { + "External id": 8655,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8654 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941401, "dur": 6, + "args": { + "External id": 8656,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8655 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791941411, "dur": 24, + "args": { + "External id": 8657,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8656 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941413, "dur": 4, + "args": { + "External id": 8658,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8657 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791941419, "dur": 16, + "args": { + "External id": 8659,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8658 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791941420, "dur": 14, + "args": { + "External id": 8660,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8659 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941455, "dur": 5, + "args": { + "External id": 8661,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8660 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791941464, "dur": 22, + "args": { + "External id": 8662,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8661 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941465, "dur": 5, + "args": { + "External id": 8663,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8662 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791941471, "dur": 14, + "args": { + "External id": 8664,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8663 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791941473, "dur": 11, + "args": { + "External id": 8665,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8664 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941503, "dur": 6, + "args": { + "External id": 8666,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8665 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791941512, "dur": 21, + "args": { + "External id": 8667,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8666 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941513, "dur": 5, + "args": { + "External id": 8668,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8667 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791941519, "dur": 14, + "args": { + "External id": 8669,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8668 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791941521, "dur": 11, + "args": { + "External id": 8670,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8669 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941551, "dur": 6, + "args": { + "External id": 8671,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8670 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791941560, "dur": 21, + "args": { + "External id": 8672,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8671 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941561, "dur": 5, + "args": { + "External id": 8673,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8672 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791941567, "dur": 14, + "args": { + "External id": 8674,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8673 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791941569, "dur": 11, + "args": { + "External id": 8675,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8674 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941599, "dur": 5, + "args": { + "External id": 8676,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8675 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791941608, "dur": 21, + "args": { + "External id": 8677,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8676 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941609, "dur": 5, + "args": { + "External id": 8678,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8677 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791941615, "dur": 14, + "args": { + "External id": 8679,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8678 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791941616, "dur": 11, + "args": { + "External id": 8680,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8679 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941647, "dur": 5, + "args": { + "External id": 8681,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8680 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791941656, "dur": 21, + "args": { + "External id": 8682,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8681 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941657, "dur": 4, + "args": { + "External id": 8683,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8682 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791941663, "dur": 13, + "args": { + "External id": 8684,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8683 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791941664, "dur": 11, + "args": { + "External id": 8685,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8684 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941694, "dur": 5, + "args": { + "External id": 8686,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8685 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791941703, "dur": 21, + "args": { + "External id": 8687,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8686 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941704, "dur": 5, + "args": { + "External id": 8688,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8687 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791941710, "dur": 14, + "args": { + "External id": 8689,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8688 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791941712, "dur": 11, + "args": { + "External id": 8690,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8689 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941742, "dur": 5, + "args": { + "External id": 8691,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8690 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791941751, "dur": 21, + "args": { + "External id": 8692,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8691 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941752, "dur": 4, + "args": { + "External id": 8693,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8692 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791941757, "dur": 14, + "args": { + "External id": 8694,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8693 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791941759, "dur": 11, + "args": { + "External id": 8695,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8694 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941789, "dur": 6, + "args": { + "External id": 8696,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8695 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791941798, "dur": 21, + "args": { + "External id": 8697,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8696 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941800, "dur": 4, + "args": { + "External id": 8698,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8697 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791941805, "dur": 14, + "args": { + "External id": 8699,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8698 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791941807, "dur": 11, + "args": { + "External id": 8700,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8699 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941837, "dur": 5, + "args": { + "External id": 8701,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8700 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791941846, "dur": 31, + "args": { + "External id": 8702,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8701 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941847, "dur": 4, + "args": { + "External id": 8703,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8702 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791941853, "dur": 24, + "args": { + "External id": 8704,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8703 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791941864, "dur": 11, + "args": { + "External id": 8705,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8704 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941896, "dur": 5, + "args": { + "External id": 8706,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8705 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791941905, "dur": 21, + "args": { + "External id": 8707,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8706 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941906, "dur": 4, + "args": { + "External id": 8708,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8707 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791941912, "dur": 14, + "args": { + "External id": 8709,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8708 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791941913, "dur": 12, + "args": { + "External id": 8710,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8709 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941944, "dur": 5, + "args": { + "External id": 8711,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8710 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791941953, "dur": 21, + "args": { + "External id": 8712,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8711 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941954, "dur": 5, + "args": { + "External id": 8713,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8712 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791941960, "dur": 14, + "args": { + "External id": 8714,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8713 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791941962, "dur": 11, + "args": { + "External id": 8715,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8714 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791941993, "dur": 5, + "args": { + "External id": 8716,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8715 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791942002, "dur": 21, + "args": { + "External id": 8717,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8716 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942003, "dur": 4, + "args": { + "External id": 8718,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8717 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791942009, "dur": 14, + "args": { + "External id": 8719,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8718 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791942010, "dur": 12, + "args": { + "External id": 8720,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8719 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942041, "dur": 5, + "args": { + "External id": 8721,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8720 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791942050, "dur": 21, + "args": { + "External id": 8722,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8721 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942051, "dur": 5, + "args": { + "External id": 8723,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8722 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791942057, "dur": 14, + "args": { + "External id": 8724,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8723 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791942059, "dur": 11, + "args": { + "External id": 8725,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8724 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942089, "dur": 5, + "args": { + "External id": 8726,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8725 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791942098, "dur": 21, + "args": { + "External id": 8727,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8726 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942099, "dur": 5, + "args": { + "External id": 8728,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8727 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791942105, "dur": 14, + "args": { + "External id": 8729,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8728 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791942107, "dur": 11, + "args": { + "External id": 8730,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8729 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942137, "dur": 5, + "args": { + "External id": 8731,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8730 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791942146, "dur": 21, + "args": { + "External id": 8732,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8731 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942147, "dur": 5, + "args": { + "External id": 8733,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8732 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791942153, "dur": 14, + "args": { + "External id": 8734,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8733 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791942155, "dur": 11, + "args": { + "External id": 8735,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8734 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942185, "dur": 5, + "args": { + "External id": 8736,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8735 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791942194, "dur": 22, + "args": { + "External id": 8737,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8736 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942196, "dur": 4, + "args": { + "External id": 8738,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8737 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791942201, "dur": 14, + "args": { + "External id": 8739,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8738 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791942203, "dur": 11, + "args": { + "External id": 8740,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8739 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942233, "dur": 5, + "args": { + "External id": 8741,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8740 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791942242, "dur": 22, + "args": { + "External id": 8742,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8741 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942243, "dur": 5, + "args": { + "External id": 8743,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8742 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791942249, "dur": 14, + "args": { + "External id": 8744,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8743 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791942251, "dur": 11, + "args": { + "External id": 8745,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8744 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942281, "dur": 6, + "args": { + "External id": 8746,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8745 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791942290, "dur": 21, + "args": { + "External id": 8747,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8746 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942292, "dur": 4, + "args": { + "External id": 8748,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8747 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791942297, "dur": 14, + "args": { + "External id": 8749,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8748 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791942299, "dur": 11, + "args": { + "External id": 8750,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8749 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942330, "dur": 5, + "args": { + "External id": 8751,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8750 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791942339, "dur": 21, + "args": { + "External id": 8752,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8751 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942340, "dur": 4, + "args": { + "External id": 8753,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8752 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791942345, "dur": 14, + "args": { + "External id": 8754,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8753 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791942347, "dur": 11, + "args": { + "External id": 8755,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8754 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942377, "dur": 5, + "args": { + "External id": 8756,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8755 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791942386, "dur": 21, + "args": { + "External id": 8757,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8756 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942388, "dur": 4, + "args": { + "External id": 8758,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8757 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791942393, "dur": 14, + "args": { + "External id": 8759,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8758 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791942395, "dur": 11, + "args": { + "External id": 8760,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8759 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942425, "dur": 5, + "args": { + "External id": 8761,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8760 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791942434, "dur": 21, + "args": { + "External id": 8762,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8761 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942435, "dur": 5, + "args": { + "External id": 8763,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8762 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791942441, "dur": 14, + "args": { + "External id": 8764,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8763 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791942443, "dur": 11, + "args": { + "External id": 8765,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8764 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942473, "dur": 5, + "args": { + "External id": 8766,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8765 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791942482, "dur": 22, + "args": { + "External id": 8767,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8766 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942484, "dur": 4, + "args": { + "External id": 8768,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8767 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791942490, "dur": 14, + "args": { + "External id": 8769,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8768 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791942491, "dur": 11, + "args": { + "External id": 8770,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8769 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942522, "dur": 5, + "args": { + "External id": 8771,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8770 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791942531, "dur": 21, + "args": { + "External id": 8772,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8771 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942532, "dur": 4, + "args": { + "External id": 8773,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8772 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791942538, "dur": 14, + "args": { + "External id": 8774,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8773 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791942539, "dur": 11, + "args": { + "External id": 8775,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8774 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942570, "dur": 5, + "args": { + "External id": 8776,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8775 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791942579, "dur": 21, + "args": { + "External id": 8777,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8776 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942580, "dur": 4, + "args": { + "External id": 8778,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8777 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791942585, "dur": 14, + "args": { + "External id": 8779,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8778 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791942587, "dur": 11, + "args": { + "External id": 8780,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8779 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942618, "dur": 5, + "args": { + "External id": 8781,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8780 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791942627, "dur": 21, + "args": { + "External id": 8782,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8781 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942628, "dur": 4, + "args": { + "External id": 8783,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8782 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791942633, "dur": 14, + "args": { + "External id": 8784,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8783 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791942635, "dur": 11, + "args": { + "External id": 8785,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8784 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942666, "dur": 5, + "args": { + "External id": 8786,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8785 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791942675, "dur": 21, + "args": { + "External id": 8787,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8786 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942676, "dur": 4, + "args": { + "External id": 8788,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8787 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791942682, "dur": 14, + "args": { + "External id": 8789,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8788 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791942683, "dur": 11, + "args": { + "External id": 8790,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8789 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942714, "dur": 5, + "args": { + "External id": 8791,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8790 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791942723, "dur": 21, + "args": { + "External id": 8792,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8791 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942724, "dur": 4, + "args": { + "External id": 8793,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8792 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791942729, "dur": 14, + "args": { + "External id": 8794,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8793 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791942731, "dur": 11, + "args": { + "External id": 8795,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8794 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942761, "dur": 5, + "args": { + "External id": 8796,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8795 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791942770, "dur": 21, + "args": { + "External id": 8797,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8796 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942771, "dur": 5, + "args": { + "External id": 8798,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8797 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791942777, "dur": 14, + "args": { + "External id": 8799,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8798 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791942779, "dur": 11, + "args": { + "External id": 8800,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8799 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942809, "dur": 5, + "args": { + "External id": 8801,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8800 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791942818, "dur": 21, + "args": { + "External id": 8802,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8801 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942819, "dur": 4, + "args": { + "External id": 8803,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8802 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791942825, "dur": 14, + "args": { + "External id": 8804,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8803 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791942826, "dur": 12, + "args": { + "External id": 8805,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8804 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942857, "dur": 5, + "args": { + "External id": 8806,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8805 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791942866, "dur": 21, + "args": { + "External id": 8807,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8806 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942867, "dur": 5, + "args": { + "External id": 8808,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8807 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791942873, "dur": 14, + "args": { + "External id": 8809,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8808 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791942875, "dur": 11, + "args": { + "External id": 8810,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8809 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942905, "dur": 5, + "args": { + "External id": 8811,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8810 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791942914, "dur": 22, + "args": { + "External id": 8812,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8811 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942916, "dur": 4, + "args": { + "External id": 8813,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8812 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791942921, "dur": 14, + "args": { + "External id": 8814,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8813 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791942923, "dur": 11, + "args": { + "External id": 8815,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8814 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942954, "dur": 5, + "args": { + "External id": 8816,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8815 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791942963, "dur": 21, + "args": { + "External id": 8817,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8816 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791942964, "dur": 4, + "args": { + "External id": 8818,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8817 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791942969, "dur": 14, + "args": { + "External id": 8819,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8818 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791942971, "dur": 11, + "args": { + "External id": 8820,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8819 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943002, "dur": 5, + "args": { + "External id": 8821,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8820 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791943011, "dur": 21, + "args": { + "External id": 8822,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8821 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943012, "dur": 4, + "args": { + "External id": 8823,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8822 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791943018, "dur": 14, + "args": { + "External id": 8824,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8823 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791943019, "dur": 11, + "args": { + "External id": 8825,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8824 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943050, "dur": 5, + "args": { + "External id": 8826,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8825 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791943059, "dur": 21, + "args": { + "External id": 8827,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8826 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943060, "dur": 4, + "args": { + "External id": 8828,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8827 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791943065, "dur": 14, + "args": { + "External id": 8829,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8828 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791943067, "dur": 11, + "args": { + "External id": 8830,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8829 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943097, "dur": 5, + "args": { + "External id": 8831,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8830 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791943106, "dur": 21, + "args": { + "External id": 8832,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8831 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943108, "dur": 4, + "args": { + "External id": 8833,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8832 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791943113, "dur": 14, + "args": { + "External id": 8834,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8833 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791943115, "dur": 11, + "args": { + "External id": 8835,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8834 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943145, "dur": 5, + "args": { + "External id": 8836,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8835 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791943154, "dur": 21, + "args": { + "External id": 8837,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8836 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943155, "dur": 5, + "args": { + "External id": 8838,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8837 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791943161, "dur": 14, + "args": { + "External id": 8839,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8838 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791943163, "dur": 11, + "args": { + "External id": 8840,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8839 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943193, "dur": 5, + "args": { + "External id": 8841,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8840 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791943202, "dur": 21, + "args": { + "External id": 8842,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8841 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943203, "dur": 5, + "args": { + "External id": 8843,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8842 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791943209, "dur": 14, + "args": { + "External id": 8844,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8843 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791943211, "dur": 11, + "args": { + "External id": 8845,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8844 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943242, "dur": 5, + "args": { + "External id": 8846,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8845 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791943251, "dur": 21, + "args": { + "External id": 8847,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8846 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943252, "dur": 5, + "args": { + "External id": 8848,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8847 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791943258, "dur": 14, + "args": { + "External id": 8849,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8848 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791943260, "dur": 10, + "args": { + "External id": 8850,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8849 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943290, "dur": 5, + "args": { + "External id": 8851,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8850 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791943299, "dur": 21, + "args": { + "External id": 8852,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8851 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943300, "dur": 4, + "args": { + "External id": 8853,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8852 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791943306, "dur": 13, + "args": { + "External id": 8854,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8853 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791943307, "dur": 11, + "args": { + "External id": 8855,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8854 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943338, "dur": 5, + "args": { + "External id": 8856,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8855 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791943347, "dur": 21, + "args": { + "External id": 8857,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8856 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943348, "dur": 4, + "args": { + "External id": 8858,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8857 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791943354, "dur": 14, + "args": { + "External id": 8859,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8858 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791943355, "dur": 11, + "args": { + "External id": 8860,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8859 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943386, "dur": 5, + "args": { + "External id": 8861,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8860 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791943395, "dur": 21, + "args": { + "External id": 8862,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8861 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943396, "dur": 5, + "args": { + "External id": 8863,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8862 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791943402, "dur": 14, + "args": { + "External id": 8864,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8863 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791943403, "dur": 11, + "args": { + "External id": 8865,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8864 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943434, "dur": 5, + "args": { + "External id": 8866,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8865 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791943443, "dur": 21, + "args": { + "External id": 8867,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8866 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943444, "dur": 5, + "args": { + "External id": 8868,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8867 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791943450, "dur": 14, + "args": { + "External id": 8869,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8868 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791943451, "dur": 12, + "args": { + "External id": 8870,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8869 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943482, "dur": 6, + "args": { + "External id": 8871,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8870 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791943491, "dur": 22, + "args": { + "External id": 8872,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8871 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943493, "dur": 4, + "args": { + "External id": 8873,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8872 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791943498, "dur": 14, + "args": { + "External id": 8874,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8873 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791943500, "dur": 11, + "args": { + "External id": 8875,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8874 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943530, "dur": 5, + "args": { + "External id": 8876,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8875 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791943539, "dur": 21, + "args": { + "External id": 8877,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8876 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943540, "dur": 5, + "args": { + "External id": 8878,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8877 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791943546, "dur": 14, + "args": { + "External id": 8879,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8878 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791943548, "dur": 11, + "args": { + "External id": 8880,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8879 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943578, "dur": 5, + "args": { + "External id": 8881,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8880 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791943587, "dur": 21, + "args": { + "External id": 8882,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8881 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943588, "dur": 5, + "args": { + "External id": 8883,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8882 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791943594, "dur": 14, + "args": { + "External id": 8884,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8883 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791943596, "dur": 11, + "args": { + "External id": 8885,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8884 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943626, "dur": 5, + "args": { + "External id": 8886,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8885 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791943635, "dur": 21, + "args": { + "External id": 8887,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8886 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943636, "dur": 5, + "args": { + "External id": 8888,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8887 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791943642, "dur": 14, + "args": { + "External id": 8889,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8888 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791943643, "dur": 11, + "args": { + "External id": 8890,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8889 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943674, "dur": 5, + "args": { + "External id": 8891,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8890 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791943683, "dur": 21, + "args": { + "External id": 8892,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8891 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943684, "dur": 5, + "args": { + "External id": 8893,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8892 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791943690, "dur": 14, + "args": { + "External id": 8894,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8893 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791943691, "dur": 11, + "args": { + "External id": 8895,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8894 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943722, "dur": 5, + "args": { + "External id": 8896,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8895 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791943730, "dur": 21, + "args": { + "External id": 8897,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8896 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943732, "dur": 4, + "args": { + "External id": 8898,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8897 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791943737, "dur": 14, + "args": { + "External id": 8899,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8898 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791943739, "dur": 11, + "args": { + "External id": 8900,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8899 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943769, "dur": 5, + "args": { + "External id": 8901,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8900 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791943777, "dur": 22, + "args": { + "External id": 8902,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8901 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943779, "dur": 4, + "args": { + "External id": 8903,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8902 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791943784, "dur": 14, + "args": { + "External id": 8904,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8903 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791943786, "dur": 11, + "args": { + "External id": 8905,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8904 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943875, "dur": 7, + "args": { + "External id": 8906,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8905 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791943886, "dur": 25, + "args": { + "External id": 8907,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8906 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943888, "dur": 4, + "args": { + "External id": 8908,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8907 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791943894, "dur": 17, + "args": { + "External id": 8909,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8908 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791943895, "dur": 14, + "args": { + "External id": 8910,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8909 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943932, "dur": 5, + "args": { + "External id": 8911,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8910 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791943941, "dur": 22, + "args": { + "External id": 8912,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8911 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943942, "dur": 4, + "args": { + "External id": 8913,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8912 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791943948, "dur": 14, + "args": { + "External id": 8914,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8913 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791943949, "dur": 12, + "args": { + "External id": 8915,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8914 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943981, "dur": 5, + "args": { + "External id": 8916,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8915 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791943990, "dur": 23, + "args": { + "External id": 8917,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8916 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791943991, "dur": 5, + "args": { + "External id": 8918,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8917 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791943997, "dur": 15, + "args": { + "External id": 8919,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8918 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791943999, "dur": 12, + "args": { + "External id": 8920,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8919 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944034, "dur": 5, + "args": { + "External id": 8921,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8920 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791944043, "dur": 22, + "args": { + "External id": 8922,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8921 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944044, "dur": 5, + "args": { + "External id": 8923,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8922 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791944050, "dur": 14, + "args": { + "External id": 8924,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8923 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791944052, "dur": 11, + "args": { + "External id": 8925,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8924 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944083, "dur": 5, + "args": { + "External id": 8926,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8925 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791944092, "dur": 21, + "args": { + "External id": 8927,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8926 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944093, "dur": 4, + "args": { + "External id": 8928,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8927 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791944099, "dur": 13, + "args": { + "External id": 8929,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8928 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791944100, "dur": 11, + "args": { + "External id": 8930,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8929 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944131, "dur": 5, + "args": { + "External id": 8931,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8930 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791944140, "dur": 21, + "args": { + "External id": 8932,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8931 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944141, "dur": 4, + "args": { + "External id": 8933,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8932 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791944147, "dur": 13, + "args": { + "External id": 8934,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8933 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791944148, "dur": 11, + "args": { + "External id": 8935,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8934 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944179, "dur": 5, + "args": { + "External id": 8936,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8935 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791944188, "dur": 21, + "args": { + "External id": 8937,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8936 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944189, "dur": 4, + "args": { + "External id": 8938,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8937 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791944195, "dur": 14, + "args": { + "External id": 8939,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8938 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791944196, "dur": 11, + "args": { + "External id": 8940,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8939 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944247, "dur": 6, + "args": { + "External id": 8941,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8940 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791944257, "dur": 22, + "args": { + "External id": 8942,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8941 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944258, "dur": 5, + "args": { + "External id": 8943,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8942 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791944264, "dur": 15, + "args": { + "External id": 8944,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8943 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791944266, "dur": 12, + "args": { + "External id": 8945,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8944 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944298, "dur": 5, + "args": { + "External id": 8946,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8945 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791944307, "dur": 22, + "args": { + "External id": 8947,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8946 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944308, "dur": 5, + "args": { + "External id": 8948,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8947 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791944314, "dur": 14, + "args": { + "External id": 8949,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8948 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791944316, "dur": 11, + "args": { + "External id": 8950,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8949 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944347, "dur": 5, + "args": { + "External id": 8951,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8950 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791944356, "dur": 21, + "args": { + "External id": 8952,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8951 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944357, "dur": 4, + "args": { + "External id": 8953,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8952 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791944362, "dur": 14, + "args": { + "External id": 8954,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8953 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791944364, "dur": 11, + "args": { + "External id": 8955,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8954 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944448, "dur": 7, + "args": { + "External id": 8956,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8955 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791944460, "dur": 24, + "args": { + "External id": 8957,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8956 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944461, "dur": 4, + "args": { + "External id": 8958,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8957 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791944467, "dur": 17, + "args": { + "External id": 8959,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8958 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791944469, "dur": 13, + "args": { + "External id": 8960,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8959 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944504, "dur": 6, + "args": { + "External id": 8961,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8960 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791944514, "dur": 21, + "args": { + "External id": 8962,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8961 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944515, "dur": 4, + "args": { + "External id": 8963,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8962 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791944520, "dur": 15, + "args": { + "External id": 8964,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8963 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791944522, "dur": 11, + "args": { + "External id": 8965,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8964 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944554, "dur": 5, + "args": { + "External id": 8966,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8965 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791944563, "dur": 21, + "args": { + "External id": 8967,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8966 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944564, "dur": 4, + "args": { + "External id": 8968,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8967 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791944570, "dur": 14, + "args": { + "External id": 8969,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8968 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791944571, "dur": 12, + "args": { + "External id": 8970,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8969 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944602, "dur": 5, + "args": { + "External id": 8971,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8970 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791944611, "dur": 22, + "args": { + "External id": 8972,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8971 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944613, "dur": 4, + "args": { + "External id": 8973,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8972 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791944619, "dur": 14, + "args": { + "External id": 8974,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8973 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791944620, "dur": 11, + "args": { + "External id": 8975,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8974 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944651, "dur": 5, + "args": { + "External id": 8976,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8975 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791944660, "dur": 22, + "args": { + "External id": 8977,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8976 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944661, "dur": 5, + "args": { + "External id": 8978,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8977 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791944667, "dur": 14, + "args": { + "External id": 8979,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8978 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791944669, "dur": 11, + "args": { + "External id": 8980,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8979 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944700, "dur": 5, + "args": { + "External id": 8981,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8980 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791944709, "dur": 21, + "args": { + "External id": 8982,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8981 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944710, "dur": 4, + "args": { + "External id": 8983,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8982 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791944716, "dur": 14, + "args": { + "External id": 8984,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8983 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791944717, "dur": 12, + "args": { + "External id": 8985,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8984 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944748, "dur": 6, + "args": { + "External id": 8986,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8985 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791944758, "dur": 21, + "args": { + "External id": 8987,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8986 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944759, "dur": 4, + "args": { + "External id": 8988,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8987 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791944764, "dur": 14, + "args": { + "External id": 8989,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8988 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791944766, "dur": 11, + "args": { + "External id": 8990,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8989 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944797, "dur": 5, + "args": { + "External id": 8991,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8990 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791944806, "dur": 21, + "args": { + "External id": 8992,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8991 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944807, "dur": 4, + "args": { + "External id": 8993,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8992 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791944813, "dur": 14, + "args": { + "External id": 8994,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8993 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791944814, "dur": 11, + "args": { + "External id": 8995,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8994 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944845, "dur": 5, + "args": { + "External id": 8996,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8995 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791944854, "dur": 21, + "args": { + "External id": 8997,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 8996 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944855, "dur": 5, + "args": { + "External id": 8998,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8997 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791944861, "dur": 14, + "args": { + "External id": 8999,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 8998 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791944863, "dur": 11, + "args": { + "External id": 9000,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 8999 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944893, "dur": 5, + "args": { + "External id": 9001,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9000 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791944902, "dur": 21, + "args": { + "External id": 9002,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9001 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944903, "dur": 5, + "args": { + "External id": 9003,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9002 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791944909, "dur": 14, + "args": { + "External id": 9004,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9003 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791944911, "dur": 10, + "args": { + "External id": 9005,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9004 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944941, "dur": 5, + "args": { + "External id": 9006,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9005 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791944950, "dur": 21, + "args": { + "External id": 9007,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9006 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944951, "dur": 4, + "args": { + "External id": 9008,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9007 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791944957, "dur": 14, + "args": { + "External id": 9009,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9008 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791944958, "dur": 11, + "args": { + "External id": 9010,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9009 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944989, "dur": 5, + "args": { + "External id": 9011,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9010 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791944998, "dur": 21, + "args": { + "External id": 9012,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9011 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791944999, "dur": 5, + "args": { + "External id": 9013,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9012 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791945005, "dur": 14, + "args": { + "External id": 9014,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9013 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791945007, "dur": 11, + "args": { + "External id": 9015,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9014 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945037, "dur": 5, + "args": { + "External id": 9016,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9015 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791945046, "dur": 21, + "args": { + "External id": 9017,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9016 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945047, "dur": 5, + "args": { + "External id": 9018,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9017 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791945053, "dur": 14, + "args": { + "External id": 9019,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9018 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791945055, "dur": 10, + "args": { + "External id": 9020,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9019 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945085, "dur": 5, + "args": { + "External id": 9021,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9020 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791945094, "dur": 22, + "args": { + "External id": 9022,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9021 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945095, "dur": 5, + "args": { + "External id": 9023,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9022 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791945101, "dur": 14, + "args": { + "External id": 9024,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9023 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791945103, "dur": 11, + "args": { + "External id": 9025,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9024 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945133, "dur": 5, + "args": { + "External id": 9026,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9025 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791945142, "dur": 22, + "args": { + "External id": 9027,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9026 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945143, "dur": 5, + "args": { + "External id": 9028,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9027 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791945149, "dur": 14, + "args": { + "External id": 9029,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9028 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791945151, "dur": 11, + "args": { + "External id": 9030,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9029 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945182, "dur": 5, + "args": { + "External id": 9031,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9030 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791945191, "dur": 21, + "args": { + "External id": 9032,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9031 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945192, "dur": 4, + "args": { + "External id": 9033,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9032 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791945197, "dur": 14, + "args": { + "External id": 9034,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9033 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791945199, "dur": 11, + "args": { + "External id": 9035,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9034 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945230, "dur": 5, + "args": { + "External id": 9036,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9035 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791945239, "dur": 21, + "args": { + "External id": 9037,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9036 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945240, "dur": 4, + "args": { + "External id": 9038,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9037 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791945245, "dur": 14, + "args": { + "External id": 9039,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9038 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791945247, "dur": 11, + "args": { + "External id": 9040,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9039 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945277, "dur": 6, + "args": { + "External id": 9041,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9040 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791945287, "dur": 21, + "args": { + "External id": 9042,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9041 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945288, "dur": 4, + "args": { + "External id": 9043,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9042 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791945293, "dur": 14, + "args": { + "External id": 9044,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9043 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791945295, "dur": 11, + "args": { + "External id": 9045,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9044 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945326, "dur": 5, + "args": { + "External id": 9046,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9045 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791945335, "dur": 21, + "args": { + "External id": 9047,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9046 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945336, "dur": 4, + "args": { + "External id": 9048,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9047 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791945342, "dur": 14, + "args": { + "External id": 9049,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9048 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791945343, "dur": 12, + "args": { + "External id": 9050,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9049 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945374, "dur": 5, + "args": { + "External id": 9051,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9050 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791945383, "dur": 21, + "args": { + "External id": 9052,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9051 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945384, "dur": 4, + "args": { + "External id": 9053,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9052 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791945390, "dur": 14, + "args": { + "External id": 9054,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9053 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791945391, "dur": 11, + "args": { + "External id": 9055,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9054 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945422, "dur": 5, + "args": { + "External id": 9056,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9055 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791945431, "dur": 21, + "args": { + "External id": 9057,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9056 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945432, "dur": 4, + "args": { + "External id": 9058,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9057 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791945437, "dur": 14, + "args": { + "External id": 9059,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9058 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791945439, "dur": 11, + "args": { + "External id": 9060,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9059 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945469, "dur": 6, + "args": { + "External id": 9061,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9060 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791945479, "dur": 21, + "args": { + "External id": 9062,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9061 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945480, "dur": 4, + "args": { + "External id": 9063,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9062 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791945485, "dur": 14, + "args": { + "External id": 9064,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9063 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791945487, "dur": 11, + "args": { + "External id": 9065,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9064 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945517, "dur": 6, + "args": { + "External id": 9066,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9065 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791945526, "dur": 22, + "args": { + "External id": 9067,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9066 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945528, "dur": 4, + "args": { + "External id": 9068,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9067 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791945534, "dur": 14, + "args": { + "External id": 9069,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9068 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791945535, "dur": 11, + "args": { + "External id": 9070,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9069 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945566, "dur": 5, + "args": { + "External id": 9071,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9070 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791945575, "dur": 23, + "args": { + "External id": 9072,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9071 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945578, "dur": 4, + "args": { + "External id": 9073,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9072 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791945583, "dur": 14, + "args": { + "External id": 9074,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9073 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791945585, "dur": 11, + "args": { + "External id": 9075,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9074 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945616, "dur": 5, + "args": { + "External id": 9076,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9075 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791945625, "dur": 21, + "args": { + "External id": 9077,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9076 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945626, "dur": 4, + "args": { + "External id": 9078,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9077 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791945631, "dur": 15, + "args": { + "External id": 9079,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9078 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791945633, "dur": 11, + "args": { + "External id": 9080,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9079 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945664, "dur": 5, + "args": { + "External id": 9081,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9080 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791945673, "dur": 21, + "args": { + "External id": 9082,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9081 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945674, "dur": 5, + "args": { + "External id": 9083,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9082 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791945680, "dur": 14, + "args": { + "External id": 9084,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9083 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791945682, "dur": 11, + "args": { + "External id": 9085,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9084 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945712, "dur": 6, + "args": { + "External id": 9086,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9085 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791945721, "dur": 21, + "args": { + "External id": 9087,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9086 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945723, "dur": 4, + "args": { + "External id": 9088,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9087 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791945728, "dur": 14, + "args": { + "External id": 9089,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9088 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791945730, "dur": 11, + "args": { + "External id": 9090,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9089 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945760, "dur": 5, + "args": { + "External id": 9091,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9090 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791945769, "dur": 21, + "args": { + "External id": 9092,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9091 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945770, "dur": 5, + "args": { + "External id": 9093,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9092 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791945776, "dur": 14, + "args": { + "External id": 9094,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9093 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791945778, "dur": 11, + "args": { + "External id": 9095,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9094 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945808, "dur": 5, + "args": { + "External id": 9096,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9095 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791945817, "dur": 21, + "args": { + "External id": 9097,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9096 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945818, "dur": 4, + "args": { + "External id": 9098,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9097 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791945823, "dur": 14, + "args": { + "External id": 9099,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9098 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791945825, "dur": 11, + "args": { + "External id": 9100,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9099 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945855, "dur": 5, + "args": { + "External id": 9101,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9100 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791945864, "dur": 25, + "args": { + "External id": 9102,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9101 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945865, "dur": 5, + "args": { + "External id": 9103,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9102 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791945871, "dur": 17, + "args": { + "External id": 9104,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9103 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791945876, "dur": 11, + "args": { + "External id": 9105,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9104 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945907, "dur": 5, + "args": { + "External id": 9106,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9105 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791945916, "dur": 21, + "args": { + "External id": 9107,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9106 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945917, "dur": 5, + "args": { + "External id": 9108,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9107 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791945923, "dur": 14, + "args": { + "External id": 9109,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9108 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791945925, "dur": 11, + "args": { + "External id": 9110,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9109 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945955, "dur": 5, + "args": { + "External id": 9111,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9110 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791945964, "dur": 21, + "args": { + "External id": 9112,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9111 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791945965, "dur": 4, + "args": { + "External id": 9113,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9112 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791945970, "dur": 14, + "args": { + "External id": 9114,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9113 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791945972, "dur": 11, + "args": { + "External id": 9115,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9114 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946002, "dur": 5, + "args": { + "External id": 9116,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9115 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791946011, "dur": 22, + "args": { + "External id": 9117,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9116 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946012, "dur": 5, + "args": { + "External id": 9118,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9117 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791946018, "dur": 14, + "args": { + "External id": 9119,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9118 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791946020, "dur": 11, + "args": { + "External id": 9120,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9119 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946050, "dur": 5, + "args": { + "External id": 9121,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9120 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791946059, "dur": 21, + "args": { + "External id": 9122,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9121 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946060, "dur": 5, + "args": { + "External id": 9123,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9122 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791946066, "dur": 14, + "args": { + "External id": 9124,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9123 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791946068, "dur": 11, + "args": { + "External id": 9125,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9124 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946098, "dur": 6, + "args": { + "External id": 9126,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9125 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791946108, "dur": 21, + "args": { + "External id": 9127,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9126 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946109, "dur": 4, + "args": { + "External id": 9128,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9127 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791946114, "dur": 14, + "args": { + "External id": 9129,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9128 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791946116, "dur": 11, + "args": { + "External id": 9130,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9129 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946146, "dur": 5, + "args": { + "External id": 9131,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9130 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791946155, "dur": 22, + "args": { + "External id": 9132,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9131 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946156, "dur": 5, + "args": { + "External id": 9133,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9132 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791946162, "dur": 14, + "args": { + "External id": 9134,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9133 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791946164, "dur": 11, + "args": { + "External id": 9135,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9134 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946194, "dur": 5, + "args": { + "External id": 9136,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9135 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791946203, "dur": 22, + "args": { + "External id": 9137,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9136 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946204, "dur": 5, + "args": { + "External id": 9138,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9137 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791946210, "dur": 14, + "args": { + "External id": 9139,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9138 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791946212, "dur": 11, + "args": { + "External id": 9140,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9139 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946243, "dur": 5, + "args": { + "External id": 9141,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9140 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791946252, "dur": 21, + "args": { + "External id": 9142,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9141 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946253, "dur": 4, + "args": { + "External id": 9143,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9142 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791946258, "dur": 14, + "args": { + "External id": 9144,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9143 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791946260, "dur": 11, + "args": { + "External id": 9145,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9144 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946291, "dur": 5, + "args": { + "External id": 9146,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9145 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791946300, "dur": 21, + "args": { + "External id": 9147,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9146 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946301, "dur": 4, + "args": { + "External id": 9148,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9147 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791946306, "dur": 15, + "args": { + "External id": 9149,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9148 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791946308, "dur": 11, + "args": { + "External id": 9150,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9149 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946339, "dur": 5, + "args": { + "External id": 9151,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9150 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791946348, "dur": 21, + "args": { + "External id": 9152,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9151 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946349, "dur": 4, + "args": { + "External id": 9153,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9152 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791946355, "dur": 14, + "args": { + "External id": 9154,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9153 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791946356, "dur": 11, + "args": { + "External id": 9155,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9154 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946387, "dur": 5, + "args": { + "External id": 9156,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9155 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791946396, "dur": 21, + "args": { + "External id": 9157,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9156 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946397, "dur": 5, + "args": { + "External id": 9158,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9157 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791946403, "dur": 14, + "args": { + "External id": 9159,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9158 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791946404, "dur": 11, + "args": { + "External id": 9160,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9159 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946435, "dur": 5, + "args": { + "External id": 9161,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9160 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791946444, "dur": 21, + "args": { + "External id": 9162,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9161 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946445, "dur": 4, + "args": { + "External id": 9163,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9162 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791946450, "dur": 15, + "args": { + "External id": 9164,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9163 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791946452, "dur": 11, + "args": { + "External id": 9165,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9164 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946483, "dur": 5, + "args": { + "External id": 9166,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9165 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791946492, "dur": 21, + "args": { + "External id": 9167,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9166 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946493, "dur": 4, + "args": { + "External id": 9168,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9167 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791946498, "dur": 14, + "args": { + "External id": 9169,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9168 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791946500, "dur": 11, + "args": { + "External id": 9170,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9169 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946530, "dur": 5, + "args": { + "External id": 9171,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9170 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791946539, "dur": 21, + "args": { + "External id": 9172,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9171 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946540, "dur": 5, + "args": { + "External id": 9173,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9172 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791946546, "dur": 14, + "args": { + "External id": 9174,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9173 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791946548, "dur": 11, + "args": { + "External id": 9175,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9174 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946578, "dur": 5, + "args": { + "External id": 9176,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9175 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791946587, "dur": 22, + "args": { + "External id": 9177,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9176 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946589, "dur": 4, + "args": { + "External id": 9178,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9177 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791946594, "dur": 15, + "args": { + "External id": 9179,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9178 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791946596, "dur": 11, + "args": { + "External id": 9180,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9179 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946627, "dur": 5, + "args": { + "External id": 9181,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9180 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791946636, "dur": 21, + "args": { + "External id": 9182,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9181 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946637, "dur": 5, + "args": { + "External id": 9183,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9182 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791946643, "dur": 14, + "args": { + "External id": 9184,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9183 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791946644, "dur": 12, + "args": { + "External id": 9185,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9184 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946675, "dur": 5, + "args": { + "External id": 9186,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9185 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791946684, "dur": 21, + "args": { + "External id": 9187,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9186 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946685, "dur": 5, + "args": { + "External id": 9188,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9187 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791946691, "dur": 14, + "args": { + "External id": 9189,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9188 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791946692, "dur": 11, + "args": { + "External id": 9190,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9189 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946723, "dur": 5, + "args": { + "External id": 9191,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9190 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791946732, "dur": 22, + "args": { + "External id": 9192,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9191 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946733, "dur": 5, + "args": { + "External id": 9193,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9192 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791946739, "dur": 15, + "args": { + "External id": 9194,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9193 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791946741, "dur": 11, + "args": { + "External id": 9195,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9194 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946772, "dur": 6, + "args": { + "External id": 9196,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9195 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791946782, "dur": 21, + "args": { + "External id": 9197,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9196 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946783, "dur": 4, + "args": { + "External id": 9198,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9197 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791946789, "dur": 14, + "args": { + "External id": 9199,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9198 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791946790, "dur": 12, + "args": { + "External id": 9200,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9199 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946821, "dur": 5, + "args": { + "External id": 9201,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9200 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791946830, "dur": 22, + "args": { + "External id": 9202,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9201 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946831, "dur": 5, + "args": { + "External id": 9203,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9202 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791946837, "dur": 14, + "args": { + "External id": 9204,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9203 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791946839, "dur": 11, + "args": { + "External id": 9205,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9204 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946930, "dur": 6, + "args": { + "External id": 9206,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9205 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791946941, "dur": 25, + "args": { + "External id": 9207,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9206 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946942, "dur": 5, + "args": { + "External id": 9208,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9207 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791946948, "dur": 18, + "args": { + "External id": 9209,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9208 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791946950, "dur": 14, + "args": { + "External id": 9210,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9209 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946987, "dur": 5, + "args": { + "External id": 9211,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9210 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791946996, "dur": 22, + "args": { + "External id": 9212,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9211 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791946998, "dur": 4, + "args": { + "External id": 9213,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9212 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791947003, "dur": 15, + "args": { + "External id": 9214,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9213 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791947005, "dur": 11, + "args": { + "External id": 9215,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9214 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947036, "dur": 5, + "args": { + "External id": 9216,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9215 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791947056, "dur": 22, + "args": { + "External id": 9217,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9216 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947057, "dur": 5, + "args": { + "External id": 9218,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9217 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791947063, "dur": 15, + "args": { + "External id": 9219,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9218 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791947065, "dur": 11, + "args": { + "External id": 9220,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9219 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947096, "dur": 11, + "args": { + "External id": 9221,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9220 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791947112, "dur": 23, + "args": { + "External id": 9222,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9221 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947114, "dur": 4, + "args": { + "External id": 9223,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9222 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791947120, "dur": 14, + "args": { + "External id": 9224,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9223 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791947121, "dur": 12, + "args": { + "External id": 9225,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9224 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947153, "dur": 5, + "args": { + "External id": 9226,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9225 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791947162, "dur": 22, + "args": { + "External id": 9227,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9226 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947164, "dur": 4, + "args": { + "External id": 9228,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9227 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791947169, "dur": 15, + "args": { + "External id": 9229,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9228 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791947171, "dur": 11, + "args": { + "External id": 9230,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9229 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947202, "dur": 5, + "args": { + "External id": 9231,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9230 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791947211, "dur": 21, + "args": { + "External id": 9232,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9231 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947212, "dur": 5, + "args": { + "External id": 9233,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9232 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791947218, "dur": 14, + "args": { + "External id": 9234,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9233 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791947220, "dur": 11, + "args": { + "External id": 9235,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9234 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947250, "dur": 5, + "args": { + "External id": 9236,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9235 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791947259, "dur": 21, + "args": { + "External id": 9237,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9236 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947260, "dur": 5, + "args": { + "External id": 9238,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9237 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791947266, "dur": 14, + "args": { + "External id": 9239,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9238 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791947268, "dur": 11, + "args": { + "External id": 9240,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9239 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947298, "dur": 5, + "args": { + "External id": 9241,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9240 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791947307, "dur": 21, + "args": { + "External id": 9242,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9241 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947308, "dur": 5, + "args": { + "External id": 9243,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9242 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791947314, "dur": 14, + "args": { + "External id": 9244,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9243 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791947315, "dur": 11, + "args": { + "External id": 9245,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9244 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947346, "dur": 5, + "args": { + "External id": 9246,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9245 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791947355, "dur": 21, + "args": { + "External id": 9247,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9246 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947356, "dur": 4, + "args": { + "External id": 9248,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9247 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791947362, "dur": 14, + "args": { + "External id": 9249,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9248 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791947363, "dur": 11, + "args": { + "External id": 9250,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9249 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947393, "dur": 6, + "args": { + "External id": 9251,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9250 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791947402, "dur": 22, + "args": { + "External id": 9252,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9251 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947404, "dur": 4, + "args": { + "External id": 9253,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9252 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791947409, "dur": 14, + "args": { + "External id": 9254,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9253 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791947411, "dur": 11, + "args": { + "External id": 9255,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9254 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947494, "dur": 7, + "args": { + "External id": 9256,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9255 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791947505, "dur": 25, + "args": { + "External id": 9257,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9256 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947507, "dur": 4, + "args": { + "External id": 9258,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9257 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791947513, "dur": 16, + "args": { + "External id": 9259,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9258 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791947514, "dur": 14, + "args": { + "External id": 9260,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9259 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947550, "dur": 5, + "args": { + "External id": 9261,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9260 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791947560, "dur": 21, + "args": { + "External id": 9262,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9261 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947561, "dur": 4, + "args": { + "External id": 9263,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9262 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791947567, "dur": 14, + "args": { + "External id": 9264,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9263 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791947568, "dur": 12, + "args": { + "External id": 9265,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9264 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947599, "dur": 6, + "args": { + "External id": 9266,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9265 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791947609, "dur": 21, + "args": { + "External id": 9267,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9266 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947610, "dur": 4, + "args": { + "External id": 9268,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9267 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791947615, "dur": 15, + "args": { + "External id": 9269,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9268 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791947617, "dur": 12, + "args": { + "External id": 9270,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9269 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947648, "dur": 5, + "args": { + "External id": 9271,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9270 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791947657, "dur": 22, + "args": { + "External id": 9272,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9271 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947658, "dur": 5, + "args": { + "External id": 9273,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9272 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791947664, "dur": 14, + "args": { + "External id": 9274,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9273 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791947666, "dur": 11, + "args": { + "External id": 9275,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9274 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947697, "dur": 5, + "args": { + "External id": 9276,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9275 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791947706, "dur": 21, + "args": { + "External id": 9277,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9276 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947707, "dur": 4, + "args": { + "External id": 9278,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9277 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791947713, "dur": 14, + "args": { + "External id": 9279,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9278 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791947714, "dur": 11, + "args": { + "External id": 9280,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9279 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947745, "dur": 6, + "args": { + "External id": 9281,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9280 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791947754, "dur": 22, + "args": { + "External id": 9282,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9281 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947756, "dur": 4, + "args": { + "External id": 9283,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9282 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791947761, "dur": 15, + "args": { + "External id": 9284,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9283 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791947763, "dur": 11, + "args": { + "External id": 9285,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9284 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947794, "dur": 5, + "args": { + "External id": 9286,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9285 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791947803, "dur": 22, + "args": { + "External id": 9287,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9286 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947805, "dur": 4, + "args": { + "External id": 9288,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9287 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791947810, "dur": 14, + "args": { + "External id": 9289,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9288 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791947812, "dur": 11, + "args": { + "External id": 9290,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9289 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947843, "dur": 5, + "args": { + "External id": 9291,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9290 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791947851, "dur": 22, + "args": { + "External id": 9292,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9291 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947853, "dur": 4, + "args": { + "External id": 9293,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9292 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791947858, "dur": 14, + "args": { + "External id": 9294,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9293 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791947860, "dur": 11, + "args": { + "External id": 9295,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9294 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947890, "dur": 5, + "args": { + "External id": 9296,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9295 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791947899, "dur": 21, + "args": { + "External id": 9297,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9296 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947901, "dur": 4, + "args": { + "External id": 9298,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9297 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791947906, "dur": 14, + "args": { + "External id": 9299,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9298 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791947908, "dur": 11, + "args": { + "External id": 9300,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9299 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947938, "dur": 5, + "args": { + "External id": 9301,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9300 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791947947, "dur": 22, + "args": { + "External id": 9302,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9301 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947948, "dur": 5, + "args": { + "External id": 9303,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9302 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791947954, "dur": 14, + "args": { + "External id": 9304,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9303 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791947956, "dur": 11, + "args": { + "External id": 9305,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9304 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947987, "dur": 5, + "args": { + "External id": 9306,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9305 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791947996, "dur": 21, + "args": { + "External id": 9307,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9306 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791947997, "dur": 5, + "args": { + "External id": 9308,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9307 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791948003, "dur": 14, + "args": { + "External id": 9309,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9308 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791948005, "dur": 11, + "args": { + "External id": 9310,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9309 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948035, "dur": 5, + "args": { + "External id": 9311,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9310 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791948044, "dur": 23, + "args": { + "External id": 9312,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9311 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948045, "dur": 5, + "args": { + "External id": 9313,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9312 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791948051, "dur": 16, + "args": { + "External id": 9314,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9313 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791948053, "dur": 13, + "args": { + "External id": 9315,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9314 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948086, "dur": 5, + "args": { + "External id": 9316,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9315 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791948095, "dur": 21, + "args": { + "External id": 9317,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9316 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948096, "dur": 4, + "args": { + "External id": 9318,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9317 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791948101, "dur": 14, + "args": { + "External id": 9319,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9318 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791948103, "dur": 11, + "args": { + "External id": 9320,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9319 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948133, "dur": 5, + "args": { + "External id": 9321,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9320 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791948142, "dur": 21, + "args": { + "External id": 9322,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9321 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948143, "dur": 5, + "args": { + "External id": 9323,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9322 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791948149, "dur": 14, + "args": { + "External id": 9324,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9323 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791948151, "dur": 11, + "args": { + "External id": 9325,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9324 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948181, "dur": 5, + "args": { + "External id": 9326,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9325 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791948190, "dur": 21, + "args": { + "External id": 9327,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9326 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948191, "dur": 4, + "args": { + "External id": 9328,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9327 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791948197, "dur": 13, + "args": { + "External id": 9329,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9328 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791948198, "dur": 11, + "args": { + "External id": 9330,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9329 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948250, "dur": 6, + "args": { + "External id": 9331,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9330 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791948260, "dur": 23, + "args": { + "External id": 9332,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9331 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948262, "dur": 4, + "args": { + "External id": 9333,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9332 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791948267, "dur": 15, + "args": { + "External id": 9334,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9333 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791948269, "dur": 12, + "args": { + "External id": 9335,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9334 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948301, "dur": 5, + "args": { + "External id": 9336,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9335 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791948310, "dur": 21, + "args": { + "External id": 9337,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9336 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948311, "dur": 5, + "args": { + "External id": 9338,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9337 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791948317, "dur": 14, + "args": { + "External id": 9339,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9338 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791948319, "dur": 10, + "args": { + "External id": 9340,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9339 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948348, "dur": 6, + "args": { + "External id": 9341,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9340 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791948357, "dur": 21, + "args": { + "External id": 9342,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9341 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948359, "dur": 4, + "args": { + "External id": 9343,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9342 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791948364, "dur": 14, + "args": { + "External id": 9344,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9343 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791948366, "dur": 11, + "args": { + "External id": 9345,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9344 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948396, "dur": 5, + "args": { + "External id": 9346,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9345 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791948405, "dur": 21, + "args": { + "External id": 9347,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9346 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948406, "dur": 4, + "args": { + "External id": 9348,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9347 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791948411, "dur": 14, + "args": { + "External id": 9349,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9348 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791948413, "dur": 11, + "args": { + "External id": 9350,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9349 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948443, "dur": 5, + "args": { + "External id": 9351,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9350 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791948452, "dur": 21, + "args": { + "External id": 9352,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9351 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948453, "dur": 5, + "args": { + "External id": 9353,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9352 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791948459, "dur": 14, + "args": { + "External id": 9354,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9353 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791948461, "dur": 11, + "args": { + "External id": 9355,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9354 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948491, "dur": 5, + "args": { + "External id": 9356,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9355 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791948500, "dur": 21, + "args": { + "External id": 9357,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9356 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948501, "dur": 4, + "args": { + "External id": 9358,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9357 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791948507, "dur": 13, + "args": { + "External id": 9359,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9358 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791948508, "dur": 11, + "args": { + "External id": 9360,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9359 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948539, "dur": 4, + "args": { + "External id": 9361,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9360 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791948547, "dur": 21, + "args": { + "External id": 9362,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9361 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948549, "dur": 4, + "args": { + "External id": 9363,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9362 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791948554, "dur": 14, + "args": { + "External id": 9364,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9363 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791948556, "dur": 11, + "args": { + "External id": 9365,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9364 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948586, "dur": 5, + "args": { + "External id": 9366,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9365 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791948595, "dur": 21, + "args": { + "External id": 9367,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9366 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948596, "dur": 5, + "args": { + "External id": 9368,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9367 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791948602, "dur": 14, + "args": { + "External id": 9369,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9368 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791948603, "dur": 11, + "args": { + "External id": 9370,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9369 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948633, "dur": 5, + "args": { + "External id": 9371,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9370 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791948642, "dur": 21, + "args": { + "External id": 9372,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9371 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948643, "dur": 5, + "args": { + "External id": 9373,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9372 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791948649, "dur": 14, + "args": { + "External id": 9374,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9373 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791948651, "dur": 11, + "args": { + "External id": 9375,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9374 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948681, "dur": 5, + "args": { + "External id": 9376,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9375 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791948690, "dur": 21, + "args": { + "External id": 9377,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9376 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948691, "dur": 4, + "args": { + "External id": 9378,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9377 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791948696, "dur": 14, + "args": { + "External id": 9379,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9378 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791948698, "dur": 11, + "args": { + "External id": 9380,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9379 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948728, "dur": 5, + "args": { + "External id": 9381,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9380 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791948737, "dur": 21, + "args": { + "External id": 9382,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9381 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948739, "dur": 4, + "args": { + "External id": 9383,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9382 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791948744, "dur": 14, + "args": { + "External id": 9384,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9383 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791948746, "dur": 11, + "args": { + "External id": 9385,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9384 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948776, "dur": 5, + "args": { + "External id": 9386,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9385 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791948785, "dur": 21, + "args": { + "External id": 9387,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9386 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948786, "dur": 5, + "args": { + "External id": 9388,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9387 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791948792, "dur": 14, + "args": { + "External id": 9389,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9388 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791948793, "dur": 11, + "args": { + "External id": 9390,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9389 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948823, "dur": 5, + "args": { + "External id": 9391,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9390 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791948832, "dur": 21, + "args": { + "External id": 9392,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9391 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948833, "dur": 5, + "args": { + "External id": 9393,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9392 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791948839, "dur": 14, + "args": { + "External id": 9394,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9393 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791948840, "dur": 11, + "args": { + "External id": 9395,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9394 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948870, "dur": 5, + "args": { + "External id": 9396,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9395 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791948879, "dur": 21, + "args": { + "External id": 9397,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9396 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948880, "dur": 5, + "args": { + "External id": 9398,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9397 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791948886, "dur": 14, + "args": { + "External id": 9399,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9398 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791948887, "dur": 11, + "args": { + "External id": 9400,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9399 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948917, "dur": 5, + "args": { + "External id": 9401,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9400 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791948926, "dur": 21, + "args": { + "External id": 9402,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9401 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948927, "dur": 5, + "args": { + "External id": 9403,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9402 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791948933, "dur": 14, + "args": { + "External id": 9404,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9403 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791948934, "dur": 11, + "args": { + "External id": 9405,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9404 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948964, "dur": 5, + "args": { + "External id": 9406,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9405 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791948973, "dur": 21, + "args": { + "External id": 9407,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9406 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791948974, "dur": 4, + "args": { + "External id": 9408,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9407 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791948980, "dur": 13, + "args": { + "External id": 9409,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9408 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791948981, "dur": 11, + "args": { + "External id": 9410,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9409 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949011, "dur": 5, + "args": { + "External id": 9411,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9410 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791949020, "dur": 21, + "args": { + "External id": 9412,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9411 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949021, "dur": 5, + "args": { + "External id": 9413,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9412 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791949027, "dur": 14, + "args": { + "External id": 9414,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9413 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791949029, "dur": 10, + "args": { + "External id": 9415,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9414 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949059, "dur": 5, + "args": { + "External id": 9416,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9415 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791949068, "dur": 20, + "args": { + "External id": 9417,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9416 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949069, "dur": 4, + "args": { + "External id": 9418,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9417 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791949074, "dur": 14, + "args": { + "External id": 9419,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9418 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791949076, "dur": 11, + "args": { + "External id": 9420,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9419 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949106, "dur": 5, + "args": { + "External id": 9421,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9420 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791949115, "dur": 20, + "args": { + "External id": 9422,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9421 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949116, "dur": 4, + "args": { + "External id": 9423,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9422 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791949121, "dur": 14, + "args": { + "External id": 9424,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9423 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791949123, "dur": 11, + "args": { + "External id": 9425,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9424 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949153, "dur": 5, + "args": { + "External id": 9426,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9425 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791949162, "dur": 21, + "args": { + "External id": 9427,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9426 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949163, "dur": 4, + "args": { + "External id": 9428,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9427 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791949168, "dur": 14, + "args": { + "External id": 9429,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9428 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791949170, "dur": 11, + "args": { + "External id": 9430,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9429 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949200, "dur": 5, + "args": { + "External id": 9431,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9430 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791949209, "dur": 21, + "args": { + "External id": 9432,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9431 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949210, "dur": 5, + "args": { + "External id": 9433,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9432 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791949216, "dur": 14, + "args": { + "External id": 9434,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9433 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791949218, "dur": 10, + "args": { + "External id": 9435,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9434 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949248, "dur": 5, + "args": { + "External id": 9436,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9435 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791949257, "dur": 21, + "args": { + "External id": 9437,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9436 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949258, "dur": 4, + "args": { + "External id": 9438,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9437 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791949264, "dur": 13, + "args": { + "External id": 9439,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9438 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791949265, "dur": 11, + "args": { + "External id": 9440,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9439 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949295, "dur": 5, + "args": { + "External id": 9441,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9440 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791949304, "dur": 20, + "args": { + "External id": 9442,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9441 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949305, "dur": 4, + "args": { + "External id": 9443,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9442 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791949310, "dur": 14, + "args": { + "External id": 9444,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9443 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791949312, "dur": 11, + "args": { + "External id": 9445,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9444 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949342, "dur": 5, + "args": { + "External id": 9446,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9445 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791949351, "dur": 21, + "args": { + "External id": 9447,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9446 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949352, "dur": 5, + "args": { + "External id": 9448,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9447 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791949358, "dur": 14, + "args": { + "External id": 9449,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9448 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791949360, "dur": 11, + "args": { + "External id": 9450,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9449 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949390, "dur": 5, + "args": { + "External id": 9451,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9450 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791949399, "dur": 20, + "args": { + "External id": 9452,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9451 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949400, "dur": 4, + "args": { + "External id": 9453,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9452 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791949405, "dur": 14, + "args": { + "External id": 9454,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9453 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791949407, "dur": 11, + "args": { + "External id": 9455,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9454 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949437, "dur": 5, + "args": { + "External id": 9456,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9455 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791949446, "dur": 21, + "args": { + "External id": 9457,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9456 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949447, "dur": 4, + "args": { + "External id": 9458,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9457 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791949453, "dur": 13, + "args": { + "External id": 9459,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9458 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791949454, "dur": 11, + "args": { + "External id": 9460,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9459 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949484, "dur": 5, + "args": { + "External id": 9461,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9460 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791949493, "dur": 21, + "args": { + "External id": 9462,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9461 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949494, "dur": 5, + "args": { + "External id": 9463,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9462 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791949500, "dur": 14, + "args": { + "External id": 9464,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9463 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791949501, "dur": 11, + "args": { + "External id": 9465,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9464 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949532, "dur": 5, + "args": { + "External id": 9466,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9465 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791949540, "dur": 21, + "args": { + "External id": 9467,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9466 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949542, "dur": 4, + "args": { + "External id": 9468,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9467 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791949547, "dur": 14, + "args": { + "External id": 9469,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9468 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791949549, "dur": 11, + "args": { + "External id": 9470,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9469 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949579, "dur": 5, + "args": { + "External id": 9471,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9470 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791949588, "dur": 20, + "args": { + "External id": 9472,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9471 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949589, "dur": 4, + "args": { + "External id": 9473,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9472 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791949594, "dur": 14, + "args": { + "External id": 9474,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9473 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791949596, "dur": 11, + "args": { + "External id": 9475,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9474 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949626, "dur": 5, + "args": { + "External id": 9476,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9475 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791949635, "dur": 20, + "args": { + "External id": 9477,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9476 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949636, "dur": 4, + "args": { + "External id": 9478,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9477 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791949641, "dur": 14, + "args": { + "External id": 9479,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9478 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791949643, "dur": 11, + "args": { + "External id": 9480,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9479 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949673, "dur": 5, + "args": { + "External id": 9481,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9480 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791949682, "dur": 20, + "args": { + "External id": 9482,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9481 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949683, "dur": 4, + "args": { + "External id": 9483,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9482 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791949689, "dur": 13, + "args": { + "External id": 9484,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9483 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791949690, "dur": 11, + "args": { + "External id": 9485,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9484 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949720, "dur": 5, + "args": { + "External id": 9486,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9485 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791949729, "dur": 20, + "args": { + "External id": 9487,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9486 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949730, "dur": 4, + "args": { + "External id": 9488,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9487 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791949736, "dur": 13, + "args": { + "External id": 9489,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9488 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791949737, "dur": 11, + "args": { + "External id": 9490,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9489 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949767, "dur": 5, + "args": { + "External id": 9491,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9490 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791949776, "dur": 21, + "args": { + "External id": 9492,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9491 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949777, "dur": 4, + "args": { + "External id": 9493,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9492 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791949783, "dur": 13, + "args": { + "External id": 9494,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9493 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791949784, "dur": 11, + "args": { + "External id": 9495,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9494 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949814, "dur": 5, + "args": { + "External id": 9496,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9495 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791949823, "dur": 21, + "args": { + "External id": 9497,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9496 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949824, "dur": 4, + "args": { + "External id": 9498,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9497 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791949830, "dur": 13, + "args": { + "External id": 9499,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9498 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791949831, "dur": 11, + "args": { + "External id": 9500,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9499 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949861, "dur": 5, + "args": { + "External id": 9501,"Record function id": 0, "Concrete Inputs": ["[1, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9500 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649791949870, "dur": 20, + "args": { + "External id": 9502,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9501 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791949871, "dur": 4, + "args": { + "External id": 9503,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9502 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649791949876, "dur": 14, + "args": { + "External id": 9504,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 9503 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649791949878, "dur": 11, + "args": { + "External id": 9505,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 9504 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::randn", "pid": 435, "tid": 435, + "ts": 1776649791950206, "dur": 73, + "args": { + "External id": 9506,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 9505 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791950211, "dur": 9, + "args": { + "External id": 9507,"Record function id": 0, "Concrete Inputs": ["[4, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9506 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::normal_", "pid": 435, "tid": 435, + "ts": 1776649791950232, "dur": 47, + "args": { + "External id": 9508,"Record function id": 0, "Concrete Inputs": ["", "0.", "1.", ""], "Input type": ["c10::Half", "Scalar", "Scalar", ""], "Input Dims": [[4, 4096], [], [], []], "Ev Idx": 9507 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::randn", "pid": 435, "tid": 435, + "ts": 1776649791950345, "dur": 32, + "args": { + "External id": 9509,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "", "5", "", "", "False"], "Input type": ["ScalarList", "", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9508 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791950349, "dur": 6, + "args": { + "External id": 9510,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9509 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::normal_", "pid": 435, "tid": 435, + "ts": 1776649791950357, "dur": 19, + "args": { + "External id": 9511,"Record function id": 0, "Concrete Inputs": ["", "0.", "1.", ""], "Input type": ["c10::Half", "Scalar", "Scalar", ""], "Input Dims": [[4096, 8192], [], [], []], "Ev Idx": 9510 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 435, "tid": 435, + "ts": 1776649791950434, "dur": 42, + "args": { + "External id": 9512,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "double"], "Input Dims": [[4096, 8192], []], "Ev Idx": 9511 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791950507, "dur": 14, + "args": { + "External id": 9513,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 8192], []], "Ev Idx": 9512 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791950511, "dur": 10, + "args": { + "External id": 9514,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 8192], []], "Ev Idx": 9513 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 435, "tid": 435, + "ts": 1776649791950527, "dur": 60, + "args": { + "External id": 9515,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[32, 128, 8192]], "Ev Idx": 9514 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791950532, "dur": 6, + "args": { + "External id": 9516,"Record function id": 0, "Concrete Inputs": ["[0]", "5", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9515 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 435, "tid": 435, + "ts": 1776649791950541, "dur": 45, + "args": { + "External id": 9517,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[32, 128, 8192], [0]], "Ev Idx": 9516 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 435, "tid": 435, + "ts": 1776649791950551, "dur": 8, + "args": { + "External id": 9518,"Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]", ""], "Input type": ["c10::Half", "ScalarList", ""], "Input Dims": [[0], [], []], "Ev Idx": 9517 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 435, "tid": 435, + "ts": 1776649791950620, "dur": 48, + "args": { + "External id": 9519,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1]", "False"], "Input type": ["c10::Half", "ScalarList", "Scalar"], "Input Dims": [[32, 128, 8192], [], []], "Ev Idx": 9518 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791950637, "dur": 2, + "args": { + "External id": 9520,"Record function id": 0, "Concrete Inputs": ["", "[32, 1, 8192]", "[8192, 0, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 9519 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 435, "tid": 435, + "ts": 1776649791950682, "dur": 27, + "args": { + "External id": 9521,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "double"], "Input Dims": [[32, 8192], []], "Ev Idx": 9520 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 435, "tid": 435, + "ts": 1776649791950735, "dur": 37, + "args": { + "External id": 9522,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "9.9999999999999995e-07", ""], "Input type": ["c10::Half", "Scalar", ""], "Input Dims": [[32, 8192], [], []], "Ev Idx": 9521 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791950740, "dur": 1, + "args": { + "External id": 9523,"Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 8192], [], [], [], []], "Ev Idx": 9522 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791950786, "dur": 0, + "args": { + "External id": 9524,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 8192], [], [], [], []], "Ev Idx": 9523 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649791950797, "dur": 8, + "args": { + "External id": 9525,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 8192], []], "Ev Idx": 9524 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791950803, "dur": 1, + "args": { + "External id": 9526,"Record function id": 0, "Concrete Inputs": ["", "[32, 1, 8192]", "[8192, 8192, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 9525 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 435, "tid": 435, + "ts": 1776649791950808, "dur": 28, + "args": { + "External id": 9527,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[32, 128, 8192], [32, 1, 8192]], "Ev Idx": 9526 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 435, "tid": 435, + "ts": 1776649791950843, "dur": 23, + "args": { + "External id": 9528,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[32, 128, 8192]], "Ev Idx": 9527 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 435, "tid": 435, + "ts": 1776649791950873, "dur": 16, + "args": { + "External id": 9529,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "-8", "7"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[32, 128, 8192], [], []], "Ev Idx": 9528 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791950874, "dur": 0, + "args": { + "External id": 9530,"Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 128, 8192], [], [], [], []], "Ev Idx": 9529 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791950902, "dur": 45, + "args": { + "External id": 9531,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 128, 8192], [], [], [], []], "Ev Idx": 9530 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791950905, "dur": 41, + "args": { + "External id": 9532,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["c10::Half", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[32, 128, 8192], [], [], [], [], [], []], "Ev Idx": 9531 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791950913, "dur": 8, + "args": { + "External id": 9533,"Record function id": 0, "Concrete Inputs": ["[32, 128, 8192]", "[1048576, 8192, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9532 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791950924, "dur": 22, + "args": { + "External id": 9534,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "c10::Half", "Scalar"], "Input Dims": [[32, 128, 8192], [32, 128, 8192], []], "Ev Idx": 9533 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 435, "tid": 435, + "ts": 1776649791950961, "dur": 29, + "args": { + "External id": 9535,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["signed char", "long int", "Scalar"], "Input Dims": [[32, 128, 8192], [], []], "Ev Idx": 9534 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791950995, "dur": 25, + "args": { + "External id": 9536,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "False", "False", ""], "Input type": ["signed char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 128, 8192], [], [], [], []], "Ev Idx": 9535 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791950996, "dur": 24, + "args": { + "External id": 9537,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "", "", "", "False", ""], "Input type": ["signed char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[32, 128, 8192], [], [], [], [], [], []], "Ev Idx": 9536 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791950998, "dur": 6, + "args": { + "External id": 9538,"Record function id": 0, "Concrete Inputs": ["[32, 128, 8192]", "[1048576, 8192, 1]", "0", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9537 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791951005, "dur": 14, + "args": { + "External id": 9539,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["unsigned char", "signed char", "Scalar"], "Input Dims": [[32, 128, 8192], [32, 128, 8192], []], "Ev Idx": 9538 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791951027, "dur": 5, + "args": { + "External id": 9540,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["unsigned char", "ScalarList"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 9539 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791951029, "dur": 2, + "args": { + "External id": 9541,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["unsigned char", "ScalarList"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 9540 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649791951042, "dur": 8, + "args": { + "External id": 9542,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 9541 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791951046, "dur": 1, + "args": { + "External id": 9543,"Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]", "[8192, 1]", "0"], "Input type": ["unsigned char", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 8192], [], [], []], "Ev Idx": 9542 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649791951052, "dur": 4, + "args": { + "External id": 9544,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "2"], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 9543 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791951053, "dur": 2, + "args": { + "External id": 9545,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]", "[8192, 2]", "0"], "Input type": ["unsigned char", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 8192], [], [], []], "Ev Idx": 9544 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649791951062, "dur": 3, + "args": { + "External id": 9546,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 9545 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791951064, "dur": 1, + "args": { + "External id": 9547,"Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]", "[8192, 1]", "0"], "Input type": ["unsigned char", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 8192], [], [], []], "Ev Idx": 9546 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649791951067, "dur": 3, + "args": { + "External id": 9548,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "1", "9223372036854775807", "2"], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 9547 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791951069, "dur": 0, + "args": { + "External id": 9549,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]", "[8192, 2]", "1"], "Input type": ["unsigned char", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 8192], [], [], []], "Ev Idx": 9548 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__lshift__", "pid": 435, "tid": 435, + "ts": 1776649791951077, "dur": 47, + "args": { + "External id": 9550,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 9549 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791951089, "dur": 6, + "args": { + "External id": 9551,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "0", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9550 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__or__", "pid": 435, "tid": 435, + "ts": 1776649791951130, "dur": 33, + "args": { + "External id": 9552,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["unsigned char", "unsigned char"], "Input Dims": [[4096, 4096], [4096, 4096]], "Ev Idx": 9551 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_or", "pid": 435, "tid": 435, + "ts": 1776649791951135, "dur": 27, + "args": { + "External id": 9553,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["unsigned char", "unsigned char"], "Input Dims": [[4096, 4096], [4096, 4096]], "Ev Idx": 9552 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791951168, "dur": 1, + "args": { + "External id": 9554,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 9553 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__and__", "pid": 435, "tid": 435, + "ts": 1776649791951187, "dur": 37, + "args": { + "External id": 9555,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "15"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 9554 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_and", "pid": 435, "tid": 435, + "ts": 1776649791951190, "dur": 33, + "args": { + "External id": 9556,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "15"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 9555 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_and", "pid": 435, "tid": 435, + "ts": 1776649791951196, "dur": 26, + "args": { + "External id": 9557,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["unsigned char", "long int"], "Input Dims": [[4096, 4096], []], "Ev Idx": 9556 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791951228, "dur": 26, + "args": { + "External id": 9558,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 9557 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791951229, "dur": 25, + "args": { + "External id": 9559,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["unsigned char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], [], []], "Ev Idx": 9558 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791951232, "dur": 6, + "args": { + "External id": 9560,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "[4096, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9559 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791951239, "dur": 14, + "args": { + "External id": 9561,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "unsigned char", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 9560 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__rshift__", "pid": 435, "tid": 435, + "ts": 1776649791951262, "dur": 29, + "args": { + "External id": 9562,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 9561 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791951269, "dur": 5, + "args": { + "External id": 9563,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "0", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9562 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791951295, "dur": 22, + "args": { + "External id": 9564,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 9563 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791951296, "dur": 20, + "args": { + "External id": 9565,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["unsigned char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], [], []], "Ev Idx": 9564 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791951298, "dur": 5, + "args": { + "External id": 9566,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "[4096, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9565 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791951304, "dur": 12, + "args": { + "External id": 9567,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "unsigned char", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 9566 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::stack", "pid": 435, "tid": 435, + "ts": 1776649791951348, "dur": 72, + "args": { + "External id": 9568,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["TensorList", "Scalar"], "Input Dims": [[[4096, 4096], [4096, 4096]], []], "Ev Idx": 9567 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649791951354, "dur": 3, + "args": { + "External id": 9569,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["signed char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 9568 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791951356, "dur": 1, + "args": { + "External id": 9570,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096, 1]", "[4096, 1, 1]", ""], "Input type": ["signed char", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 9569 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649791951358, "dur": 2, + "args": { + "External id": 9571,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["signed char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 9570 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791951360, "dur": 0, + "args": { + "External id": 9572,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096, 1]", "[4096, 1, 1]", ""], "Input type": ["signed char", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 9571 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 435, "tid": 435, + "ts": 1776649791951363, "dur": 55, + "args": { + "External id": 9573,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["TensorList", "Scalar"], "Input Dims": [[[4096, 4096, 1], [4096, 4096, 1]], []], "Ev Idx": 9572 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791951428, "dur": 4, + "args": { + "External id": 9574,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["signed char", "ScalarList"], "Input Dims": [[4096, 4096, 2], []], "Ev Idx": 9573 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791951429, "dur": 3, + "args": { + "External id": 9575,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["signed char", "ScalarList"], "Input Dims": [[4096, 4096, 2], []], "Ev Idx": 9574 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::repeat_interleave", "pid": 435, "tid": 435, + "ts": 1776649791951442, "dur": 73, + "args": { + "External id": 9576,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "128", "0", ""], "Input type": ["c10::Half", "Scalar", "Scalar", ""], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 9575 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649791951445, "dur": 3, + "args": { + "External id": 9577,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 8192], []], "Ev Idx": 9576 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791951446, "dur": 1, + "args": { + "External id": 9578,"Record function id": 0, "Concrete Inputs": ["", "[32, 1, 8192]", "[8192, 8192, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 9577 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 435, "tid": 435, + "ts": 1776649791951451, "dur": 7, + "args": { + "External id": 9579,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]", "False"], "Input type": ["c10::Half", "ScalarList", "Scalar"], "Input Dims": [[32, 1, 8192], [], []], "Ev Idx": 9578 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791951456, "dur": 1, + "args": { + "External id": 9580,"Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]", "[8192, 0, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 1, 8192], [], [], []], "Ev Idx": 9579 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649791951460, "dur": 46, + "args": { + "External id": 9581,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 9580 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649791951466, "dur": 18, + "args": { + "External id": 9582,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[32, 128, 8192], [], [], [], [], []], "Ev Idx": 9581 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791951469, "dur": 7, + "args": { + "External id": 9583,"Record function id": 0, "Concrete Inputs": ["[32, 128, 8192]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9582 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791951485, "dur": 21, + "args": { + "External id": 9584,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[32, 128, 8192], [32, 128, 8192], []], "Ev Idx": 9583 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::flatten", "pid": 435, "tid": 435, + "ts": 1776649791951509, "dur": 6, + "args": { + "External id": 9585,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[32, 128, 8192], [], []], "Ev Idx": 9584 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791951512, "dur": 2, + "args": { + "External id": 9586,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 9585 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 435, "tid": 435, + "ts": 1776649791951527, "dur": 32, + "args": { + "External id": 9587,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["signed char", "long int", "Scalar"], "Input Dims": [[4096, 8192], [], []], "Ev Idx": 9586 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791951567, "dur": 27, + "args": { + "External id": 9588,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["signed char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 9587 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791951568, "dur": 26, + "args": { + "External id": 9589,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "", "", "", "False", ""], "Input type": ["signed char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 8192], [], [], [], [], [], []], "Ev Idx": 9588 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791951572, "dur": 6, + "args": { + "External id": 9590,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9589 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791951579, "dur": 14, + "args": { + "External id": 9591,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "signed char", "Scalar"], "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 9590 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 435, "tid": 435, + "ts": 1776649791951600, "dur": 19, + "args": { + "External id": 9592,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4096, 8192], [4096, 8192]], "Ev Idx": 9591 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791951682, "dur": 224, + "args": { + "External id": 9593,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9592 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791951688, "dur": 216, + "args": { + "External id": 9594,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9593 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791951915, "dur": 33, + "args": { + "External id": 9595,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9594 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791951918, "dur": 29, + "args": { + "External id": 9596,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9595 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791951953, "dur": 27, + "args": { + "External id": 9597,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9596 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791951955, "dur": 24, + "args": { + "External id": 9598,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9597 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791951985, "dur": 26, + "args": { + "External id": 9599,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9598 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791951987, "dur": 24, + "args": { + "External id": 9600,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9599 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791952016, "dur": 26, + "args": { + "External id": 9601,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9600 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791952019, "dur": 23, + "args": { + "External id": 9602,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9601 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791952047, "dur": 26, + "args": { + "External id": 9603,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9602 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791952049, "dur": 23, + "args": { + "External id": 9604,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9603 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791952077, "dur": 26, + "args": { + "External id": 9605,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9604 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791952080, "dur": 22, + "args": { + "External id": 9606,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9605 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791952108, "dur": 25, + "args": { + "External id": 9607,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9606 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791952110, "dur": 23, + "args": { + "External id": 9608,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9607 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791952138, "dur": 26, + "args": { + "External id": 9609,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9608 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791952140, "dur": 23, + "args": { + "External id": 9610,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9609 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791952168, "dur": 26, + "args": { + "External id": 9611,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9610 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791952171, "dur": 22, + "args": { + "External id": 9612,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9611 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953085, "dur": 37, + "args": { + "External id": 9613,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9612 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953089, "dur": 32, + "args": { + "External id": 9614,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9613 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953128, "dur": 27, + "args": { + "External id": 9615,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9614 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953131, "dur": 24, + "args": { + "External id": 9616,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9615 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953160, "dur": 26, + "args": { + "External id": 9617,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9616 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953162, "dur": 23, + "args": { + "External id": 9618,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9617 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953191, "dur": 26, + "args": { + "External id": 9619,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9618 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953193, "dur": 23, + "args": { + "External id": 9620,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9619 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953221, "dur": 26, + "args": { + "External id": 9621,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9620 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953223, "dur": 23, + "args": { + "External id": 9622,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9621 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953251, "dur": 26, + "args": { + "External id": 9623,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9622 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953253, "dur": 23, + "args": { + "External id": 9624,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9623 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953281, "dur": 26, + "args": { + "External id": 9625,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9624 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953284, "dur": 22, + "args": { + "External id": 9626,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9625 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953312, "dur": 25, + "args": { + "External id": 9627,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9626 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953314, "dur": 22, + "args": { + "External id": 9628,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9627 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953341, "dur": 26, + "args": { + "External id": 9629,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9628 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953344, "dur": 22, + "args": { + "External id": 9630,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9629 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953371, "dur": 26, + "args": { + "External id": 9631,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9630 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953374, "dur": 23, + "args": { + "External id": 9632,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9631 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953402, "dur": 25, + "args": { + "External id": 9633,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9632 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953404, "dur": 23, + "args": { + "External id": 9634,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9633 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953432, "dur": 25, + "args": { + "External id": 9635,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9634 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953434, "dur": 23, + "args": { + "External id": 9636,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9635 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953462, "dur": 26, + "args": { + "External id": 9637,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9636 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953464, "dur": 23, + "args": { + "External id": 9638,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9637 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953493, "dur": 27, + "args": { + "External id": 9639,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9638 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953496, "dur": 23, + "args": { + "External id": 9640,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9639 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953524, "dur": 26, + "args": { + "External id": 9641,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9640 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953527, "dur": 23, + "args": { + "External id": 9642,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9641 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953555, "dur": 26, + "args": { + "External id": 9643,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9642 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953557, "dur": 23, + "args": { + "External id": 9644,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9643 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953585, "dur": 26, + "args": { + "External id": 9645,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9644 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953588, "dur": 22, + "args": { + "External id": 9646,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9645 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953616, "dur": 26, + "args": { + "External id": 9647,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9646 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953618, "dur": 23, + "args": { + "External id": 9648,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9647 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953646, "dur": 27, + "args": { + "External id": 9649,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9648 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953649, "dur": 23, + "args": { + "External id": 9650,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9649 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953677, "dur": 26, + "args": { + "External id": 9651,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9650 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953680, "dur": 22, + "args": { + "External id": 9652,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9651 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953708, "dur": 26, + "args": { + "External id": 9653,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9652 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953710, "dur": 23, + "args": { + "External id": 9654,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9653 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953738, "dur": 26, + "args": { + "External id": 9655,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9654 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953741, "dur": 23, + "args": { + "External id": 9656,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9655 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953769, "dur": 26, + "args": { + "External id": 9657,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9656 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953772, "dur": 22, + "args": { + "External id": 9658,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9657 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953800, "dur": 26, + "args": { + "External id": 9659,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9658 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953802, "dur": 23, + "args": { + "External id": 9660,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9659 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953830, "dur": 26, + "args": { + "External id": 9661,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9660 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953833, "dur": 22, + "args": { + "External id": 9662,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9661 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953861, "dur": 25, + "args": { + "External id": 9663,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9662 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953863, "dur": 23, + "args": { + "External id": 9664,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9663 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953891, "dur": 26, + "args": { + "External id": 9665,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9664 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953893, "dur": 23, + "args": { + "External id": 9666,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9665 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953921, "dur": 26, + "args": { + "External id": 9667,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9666 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953924, "dur": 23, + "args": { + "External id": 9668,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9667 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953952, "dur": 26, + "args": { + "External id": 9669,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9668 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953954, "dur": 23, + "args": { + "External id": 9670,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9669 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791953982, "dur": 26, + "args": { + "External id": 9671,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9670 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791953984, "dur": 23, + "args": { + "External id": 9672,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9671 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791954013, "dur": 26, + "args": { + "External id": 9673,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9672 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791954015, "dur": 23, + "args": { + "External id": 9674,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9673 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791954043, "dur": 26, + "args": { + "External id": 9675,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9674 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791954046, "dur": 23, + "args": { + "External id": 9676,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9675 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791954074, "dur": 26, + "args": { + "External id": 9677,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9676 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791954076, "dur": 23, + "args": { + "External id": 9678,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9677 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791954104, "dur": 27, + "args": { + "External id": 9679,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9678 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791954107, "dur": 23, + "args": { + "External id": 9680,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9679 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791954135, "dur": 26, + "args": { + "External id": 9681,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9680 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791954138, "dur": 22, + "args": { + "External id": 9682,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9681 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791954166, "dur": 26, + "args": { + "External id": 9683,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9682 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791954168, "dur": 23, + "args": { + "External id": 9684,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9683 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791954196, "dur": 26, + "args": { + "External id": 9685,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9684 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791954199, "dur": 22, + "args": { + "External id": 9686,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9685 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791954227, "dur": 26, + "args": { + "External id": 9687,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9686 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791954229, "dur": 23, + "args": { + "External id": 9688,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9687 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791954257, "dur": 26, + "args": { + "External id": 9689,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9688 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791954260, "dur": 22, + "args": { + "External id": 9690,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9689 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791954288, "dur": 26, + "args": { + "External id": 9691,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9690 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791954290, "dur": 23, + "args": { + "External id": 9692,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9691 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791954319, "dur": 26, + "args": { + "External id": 9693,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9692 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791954321, "dur": 23, + "args": { + "External id": 9694,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9693 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791954350, "dur": 26, + "args": { + "External id": 9695,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9694 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791954352, "dur": 23, + "args": { + "External id": 9696,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9695 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791954380, "dur": 26, + "args": { + "External id": 9697,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9696 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791954382, "dur": 23, + "args": { + "External id": 9698,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9697 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791954410, "dur": 27, + "args": { + "External id": 9699,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9698 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791954413, "dur": 23, + "args": { + "External id": 9700,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9699 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791954441, "dur": 26, + "args": { + "External id": 9701,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9700 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791954444, "dur": 22, + "args": { + "External id": 9702,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9701 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791954471, "dur": 26, + "args": { + "External id": 9703,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9702 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791954474, "dur": 22, + "args": { + "External id": 9704,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9703 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791954502, "dur": 26, + "args": { + "External id": 9705,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9704 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791954505, "dur": 22, + "args": { + "External id": 9706,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9705 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791954532, "dur": 26, + "args": { + "External id": 9707,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9706 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791954535, "dur": 23, + "args": { + "External id": 9708,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9707 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791954563, "dur": 26, + "args": { + "External id": 9709,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9708 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791954565, "dur": 23, + "args": { + "External id": 9710,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9709 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791954593, "dur": 26, + "args": { + "External id": 9711,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9710 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791954595, "dur": 23, + "args": { + "External id": 9712,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9711 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791955577, "dur": 35, + "args": { + "External id": 9713,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9712 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791955581, "dur": 30, + "args": { + "External id": 9714,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9713 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791955618, "dur": 27, + "args": { + "External id": 9715,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9714 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791955620, "dur": 24, + "args": { + "External id": 9716,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9715 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791955650, "dur": 27, + "args": { + "External id": 9717,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9716 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791955653, "dur": 23, + "args": { + "External id": 9718,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9717 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791955681, "dur": 27, + "args": { + "External id": 9719,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9718 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791955684, "dur": 23, + "args": { + "External id": 9720,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9719 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791955712, "dur": 27, + "args": { + "External id": 9721,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9720 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791955715, "dur": 23, + "args": { + "External id": 9722,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9721 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791955743, "dur": 26, + "args": { + "External id": 9723,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9722 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791955746, "dur": 23, + "args": { + "External id": 9724,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9723 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791955774, "dur": 26, + "args": { + "External id": 9725,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9724 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791955776, "dur": 23, + "args": { + "External id": 9726,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9725 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791955805, "dur": 26, + "args": { + "External id": 9727,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9726 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791955807, "dur": 23, + "args": { + "External id": 9728,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9727 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791955847, "dur": 26, + "args": { + "External id": 9729,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9728 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791955849, "dur": 24, + "args": { + "External id": 9730,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9729 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791955878, "dur": 27, + "args": { + "External id": 9731,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9730 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791955881, "dur": 23, + "args": { + "External id": 9732,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9731 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956122, "dur": 34, + "args": { + "External id": 9733,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9732 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956126, "dur": 29, + "args": { + "External id": 9734,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9733 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956161, "dur": 27, + "args": { + "External id": 9735,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9734 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956164, "dur": 23, + "args": { + "External id": 9736,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9735 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956193, "dur": 46, + "args": { + "External id": 9737,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9736 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956195, "dur": 42, + "args": { + "External id": 9738,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9737 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956245, "dur": 28, + "args": { + "External id": 9739,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9738 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956247, "dur": 26, + "args": { + "External id": 9740,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9739 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956278, "dur": 27, + "args": { + "External id": 9741,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9740 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956281, "dur": 23, + "args": { + "External id": 9742,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9741 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956309, "dur": 27, + "args": { + "External id": 9743,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9742 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956312, "dur": 23, + "args": { + "External id": 9744,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9743 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956341, "dur": 26, + "args": { + "External id": 9745,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9744 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956343, "dur": 23, + "args": { + "External id": 9746,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9745 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956371, "dur": 27, + "args": { + "External id": 9747,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9746 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956374, "dur": 23, + "args": { + "External id": 9748,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9747 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956402, "dur": 26, + "args": { + "External id": 9749,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9748 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956405, "dur": 23, + "args": { + "External id": 9750,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9749 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956433, "dur": 27, + "args": { + "External id": 9751,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9750 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956435, "dur": 24, + "args": { + "External id": 9752,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9751 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956464, "dur": 26, + "args": { + "External id": 9753,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9752 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956467, "dur": 23, + "args": { + "External id": 9754,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9753 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956495, "dur": 26, + "args": { + "External id": 9755,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9754 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956498, "dur": 23, + "args": { + "External id": 9756,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9755 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956526, "dur": 26, + "args": { + "External id": 9757,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9756 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956529, "dur": 22, + "args": { + "External id": 9758,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9757 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956557, "dur": 26, + "args": { + "External id": 9759,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9758 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956559, "dur": 24, + "args": { + "External id": 9760,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9759 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956588, "dur": 26, + "args": { + "External id": 9761,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9760 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956590, "dur": 23, + "args": { + "External id": 9762,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9761 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956618, "dur": 26, + "args": { + "External id": 9763,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9762 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956621, "dur": 22, + "args": { + "External id": 9764,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9763 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956648, "dur": 26, + "args": { + "External id": 9765,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9764 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956651, "dur": 23, + "args": { + "External id": 9766,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9765 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956681, "dur": 26, + "args": { + "External id": 9767,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9766 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956683, "dur": 23, + "args": { + "External id": 9768,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9767 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956711, "dur": 26, + "args": { + "External id": 9769,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9768 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956714, "dur": 23, + "args": { + "External id": 9770,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9769 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956742, "dur": 26, + "args": { + "External id": 9771,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9770 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956744, "dur": 23, + "args": { + "External id": 9772,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9771 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956772, "dur": 26, + "args": { + "External id": 9773,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9772 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956774, "dur": 23, + "args": { + "External id": 9774,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9773 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956802, "dur": 26, + "args": { + "External id": 9775,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9774 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956805, "dur": 22, + "args": { + "External id": 9776,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9775 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956833, "dur": 25, + "args": { + "External id": 9777,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9776 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956835, "dur": 23, + "args": { + "External id": 9778,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9777 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956863, "dur": 26, + "args": { + "External id": 9779,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9778 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956865, "dur": 23, + "args": { + "External id": 9780,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9779 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956893, "dur": 26, + "args": { + "External id": 9781,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9780 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956896, "dur": 22, + "args": { + "External id": 9782,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9781 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956924, "dur": 26, + "args": { + "External id": 9783,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9782 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956926, "dur": 24, + "args": { + "External id": 9784,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9783 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956955, "dur": 26, + "args": { + "External id": 9785,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9784 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956957, "dur": 23, + "args": { + "External id": 9786,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9785 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791956986, "dur": 26, + "args": { + "External id": 9787,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9786 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791956988, "dur": 23, + "args": { + "External id": 9788,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9787 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791957016, "dur": 26, + "args": { + "External id": 9789,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9788 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791957018, "dur": 23, + "args": { + "External id": 9790,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9789 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791957047, "dur": 26, + "args": { + "External id": 9791,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9790 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791957049, "dur": 23, + "args": { + "External id": 9792,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9791 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791957077, "dur": 31, + "args": { + "External id": 9793,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9792 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791957079, "dur": 29, + "args": { + "External id": 9794,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9793 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791957113, "dur": 28, + "args": { + "External id": 9795,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9794 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791957116, "dur": 24, + "args": { + "External id": 9796,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9795 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791957146, "dur": 26, + "args": { + "External id": 9797,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9796 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791957148, "dur": 23, + "args": { + "External id": 9798,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9797 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791957177, "dur": 26, + "args": { + "External id": 9799,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9798 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791957179, "dur": 23, + "args": { + "External id": 9800,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9799 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791957207, "dur": 26, + "args": { + "External id": 9801,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9800 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791957210, "dur": 22, + "args": { + "External id": 9802,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9801 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791957238, "dur": 26, + "args": { + "External id": 9803,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9802 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791957240, "dur": 23, + "args": { + "External id": 9804,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9803 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791957269, "dur": 26, + "args": { + "External id": 9805,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9804 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791957271, "dur": 23, + "args": { + "External id": 9806,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9805 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791957299, "dur": 27, + "args": { + "External id": 9807,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9806 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791957302, "dur": 23, + "args": { + "External id": 9808,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9807 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791957330, "dur": 26, + "args": { + "External id": 9809,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9808 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791957333, "dur": 22, + "args": { + "External id": 9810,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9809 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791957361, "dur": 25, + "args": { + "External id": 9811,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9810 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791957363, "dur": 23, + "args": { + "External id": 9812,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9811 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791957391, "dur": 25, + "args": { + "External id": 9813,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9812 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791957393, "dur": 23, + "args": { + "External id": 9814,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9813 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791957421, "dur": 25, + "args": { + "External id": 9815,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9814 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791957423, "dur": 23, + "args": { + "External id": 9816,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9815 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791957451, "dur": 26, + "args": { + "External id": 9817,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9816 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791957453, "dur": 23, + "args": { + "External id": 9818,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9817 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791957481, "dur": 26, + "args": { + "External id": 9819,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9818 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791957483, "dur": 23, + "args": { + "External id": 9820,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9819 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791957511, "dur": 26, + "args": { + "External id": 9821,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9820 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791957514, "dur": 22, + "args": { + "External id": 9822,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9821 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791957542, "dur": 25, + "args": { + "External id": 9823,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9822 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791957544, "dur": 23, + "args": { + "External id": 9824,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9823 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791957572, "dur": 25, + "args": { + "External id": 9825,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9824 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791957574, "dur": 23, + "args": { + "External id": 9826,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9825 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791957602, "dur": 26, + "args": { + "External id": 9827,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9826 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791957604, "dur": 23, + "args": { + "External id": 9828,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9827 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791957632, "dur": 26, + "args": { + "External id": 9829,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9828 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791957635, "dur": 22, + "args": { + "External id": 9830,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9829 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791957663, "dur": 25, + "args": { + "External id": 9831,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9830 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791957665, "dur": 22, + "args": { + "External id": 9832,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9831 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791958614, "dur": 35, + "args": { + "External id": 9833,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9832 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791958617, "dur": 31, + "args": { + "External id": 9834,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9833 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791958654, "dur": 28, + "args": { + "External id": 9835,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9834 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791958657, "dur": 24, + "args": { + "External id": 9836,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9835 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791958686, "dur": 26, + "args": { + "External id": 9837,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9836 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791958689, "dur": 22, + "args": { + "External id": 9838,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9837 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791958717, "dur": 26, + "args": { + "External id": 9839,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9838 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791958719, "dur": 23, + "args": { + "External id": 9840,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9839 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791958747, "dur": 26, + "args": { + "External id": 9841,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9840 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791958750, "dur": 22, + "args": { + "External id": 9842,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9841 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791958777, "dur": 26, + "args": { + "External id": 9843,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9842 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791958780, "dur": 23, + "args": { + "External id": 9844,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9843 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791958808, "dur": 26, + "args": { + "External id": 9845,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9844 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791958810, "dur": 23, + "args": { + "External id": 9846,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9845 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791958839, "dur": 26, + "args": { + "External id": 9847,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9846 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791958841, "dur": 23, + "args": { + "External id": 9848,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9847 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791958869, "dur": 26, + "args": { + "External id": 9849,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9848 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791958871, "dur": 23, + "args": { + "External id": 9850,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9849 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791958900, "dur": 25, + "args": { + "External id": 9851,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9850 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791958902, "dur": 23, + "args": { + "External id": 9852,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9851 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959162, "dur": 33, + "args": { + "External id": 9853,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9852 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959165, "dur": 29, + "args": { + "External id": 9854,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9853 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959201, "dur": 27, + "args": { + "External id": 9855,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9854 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959203, "dur": 24, + "args": { + "External id": 9856,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9855 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959233, "dur": 26, + "args": { + "External id": 9857,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9856 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959235, "dur": 23, + "args": { + "External id": 9858,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9857 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959264, "dur": 26, + "args": { + "External id": 9859,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9858 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959266, "dur": 23, + "args": { + "External id": 9860,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9859 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959294, "dur": 26, + "args": { + "External id": 9861,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9860 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959297, "dur": 23, + "args": { + "External id": 9862,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9861 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959325, "dur": 26, + "args": { + "External id": 9863,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9862 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959327, "dur": 23, + "args": { + "External id": 9864,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9863 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959355, "dur": 26, + "args": { + "External id": 9865,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9864 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959358, "dur": 22, + "args": { + "External id": 9866,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9865 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959386, "dur": 26, + "args": { + "External id": 9867,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9866 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959388, "dur": 23, + "args": { + "External id": 9868,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9867 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959416, "dur": 27, + "args": { + "External id": 9869,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9868 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959419, "dur": 23, + "args": { + "External id": 9870,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9869 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959447, "dur": 26, + "args": { + "External id": 9871,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9870 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959449, "dur": 23, + "args": { + "External id": 9872,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9871 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959477, "dur": 26, + "args": { + "External id": 9873,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9872 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959479, "dur": 23, + "args": { + "External id": 9874,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9873 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959508, "dur": 25, + "args": { + "External id": 9875,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9874 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959510, "dur": 23, + "args": { + "External id": 9876,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9875 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959538, "dur": 26, + "args": { + "External id": 9877,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9876 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959540, "dur": 23, + "args": { + "External id": 9878,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9877 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959568, "dur": 26, + "args": { + "External id": 9879,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9878 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959570, "dur": 23, + "args": { + "External id": 9880,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9879 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959598, "dur": 25, + "args": { + "External id": 9881,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9880 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959600, "dur": 23, + "args": { + "External id": 9882,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9881 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959628, "dur": 26, + "args": { + "External id": 9883,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9882 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959630, "dur": 23, + "args": { + "External id": 9884,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9883 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959658, "dur": 26, + "args": { + "External id": 9885,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9884 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959661, "dur": 22, + "args": { + "External id": 9886,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9885 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959688, "dur": 26, + "args": { + "External id": 9887,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9886 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959690, "dur": 23, + "args": { + "External id": 9888,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9887 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959718, "dur": 26, + "args": { + "External id": 9889,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9888 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959721, "dur": 22, + "args": { + "External id": 9890,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9889 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959748, "dur": 26, + "args": { + "External id": 9891,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9890 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959751, "dur": 22, + "args": { + "External id": 9892,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9891 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959778, "dur": 26, + "args": { + "External id": 9893,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9892 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959780, "dur": 23, + "args": { + "External id": 9894,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9893 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959810, "dur": 26, + "args": { + "External id": 9895,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9894 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959812, "dur": 23, + "args": { + "External id": 9896,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9895 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959841, "dur": 26, + "args": { + "External id": 9897,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9896 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959843, "dur": 23, + "args": { + "External id": 9898,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9897 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959871, "dur": 26, + "args": { + "External id": 9899,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9898 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959874, "dur": 23, + "args": { + "External id": 9900,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9899 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959902, "dur": 26, + "args": { + "External id": 9901,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9900 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959904, "dur": 23, + "args": { + "External id": 9902,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9901 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959932, "dur": 26, + "args": { + "External id": 9903,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9902 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959935, "dur": 22, + "args": { + "External id": 9904,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9903 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959963, "dur": 25, + "args": { + "External id": 9905,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9904 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959965, "dur": 23, + "args": { + "External id": 9906,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9905 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791959993, "dur": 26, + "args": { + "External id": 9907,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9906 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791959995, "dur": 23, + "args": { + "External id": 9908,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9907 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791960023, "dur": 26, + "args": { + "External id": 9909,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9908 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791960026, "dur": 22, + "args": { + "External id": 9910,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9909 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791960053, "dur": 26, + "args": { + "External id": 9911,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9910 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791960055, "dur": 23, + "args": { + "External id": 9912,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9911 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791960084, "dur": 25, + "args": { + "External id": 9913,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9912 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791960086, "dur": 22, + "args": { + "External id": 9914,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9913 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791960113, "dur": 26, + "args": { + "External id": 9915,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9914 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791960116, "dur": 22, + "args": { + "External id": 9916,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9915 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791960143, "dur": 25, + "args": { + "External id": 9917,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9916 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791960145, "dur": 23, + "args": { + "External id": 9918,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9917 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791960173, "dur": 25, + "args": { + "External id": 9919,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9918 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791960175, "dur": 23, + "args": { + "External id": 9920,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9919 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791960203, "dur": 46, + "args": { + "External id": 9921,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9920 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791960205, "dur": 43, + "args": { + "External id": 9922,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9921 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791960254, "dur": 28, + "args": { + "External id": 9923,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9922 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791960257, "dur": 24, + "args": { + "External id": 9924,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9923 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791960286, "dur": 26, + "args": { + "External id": 9925,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9924 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791960288, "dur": 23, + "args": { + "External id": 9926,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9925 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791960316, "dur": 26, + "args": { + "External id": 9927,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9926 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791960318, "dur": 23, + "args": { + "External id": 9928,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9927 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791960347, "dur": 26, + "args": { + "External id": 9929,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9928 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791960350, "dur": 22, + "args": { + "External id": 9930,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9929 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791960378, "dur": 25, + "args": { + "External id": 9931,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9930 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791960380, "dur": 23, + "args": { + "External id": 9932,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9931 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791960408, "dur": 26, + "args": { + "External id": 9933,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9932 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791960410, "dur": 23, + "args": { + "External id": 9934,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9933 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791960438, "dur": 26, + "args": { + "External id": 9935,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9934 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791960440, "dur": 23, + "args": { + "External id": 9936,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9935 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791960468, "dur": 26, + "args": { + "External id": 9937,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9936 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791960471, "dur": 22, + "args": { + "External id": 9938,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9937 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791960498, "dur": 26, + "args": { + "External id": 9939,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9938 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791960500, "dur": 23, + "args": { + "External id": 9940,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9939 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791960528, "dur": 25, + "args": { + "External id": 9941,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9940 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791960530, "dur": 23, + "args": { + "External id": 9942,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9941 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791960558, "dur": 26, + "args": { + "External id": 9943,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9942 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791960561, "dur": 22, + "args": { + "External id": 9944,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9943 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791960588, "dur": 26, + "args": { + "External id": 9945,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9944 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791960590, "dur": 23, + "args": { + "External id": 9946,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9945 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791960618, "dur": 25, + "args": { + "External id": 9947,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9946 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791960620, "dur": 23, + "args": { + "External id": 9948,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9947 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791960648, "dur": 25, + "args": { + "External id": 9949,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9948 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791960650, "dur": 23, + "args": { + "External id": 9950,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9949 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791960678, "dur": 25, + "args": { + "External id": 9951,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9950 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791960680, "dur": 22, + "args": { + "External id": 9952,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9951 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791961662, "dur": 34, + "args": { + "External id": 9953,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9952 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791961666, "dur": 29, + "args": { + "External id": 9954,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9953 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791961702, "dur": 27, + "args": { + "External id": 9955,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9954 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791961704, "dur": 24, + "args": { + "External id": 9956,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9955 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791961734, "dur": 25, + "args": { + "External id": 9957,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9956 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791961736, "dur": 22, + "args": { + "External id": 9958,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9957 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791961764, "dur": 25, + "args": { + "External id": 9959,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9958 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791961766, "dur": 22, + "args": { + "External id": 9960,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9959 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791961794, "dur": 25, + "args": { + "External id": 9961,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9960 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791961796, "dur": 22, + "args": { + "External id": 9962,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9961 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791961824, "dur": 25, + "args": { + "External id": 9963,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9962 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791961826, "dur": 22, + "args": { + "External id": 9964,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9963 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791961853, "dur": 26, + "args": { + "External id": 9965,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9964 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791961856, "dur": 22, + "args": { + "External id": 9966,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9965 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791961883, "dur": 26, + "args": { + "External id": 9967,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9966 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791961885, "dur": 23, + "args": { + "External id": 9968,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9967 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791961913, "dur": 25, + "args": { + "External id": 9969,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9968 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791961915, "dur": 23, + "args": { + "External id": 9970,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9969 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791961943, "dur": 25, + "args": { + "External id": 9971,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9970 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791961945, "dur": 22, + "args": { + "External id": 9972,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9971 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962204, "dur": 34, + "args": { + "External id": 9973,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9972 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962208, "dur": 29, + "args": { + "External id": 9974,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9973 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962243, "dur": 27, + "args": { + "External id": 9975,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9974 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962245, "dur": 24, + "args": { + "External id": 9976,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9975 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962274, "dur": 30, + "args": { + "External id": 9977,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9976 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962280, "dur": 23, + "args": { + "External id": 9978,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9977 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962309, "dur": 26, + "args": { + "External id": 9979,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9978 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962311, "dur": 23, + "args": { + "External id": 9980,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9979 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962339, "dur": 26, + "args": { + "External id": 9981,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9980 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962341, "dur": 23, + "args": { + "External id": 9982,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9981 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962369, "dur": 25, + "args": { + "External id": 9983,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9982 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962371, "dur": 23, + "args": { + "External id": 9984,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9983 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962399, "dur": 25, + "args": { + "External id": 9985,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9984 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962401, "dur": 23, + "args": { + "External id": 9986,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9985 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962429, "dur": 25, + "args": { + "External id": 9987,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9986 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962431, "dur": 23, + "args": { + "External id": 9988,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9987 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962459, "dur": 25, + "args": { + "External id": 9989,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9988 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962461, "dur": 23, + "args": { + "External id": 9990,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9989 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962489, "dur": 25, + "args": { + "External id": 9991,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9990 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962491, "dur": 22, + "args": { + "External id": 9992,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9991 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962519, "dur": 25, + "args": { + "External id": 9993,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9992 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962521, "dur": 23, + "args": { + "External id": 9994,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9993 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962549, "dur": 25, + "args": { + "External id": 9995,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9994 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962551, "dur": 22, + "args": { + "External id": 9996,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9995 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962578, "dur": 26, + "args": { + "External id": 9997,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9996 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962581, "dur": 22, + "args": { + "External id": 9998,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9997 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962608, "dur": 26, + "args": { + "External id": 9999,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9998 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962610, "dur": 23, + "args": { + "External id": 10000,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 9999 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962638, "dur": 25, + "args": { + "External id": 10001,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10000 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962640, "dur": 23, + "args": { + "External id": 10002,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10001 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962668, "dur": 25, + "args": { + "External id": 10003,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10002 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962670, "dur": 22, + "args": { + "External id": 10004,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10003 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962697, "dur": 26, + "args": { + "External id": 10005,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10004 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962700, "dur": 22, + "args": { + "External id": 10006,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10005 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962728, "dur": 25, + "args": { + "External id": 10007,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10006 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962730, "dur": 22, + "args": { + "External id": 10008,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10007 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962757, "dur": 25, + "args": { + "External id": 10009,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10008 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962759, "dur": 23, + "args": { + "External id": 10010,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10009 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962787, "dur": 25, + "args": { + "External id": 10011,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10010 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962789, "dur": 23, + "args": { + "External id": 10012,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10011 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962817, "dur": 25, + "args": { + "External id": 10013,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10012 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962819, "dur": 22, + "args": { + "External id": 10014,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10013 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962846, "dur": 26, + "args": { + "External id": 10015,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10014 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962849, "dur": 22, + "args": { + "External id": 10016,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10015 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962876, "dur": 26, + "args": { + "External id": 10017,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10016 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962879, "dur": 22, + "args": { + "External id": 10018,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10017 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962906, "dur": 26, + "args": { + "External id": 10019,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10018 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962909, "dur": 22, + "args": { + "External id": 10020,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10019 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962936, "dur": 26, + "args": { + "External id": 10021,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10020 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962938, "dur": 23, + "args": { + "External id": 10022,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10021 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962967, "dur": 25, + "args": { + "External id": 10023,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10022 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962969, "dur": 22, + "args": { + "External id": 10024,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10023 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791962996, "dur": 26, + "args": { + "External id": 10025,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10024 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791962999, "dur": 22, + "args": { + "External id": 10026,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10025 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791963027, "dur": 27, + "args": { + "External id": 10027,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10026 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791963029, "dur": 25, + "args": { + "External id": 10028,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10027 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791963059, "dur": 28, + "args": { + "External id": 10029,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10028 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791963062, "dur": 24, + "args": { + "External id": 10030,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10029 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791963091, "dur": 25, + "args": { + "External id": 10031,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10030 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791963093, "dur": 23, + "args": { + "External id": 10032,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10031 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791963121, "dur": 25, + "args": { + "External id": 10033,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10032 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791963123, "dur": 22, + "args": { + "External id": 10034,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10033 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791963150, "dur": 26, + "args": { + "External id": 10035,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10034 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791963153, "dur": 22, + "args": { + "External id": 10036,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10035 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791963180, "dur": 25, + "args": { + "External id": 10037,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10036 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791963182, "dur": 23, + "args": { + "External id": 10038,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10037 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791963210, "dur": 25, + "args": { + "External id": 10039,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10038 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791963212, "dur": 22, + "args": { + "External id": 10040,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10039 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791963240, "dur": 25, + "args": { + "External id": 10041,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10040 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791963242, "dur": 22, + "args": { + "External id": 10042,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10041 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791963269, "dur": 26, + "args": { + "External id": 10043,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10042 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791963271, "dur": 23, + "args": { + "External id": 10044,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10043 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791963299, "dur": 26, + "args": { + "External id": 10045,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10044 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791963301, "dur": 23, + "args": { + "External id": 10046,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10045 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791963329, "dur": 25, + "args": { + "External id": 10047,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10046 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791963331, "dur": 23, + "args": { + "External id": 10048,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10047 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791963359, "dur": 25, + "args": { + "External id": 10049,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10048 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791963361, "dur": 23, + "args": { + "External id": 10050,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10049 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791963388, "dur": 26, + "args": { + "External id": 10051,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10050 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791963391, "dur": 22, + "args": { + "External id": 10052,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10051 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791963418, "dur": 25, + "args": { + "External id": 10053,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10052 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791963420, "dur": 23, + "args": { + "External id": 10054,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10053 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791963448, "dur": 25, + "args": { + "External id": 10055,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10054 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791963450, "dur": 22, + "args": { + "External id": 10056,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10055 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791963477, "dur": 26, + "args": { + "External id": 10057,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10056 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791963480, "dur": 22, + "args": { + "External id": 10058,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10057 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791963507, "dur": 26, + "args": { + "External id": 10059,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10058 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791963510, "dur": 22, + "args": { + "External id": 10060,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10059 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791963537, "dur": 25, + "args": { + "External id": 10061,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10060 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791963539, "dur": 23, + "args": { + "External id": 10062,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10061 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791963567, "dur": 25, + "args": { + "External id": 10063,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10062 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791963569, "dur": 22, + "args": { + "External id": 10064,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10063 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791963596, "dur": 26, + "args": { + "External id": 10065,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10064 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791963599, "dur": 22, + "args": { + "External id": 10066,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10065 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791963626, "dur": 25, + "args": { + "External id": 10067,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10066 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791963628, "dur": 23, + "args": { + "External id": 10068,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10067 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791963656, "dur": 25, + "args": { + "External id": 10069,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10068 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791963658, "dur": 22, + "args": { + "External id": 10070,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10069 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791963685, "dur": 26, + "args": { + "External id": 10071,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10070 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791963688, "dur": 22, + "args": { + "External id": 10072,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10071 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791964690, "dur": 38, + "args": { + "External id": 10073,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10072 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791964694, "dur": 33, + "args": { + "External id": 10074,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10073 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791964733, "dur": 28, + "args": { + "External id": 10075,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10074 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791964736, "dur": 24, + "args": { + "External id": 10076,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10075 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791964765, "dur": 26, + "args": { + "External id": 10077,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10076 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791964767, "dur": 24, + "args": { + "External id": 10078,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10077 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791964796, "dur": 26, + "args": { + "External id": 10079,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10078 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791964798, "dur": 23, + "args": { + "External id": 10080,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10079 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791964826, "dur": 26, + "args": { + "External id": 10081,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10080 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791964829, "dur": 23, + "args": { + "External id": 10082,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10081 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791964857, "dur": 25, + "args": { + "External id": 10083,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10082 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791964859, "dur": 22, + "args": { + "External id": 10084,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10083 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791964887, "dur": 25, + "args": { + "External id": 10085,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10084 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791964889, "dur": 23, + "args": { + "External id": 10086,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10085 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791964917, "dur": 26, + "args": { + "External id": 10087,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10086 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791964919, "dur": 23, + "args": { + "External id": 10088,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10087 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791964947, "dur": 26, + "args": { + "External id": 10089,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10088 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791964949, "dur": 23, + "args": { + "External id": 10090,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10089 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791964977, "dur": 25, + "args": { + "External id": 10091,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10090 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791964979, "dur": 23, + "args": { + "External id": 10092,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10091 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791965231, "dur": 34, + "args": { + "External id": 10093,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10092 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791965235, "dur": 29, + "args": { + "External id": 10094,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10093 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791965270, "dur": 27, + "args": { + "External id": 10095,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10094 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791965273, "dur": 24, + "args": { + "External id": 10096,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10095 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791965302, "dur": 26, + "args": { + "External id": 10097,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10096 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791965304, "dur": 23, + "args": { + "External id": 10098,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10097 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791965332, "dur": 26, + "args": { + "External id": 10099,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10098 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791965335, "dur": 22, + "args": { + "External id": 10100,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10099 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791965362, "dur": 26, + "args": { + "External id": 10101,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10100 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791965365, "dur": 22, + "args": { + "External id": 10102,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10101 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791965392, "dur": 26, + "args": { + "External id": 10103,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10102 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791965395, "dur": 22, + "args": { + "External id": 10104,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10103 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791965422, "dur": 26, + "args": { + "External id": 10105,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10104 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791965425, "dur": 22, + "args": { + "External id": 10106,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10105 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791965452, "dur": 26, + "args": { + "External id": 10107,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10106 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791965454, "dur": 23, + "args": { + "External id": 10108,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10107 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791965482, "dur": 25, + "args": { + "External id": 10109,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10108 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791965484, "dur": 23, + "args": { + "External id": 10110,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10109 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791965512, "dur": 26, + "args": { + "External id": 10111,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10110 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791965514, "dur": 23, + "args": { + "External id": 10112,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10111 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791965542, "dur": 25, + "args": { + "External id": 10113,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10112 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791965544, "dur": 23, + "args": { + "External id": 10114,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10113 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791965572, "dur": 25, + "args": { + "External id": 10115,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10114 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791965574, "dur": 22, + "args": { + "External id": 10116,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10115 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791965601, "dur": 26, + "args": { + "External id": 10117,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10116 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791965603, "dur": 23, + "args": { + "External id": 10118,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10117 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791965631, "dur": 25, + "args": { + "External id": 10119,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10118 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791965633, "dur": 23, + "args": { + "External id": 10120,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10119 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791965661, "dur": 25, + "args": { + "External id": 10121,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10120 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791965663, "dur": 23, + "args": { + "External id": 10122,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10121 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791965691, "dur": 25, + "args": { + "External id": 10123,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10122 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791965693, "dur": 23, + "args": { + "External id": 10124,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10123 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791965721, "dur": 25, + "args": { + "External id": 10125,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10124 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791965723, "dur": 22, + "args": { + "External id": 10126,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10125 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791965751, "dur": 25, + "args": { + "External id": 10127,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10126 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791965753, "dur": 22, + "args": { + "External id": 10128,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10127 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791965781, "dur": 26, + "args": { + "External id": 10129,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10128 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791965783, "dur": 23, + "args": { + "External id": 10130,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10129 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791965811, "dur": 26, + "args": { + "External id": 10131,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10130 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791965813, "dur": 23, + "args": { + "External id": 10132,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10131 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791965841, "dur": 25, + "args": { + "External id": 10133,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10132 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791965843, "dur": 23, + "args": { + "External id": 10134,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10133 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791965871, "dur": 25, + "args": { + "External id": 10135,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10134 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791965873, "dur": 23, + "args": { + "External id": 10136,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10135 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791965901, "dur": 25, + "args": { + "External id": 10137,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10136 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791965903, "dur": 23, + "args": { + "External id": 10138,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10137 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791965931, "dur": 25, + "args": { + "External id": 10139,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10138 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791965933, "dur": 22, + "args": { + "External id": 10140,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10139 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791965960, "dur": 26, + "args": { + "External id": 10141,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10140 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791965963, "dur": 22, + "args": { + "External id": 10142,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10141 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791965990, "dur": 25, + "args": { + "External id": 10143,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10142 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791965992, "dur": 23, + "args": { + "External id": 10144,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10143 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791966020, "dur": 26, + "args": { + "External id": 10145,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10144 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791966022, "dur": 23, + "args": { + "External id": 10146,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10145 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791966050, "dur": 25, + "args": { + "External id": 10147,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10146 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791966052, "dur": 23, + "args": { + "External id": 10148,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10147 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791966080, "dur": 25, + "args": { + "External id": 10149,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10148 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791966082, "dur": 23, + "args": { + "External id": 10150,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10149 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791966110, "dur": 26, + "args": { + "External id": 10151,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10150 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791966113, "dur": 22, + "args": { + "External id": 10152,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10151 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791966140, "dur": 25, + "args": { + "External id": 10153,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10152 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791966142, "dur": 22, + "args": { + "External id": 10154,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10153 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791966170, "dur": 25, + "args": { + "External id": 10155,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10154 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791966172, "dur": 22, + "args": { + "External id": 10156,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10155 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791966199, "dur": 26, + "args": { + "External id": 10157,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10156 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791966201, "dur": 23, + "args": { + "External id": 10158,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10157 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791966230, "dur": 25, + "args": { + "External id": 10159,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10158 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791966232, "dur": 22, + "args": { + "External id": 10160,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10159 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791966259, "dur": 26, + "args": { + "External id": 10161,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10160 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791966262, "dur": 22, + "args": { + "External id": 10162,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10161 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791966290, "dur": 25, + "args": { + "External id": 10163,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10162 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791966292, "dur": 22, + "args": { + "External id": 10164,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10163 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791966319, "dur": 26, + "args": { + "External id": 10165,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10164 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791966322, "dur": 22, + "args": { + "External id": 10166,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10165 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791966350, "dur": 25, + "args": { + "External id": 10167,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10166 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791966352, "dur": 22, + "args": { + "External id": 10168,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10167 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791966379, "dur": 26, + "args": { + "External id": 10169,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10168 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791966381, "dur": 23, + "args": { + "External id": 10170,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10169 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791966409, "dur": 25, + "args": { + "External id": 10171,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10170 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791966411, "dur": 23, + "args": { + "External id": 10172,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10171 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791966439, "dur": 25, + "args": { + "External id": 10173,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10172 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791966441, "dur": 22, + "args": { + "External id": 10174,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10173 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791966468, "dur": 26, + "args": { + "External id": 10175,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10174 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791966471, "dur": 22, + "args": { + "External id": 10176,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10175 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791966498, "dur": 26, + "args": { + "External id": 10177,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10176 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791966500, "dur": 23, + "args": { + "External id": 10178,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10177 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791966528, "dur": 25, + "args": { + "External id": 10179,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10178 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791966530, "dur": 23, + "args": { + "External id": 10180,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10179 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791966558, "dur": 25, + "args": { + "External id": 10181,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10180 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791966560, "dur": 22, + "args": { + "External id": 10182,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10181 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791966588, "dur": 25, + "args": { + "External id": 10183,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10182 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791966590, "dur": 22, + "args": { + "External id": 10184,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10183 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791966618, "dur": 25, + "args": { + "External id": 10185,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10184 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791966620, "dur": 23, + "args": { + "External id": 10186,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10185 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791966648, "dur": 25, + "args": { + "External id": 10187,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10186 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791966650, "dur": 22, + "args": { + "External id": 10188,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10187 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791966678, "dur": 26, + "args": { + "External id": 10189,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10188 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791966680, "dur": 23, + "args": { + "External id": 10190,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10189 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649791966708, "dur": 29, + "args": { + "External id": 10191,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10190 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649791966710, "dur": 26, + "args": { + "External id": 10192,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4, 4096], [4096, 8192]], "Ev Idx": 10191 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__and__", "pid": 435, "tid": 435, + "ts": 1776649791967767, "dur": 40, + "args": { + "External id": 10193,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "15"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 10192 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_and", "pid": 435, "tid": 435, + "ts": 1776649791967769, "dur": 38, + "args": { + "External id": 10194,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "15"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 10193 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_and", "pid": 435, "tid": 435, + "ts": 1776649791967775, "dur": 30, + "args": { + "External id": 10195,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["unsigned char", "long int"], "Input Dims": [[4096, 4096], []], "Ev Idx": 10194 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791967813, "dur": 31, + "args": { + "External id": 10196,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 10195 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791967814, "dur": 29, + "args": { + "External id": 10197,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["unsigned char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], [], []], "Ev Idx": 10196 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791967819, "dur": 6, + "args": { + "External id": 10198,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "[4096, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10197 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791967827, "dur": 16, + "args": { + "External id": 10199,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "unsigned char", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 10198 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__rshift__", "pid": 435, "tid": 435, + "ts": 1776649791967850, "dur": 26, + "args": { + "External id": 10200,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 10199 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791967857, "dur": 5, + "args": { + "External id": 10201,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "0", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10200 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791967880, "dur": 23, + "args": { + "External id": 10202,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 10201 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791967881, "dur": 21, + "args": { + "External id": 10203,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["unsigned char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], [], []], "Ev Idx": 10202 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791967883, "dur": 5, + "args": { + "External id": 10204,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "[4096, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10203 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791967890, "dur": 12, + "args": { + "External id": 10205,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "unsigned char", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 10204 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::stack", "pid": 435, "tid": 435, + "ts": 1776649791967928, "dur": 37, + "args": { + "External id": 10206,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["TensorList", "Scalar"], "Input Dims": [[[4096, 4096], [4096, 4096]], []], "Ev Idx": 10205 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649791967931, "dur": 4, + "args": { + "External id": 10207,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["signed char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 10206 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791967933, "dur": 2, + "args": { + "External id": 10208,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096, 1]", "[4096, 1, 1]", ""], "Input type": ["signed char", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 10207 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649791967936, "dur": 2, + "args": { + "External id": 10209,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["signed char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 10208 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791967937, "dur": 1, + "args": { + "External id": 10210,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096, 1]", "[4096, 1, 1]", ""], "Input type": ["signed char", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 10209 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 435, "tid": 435, + "ts": 1776649791967940, "dur": 24, + "args": { + "External id": 10211,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["TensorList", "Scalar"], "Input Dims": [[[4096, 4096, 1], [4096, 4096, 1]], []], "Ev Idx": 10210 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649791967973, "dur": 5, + "args": { + "External id": 10212,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["signed char", "ScalarList"], "Input Dims": [[4096, 4096, 2], []], "Ev Idx": 10211 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791967974, "dur": 4, + "args": { + "External id": 10213,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["signed char", "ScalarList"], "Input Dims": [[4096, 4096, 2], []], "Ev Idx": 10212 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::repeat_interleave", "pid": 435, "tid": 435, + "ts": 1776649791967983, "dur": 49, + "args": { + "External id": 10214,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "128", "0", ""], "Input type": ["c10::Half", "Scalar", "Scalar", ""], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 10213 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649791967985, "dur": 3, + "args": { + "External id": 10215,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 8192], []], "Ev Idx": 10214 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791967987, "dur": 0, + "args": { + "External id": 10216,"Record function id": 0, "Concrete Inputs": ["", "[32, 1, 8192]", "[8192, 8192, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 10215 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 435, "tid": 435, + "ts": 1776649791967990, "dur": 4, + "args": { + "External id": 10217,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]", "False"], "Input type": ["c10::Half", "ScalarList", "Scalar"], "Input Dims": [[32, 1, 8192], [], []], "Ev Idx": 10216 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791967993, "dur": 0, + "args": { + "External id": 10218,"Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]", "[8192, 0, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 1, 8192], [], [], []], "Ev Idx": 10217 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649791967996, "dur": 30, + "args": { + "External id": 10219,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 10218 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649791967998, "dur": 9, + "args": { + "External id": 10220,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[32, 128, 8192], [], [], [], [], []], "Ev Idx": 10219 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791968000, "dur": 7, + "args": { + "External id": 10221,"Record function id": 0, "Concrete Inputs": ["[32, 128, 8192]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10220 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791968009, "dur": 16, + "args": { + "External id": 10222,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[32, 128, 8192], [32, 128, 8192], []], "Ev Idx": 10221 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::flatten", "pid": 435, "tid": 435, + "ts": 1776649791968028, "dur": 4, + "args": { + "External id": 10223,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[32, 128, 8192], [], []], "Ev Idx": 10222 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649791968029, "dur": 2, + "args": { + "External id": 10224,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 10223 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 435, "tid": 435, + "ts": 1776649791968040, "dur": 19, + "args": { + "External id": 10225,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["signed char", "long int", "Scalar"], "Input Dims": [[4096, 8192], [], []], "Ev Idx": 10224 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649791968064, "dur": 25, + "args": { + "External id": 10226,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["signed char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 10225 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649791968065, "dur": 24, + "args": { + "External id": 10227,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "", "", "", "False", ""], "Input type": ["signed char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 8192], [], [], [], [], [], []], "Ev Idx": 10226 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649791968068, "dur": 6, + "args": { + "External id": 10228,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10227 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791968075, "dur": 13, + "args": { + "External id": 10229,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "signed char", "Scalar"], "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 10228 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 435, "tid": 435, + "ts": 1776649791968095, "dur": 17, + "args": { + "External id": 10230,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4096, 8192], [4096, 8192]], "Ev Idx": 10229 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 435, "tid": 435, + "ts": 1776649791968124, "dur": 12, + "args": { + "External id": 10231,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[4096, 8192]], "Ev Idx": 10230 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 435, "tid": 435, + "ts": 1776649791968129, "dur": 6, + "args": { + "External id": 10232,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[4096, 8192], [], []], "Ev Idx": 10231 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649791968133, "dur": 1, + "args": { + "External id": 10233,"Record function id": 0, "Concrete Inputs": ["", "[8192, 4096]", "[1, 8192]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 8192], [], [], []], "Ev Idx": 10232 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 435, "tid": 435, + "ts": 1776649791968154, "dur": 29, + "args": { + "External id": 10234,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[8192, 4096], []], "Ev Idx": 10233 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649791968156, "dur": 27, + "args": { + "External id": 10235,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[8192, 4096], []], "Ev Idx": 10234 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649791968158, "dur": 7, + "args": { + "External id": 10236,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[8192, 4096], [], [], [], [], []], "Ev Idx": 10235 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791968159, "dur": 6, + "args": { + "External id": 10237,"Record function id": 0, "Concrete Inputs": ["[8192, 4096]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10236 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649791968167, "dur": 15, + "args": { + "External id": 10238,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[8192, 4096], [8192, 4096], []], "Ev Idx": 10237 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649791968282, "dur": 43, + "args": { + "External id": 10239,"Record function id": 0, "Concrete Inputs": ["[8192, 4096]", "", "", "", "False", ""], "Input type": ["ScalarList", "", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10238 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 435, "tid": 435, + "ts": 1776649791968342, "dur": 18, + "args": { + "External id": 10240,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Dims": [[8192, 4096]], "Ev Idx": 10239 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 435, "tid": 435, + "ts": 1776649791968355, "dur": 4, + "args": { + "External id": 10241,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Dims": [[8192, 4096]], "Ev Idx": 10240 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::uniform_", "pid": 435, "tid": 435, + "ts": 1776649791968490, "dur": 212354, + "args": { + "External id": 10242,"Record function id": 0, "Concrete Inputs": ["", "-0.015624999999999998", "0.015624999999999998", ""], "Input type": ["float", "Scalar", "Scalar", ""], "Input Dims": [[8192, 4096], [], [], []], "Ev Idx": 10241 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792181086, "dur": 200646, + "args": { + "External id": 10243,"Record function id": 0, "Concrete Inputs": ["", "", "5", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[8192, 4096], [], [], [], [], []], "Ev Idx": 10242 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649792181094, "dur": 200634, + "args": { + "External id": 10244,"Record function id": 0, "Concrete Inputs": ["", "5", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[8192, 4096], [], [], [], [], [], []], "Ev Idx": 10243 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649792181123, "dur": 58, + "args": { + "External id": 10245,"Record function id": 0, "Concrete Inputs": ["[8192, 4096]", "[4096, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10244 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792181185, "dur": 200525, + "args": { + "External id": 10246,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "float", "Scalar"], "Input Dims": [[8192, 4096], [8192, 4096], []], "Ev Idx": 10245 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792181206, "dur": 95686, + "args": { + "External id": 10247,"Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[8192, 4096], [], [], [], []], "Ev Idx": 10246 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649792181208, "dur": 95681, + "args": { + "External id": 10248,"Record function id": 0, "Concrete Inputs": ["", "5", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[8192, 4096], [], [], [], [], [], []], "Ev Idx": 10247 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649792181210, "dur": 40, + "args": { + "External id": 10249,"Record function id": 0, "Concrete Inputs": ["[8192, 4096]", "[4096, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10248 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792181251, "dur": 95622, + "args": { + "External id": 10250,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "float", "Scalar"], "Input Dims": [[8192, 4096], [8192, 4096], []], "Ev Idx": 10249 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::expand_as", "pid": 435, "tid": 435, + "ts": 1776649792276925, "dur": 35, + "args": { + "External id": 10251,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8192, 4096], [8192, 4096]], "Ev Idx": 10250 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 435, "tid": 435, + "ts": 1776649792276941, "dur": 19, + "args": { + "External id": 10252,"Record function id": 0, "Concrete Inputs": ["", "[8192, 4096]", "False"], "Input type": ["c10::Half", "ScalarList", "Scalar"], "Input Dims": [[8192, 4096], [], []], "Ev Idx": 10251 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792276951, "dur": 8, + "args": { + "External id": 10253,"Record function id": 0, "Concrete Inputs": ["", "[8192, 4096]", "[4096, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[8192, 4096], [], [], []], "Ev Idx": 10252 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792276966, "dur": 7311, + "args": { + "External id": 10254,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[8192, 4096], [8192, 4096], []], "Ev Idx": 10253 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_has_compatible_shallow_copy_type", "pid": 435, "tid": 435, + "ts": 1776649792381859, "dur": 2, + "args": { + "External id": 10255,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "c10::Half"], "Input Dims": [[8192, 4096], [8192, 4096]], "Ev Idx": 10254 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_has_compatible_shallow_copy_type", "pid": 435, "tid": 435, + "ts": 1776649792381884, "dur": 0, + "args": { + "External id": 10256,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "c10::Half"], "Input Dims": [[8192, 4096], [8192, 4096]], "Ev Idx": 10255 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792397168, "dur": 125, + "args": { + "External id": 10257,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[8192, 4096], [8192, 4096], []], "Ev Idx": 10256 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792397416, "dur": 43, + "args": { + "External id": 10258,"Record function id": 0, "Concrete Inputs": ["[256, 16384]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10257 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792397486, "dur": 2, + "args": { + "External id": 10259,"Record function id": 0, "Concrete Inputs": ["[32, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10258 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792397503, "dur": 11, + "args": { + "External id": 10260,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10259 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792397506, "dur": 1, + "args": { + "External id": 10261,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10260 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792397511, "dur": 2, + "args": { + "External id": 10262,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10261 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792397551, "dur": 4809, + "args": { + "External id": 10263,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 16384], [], [], [], [], [], [], []], "Ev Idx": 10262 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649792397555, "dur": 4804, + "args": { + "External id": 10264,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[256, 16384], [], [], [], [], [], []], "Ev Idx": 10263 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649792397566, "dur": 23, + "args": { + "External id": 10265,"Record function id": 0, "Concrete Inputs": ["[256, 16384]", "[16384, 1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10264 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792397591, "dur": 4766, + "args": { + "External id": 10266,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 16384], [256, 16384], []], "Ev Idx": 10265 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792402375, "dur": 97, + "args": { + "External id": 10267,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[32, 8192], [], [], [], [], [], [], []], "Ev Idx": 10266 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649792402377, "dur": 95, + "args": { + "External id": 10268,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[32, 8192], [], [], [], [], [], []], "Ev Idx": 10267 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649792402380, "dur": 9, + "args": { + "External id": 10269,"Record function id": 0, "Concrete Inputs": ["[32, 8192]", "[8192, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10268 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792402391, "dur": 80, + "args": { + "External id": 10270,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[32, 8192], [32, 8192], []], "Ev Idx": 10269 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792402479, "dur": 32, + "args": { + "External id": 10271,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[1024], [], [], [], [], [], [], []], "Ev Idx": 10270 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649792402480, "dur": 30, + "args": { + "External id": 10272,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[1024], [], [], [], [], [], []], "Ev Idx": 10271 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649792402482, "dur": 6, + "args": { + "External id": 10273,"Record function id": 0, "Concrete Inputs": ["[1024]", "[1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10272 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792402489, "dur": 21, + "args": { + "External id": 10274,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[1024], [1024], []], "Ev Idx": 10273 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 435, "tid": 435, + "ts": 1776649792402538, "dur": 26, + "args": { + "External id": 10275,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[32, 8192]], "Ev Idx": 10274 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 435, "tid": 435, + "ts": 1776649792402545, "dur": 15, + "args": { + "External id": 10276,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[32, 8192], [], []], "Ev Idx": 10275 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792402554, "dur": 5, + "args": { + "External id": 10277,"Record function id": 0, "Concrete Inputs": ["", "[8192, 32]", "[1, 8192]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 10276 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 435, "tid": 435, + "ts": 1776649792402568, "dur": 4, + "args": { + "External id": 10278,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[8192, 4096]], "Ev Idx": 10277 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 435, "tid": 435, + "ts": 1776649792402569, "dur": 2, + "args": { + "External id": 10279,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[8192, 4096], [], []], "Ev Idx": 10278 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792402571, "dur": 0, + "args": { + "External id": 10280,"Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]", "[1, 4096]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[8192, 4096], [], [], []], "Ev Idx": 10279 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649792402581, "dur": 24, + "args": { + "External id": 10281,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 128, 8192]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 8192], []], "Ev Idx": 10280 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 435, "tid": 435, + "ts": 1776649792402598, "dur": 7, + "args": { + "External id": 10282,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]", "[128, 1, 4096]"], "Input type": ["c10::Half", "ScalarList", "ScalarList"], "Input Dims": [[4096, 8192], [], []], "Ev Idx": 10281 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 435, "tid": 435, + "ts": 1776649792402617, "dur": 8, + "args": { + "External id": 10283,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1, 0, 2]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 10282 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792402624, "dur": 0, + "args": { + "External id": 10284,"Record function id": 0, "Concrete Inputs": ["", "[128, 32, 8192]", "[1, 128, 4096]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 128, 8192], [], [], []], "Ev Idx": 10283 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649792402629, "dur": 75, + "args": { + "External id": 10285,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[128, -1]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[128, 32, 8192], []], "Ev Idx": 10284 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649792402632, "dur": 63, + "args": { + "External id": 10286,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[128, 32, 8192], []], "Ev Idx": 10285 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649792402637, "dur": 16, + "args": { + "External id": 10287,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[128, 32, 8192], [], [], [], [], []], "Ev Idx": 10286 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792402641, "dur": 11, + "args": { + "External id": 10288,"Record function id": 0, "Concrete Inputs": ["[128, 32, 8192]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10287 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792402655, "dur": 39, + "args": { + "External id": 10289,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[128, 32, 8192], [128, 32, 8192], []], "Ev Idx": 10288 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 435, "tid": 435, + "ts": 1776649792402698, "dur": 6, + "args": { + "External id": 10290,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[128, 262144]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[128, 32, 8192], []], "Ev Idx": 10289 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649792402712, "dur": 31, + "args": { + "External id": 10291,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1, -1]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[8192, 32], []], "Ev Idx": 10290 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649792402714, "dur": 25, + "args": { + "External id": 10292,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[8192, 32], []], "Ev Idx": 10291 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649792402716, "dur": 8, + "args": { + "External id": 10293,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[8192, 32], [], [], [], [], []], "Ev Idx": 10292 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792402717, "dur": 6, + "args": { + "External id": 10294,"Record function id": 0, "Concrete Inputs": ["[8192, 32]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10293 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792402725, "dur": 14, + "args": { + "External id": 10295,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[8192, 32], [8192, 32], []], "Ev Idx": 10294 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 435, "tid": 435, + "ts": 1776649792402741, "dur": 1, + "args": { + "External id": 10296,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1, 262144]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[8192, 32], []], "Ev Idx": 10295 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 435, "tid": 435, + "ts": 1776649792402752, "dur": 35, + "args": { + "External id": 10297,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[128, 262144], [1, 262144]], "Ev Idx": 10296 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 435, "tid": 435, + "ts": 1776649792402793, "dur": 24, + "args": { + "External id": 10298,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[128, 262144]], "Ev Idx": 10297 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792402824, "dur": 30, + "args": { + "External id": 10299,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[128, 262144], [], [], [], []], "Ev Idx": 10298 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649792402826, "dur": 28, + "args": { + "External id": 10300,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "", "", "", "False", ""], "Input type": ["c10::Half", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[128, 262144], [], [], [], [], [], []], "Ev Idx": 10299 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649792402829, "dur": 6, + "args": { + "External id": 10301,"Record function id": 0, "Concrete Inputs": ["[128, 262144]", "[262144, 1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10300 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792402836, "dur": 17, + "args": { + "External id": 10302,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "c10::Half", "Scalar"], "Input Dims": [[128, 262144], [128, 262144], []], "Ev Idx": 10301 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 435, "tid": 435, + "ts": 1776649792402870, "dur": 32, + "args": { + "External id": 10303,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["int", "long int", "Scalar"], "Input Dims": [[128, 262144], [], []], "Ev Idx": 10302 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 435, "tid": 435, + "ts": 1776649792402915, "dur": 32, + "args": { + "External id": 10304,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "15"], "Input type": ["int", "Scalar", "Scalar"], "Input Dims": [[128, 262144], [], []], "Ev Idx": 10303 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792402920, "dur": 0, + "args": { + "External id": 10305,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[128, 262144], [], [], [], []], "Ev Idx": 10304 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649792402956, "dur": 8, + "args": { + "External id": 10306,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[128, -1, 8192]"], "Input type": ["int", "ScalarList"], "Input Dims": [[128, 262144], []], "Ev Idx": 10305 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649792402958, "dur": 6, + "args": { + "External id": 10307,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[128, -1, 8192]"], "Input type": ["int", "ScalarList"], "Input Dims": [[128, 262144], []], "Ev Idx": 10306 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 435, "tid": 435, + "ts": 1776649792402968, "dur": 4, + "args": { + "External id": 10308,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1, 0, 2]"], "Input type": ["int", "ScalarList"], "Input Dims": [[128, 32, 8192], []], "Ev Idx": 10307 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792402970, "dur": 1, + "args": { + "External id": 10309,"Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]", "[8192, 262144, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Dims": [[128, 32, 8192], [], [], []], "Ev Idx": 10308 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649792402975, "dur": 37, + "args": { + "External id": 10310,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["int", "ScalarList"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 10309 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649792402978, "dur": 30, + "args": { + "External id": 10311,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 10310 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649792402980, "dur": 9, + "args": { + "External id": 10312,"Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "0"], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[32, 128, 8192], [], [], [], [], []], "Ev Idx": 10311 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792402981, "dur": 7, + "args": { + "External id": 10313,"Record function id": 0, "Concrete Inputs": ["[32, 128, 8192]", "3", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10312 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792402990, "dur": 17, + "args": { + "External id": 10314,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[32, 128, 8192], [32, 128, 8192], []], "Ev Idx": 10313 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 435, "tid": 435, + "ts": 1776649792403010, "dur": 2, + "args": { + "External id": 10315,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["int", "ScalarList"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 10314 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649792403020, "dur": 3, + "args": { + "External id": 10316,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 64]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[1, 262144], []], "Ev Idx": 10315 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649792403021, "dur": 2, + "args": { + "External id": 10317,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 64]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[1, 262144], []], "Ev Idx": 10316 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649792403038, "dur": 10, + "args": { + "External id": 10318,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 64], [], [], [], []], "Ev Idx": 10317 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792403045, "dur": 1, + "args": { + "External id": 10319,"Record function id": 0, "Concrete Inputs": ["", "[4096, 64]", "[64, 1]", "0"], "Input type": ["c10::Half", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 64], [], [], []], "Ev Idx": 10318 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792403070, "dur": 2, + "args": { + "External id": 10320,"Record function id": 0, "Concrete Inputs": ["[64]", "4", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10319 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792403080, "dur": 1704, + "args": { + "External id": 10321,"Record function id": 0, "Concrete Inputs": ["", "", "4", "False", "False", ""], "Input type": ["long int", "", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[64], [], [], [], [], []], "Ev Idx": 10320 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649792403082, "dur": 1702, + "args": { + "External id": 10322,"Record function id": 0, "Concrete Inputs": ["", "4", "", "", "", "False", ""], "Input type": ["long int", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[64], [], [], [], [], [], []], "Ev Idx": 10321 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649792403084, "dur": 7, + "args": { + "External id": 10323,"Record function id": 0, "Concrete Inputs": ["[64]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10322 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792403093, "dur": 1690, + "args": { + "External id": 10324,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Dims": [[64], [64], []], "Ev Idx": 10323 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 435, "tid": 435, + "ts": 1776649792404788, "dur": 1, + "args": { + "External id": 10325,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Dims": [[64]], "Ev Idx": 10324 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::index", "pid": 435, "tid": 435, + "ts": 1776649792404795, "dur": 52, + "args": { + "External id": 10326,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", ""], "Input Dims": [[4096, 64], []], "Ev Idx": 10325 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792404809, "dur": 1, + "args": { + "External id": 10327,"Record function id": 0, "Concrete Inputs": ["", "[4096, 64]", "[64, 0]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 64], [], [], []], "Ev Idx": 10326 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649792404812, "dur": 2, + "args": { + "External id": 10328,"Record function id": 0, "Concrete Inputs": ["", "[1, 64]"], "Input type": ["long int", "ScalarList"], "Input Dims": [[64], []], "Ev Idx": 10327 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649792404813, "dur": 1, + "args": { + "External id": 10329,"Record function id": 0, "Concrete Inputs": ["", "[1, 64]"], "Input type": ["long int", "ScalarList"], "Input Dims": [[64], []], "Ev Idx": 10328 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649792404860, "dur": 4, + "args": { + "External id": 10330,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 8192]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 64], []], "Ev Idx": 10329 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649792404861, "dur": 2, + "args": { + "External id": 10331,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 8192]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 64], []], "Ev Idx": 10330 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649792404867, "dur": 3, + "args": { + "External id": 10332,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 16, 512, 16]"], "Input type": ["int", "ScalarList"], "Input Dims": [[4096, 8192], []], "Ev Idx": 10331 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649792404868, "dur": 2, + "args": { + "External id": 10333,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 16, 512, 16]"], "Input type": ["int", "ScalarList"], "Input Dims": [[4096, 8192], []], "Ev Idx": 10332 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 435, "tid": 435, + "ts": 1776649792404873, "dur": 5, + "args": { + "External id": 10334,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[0, 2, 1, 3]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 16, 512, 16], []], "Ev Idx": 10333 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792404876, "dur": 1, + "args": { + "External id": 10335,"Record function id": 0, "Concrete Inputs": ["", "[256, 512, 16, 16]", "[131072, 16, 8192, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Dims": [[256, 16, 512, 16], [], [], []], "Ev Idx": 10334 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649792404881, "dur": 35, + "args": { + "External id": 10336,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 131072]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 512, 16, 16], []], "Ev Idx": 10335 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649792404883, "dur": 29, + "args": { + "External id": 10337,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[256, 512, 16, 16], []], "Ev Idx": 10336 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649792404885, "dur": 9, + "args": { + "External id": 10338,"Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "0"], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[256, 512, 16, 16], [], [], [], [], []], "Ev Idx": 10337 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792404887, "dur": 7, + "args": { + "External id": 10339,"Record function id": 0, "Concrete Inputs": ["[256, 512, 16, 16]", "3", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10338 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792404896, "dur": 16, + "args": { + "External id": 10340,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 512, 16, 16], [256, 512, 16, 16], []], "Ev Idx": 10339 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 435, "tid": 435, + "ts": 1776649792404915, "dur": 1, + "args": { + "External id": 10341,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 131072]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 512, 16, 16], []], "Ev Idx": 10340 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649792404925, "dur": 3, + "args": { + "External id": 10342,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 131072], []], "Ev Idx": 10341 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649792404926, "dur": 1, + "args": { + "External id": 10343,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 131072], []], "Ev Idx": 10342 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649792404933, "dur": 5, + "args": { + "External id": 10344,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[32768, 1024], [], [], [], []], "Ev Idx": 10343 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792404936, "dur": 1, + "args": { + "External id": 10345,"Record function id": 0, "Concrete Inputs": ["", "[32768, 1024]", "[1024, 1]", "0"], "Input type": ["int", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[32768, 1024], [], [], []], "Ev Idx": 10344 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::index", "pid": 435, "tid": 435, + "ts": 1776649792404941, "dur": 229, + "args": { + "External id": 10346,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["int", ""], "Input Dims": [[32768, 1024], []], "Ev Idx": 10345 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792404945, "dur": 195, + "args": { + "External id": 10347,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "False", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Dims": [[1024], [], [], [], [], [], [], []], "Ev Idx": 10346 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649792404947, "dur": 193, + "args": { + "External id": 10348,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Dims": [[1024], [], [], [], [], [], []], "Ev Idx": 10347 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649792404950, "dur": 5, + "args": { + "External id": 10349,"Record function id": 0, "Concrete Inputs": ["[1024]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10348 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792404957, "dur": 182, + "args": { + "External id": 10350,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Dims": [[1024], [1024], []], "Ev Idx": 10349 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792405143, "dur": 1, + "args": { + "External id": 10351,"Record function id": 0, "Concrete Inputs": ["", "[32768, 1024]", "[1024, 0]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Dims": [[32768, 1024], [], [], []], "Ev Idx": 10350 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649792405145, "dur": 2, + "args": { + "External id": 10352,"Record function id": 0, "Concrete Inputs": ["", "[1, 1024]"], "Input type": ["long int", "ScalarList"], "Input Dims": [[1024], []], "Ev Idx": 10351 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649792405146, "dur": 1, + "args": { + "External id": 10353,"Record function id": 0, "Concrete Inputs": ["", "[1, 1024]"], "Input type": ["long int", "ScalarList"], "Input Dims": [[1024], []], "Ev Idx": 10352 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649792405181, "dur": 3, + "args": { + "External id": 10354,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 131072]"], "Input type": ["int", "ScalarList"], "Input Dims": [[32768, 1024], []], "Ev Idx": 10353 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649792405182, "dur": 2, + "args": { + "External id": 10355,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 131072]"], "Input type": ["int", "ScalarList"], "Input Dims": [[32768, 1024], []], "Ev Idx": 10354 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792427762, "dur": 92532, + "args": { + "External id": 10356,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 131072], [], [], [], [], [], [], []], "Ev Idx": 10355 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649792427768, "dur": 92524, + "args": { + "External id": 10357,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[256, 131072], [], [], [], [], [], []], "Ev Idx": 10356 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649792427787, "dur": 55, + "args": { + "External id": 10358,"Record function id": 0, "Concrete Inputs": ["[256, 131072]", "[131072, 1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10357 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792427846, "dur": 92438, + "args": { + "External id": 10359,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 131072], [256, 131072], []], "Ev Idx": 10358 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 435, "tid": 435, + "ts": 1776649792520344, "dur": 21, + "args": { + "External id": 10360,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 131072]], "Ev Idx": 10359 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 435, "tid": 435, + "ts": 1776649792520349, "dur": 16, + "args": { + "External id": 10361,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 131072]], "Ev Idx": 10360 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792520369, "dur": 1, + "args": { + "External id": 10362,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 131072], [], [], [], [], [], [], []], "Ev Idx": 10361 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::resolve_conj", "pid": 435, "tid": 435, + "ts": 1776649792520374, "dur": 1, + "args": { + "External id": 10363,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 131072]], "Ev Idx": 10362 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::resolve_neg", "pid": 435, "tid": 435, + "ts": 1776649792520376, "dur": 1, + "args": { + "External id": 10364,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 131072]], "Ev Idx": 10363 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 435, "tid": 435, + "ts": 1776649792641487, "dur": 9, + "args": { + "External id": 10365,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 16384]], "Ev Idx": 10364 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792641534, "dur": 6860, + "args": { + "External id": 10366,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 16384], [], [], [], [], [], [], []], "Ev Idx": 10365 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649792641539, "dur": 6853, + "args": { + "External id": 10367,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[256, 16384], [], [], [], [], [], []], "Ev Idx": 10366 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649792641561, "dur": 63, + "args": { + "External id": 10368,"Record function id": 0, "Concrete Inputs": ["[256, 16384]", "[16384, 1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10367 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792641628, "dur": 6754, + "args": { + "External id": 10369,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 16384], [256, 16384], []], "Ev Idx": 10368 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792648494, "dur": 1, + "args": { + "External id": 10370,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 16384], [], [], [], [], [], [], []], "Ev Idx": 10369 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649792648527, "dur": 40, + "args": { + "External id": 10371,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[256, 16384], [], [], [], []], "Ev Idx": 10370 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792648552, "dur": 6, + "args": { + "External id": 10372,"Record function id": 0, "Concrete Inputs": ["", "[256, 16384]", "[16384, 1]", "0"], "Input type": ["int", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[256, 16384], [], [], []], "Ev Idx": 10371 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649792648569, "dur": 7, + "args": { + "External id": 10373,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[256, 16384], [], [], [], []], "Ev Idx": 10372 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792648573, "dur": 1, + "args": { + "External id": 10374,"Record function id": 0, "Concrete Inputs": ["", "[256, 16384]", "[16384, 1]", "0"], "Input type": ["int", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[256, 16384], [], [], []], "Ev Idx": 10373 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792648582, "dur": 75, + "args": { + "External id": 10375,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 16384], [256, 16384], []], "Ev Idx": 10374 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792648668, "dur": 1, + "args": { + "External id": 10376,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[32, 8192], [], [], [], [], [], [], []], "Ev Idx": 10375 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649792648674, "dur": 6, + "args": { + "External id": 10377,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[32, 8192], [], [], [], []], "Ev Idx": 10376 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792648677, "dur": 1, + "args": { + "External id": 10378,"Record function id": 0, "Concrete Inputs": ["", "[32, 8192]", "[8192, 1]", "0"], "Input type": ["c10::Half", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 10377 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649792648681, "dur": 4, + "args": { + "External id": 10379,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[32, 8192], [], [], [], []], "Ev Idx": 10378 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792648683, "dur": 1, + "args": { + "External id": 10380,"Record function id": 0, "Concrete Inputs": ["", "[32, 8192]", "[8192, 1]", "0"], "Input type": ["c10::Half", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 10379 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792648686, "dur": 22, + "args": { + "External id": 10381,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[32, 8192], [32, 8192], []], "Ev Idx": 10380 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792649822, "dur": 25, + "args": { + "External id": 10382,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10381 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792649860, "dur": 60, + "args": { + "External id": 10383,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10382 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792649864, "dur": 5, + "args": { + "External id": 10384,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10383 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792649872, "dur": 47, + "args": { + "External id": 10385,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10384 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792649877, "dur": 41, + "args": { + "External id": 10386,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10385 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792649963, "dur": 7, + "args": { + "External id": 10387,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10386 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792649974, "dur": 22, + "args": { + "External id": 10388,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10387 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792649975, "dur": 5, + "args": { + "External id": 10389,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10388 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792649981, "dur": 15, + "args": { + "External id": 10390,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10389 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792649983, "dur": 12, + "args": { + "External id": 10391,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10390 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650016, "dur": 5, + "args": { + "External id": 10392,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10391 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792650026, "dur": 22, + "args": { + "External id": 10393,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10392 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650027, "dur": 4, + "args": { + "External id": 10394,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10393 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792650033, "dur": 15, + "args": { + "External id": 10395,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10394 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792650034, "dur": 12, + "args": { + "External id": 10396,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10395 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650067, "dur": 5, + "args": { + "External id": 10397,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10396 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792650076, "dur": 22, + "args": { + "External id": 10398,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10397 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650077, "dur": 5, + "args": { + "External id": 10399,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10398 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792650083, "dur": 14, + "args": { + "External id": 10400,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10399 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792650085, "dur": 11, + "args": { + "External id": 10401,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10400 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650116, "dur": 5, + "args": { + "External id": 10402,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10401 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792650125, "dur": 22, + "args": { + "External id": 10403,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10402 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650126, "dur": 5, + "args": { + "External id": 10404,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10403 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792650132, "dur": 15, + "args": { + "External id": 10405,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10404 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792650134, "dur": 11, + "args": { + "External id": 10406,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10405 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650165, "dur": 6, + "args": { + "External id": 10407,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10406 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792650175, "dur": 21, + "args": { + "External id": 10408,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10407 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650176, "dur": 4, + "args": { + "External id": 10409,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10408 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792650182, "dur": 14, + "args": { + "External id": 10410,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10409 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792650183, "dur": 12, + "args": { + "External id": 10411,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10410 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650214, "dur": 6, + "args": { + "External id": 10412,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10411 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792650224, "dur": 21, + "args": { + "External id": 10413,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10412 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650225, "dur": 4, + "args": { + "External id": 10414,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10413 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792650231, "dur": 14, + "args": { + "External id": 10415,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10414 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792650232, "dur": 11, + "args": { + "External id": 10416,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10415 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650263, "dur": 5, + "args": { + "External id": 10417,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10416 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792650272, "dur": 22, + "args": { + "External id": 10418,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10417 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650274, "dur": 4, + "args": { + "External id": 10419,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10418 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792650279, "dur": 14, + "args": { + "External id": 10420,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10419 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792650281, "dur": 11, + "args": { + "External id": 10421,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10420 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650312, "dur": 5, + "args": { + "External id": 10422,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10421 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792650321, "dur": 21, + "args": { + "External id": 10423,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10422 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650322, "dur": 4, + "args": { + "External id": 10424,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10423 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792650328, "dur": 14, + "args": { + "External id": 10425,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10424 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792650329, "dur": 11, + "args": { + "External id": 10426,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10425 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650360, "dur": 5, + "args": { + "External id": 10427,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10426 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792650369, "dur": 22, + "args": { + "External id": 10428,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10427 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650370, "dur": 5, + "args": { + "External id": 10429,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10428 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792650376, "dur": 14, + "args": { + "External id": 10430,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10429 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792650378, "dur": 11, + "args": { + "External id": 10431,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10430 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650514, "dur": 6, + "args": { + "External id": 10432,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10431 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792650526, "dur": 29, + "args": { + "External id": 10433,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10432 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650527, "dur": 6, + "args": { + "External id": 10434,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10433 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792650535, "dur": 19, + "args": { + "External id": 10435,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10434 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792650537, "dur": 16, + "args": { + "External id": 10436,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10435 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650576, "dur": 5, + "args": { + "External id": 10437,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10436 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792650586, "dur": 21, + "args": { + "External id": 10438,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10437 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650587, "dur": 4, + "args": { + "External id": 10439,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10438 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792650593, "dur": 14, + "args": { + "External id": 10440,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10439 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792650594, "dur": 12, + "args": { + "External id": 10441,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10440 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650626, "dur": 5, + "args": { + "External id": 10442,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10441 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792650635, "dur": 22, + "args": { + "External id": 10443,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10442 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650637, "dur": 4, + "args": { + "External id": 10444,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10443 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792650642, "dur": 15, + "args": { + "External id": 10445,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10444 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792650644, "dur": 11, + "args": { + "External id": 10446,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10445 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650675, "dur": 6, + "args": { + "External id": 10447,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10446 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792650686, "dur": 22, + "args": { + "External id": 10448,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10447 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650687, "dur": 5, + "args": { + "External id": 10449,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10448 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792650693, "dur": 14, + "args": { + "External id": 10450,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10449 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792650694, "dur": 12, + "args": { + "External id": 10451,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10450 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650726, "dur": 5, + "args": { + "External id": 10452,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10451 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792650735, "dur": 21, + "args": { + "External id": 10453,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10452 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650736, "dur": 4, + "args": { + "External id": 10454,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10453 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792650742, "dur": 14, + "args": { + "External id": 10455,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10454 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792650743, "dur": 12, + "args": { + "External id": 10456,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10455 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650774, "dur": 6, + "args": { + "External id": 10457,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10456 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792650784, "dur": 21, + "args": { + "External id": 10458,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10457 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650785, "dur": 4, + "args": { + "External id": 10459,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10458 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792650791, "dur": 14, + "args": { + "External id": 10460,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10459 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792650792, "dur": 11, + "args": { + "External id": 10461,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10460 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650823, "dur": 5, + "args": { + "External id": 10462,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10461 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792650832, "dur": 22, + "args": { + "External id": 10463,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10462 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650834, "dur": 4, + "args": { + "External id": 10464,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10463 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792650839, "dur": 15, + "args": { + "External id": 10465,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10464 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792650841, "dur": 11, + "args": { + "External id": 10466,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10465 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650872, "dur": 5, + "args": { + "External id": 10467,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10466 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792650881, "dur": 22, + "args": { + "External id": 10468,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10467 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650882, "dur": 5, + "args": { + "External id": 10469,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10468 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792650888, "dur": 14, + "args": { + "External id": 10470,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10469 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792650890, "dur": 11, + "args": { + "External id": 10471,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10470 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650921, "dur": 5, + "args": { + "External id": 10472,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10471 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792650930, "dur": 22, + "args": { + "External id": 10473,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10472 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650931, "dur": 5, + "args": { + "External id": 10474,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10473 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792650937, "dur": 14, + "args": { + "External id": 10475,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10474 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792650939, "dur": 11, + "args": { + "External id": 10476,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10475 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650970, "dur": 5, + "args": { + "External id": 10477,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10476 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792650979, "dur": 21, + "args": { + "External id": 10478,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10477 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792650980, "dur": 4, + "args": { + "External id": 10479,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10478 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792650986, "dur": 14, + "args": { + "External id": 10480,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10479 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792650987, "dur": 11, + "args": { + "External id": 10481,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10480 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651018, "dur": 5, + "args": { + "External id": 10482,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10481 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792651027, "dur": 21, + "args": { + "External id": 10483,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10482 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651028, "dur": 5, + "args": { + "External id": 10484,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10483 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792651034, "dur": 14, + "args": { + "External id": 10485,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10484 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792651035, "dur": 12, + "args": { + "External id": 10486,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10485 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651067, "dur": 5, + "args": { + "External id": 10487,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10486 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792651076, "dur": 21, + "args": { + "External id": 10488,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10487 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651077, "dur": 4, + "args": { + "External id": 10489,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10488 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792651083, "dur": 14, + "args": { + "External id": 10490,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10489 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792651084, "dur": 11, + "args": { + "External id": 10491,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10490 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651115, "dur": 5, + "args": { + "External id": 10492,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10491 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792651124, "dur": 21, + "args": { + "External id": 10493,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10492 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651125, "dur": 5, + "args": { + "External id": 10494,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10493 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792651131, "dur": 14, + "args": { + "External id": 10495,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10494 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792651133, "dur": 11, + "args": { + "External id": 10496,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10495 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651163, "dur": 5, + "args": { + "External id": 10497,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10496 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792651172, "dur": 22, + "args": { + "External id": 10498,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10497 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651173, "dur": 5, + "args": { + "External id": 10499,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10498 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792651179, "dur": 15, + "args": { + "External id": 10500,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10499 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792651181, "dur": 11, + "args": { + "External id": 10501,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10500 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651212, "dur": 5, + "args": { + "External id": 10502,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10501 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792651221, "dur": 21, + "args": { + "External id": 10503,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10502 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651222, "dur": 4, + "args": { + "External id": 10504,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10503 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792651228, "dur": 14, + "args": { + "External id": 10505,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10504 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792651229, "dur": 11, + "args": { + "External id": 10506,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10505 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651260, "dur": 5, + "args": { + "External id": 10507,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10506 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792651269, "dur": 21, + "args": { + "External id": 10508,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10507 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651270, "dur": 5, + "args": { + "External id": 10509,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10508 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792651276, "dur": 14, + "args": { + "External id": 10510,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10509 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792651278, "dur": 11, + "args": { + "External id": 10511,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10510 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651308, "dur": 5, + "args": { + "External id": 10512,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10511 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792651317, "dur": 24, + "args": { + "External id": 10513,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10512 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651318, "dur": 5, + "args": { + "External id": 10514,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10513 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792651324, "dur": 16, + "args": { + "External id": 10515,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10514 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792651326, "dur": 12, + "args": { + "External id": 10516,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10515 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651361, "dur": 5, + "args": { + "External id": 10517,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10516 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792651370, "dur": 21, + "args": { + "External id": 10518,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10517 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651371, "dur": 5, + "args": { + "External id": 10519,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10518 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792651377, "dur": 14, + "args": { + "External id": 10520,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10519 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792651378, "dur": 11, + "args": { + "External id": 10521,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10520 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651409, "dur": 5, + "args": { + "External id": 10522,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10521 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792651418, "dur": 21, + "args": { + "External id": 10523,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10522 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651419, "dur": 4, + "args": { + "External id": 10524,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10523 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792651425, "dur": 14, + "args": { + "External id": 10525,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10524 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792651426, "dur": 11, + "args": { + "External id": 10526,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10525 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651456, "dur": 6, + "args": { + "External id": 10527,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10526 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792651466, "dur": 21, + "args": { + "External id": 10528,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10527 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651467, "dur": 4, + "args": { + "External id": 10529,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10528 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792651472, "dur": 14, + "args": { + "External id": 10530,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10529 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792651474, "dur": 11, + "args": { + "External id": 10531,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10530 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651504, "dur": 5, + "args": { + "External id": 10532,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10531 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792651513, "dur": 21, + "args": { + "External id": 10533,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10532 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651514, "dur": 5, + "args": { + "External id": 10534,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10533 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792651520, "dur": 14, + "args": { + "External id": 10535,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10534 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792651522, "dur": 11, + "args": { + "External id": 10536,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10535 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651553, "dur": 5, + "args": { + "External id": 10537,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10536 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792651562, "dur": 21, + "args": { + "External id": 10538,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10537 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651563, "dur": 4, + "args": { + "External id": 10539,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10538 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792651569, "dur": 13, + "args": { + "External id": 10540,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10539 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792651570, "dur": 11, + "args": { + "External id": 10541,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10540 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651601, "dur": 5, + "args": { + "External id": 10542,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10541 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792651610, "dur": 21, + "args": { + "External id": 10543,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10542 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651611, "dur": 4, + "args": { + "External id": 10544,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10543 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792651617, "dur": 14, + "args": { + "External id": 10545,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10544 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792651618, "dur": 11, + "args": { + "External id": 10546,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10545 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651649, "dur": 5, + "args": { + "External id": 10547,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10546 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792651658, "dur": 22, + "args": { + "External id": 10548,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10547 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651659, "dur": 5, + "args": { + "External id": 10549,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10548 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792651665, "dur": 14, + "args": { + "External id": 10550,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10549 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792651667, "dur": 11, + "args": { + "External id": 10551,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10550 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651698, "dur": 5, + "args": { + "External id": 10552,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10551 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792651707, "dur": 21, + "args": { + "External id": 10553,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10552 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651708, "dur": 4, + "args": { + "External id": 10554,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10553 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792651714, "dur": 14, + "args": { + "External id": 10555,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10554 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792651715, "dur": 11, + "args": { + "External id": 10556,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10555 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651746, "dur": 5, + "args": { + "External id": 10557,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10556 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792651755, "dur": 21, + "args": { + "External id": 10558,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10557 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651756, "dur": 4, + "args": { + "External id": 10559,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10558 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792651762, "dur": 14, + "args": { + "External id": 10560,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10559 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792651763, "dur": 11, + "args": { + "External id": 10561,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10560 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651794, "dur": 5, + "args": { + "External id": 10562,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10561 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792651803, "dur": 21, + "args": { + "External id": 10563,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10562 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651804, "dur": 4, + "args": { + "External id": 10564,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10563 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792651810, "dur": 14, + "args": { + "External id": 10565,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10564 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792651811, "dur": 12, + "args": { + "External id": 10566,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10565 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651842, "dur": 5, + "args": { + "External id": 10567,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10566 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792651851, "dur": 21, + "args": { + "External id": 10568,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10567 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651853, "dur": 4, + "args": { + "External id": 10569,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10568 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792651858, "dur": 14, + "args": { + "External id": 10570,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10569 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792651860, "dur": 11, + "args": { + "External id": 10571,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10570 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651891, "dur": 5, + "args": { + "External id": 10572,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10571 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792651900, "dur": 21, + "args": { + "External id": 10573,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10572 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651901, "dur": 5, + "args": { + "External id": 10574,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10573 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792651907, "dur": 14, + "args": { + "External id": 10575,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10574 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792651909, "dur": 11, + "args": { + "External id": 10576,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10575 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651939, "dur": 5, + "args": { + "External id": 10577,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10576 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792651949, "dur": 21, + "args": { + "External id": 10578,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10577 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651950, "dur": 4, + "args": { + "External id": 10579,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10578 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792651955, "dur": 14, + "args": { + "External id": 10580,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10579 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792651957, "dur": 11, + "args": { + "External id": 10581,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10580 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651988, "dur": 5, + "args": { + "External id": 10582,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10581 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792651997, "dur": 21, + "args": { + "External id": 10583,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10582 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792651998, "dur": 4, + "args": { + "External id": 10584,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10583 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792652004, "dur": 13, + "args": { + "External id": 10585,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10584 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792652005, "dur": 11, + "args": { + "External id": 10586,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10585 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792652036, "dur": 5, + "args": { + "External id": 10587,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10586 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792652045, "dur": 21, + "args": { + "External id": 10588,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10587 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792652046, "dur": 5, + "args": { + "External id": 10589,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10588 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792652052, "dur": 14, + "args": { + "External id": 10590,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10589 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792652054, "dur": 11, + "args": { + "External id": 10591,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10590 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792652084, "dur": 5, + "args": { + "External id": 10592,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10591 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792652093, "dur": 22, + "args": { + "External id": 10593,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10592 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792652094, "dur": 5, + "args": { + "External id": 10594,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10593 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792652100, "dur": 14, + "args": { + "External id": 10595,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10594 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792652102, "dur": 11, + "args": { + "External id": 10596,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10595 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792652132, "dur": 5, + "args": { + "External id": 10597,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10596 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792652141, "dur": 21, + "args": { + "External id": 10598,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10597 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792652143, "dur": 4, + "args": { + "External id": 10599,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10598 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792652148, "dur": 14, + "args": { + "External id": 10600,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10599 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792652150, "dur": 11, + "args": { + "External id": 10601,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10600 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792652180, "dur": 5, + "args": { + "External id": 10602,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10601 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792652189, "dur": 21, + "args": { + "External id": 10603,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10602 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792652190, "dur": 5, + "args": { + "External id": 10604,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10603 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792652196, "dur": 14, + "args": { + "External id": 10605,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10604 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792652198, "dur": 11, + "args": { + "External id": 10606,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10605 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792656359, "dur": 12, + "args": { + "External id": 10607,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10606 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792656380, "dur": 32, + "args": { + "External id": 10608,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10607 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792656382, "dur": 5, + "args": { + "External id": 10609,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10608 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792656389, "dur": 22, + "args": { + "External id": 10610,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10609 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792656391, "dur": 19, + "args": { + "External id": 10611,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10610 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792656436, "dur": 6, + "args": { + "External id": 10612,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10611 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792656446, "dur": 23, + "args": { + "External id": 10613,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10612 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792656448, "dur": 4, + "args": { + "External id": 10614,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10613 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792656454, "dur": 15, + "args": { + "External id": 10615,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10614 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792656455, "dur": 13, + "args": { + "External id": 10616,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10615 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792656488, "dur": 6, + "args": { + "External id": 10617,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10616 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792656498, "dur": 21, + "args": { + "External id": 10618,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10617 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792656499, "dur": 4, + "args": { + "External id": 10619,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10618 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792656505, "dur": 14, + "args": { + "External id": 10620,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10619 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792656506, "dur": 12, + "args": { + "External id": 10621,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10620 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792656538, "dur": 5, + "args": { + "External id": 10622,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10621 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792656547, "dur": 21, + "args": { + "External id": 10623,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10622 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792656548, "dur": 4, + "args": { + "External id": 10624,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10623 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792656554, "dur": 14, + "args": { + "External id": 10625,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10624 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792656555, "dur": 12, + "args": { + "External id": 10626,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10625 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792656586, "dur": 6, + "args": { + "External id": 10627,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10626 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792656595, "dur": 22, + "args": { + "External id": 10628,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10627 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792656597, "dur": 4, + "args": { + "External id": 10629,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10628 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792656603, "dur": 14, + "args": { + "External id": 10630,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10629 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792656604, "dur": 11, + "args": { + "External id": 10631,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10630 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792656635, "dur": 5, + "args": { + "External id": 10632,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10631 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792656644, "dur": 21, + "args": { + "External id": 10633,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10632 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792656645, "dur": 4, + "args": { + "External id": 10634,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10633 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792656651, "dur": 14, + "args": { + "External id": 10635,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10634 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792656652, "dur": 11, + "args": { + "External id": 10636,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10635 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792656683, "dur": 5, + "args": { + "External id": 10637,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10636 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792656692, "dur": 21, + "args": { + "External id": 10638,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10637 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792656693, "dur": 5, + "args": { + "External id": 10639,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10638 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792656699, "dur": 14, + "args": { + "External id": 10640,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10639 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792656700, "dur": 11, + "args": { + "External id": 10641,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10640 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792656731, "dur": 5, + "args": { + "External id": 10642,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10641 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792656740, "dur": 21, + "args": { + "External id": 10643,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10642 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792656741, "dur": 5, + "args": { + "External id": 10644,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10643 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792656747, "dur": 14, + "args": { + "External id": 10645,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10644 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792656749, "dur": 10, + "args": { + "External id": 10646,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10645 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792656778, "dur": 6, + "args": { + "External id": 10647,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10646 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792656793, "dur": 23, + "args": { + "External id": 10648,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10647 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792656795, "dur": 5, + "args": { + "External id": 10649,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10648 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792656801, "dur": 15, + "args": { + "External id": 10650,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10649 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792656803, "dur": 12, + "args": { + "External id": 10651,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10650 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792656835, "dur": 5, + "args": { + "External id": 10652,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10651 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792656845, "dur": 21, + "args": { + "External id": 10653,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10652 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792656846, "dur": 5, + "args": { + "External id": 10654,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10653 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792656852, "dur": 14, + "args": { + "External id": 10655,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10654 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792656854, "dur": 11, + "args": { + "External id": 10656,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10655 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792656884, "dur": 6, + "args": { + "External id": 10657,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10656 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792656894, "dur": 21, + "args": { + "External id": 10658,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10657 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792656895, "dur": 4, + "args": { + "External id": 10659,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10658 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792656900, "dur": 15, + "args": { + "External id": 10660,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10659 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792656902, "dur": 11, + "args": { + "External id": 10661,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10660 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792656933, "dur": 5, + "args": { + "External id": 10662,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10661 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792656942, "dur": 21, + "args": { + "External id": 10663,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10662 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792656943, "dur": 5, + "args": { + "External id": 10664,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10663 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792656949, "dur": 14, + "args": { + "External id": 10665,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10664 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792656951, "dur": 11, + "args": { + "External id": 10666,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10665 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792656981, "dur": 6, + "args": { + "External id": 10667,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10666 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792656990, "dur": 22, + "args": { + "External id": 10668,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10667 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792656991, "dur": 5, + "args": { + "External id": 10669,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10668 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792656997, "dur": 14, + "args": { + "External id": 10670,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10669 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792656999, "dur": 11, + "args": { + "External id": 10671,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10670 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657030, "dur": 5, + "args": { + "External id": 10672,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10671 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792657039, "dur": 21, + "args": { + "External id": 10673,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10672 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657040, "dur": 4, + "args": { + "External id": 10674,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10673 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792657046, "dur": 14, + "args": { + "External id": 10675,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10674 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792657047, "dur": 12, + "args": { + "External id": 10676,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10675 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657078, "dur": 5, + "args": { + "External id": 10677,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10676 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792657087, "dur": 21, + "args": { + "External id": 10678,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10677 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657088, "dur": 4, + "args": { + "External id": 10679,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10678 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792657093, "dur": 15, + "args": { + "External id": 10680,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10679 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792657095, "dur": 11, + "args": { + "External id": 10681,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10680 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657206, "dur": 7, + "args": { + "External id": 10682,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10681 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792657218, "dur": 26, + "args": { + "External id": 10683,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10682 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657220, "dur": 4, + "args": { + "External id": 10684,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10683 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792657226, "dur": 18, + "args": { + "External id": 10685,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10684 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792657228, "dur": 14, + "args": { + "External id": 10686,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10685 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657266, "dur": 6, + "args": { + "External id": 10687,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10686 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792657276, "dur": 22, + "args": { + "External id": 10688,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10687 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657277, "dur": 4, + "args": { + "External id": 10689,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10688 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792657283, "dur": 14, + "args": { + "External id": 10690,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10689 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792657284, "dur": 12, + "args": { + "External id": 10691,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10690 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657316, "dur": 6, + "args": { + "External id": 10692,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10691 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792657326, "dur": 21, + "args": { + "External id": 10693,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10692 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657327, "dur": 4, + "args": { + "External id": 10694,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10693 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792657332, "dur": 15, + "args": { + "External id": 10695,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10694 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792657334, "dur": 12, + "args": { + "External id": 10696,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10695 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657365, "dur": 6, + "args": { + "External id": 10697,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10696 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792657375, "dur": 21, + "args": { + "External id": 10698,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10697 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657376, "dur": 4, + "args": { + "External id": 10699,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10698 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792657382, "dur": 14, + "args": { + "External id": 10700,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10699 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792657383, "dur": 11, + "args": { + "External id": 10701,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10700 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657416, "dur": 6, + "args": { + "External id": 10702,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10701 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792657426, "dur": 21, + "args": { + "External id": 10703,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10702 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657427, "dur": 4, + "args": { + "External id": 10704,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10703 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792657432, "dur": 14, + "args": { + "External id": 10705,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10704 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792657434, "dur": 11, + "args": { + "External id": 10706,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10705 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657465, "dur": 5, + "args": { + "External id": 10707,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10706 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792657474, "dur": 21, + "args": { + "External id": 10708,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10707 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657475, "dur": 4, + "args": { + "External id": 10709,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10708 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792657481, "dur": 14, + "args": { + "External id": 10710,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10709 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792657483, "dur": 11, + "args": { + "External id": 10711,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10710 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657513, "dur": 5, + "args": { + "External id": 10712,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10711 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792657522, "dur": 21, + "args": { + "External id": 10713,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10712 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657523, "dur": 4, + "args": { + "External id": 10714,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10713 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792657529, "dur": 14, + "args": { + "External id": 10715,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10714 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792657530, "dur": 11, + "args": { + "External id": 10716,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10715 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657561, "dur": 5, + "args": { + "External id": 10717,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10716 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792657570, "dur": 21, + "args": { + "External id": 10718,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10717 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657571, "dur": 4, + "args": { + "External id": 10719,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10718 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792657577, "dur": 14, + "args": { + "External id": 10720,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10719 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792657578, "dur": 11, + "args": { + "External id": 10721,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10720 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657609, "dur": 5, + "args": { + "External id": 10722,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10721 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792657618, "dur": 22, + "args": { + "External id": 10723,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10722 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657619, "dur": 5, + "args": { + "External id": 10724,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10723 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792657625, "dur": 14, + "args": { + "External id": 10725,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10724 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792657627, "dur": 11, + "args": { + "External id": 10726,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10725 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657657, "dur": 5, + "args": { + "External id": 10727,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10726 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792657666, "dur": 21, + "args": { + "External id": 10728,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10727 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657667, "dur": 5, + "args": { + "External id": 10729,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10728 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792657673, "dur": 14, + "args": { + "External id": 10730,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10729 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792657675, "dur": 11, + "args": { + "External id": 10731,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10730 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657760, "dur": 6, + "args": { + "External id": 10732,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10731 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792657771, "dur": 25, + "args": { + "External id": 10733,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10732 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657772, "dur": 5, + "args": { + "External id": 10734,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10733 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792657778, "dur": 18, + "args": { + "External id": 10735,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10734 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792657780, "dur": 15, + "args": { + "External id": 10736,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10735 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657817, "dur": 6, + "args": { + "External id": 10737,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10736 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792657827, "dur": 22, + "args": { + "External id": 10738,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10737 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657828, "dur": 4, + "args": { + "External id": 10739,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10738 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792657834, "dur": 14, + "args": { + "External id": 10740,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10739 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792657835, "dur": 12, + "args": { + "External id": 10741,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10740 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657867, "dur": 5, + "args": { + "External id": 10742,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10741 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792657876, "dur": 22, + "args": { + "External id": 10743,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10742 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657877, "dur": 5, + "args": { + "External id": 10744,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10743 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792657883, "dur": 14, + "args": { + "External id": 10745,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10744 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792657885, "dur": 11, + "args": { + "External id": 10746,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10745 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657915, "dur": 6, + "args": { + "External id": 10747,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10746 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792657925, "dur": 21, + "args": { + "External id": 10748,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10747 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657926, "dur": 5, + "args": { + "External id": 10749,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10748 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792657932, "dur": 14, + "args": { + "External id": 10750,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10749 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792657933, "dur": 12, + "args": { + "External id": 10751,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10750 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657964, "dur": 5, + "args": { + "External id": 10752,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10751 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792657982, "dur": 22, + "args": { + "External id": 10753,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10752 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792657984, "dur": 4, + "args": { + "External id": 10754,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10753 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792657989, "dur": 15, + "args": { + "External id": 10755,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10754 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792657991, "dur": 11, + "args": { + "External id": 10756,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10755 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658022, "dur": 6, + "args": { + "External id": 10757,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10756 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792658032, "dur": 21, + "args": { + "External id": 10758,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10757 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658033, "dur": 4, + "args": { + "External id": 10759,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10758 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792658038, "dur": 14, + "args": { + "External id": 10760,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10759 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792658040, "dur": 11, + "args": { + "External id": 10761,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10760 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658071, "dur": 5, + "args": { + "External id": 10762,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10761 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792658080, "dur": 21, + "args": { + "External id": 10763,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10762 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658081, "dur": 4, + "args": { + "External id": 10764,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10763 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792658087, "dur": 13, + "args": { + "External id": 10765,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10764 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792658088, "dur": 11, + "args": { + "External id": 10766,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10765 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658119, "dur": 5, + "args": { + "External id": 10767,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10766 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792658128, "dur": 21, + "args": { + "External id": 10768,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10767 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658129, "dur": 5, + "args": { + "External id": 10769,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10768 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792658135, "dur": 14, + "args": { + "External id": 10770,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10769 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792658136, "dur": 11, + "args": { + "External id": 10771,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10770 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658167, "dur": 5, + "args": { + "External id": 10772,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10771 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792658176, "dur": 22, + "args": { + "External id": 10773,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10772 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658177, "dur": 5, + "args": { + "External id": 10774,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10773 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792658183, "dur": 14, + "args": { + "External id": 10775,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10774 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792658185, "dur": 11, + "args": { + "External id": 10776,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10775 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658215, "dur": 6, + "args": { + "External id": 10777,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10776 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792658224, "dur": 22, + "args": { + "External id": 10778,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10777 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658226, "dur": 4, + "args": { + "External id": 10779,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10778 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792658231, "dur": 14, + "args": { + "External id": 10780,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10779 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792658233, "dur": 11, + "args": { + "External id": 10781,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10780 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658263, "dur": 5, + "args": { + "External id": 10782,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10781 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792658272, "dur": 22, + "args": { + "External id": 10783,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10782 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658274, "dur": 4, + "args": { + "External id": 10784,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10783 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792658279, "dur": 15, + "args": { + "External id": 10785,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10784 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792658281, "dur": 11, + "args": { + "External id": 10786,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10785 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658312, "dur": 5, + "args": { + "External id": 10787,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10786 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792658321, "dur": 21, + "args": { + "External id": 10788,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10787 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658322, "dur": 4, + "args": { + "External id": 10789,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10788 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792658328, "dur": 14, + "args": { + "External id": 10790,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10789 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792658329, "dur": 11, + "args": { + "External id": 10791,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10790 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658360, "dur": 5, + "args": { + "External id": 10792,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10791 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792658369, "dur": 21, + "args": { + "External id": 10793,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10792 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658370, "dur": 4, + "args": { + "External id": 10794,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10793 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792658375, "dur": 15, + "args": { + "External id": 10795,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10794 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792658377, "dur": 11, + "args": { + "External id": 10796,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10795 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658408, "dur": 5, + "args": { + "External id": 10797,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10796 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792658417, "dur": 21, + "args": { + "External id": 10798,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10797 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658418, "dur": 4, + "args": { + "External id": 10799,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10798 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792658423, "dur": 15, + "args": { + "External id": 10800,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10799 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792658425, "dur": 11, + "args": { + "External id": 10801,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10800 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658456, "dur": 5, + "args": { + "External id": 10802,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10801 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792658465, "dur": 21, + "args": { + "External id": 10803,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10802 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658466, "dur": 4, + "args": { + "External id": 10804,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10803 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792658471, "dur": 15, + "args": { + "External id": 10805,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10804 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792658473, "dur": 11, + "args": { + "External id": 10806,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10805 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658504, "dur": 5, + "args": { + "External id": 10807,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10806 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792658513, "dur": 21, + "args": { + "External id": 10808,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10807 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658514, "dur": 5, + "args": { + "External id": 10809,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10808 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792658520, "dur": 14, + "args": { + "External id": 10810,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10809 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792658521, "dur": 11, + "args": { + "External id": 10811,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10810 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658552, "dur": 5, + "args": { + "External id": 10812,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10811 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792658561, "dur": 21, + "args": { + "External id": 10813,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10812 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658562, "dur": 4, + "args": { + "External id": 10814,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10813 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792658567, "dur": 14, + "args": { + "External id": 10815,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10814 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792658569, "dur": 11, + "args": { + "External id": 10816,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10815 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658599, "dur": 5, + "args": { + "External id": 10817,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10816 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792658608, "dur": 22, + "args": { + "External id": 10818,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10817 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658609, "dur": 5, + "args": { + "External id": 10819,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10818 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792658615, "dur": 15, + "args": { + "External id": 10820,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10819 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792658617, "dur": 11, + "args": { + "External id": 10821,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10820 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658647, "dur": 6, + "args": { + "External id": 10822,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10821 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792658656, "dur": 22, + "args": { + "External id": 10823,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10822 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658658, "dur": 4, + "args": { + "External id": 10824,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10823 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792658663, "dur": 14, + "args": { + "External id": 10825,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10824 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792658665, "dur": 11, + "args": { + "External id": 10826,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10825 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658695, "dur": 5, + "args": { + "External id": 10827,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10826 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792658704, "dur": 21, + "args": { + "External id": 10828,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10827 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658705, "dur": 5, + "args": { + "External id": 10829,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10828 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792658711, "dur": 14, + "args": { + "External id": 10830,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10829 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792658713, "dur": 11, + "args": { + "External id": 10831,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10830 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658743, "dur": 5, + "args": { + "External id": 10832,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10831 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792658752, "dur": 21, + "args": { + "External id": 10833,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10832 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658753, "dur": 4, + "args": { + "External id": 10834,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10833 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792658759, "dur": 14, + "args": { + "External id": 10835,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10834 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792658760, "dur": 11, + "args": { + "External id": 10836,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10835 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658791, "dur": 5, + "args": { + "External id": 10837,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10836 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792658800, "dur": 21, + "args": { + "External id": 10838,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10837 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658801, "dur": 4, + "args": { + "External id": 10839,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10838 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792658807, "dur": 14, + "args": { + "External id": 10840,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10839 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792658808, "dur": 11, + "args": { + "External id": 10841,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10840 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658839, "dur": 5, + "args": { + "External id": 10842,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10841 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792658848, "dur": 21, + "args": { + "External id": 10843,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10842 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658849, "dur": 4, + "args": { + "External id": 10844,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10843 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792658854, "dur": 14, + "args": { + "External id": 10845,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10844 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792658856, "dur": 11, + "args": { + "External id": 10846,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10845 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658887, "dur": 5, + "args": { + "External id": 10847,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10846 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792658896, "dur": 21, + "args": { + "External id": 10848,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10847 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658897, "dur": 5, + "args": { + "External id": 10849,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10848 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792658903, "dur": 14, + "args": { + "External id": 10850,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10849 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792658904, "dur": 12, + "args": { + "External id": 10851,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10850 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658935, "dur": 5, + "args": { + "External id": 10852,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10851 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792658944, "dur": 21, + "args": { + "External id": 10853,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10852 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658945, "dur": 4, + "args": { + "External id": 10854,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10853 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792658951, "dur": 14, + "args": { + "External id": 10855,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10854 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792658952, "dur": 11, + "args": { + "External id": 10856,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10855 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658984, "dur": 5, + "args": { + "External id": 10857,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10856 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792658993, "dur": 21, + "args": { + "External id": 10858,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10857 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792658994, "dur": 5, + "args": { + "External id": 10859,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10858 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792659000, "dur": 14, + "args": { + "External id": 10860,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10859 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792659002, "dur": 11, + "args": { + "External id": 10861,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10860 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659032, "dur": 5, + "args": { + "External id": 10862,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10861 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792659041, "dur": 21, + "args": { + "External id": 10863,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10862 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659042, "dur": 5, + "args": { + "External id": 10864,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10863 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792659048, "dur": 14, + "args": { + "External id": 10865,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10864 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792659049, "dur": 11, + "args": { + "External id": 10866,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10865 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659080, "dur": 5, + "args": { + "External id": 10867,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10866 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792659089, "dur": 21, + "args": { + "External id": 10868,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10867 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659090, "dur": 5, + "args": { + "External id": 10869,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10868 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792659096, "dur": 14, + "args": { + "External id": 10870,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10869 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792659098, "dur": 11, + "args": { + "External id": 10871,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10870 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659128, "dur": 5, + "args": { + "External id": 10872,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10871 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792659137, "dur": 21, + "args": { + "External id": 10873,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10872 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659138, "dur": 5, + "args": { + "External id": 10874,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10873 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792659144, "dur": 14, + "args": { + "External id": 10875,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10874 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792659146, "dur": 11, + "args": { + "External id": 10876,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10875 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659176, "dur": 5, + "args": { + "External id": 10877,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10876 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792659185, "dur": 21, + "args": { + "External id": 10878,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10877 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659186, "dur": 4, + "args": { + "External id": 10879,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10878 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792659191, "dur": 14, + "args": { + "External id": 10880,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10879 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792659193, "dur": 11, + "args": { + "External id": 10881,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10880 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659224, "dur": 5, + "args": { + "External id": 10882,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10881 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792659233, "dur": 21, + "args": { + "External id": 10883,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10882 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659234, "dur": 5, + "args": { + "External id": 10884,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10883 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792659240, "dur": 14, + "args": { + "External id": 10885,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10884 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792659241, "dur": 11, + "args": { + "External id": 10886,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10885 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659272, "dur": 6, + "args": { + "External id": 10887,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10886 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792659281, "dur": 22, + "args": { + "External id": 10888,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10887 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659283, "dur": 4, + "args": { + "External id": 10889,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10888 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792659288, "dur": 14, + "args": { + "External id": 10890,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10889 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792659290, "dur": 11, + "args": { + "External id": 10891,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10890 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659320, "dur": 5, + "args": { + "External id": 10892,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10891 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792659329, "dur": 22, + "args": { + "External id": 10893,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10892 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659330, "dur": 5, + "args": { + "External id": 10894,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10893 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792659336, "dur": 14, + "args": { + "External id": 10895,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10894 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792659338, "dur": 11, + "args": { + "External id": 10896,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10895 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659368, "dur": 6, + "args": { + "External id": 10897,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10896 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792659378, "dur": 21, + "args": { + "External id": 10898,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10897 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659379, "dur": 4, + "args": { + "External id": 10899,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10898 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792659384, "dur": 14, + "args": { + "External id": 10900,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10899 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792659386, "dur": 11, + "args": { + "External id": 10901,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10900 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659416, "dur": 5, + "args": { + "External id": 10902,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10901 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792659425, "dur": 21, + "args": { + "External id": 10903,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10902 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659426, "dur": 5, + "args": { + "External id": 10904,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10903 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792659432, "dur": 14, + "args": { + "External id": 10905,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10904 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792659434, "dur": 11, + "args": { + "External id": 10906,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10905 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659464, "dur": 5, + "args": { + "External id": 10907,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10906 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792659473, "dur": 21, + "args": { + "External id": 10908,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10907 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659474, "dur": 5, + "args": { + "External id": 10909,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10908 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792659480, "dur": 14, + "args": { + "External id": 10910,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10909 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792659482, "dur": 11, + "args": { + "External id": 10911,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10910 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659512, "dur": 5, + "args": { + "External id": 10912,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10911 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792659521, "dur": 21, + "args": { + "External id": 10913,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10912 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659522, "dur": 5, + "args": { + "External id": 10914,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10913 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792659528, "dur": 14, + "args": { + "External id": 10915,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10914 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792659529, "dur": 11, + "args": { + "External id": 10916,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10915 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659560, "dur": 6, + "args": { + "External id": 10917,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10916 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792659569, "dur": 21, + "args": { + "External id": 10918,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10917 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659571, "dur": 4, + "args": { + "External id": 10919,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10918 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792659576, "dur": 14, + "args": { + "External id": 10920,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10919 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792659578, "dur": 11, + "args": { + "External id": 10921,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10920 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659608, "dur": 5, + "args": { + "External id": 10922,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10921 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792659617, "dur": 21, + "args": { + "External id": 10923,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10922 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659618, "dur": 4, + "args": { + "External id": 10924,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10923 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792659624, "dur": 14, + "args": { + "External id": 10925,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10924 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792659625, "dur": 11, + "args": { + "External id": 10926,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10925 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659656, "dur": 5, + "args": { + "External id": 10927,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10926 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792659665, "dur": 21, + "args": { + "External id": 10928,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10927 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659666, "dur": 4, + "args": { + "External id": 10929,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10928 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792659671, "dur": 14, + "args": { + "External id": 10930,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10929 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792659673, "dur": 11, + "args": { + "External id": 10931,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10930 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659703, "dur": 5, + "args": { + "External id": 10932,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10931 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792659712, "dur": 21, + "args": { + "External id": 10933,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10932 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659714, "dur": 4, + "args": { + "External id": 10934,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10933 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792659719, "dur": 14, + "args": { + "External id": 10935,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10934 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792659721, "dur": 11, + "args": { + "External id": 10936,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10935 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659751, "dur": 5, + "args": { + "External id": 10937,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10936 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792659760, "dur": 21, + "args": { + "External id": 10938,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10937 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659761, "dur": 5, + "args": { + "External id": 10939,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10938 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792659767, "dur": 14, + "args": { + "External id": 10940,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10939 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792659768, "dur": 12, + "args": { + "External id": 10941,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10940 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659799, "dur": 5, + "args": { + "External id": 10942,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10941 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792659808, "dur": 21, + "args": { + "External id": 10943,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10942 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659809, "dur": 4, + "args": { + "External id": 10944,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10943 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792659814, "dur": 14, + "args": { + "External id": 10945,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10944 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792659816, "dur": 11, + "args": { + "External id": 10946,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10945 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659846, "dur": 5, + "args": { + "External id": 10947,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10946 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792659855, "dur": 22, + "args": { + "External id": 10948,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10947 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659856, "dur": 5, + "args": { + "External id": 10949,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10948 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792659862, "dur": 14, + "args": { + "External id": 10950,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10949 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792659864, "dur": 11, + "args": { + "External id": 10951,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10950 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659894, "dur": 6, + "args": { + "External id": 10952,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10951 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792659903, "dur": 22, + "args": { + "External id": 10953,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10952 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659905, "dur": 4, + "args": { + "External id": 10954,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10953 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792659910, "dur": 14, + "args": { + "External id": 10955,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10954 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792659912, "dur": 11, + "args": { + "External id": 10956,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10955 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659942, "dur": 5, + "args": { + "External id": 10957,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10956 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792659951, "dur": 21, + "args": { + "External id": 10958,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10957 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659952, "dur": 5, + "args": { + "External id": 10959,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10958 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792659958, "dur": 14, + "args": { + "External id": 10960,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10959 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792659960, "dur": 11, + "args": { + "External id": 10961,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10960 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792659990, "dur": 5, + "args": { + "External id": 10962,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10961 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792659999, "dur": 21, + "args": { + "External id": 10963,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10962 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660000, "dur": 5, + "args": { + "External id": 10964,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10963 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792660006, "dur": 14, + "args": { + "External id": 10965,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10964 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792660008, "dur": 11, + "args": { + "External id": 10966,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10965 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660038, "dur": 5, + "args": { + "External id": 10967,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10966 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792660047, "dur": 21, + "args": { + "External id": 10968,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10967 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660048, "dur": 5, + "args": { + "External id": 10969,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10968 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792660054, "dur": 14, + "args": { + "External id": 10970,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10969 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792660056, "dur": 11, + "args": { + "External id": 10971,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10970 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660086, "dur": 5, + "args": { + "External id": 10972,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10971 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792660095, "dur": 21, + "args": { + "External id": 10973,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10972 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660096, "dur": 5, + "args": { + "External id": 10974,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10973 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792660102, "dur": 14, + "args": { + "External id": 10975,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10974 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792660104, "dur": 11, + "args": { + "External id": 10976,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10975 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660134, "dur": 5, + "args": { + "External id": 10977,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10976 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792660143, "dur": 21, + "args": { + "External id": 10978,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10977 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660144, "dur": 5, + "args": { + "External id": 10979,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10978 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792660150, "dur": 14, + "args": { + "External id": 10980,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10979 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792660152, "dur": 11, + "args": { + "External id": 10981,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10980 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660266, "dur": 6, + "args": { + "External id": 10982,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10981 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792660277, "dur": 27, + "args": { + "External id": 10983,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10982 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660279, "dur": 5, + "args": { + "External id": 10984,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10983 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792660285, "dur": 18, + "args": { + "External id": 10985,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10984 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792660287, "dur": 15, + "args": { + "External id": 10986,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10985 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660325, "dur": 6, + "args": { + "External id": 10987,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10986 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792660335, "dur": 21, + "args": { + "External id": 10988,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10987 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660336, "dur": 5, + "args": { + "External id": 10989,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10988 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792660342, "dur": 14, + "args": { + "External id": 10990,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10989 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792660344, "dur": 11, + "args": { + "External id": 10991,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10990 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660375, "dur": 6, + "args": { + "External id": 10992,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10991 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792660384, "dur": 22, + "args": { + "External id": 10993,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10992 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660386, "dur": 4, + "args": { + "External id": 10994,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10993 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792660391, "dur": 15, + "args": { + "External id": 10995,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10994 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792660393, "dur": 11, + "args": { + "External id": 10996,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 10995 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660424, "dur": 5, + "args": { + "External id": 10997,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10996 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792660433, "dur": 21, + "args": { + "External id": 10998,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 10997 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660434, "dur": 5, + "args": { + "External id": 10999,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10998 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792660440, "dur": 14, + "args": { + "External id": 11000,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 10999 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792660442, "dur": 11, + "args": { + "External id": 11001,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11000 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660473, "dur": 5, + "args": { + "External id": 11002,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11001 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792660482, "dur": 21, + "args": { + "External id": 11003,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11002 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660483, "dur": 4, + "args": { + "External id": 11004,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11003 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792660488, "dur": 15, + "args": { + "External id": 11005,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11004 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792660490, "dur": 11, + "args": { + "External id": 11006,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11005 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660521, "dur": 5, + "args": { + "External id": 11007,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11006 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792660530, "dur": 21, + "args": { + "External id": 11008,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11007 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660531, "dur": 5, + "args": { + "External id": 11009,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11008 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792660537, "dur": 14, + "args": { + "External id": 11010,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11009 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792660538, "dur": 12, + "args": { + "External id": 11011,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11010 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660569, "dur": 5, + "args": { + "External id": 11012,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11011 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792660578, "dur": 22, + "args": { + "External id": 11013,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11012 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660579, "dur": 5, + "args": { + "External id": 11014,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11013 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792660585, "dur": 14, + "args": { + "External id": 11015,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11014 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792660587, "dur": 11, + "args": { + "External id": 11016,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11015 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660617, "dur": 6, + "args": { + "External id": 11017,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11016 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792660627, "dur": 21, + "args": { + "External id": 11018,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11017 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660628, "dur": 4, + "args": { + "External id": 11019,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11018 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792660633, "dur": 15, + "args": { + "External id": 11020,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11019 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792660635, "dur": 11, + "args": { + "External id": 11021,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11020 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660665, "dur": 5, + "args": { + "External id": 11022,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11021 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792660674, "dur": 21, + "args": { + "External id": 11023,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11022 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660675, "dur": 5, + "args": { + "External id": 11024,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11023 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792660681, "dur": 14, + "args": { + "External id": 11025,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11024 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792660683, "dur": 11, + "args": { + "External id": 11026,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11025 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660713, "dur": 5, + "args": { + "External id": 11027,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11026 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792660722, "dur": 21, + "args": { + "External id": 11028,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11027 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660723, "dur": 4, + "args": { + "External id": 11029,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11028 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792660729, "dur": 14, + "args": { + "External id": 11030,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11029 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792660730, "dur": 12, + "args": { + "External id": 11031,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11030 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660815, "dur": 6, + "args": { + "External id": 11032,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11031 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792660826, "dur": 25, + "args": { + "External id": 11033,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11032 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660827, "dur": 5, + "args": { + "External id": 11034,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11033 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792660833, "dur": 17, + "args": { + "External id": 11035,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11034 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792660835, "dur": 14, + "args": { + "External id": 11036,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11035 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660871, "dur": 6, + "args": { + "External id": 11037,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11036 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792660881, "dur": 21, + "args": { + "External id": 11038,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11037 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660882, "dur": 4, + "args": { + "External id": 11039,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11038 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792660887, "dur": 15, + "args": { + "External id": 11040,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11039 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792660889, "dur": 12, + "args": { + "External id": 11041,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11040 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660921, "dur": 5, + "args": { + "External id": 11042,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11041 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792660930, "dur": 21, + "args": { + "External id": 11043,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11042 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660931, "dur": 4, + "args": { + "External id": 11044,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11043 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792660937, "dur": 14, + "args": { + "External id": 11045,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11044 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792660938, "dur": 11, + "args": { + "External id": 11046,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11045 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660969, "dur": 5, + "args": { + "External id": 11047,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11046 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792660978, "dur": 22, + "args": { + "External id": 11048,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11047 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792660979, "dur": 5, + "args": { + "External id": 11049,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11048 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792660985, "dur": 14, + "args": { + "External id": 11050,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11049 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792660986, "dur": 12, + "args": { + "External id": 11051,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11050 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661018, "dur": 5, + "args": { + "External id": 11052,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11051 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792661027, "dur": 21, + "args": { + "External id": 11053,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11052 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661028, "dur": 4, + "args": { + "External id": 11054,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11053 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792661033, "dur": 14, + "args": { + "External id": 11055,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11054 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792661035, "dur": 11, + "args": { + "External id": 11056,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11055 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661065, "dur": 6, + "args": { + "External id": 11057,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11056 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792661074, "dur": 22, + "args": { + "External id": 11058,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11057 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661075, "dur": 5, + "args": { + "External id": 11059,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11058 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792661081, "dur": 14, + "args": { + "External id": 11060,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11059 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792661083, "dur": 11, + "args": { + "External id": 11061,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11060 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661114, "dur": 5, + "args": { + "External id": 11062,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11061 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792661123, "dur": 21, + "args": { + "External id": 11063,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11062 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661125, "dur": 4, + "args": { + "External id": 11064,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11063 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792661130, "dur": 14, + "args": { + "External id": 11065,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11064 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792661132, "dur": 11, + "args": { + "External id": 11066,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11065 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661162, "dur": 6, + "args": { + "External id": 11067,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11066 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792661172, "dur": 21, + "args": { + "External id": 11068,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11067 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661173, "dur": 4, + "args": { + "External id": 11069,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11068 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792661178, "dur": 14, + "args": { + "External id": 11070,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11069 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792661180, "dur": 11, + "args": { + "External id": 11071,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11070 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661211, "dur": 5, + "args": { + "External id": 11072,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11071 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792661220, "dur": 21, + "args": { + "External id": 11073,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11072 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661221, "dur": 4, + "args": { + "External id": 11074,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11073 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792661227, "dur": 13, + "args": { + "External id": 11075,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11074 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792661228, "dur": 11, + "args": { + "External id": 11076,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11075 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661259, "dur": 5, + "args": { + "External id": 11077,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11076 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792661268, "dur": 21, + "args": { + "External id": 11078,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11077 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661269, "dur": 5, + "args": { + "External id": 11079,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11078 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792661275, "dur": 14, + "args": { + "External id": 11080,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11079 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792661276, "dur": 11, + "args": { + "External id": 11081,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11080 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661307, "dur": 5, + "args": { + "External id": 11082,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11081 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792661316, "dur": 21, + "args": { + "External id": 11083,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11082 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661317, "dur": 4, + "args": { + "External id": 11084,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11083 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792661323, "dur": 14, + "args": { + "External id": 11085,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11084 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792661324, "dur": 11, + "args": { + "External id": 11086,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11085 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661354, "dur": 6, + "args": { + "External id": 11087,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11086 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792661364, "dur": 21, + "args": { + "External id": 11088,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11087 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661365, "dur": 4, + "args": { + "External id": 11089,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11088 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792661371, "dur": 14, + "args": { + "External id": 11090,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11089 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792661372, "dur": 11, + "args": { + "External id": 11091,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11090 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661403, "dur": 5, + "args": { + "External id": 11092,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11091 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792661412, "dur": 21, + "args": { + "External id": 11093,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11092 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661413, "dur": 4, + "args": { + "External id": 11094,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11093 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792661418, "dur": 14, + "args": { + "External id": 11095,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11094 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792661420, "dur": 11, + "args": { + "External id": 11096,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11095 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661451, "dur": 5, + "args": { + "External id": 11097,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11096 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792661460, "dur": 21, + "args": { + "External id": 11098,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11097 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661461, "dur": 4, + "args": { + "External id": 11099,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11098 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792661467, "dur": 14, + "args": { + "External id": 11100,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11099 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792661468, "dur": 11, + "args": { + "External id": 11101,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11100 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661499, "dur": 5, + "args": { + "External id": 11102,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11101 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792661508, "dur": 21, + "args": { + "External id": 11103,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11102 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661509, "dur": 4, + "args": { + "External id": 11104,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11103 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792661514, "dur": 14, + "args": { + "External id": 11105,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11104 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792661516, "dur": 11, + "args": { + "External id": 11106,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11105 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661546, "dur": 5, + "args": { + "External id": 11107,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11106 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792661555, "dur": 21, + "args": { + "External id": 11108,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11107 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661557, "dur": 4, + "args": { + "External id": 11109,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11108 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792661562, "dur": 14, + "args": { + "External id": 11110,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11109 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792661564, "dur": 11, + "args": { + "External id": 11111,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11110 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661594, "dur": 5, + "args": { + "External id": 11112,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11111 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792661603, "dur": 21, + "args": { + "External id": 11113,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11112 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661604, "dur": 5, + "args": { + "External id": 11114,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11113 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792661610, "dur": 14, + "args": { + "External id": 11115,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11114 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792661612, "dur": 11, + "args": { + "External id": 11116,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11115 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661642, "dur": 5, + "args": { + "External id": 11117,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11116 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792661651, "dur": 21, + "args": { + "External id": 11118,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11117 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661652, "dur": 4, + "args": { + "External id": 11119,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11118 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792661658, "dur": 13, + "args": { + "External id": 11120,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11119 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792661659, "dur": 11, + "args": { + "External id": 11121,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11120 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661689, "dur": 6, + "args": { + "External id": 11122,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11121 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792661698, "dur": 22, + "args": { + "External id": 11123,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11122 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661700, "dur": 4, + "args": { + "External id": 11124,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11123 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792661705, "dur": 14, + "args": { + "External id": 11125,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11124 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792661707, "dur": 11, + "args": { + "External id": 11126,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11125 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661737, "dur": 6, + "args": { + "External id": 11127,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11126 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792661746, "dur": 22, + "args": { + "External id": 11128,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11127 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661748, "dur": 4, + "args": { + "External id": 11129,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11128 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792661753, "dur": 15, + "args": { + "External id": 11130,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11129 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792661755, "dur": 11, + "args": { + "External id": 11131,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11130 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661786, "dur": 5, + "args": { + "External id": 11132,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11131 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792661795, "dur": 21, + "args": { + "External id": 11133,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11132 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661796, "dur": 4, + "args": { + "External id": 11134,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11133 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792661801, "dur": 14, + "args": { + "External id": 11135,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11134 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792661803, "dur": 11, + "args": { + "External id": 11136,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11135 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661834, "dur": 5, + "args": { + "External id": 11137,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11136 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792661843, "dur": 21, + "args": { + "External id": 11138,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11137 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661844, "dur": 4, + "args": { + "External id": 11139,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11138 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792661849, "dur": 14, + "args": { + "External id": 11140,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11139 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792661851, "dur": 11, + "args": { + "External id": 11141,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11140 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661881, "dur": 6, + "args": { + "External id": 11142,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11141 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792661890, "dur": 21, + "args": { + "External id": 11143,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11142 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661892, "dur": 4, + "args": { + "External id": 11144,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11143 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792661897, "dur": 14, + "args": { + "External id": 11145,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11144 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792661899, "dur": 11, + "args": { + "External id": 11146,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11145 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661929, "dur": 5, + "args": { + "External id": 11147,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11146 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792661938, "dur": 21, + "args": { + "External id": 11148,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11147 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661939, "dur": 5, + "args": { + "External id": 11149,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11148 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792661945, "dur": 14, + "args": { + "External id": 11150,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11149 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792661946, "dur": 11, + "args": { + "External id": 11151,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11150 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661977, "dur": 5, + "args": { + "External id": 11152,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11151 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792661986, "dur": 21, + "args": { + "External id": 11153,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11152 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792661987, "dur": 4, + "args": { + "External id": 11154,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11153 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792661993, "dur": 13, + "args": { + "External id": 11155,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11154 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792661994, "dur": 11, + "args": { + "External id": 11156,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11155 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662024, "dur": 6, + "args": { + "External id": 11157,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11156 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792662033, "dur": 21, + "args": { + "External id": 11158,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11157 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662035, "dur": 4, + "args": { + "External id": 11159,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11158 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792662040, "dur": 14, + "args": { + "External id": 11160,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11159 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792662042, "dur": 11, + "args": { + "External id": 11161,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11160 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662072, "dur": 5, + "args": { + "External id": 11162,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11161 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792662081, "dur": 21, + "args": { + "External id": 11163,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11162 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662082, "dur": 5, + "args": { + "External id": 11164,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11163 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792662088, "dur": 14, + "args": { + "External id": 11165,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11164 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792662089, "dur": 12, + "args": { + "External id": 11166,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11165 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662120, "dur": 5, + "args": { + "External id": 11167,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11166 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792662129, "dur": 21, + "args": { + "External id": 11168,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11167 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662130, "dur": 4, + "args": { + "External id": 11169,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11168 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792662135, "dur": 14, + "args": { + "External id": 11170,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11169 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792662137, "dur": 11, + "args": { + "External id": 11171,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11170 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662168, "dur": 5, + "args": { + "External id": 11172,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11171 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792662177, "dur": 21, + "args": { + "External id": 11173,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11172 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662178, "dur": 4, + "args": { + "External id": 11174,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11173 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792662184, "dur": 13, + "args": { + "External id": 11175,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11174 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792662185, "dur": 11, + "args": { + "External id": 11176,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11175 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662216, "dur": 5, + "args": { + "External id": 11177,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11176 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792662225, "dur": 21, + "args": { + "External id": 11178,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11177 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662226, "dur": 4, + "args": { + "External id": 11179,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11178 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792662231, "dur": 14, + "args": { + "External id": 11180,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11179 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792662233, "dur": 11, + "args": { + "External id": 11181,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11180 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662264, "dur": 5, + "args": { + "External id": 11182,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11181 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792662273, "dur": 21, + "args": { + "External id": 11183,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11182 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662274, "dur": 4, + "args": { + "External id": 11184,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11183 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792662279, "dur": 14, + "args": { + "External id": 11185,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11184 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792662281, "dur": 11, + "args": { + "External id": 11186,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11185 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662312, "dur": 5, + "args": { + "External id": 11187,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11186 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792662321, "dur": 21, + "args": { + "External id": 11188,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11187 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662322, "dur": 4, + "args": { + "External id": 11189,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11188 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792662328, "dur": 14, + "args": { + "External id": 11190,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11189 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792662329, "dur": 12, + "args": { + "External id": 11191,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11190 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662360, "dur": 5, + "args": { + "External id": 11192,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11191 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792662369, "dur": 21, + "args": { + "External id": 11193,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11192 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662370, "dur": 5, + "args": { + "External id": 11194,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11193 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792662376, "dur": 14, + "args": { + "External id": 11195,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11194 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792662378, "dur": 11, + "args": { + "External id": 11196,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11195 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662408, "dur": 5, + "args": { + "External id": 11197,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11196 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792662417, "dur": 21, + "args": { + "External id": 11198,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11197 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662418, "dur": 5, + "args": { + "External id": 11199,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11198 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792662424, "dur": 14, + "args": { + "External id": 11200,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11199 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792662426, "dur": 10, + "args": { + "External id": 11201,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11200 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662456, "dur": 5, + "args": { + "External id": 11202,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11201 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792662465, "dur": 21, + "args": { + "External id": 11203,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11202 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662466, "dur": 5, + "args": { + "External id": 11204,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11203 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792662472, "dur": 14, + "args": { + "External id": 11205,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11204 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792662473, "dur": 12, + "args": { + "External id": 11206,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11205 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662504, "dur": 5, + "args": { + "External id": 11207,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11206 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792662513, "dur": 21, + "args": { + "External id": 11208,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11207 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662514, "dur": 5, + "args": { + "External id": 11209,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11208 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792662520, "dur": 14, + "args": { + "External id": 11210,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11209 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792662522, "dur": 11, + "args": { + "External id": 11211,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11210 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662552, "dur": 6, + "args": { + "External id": 11212,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11211 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792662563, "dur": 22, + "args": { + "External id": 11213,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11212 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662564, "dur": 5, + "args": { + "External id": 11214,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11213 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792662570, "dur": 14, + "args": { + "External id": 11215,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11214 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792662572, "dur": 11, + "args": { + "External id": 11216,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11215 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662603, "dur": 5, + "args": { + "External id": 11217,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11216 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792662611, "dur": 22, + "args": { + "External id": 11218,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11217 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662613, "dur": 4, + "args": { + "External id": 11219,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11218 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792662618, "dur": 14, + "args": { + "External id": 11220,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11219 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792662620, "dur": 11, + "args": { + "External id": 11221,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11220 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662650, "dur": 5, + "args": { + "External id": 11222,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11221 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792662659, "dur": 21, + "args": { + "External id": 11223,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11222 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662660, "dur": 5, + "args": { + "External id": 11224,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11223 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792662666, "dur": 14, + "args": { + "External id": 11225,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11224 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792662667, "dur": 11, + "args": { + "External id": 11226,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11225 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662698, "dur": 6, + "args": { + "External id": 11227,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11226 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792662708, "dur": 21, + "args": { + "External id": 11228,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11227 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662709, "dur": 4, + "args": { + "External id": 11229,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11228 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792662715, "dur": 13, + "args": { + "External id": 11230,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11229 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792662716, "dur": 11, + "args": { + "External id": 11231,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11230 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662746, "dur": 6, + "args": { + "External id": 11232,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11231 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792662756, "dur": 20, + "args": { + "External id": 11233,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11232 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662757, "dur": 4, + "args": { + "External id": 11234,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11233 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792662762, "dur": 14, + "args": { + "External id": 11235,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11234 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792662764, "dur": 11, + "args": { + "External id": 11236,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11235 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662794, "dur": 5, + "args": { + "External id": 11237,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11236 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792662803, "dur": 21, + "args": { + "External id": 11238,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11237 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662804, "dur": 4, + "args": { + "External id": 11239,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11238 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792662810, "dur": 14, + "args": { + "External id": 11240,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11239 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792662811, "dur": 11, + "args": { + "External id": 11241,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11240 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662842, "dur": 5, + "args": { + "External id": 11242,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11241 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792662851, "dur": 20, + "args": { + "External id": 11243,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11242 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662852, "dur": 4, + "args": { + "External id": 11244,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11243 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792662857, "dur": 14, + "args": { + "External id": 11245,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11244 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792662859, "dur": 11, + "args": { + "External id": 11246,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11245 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662889, "dur": 5, + "args": { + "External id": 11247,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11246 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792662898, "dur": 21, + "args": { + "External id": 11248,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11247 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662899, "dur": 4, + "args": { + "External id": 11249,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11248 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792662905, "dur": 14, + "args": { + "External id": 11250,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11249 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792662906, "dur": 11, + "args": { + "External id": 11251,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11250 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662936, "dur": 6, + "args": { + "External id": 11252,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11251 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792662945, "dur": 22, + "args": { + "External id": 11253,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11252 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662947, "dur": 4, + "args": { + "External id": 11254,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11253 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792662952, "dur": 14, + "args": { + "External id": 11255,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11254 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792662954, "dur": 11, + "args": { + "External id": 11256,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11255 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662984, "dur": 5, + "args": { + "External id": 11257,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11256 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792662993, "dur": 21, + "args": { + "External id": 11258,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11257 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792662994, "dur": 5, + "args": { + "External id": 11259,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11258 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792663000, "dur": 14, + "args": { + "External id": 11260,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11259 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792663002, "dur": 11, + "args": { + "External id": 11261,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11260 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663032, "dur": 5, + "args": { + "External id": 11262,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11261 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792663041, "dur": 32, + "args": { + "External id": 11263,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11262 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663042, "dur": 5, + "args": { + "External id": 11264,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11263 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792663058, "dur": 14, + "args": { + "External id": 11265,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11264 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792663060, "dur": 11, + "args": { + "External id": 11266,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11265 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663091, "dur": 5, + "args": { + "External id": 11267,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11266 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792663100, "dur": 22, + "args": { + "External id": 11268,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11267 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663102, "dur": 4, + "args": { + "External id": 11269,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11268 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792663107, "dur": 14, + "args": { + "External id": 11270,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11269 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792663109, "dur": 11, + "args": { + "External id": 11271,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11270 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663139, "dur": 6, + "args": { + "External id": 11272,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11271 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792663148, "dur": 22, + "args": { + "External id": 11273,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11272 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663150, "dur": 4, + "args": { + "External id": 11274,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11273 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792663155, "dur": 14, + "args": { + "External id": 11275,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11274 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792663157, "dur": 11, + "args": { + "External id": 11276,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11275 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663188, "dur": 5, + "args": { + "External id": 11277,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11276 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792663197, "dur": 21, + "args": { + "External id": 11278,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11277 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663198, "dur": 4, + "args": { + "External id": 11279,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11278 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792663204, "dur": 14, + "args": { + "External id": 11280,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11279 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792663205, "dur": 12, + "args": { + "External id": 11281,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11280 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663296, "dur": 7, + "args": { + "External id": 11282,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11281 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792663308, "dur": 25, + "args": { + "External id": 11283,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11282 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663309, "dur": 5, + "args": { + "External id": 11284,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11283 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792663315, "dur": 18, + "args": { + "External id": 11285,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11284 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792663317, "dur": 14, + "args": { + "External id": 11286,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11285 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663354, "dur": 6, + "args": { + "External id": 11287,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11286 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792663364, "dur": 22, + "args": { + "External id": 11288,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11287 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663365, "dur": 4, + "args": { + "External id": 11289,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11288 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792663371, "dur": 14, + "args": { + "External id": 11290,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11289 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792663372, "dur": 12, + "args": { + "External id": 11291,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11290 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663404, "dur": 5, + "args": { + "External id": 11292,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11291 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792663413, "dur": 21, + "args": { + "External id": 11293,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11292 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663414, "dur": 5, + "args": { + "External id": 11294,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11293 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792663420, "dur": 14, + "args": { + "External id": 11295,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11294 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792663422, "dur": 11, + "args": { + "External id": 11296,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11295 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663452, "dur": 5, + "args": { + "External id": 11297,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11296 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792663461, "dur": 22, + "args": { + "External id": 11298,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11297 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663462, "dur": 5, + "args": { + "External id": 11299,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11298 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792663468, "dur": 14, + "args": { + "External id": 11300,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11299 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792663470, "dur": 11, + "args": { + "External id": 11301,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11300 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663501, "dur": 5, + "args": { + "External id": 11302,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11301 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792663510, "dur": 21, + "args": { + "External id": 11303,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11302 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663511, "dur": 4, + "args": { + "External id": 11304,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11303 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792663517, "dur": 13, + "args": { + "External id": 11305,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11304 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792663518, "dur": 11, + "args": { + "External id": 11306,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11305 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663549, "dur": 5, + "args": { + "External id": 11307,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11306 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792663557, "dur": 22, + "args": { + "External id": 11308,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11307 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663559, "dur": 4, + "args": { + "External id": 11309,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11308 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792663564, "dur": 15, + "args": { + "External id": 11310,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11309 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792663566, "dur": 11, + "args": { + "External id": 11311,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11310 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663597, "dur": 5, + "args": { + "External id": 11312,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11311 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792663606, "dur": 21, + "args": { + "External id": 11313,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11312 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663607, "dur": 5, + "args": { + "External id": 11314,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11313 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792663613, "dur": 14, + "args": { + "External id": 11315,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11314 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792663614, "dur": 12, + "args": { + "External id": 11316,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11315 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663645, "dur": 6, + "args": { + "External id": 11317,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11316 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792663655, "dur": 21, + "args": { + "External id": 11318,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11317 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663656, "dur": 4, + "args": { + "External id": 11319,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11318 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792663661, "dur": 14, + "args": { + "External id": 11320,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11319 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792663663, "dur": 11, + "args": { + "External id": 11321,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11320 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663693, "dur": 6, + "args": { + "External id": 11322,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11321 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792663702, "dur": 22, + "args": { + "External id": 11323,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11322 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663704, "dur": 4, + "args": { + "External id": 11324,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11323 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792663709, "dur": 14, + "args": { + "External id": 11325,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11324 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792663711, "dur": 11, + "args": { + "External id": 11326,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11325 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663741, "dur": 5, + "args": { + "External id": 11327,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11326 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792663750, "dur": 22, + "args": { + "External id": 11328,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11327 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663751, "dur": 5, + "args": { + "External id": 11329,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11328 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792663757, "dur": 14, + "args": { + "External id": 11330,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11329 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792663759, "dur": 11, + "args": { + "External id": 11331,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11330 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663842, "dur": 7, + "args": { + "External id": 11332,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11331 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792663853, "dur": 25, + "args": { + "External id": 11333,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11332 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663855, "dur": 4, + "args": { + "External id": 11334,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11333 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792663861, "dur": 17, + "args": { + "External id": 11335,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11334 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792663862, "dur": 14, + "args": { + "External id": 11336,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11335 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663898, "dur": 6, + "args": { + "External id": 11337,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11336 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792663908, "dur": 22, + "args": { + "External id": 11338,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11337 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663909, "dur": 5, + "args": { + "External id": 11339,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11338 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792663915, "dur": 14, + "args": { + "External id": 11340,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11339 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792663916, "dur": 12, + "args": { + "External id": 11341,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11340 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663948, "dur": 5, + "args": { + "External id": 11342,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11341 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792663957, "dur": 22, + "args": { + "External id": 11343,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11342 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663958, "dur": 5, + "args": { + "External id": 11344,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11343 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792663964, "dur": 14, + "args": { + "External id": 11345,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11344 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792663966, "dur": 11, + "args": { + "External id": 11346,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11345 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792663997, "dur": 5, + "args": { + "External id": 11347,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11346 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792664006, "dur": 22, + "args": { + "External id": 11348,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11347 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664007, "dur": 5, + "args": { + "External id": 11349,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11348 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792664013, "dur": 14, + "args": { + "External id": 11350,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11349 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792664015, "dur": 11, + "args": { + "External id": 11351,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11350 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664045, "dur": 5, + "args": { + "External id": 11352,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11351 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792664054, "dur": 22, + "args": { + "External id": 11353,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11352 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664056, "dur": 4, + "args": { + "External id": 11354,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11353 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792664061, "dur": 15, + "args": { + "External id": 11355,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11354 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792664063, "dur": 11, + "args": { + "External id": 11356,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11355 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664094, "dur": 5, + "args": { + "External id": 11357,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11356 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792664103, "dur": 21, + "args": { + "External id": 11358,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11357 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664104, "dur": 4, + "args": { + "External id": 11359,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11358 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792664110, "dur": 14, + "args": { + "External id": 11360,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11359 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792664111, "dur": 11, + "args": { + "External id": 11361,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11360 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664142, "dur": 5, + "args": { + "External id": 11362,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11361 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792664151, "dur": 21, + "args": { + "External id": 11363,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11362 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664152, "dur": 5, + "args": { + "External id": 11364,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11363 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792664158, "dur": 14, + "args": { + "External id": 11365,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11364 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792664160, "dur": 11, + "args": { + "External id": 11366,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11365 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664190, "dur": 5, + "args": { + "External id": 11367,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11366 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792664199, "dur": 40, + "args": { + "External id": 11368,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11367 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664200, "dur": 5, + "args": { + "External id": 11369,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11368 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792664206, "dur": 33, + "args": { + "External id": 11370,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11369 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792664208, "dur": 29, + "args": { + "External id": 11371,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11370 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664261, "dur": 5, + "args": { + "External id": 11372,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11371 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792664271, "dur": 22, + "args": { + "External id": 11373,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11372 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664272, "dur": 4, + "args": { + "External id": 11374,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11373 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792664278, "dur": 14, + "args": { + "External id": 11375,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11374 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792664279, "dur": 12, + "args": { + "External id": 11376,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11375 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664311, "dur": 5, + "args": { + "External id": 11377,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11376 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792664320, "dur": 25, + "args": { + "External id": 11378,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11377 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664321, "dur": 5, + "args": { + "External id": 11379,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11378 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792664327, "dur": 18, + "args": { + "External id": 11380,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11379 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792664332, "dur": 11, + "args": { + "External id": 11381,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11380 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664364, "dur": 6, + "args": { + "External id": 11382,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11381 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792664375, "dur": 23, + "args": { + "External id": 11383,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11382 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664376, "dur": 5, + "args": { + "External id": 11384,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11383 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792664382, "dur": 16, + "args": { + "External id": 11385,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11384 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792664384, "dur": 12, + "args": { + "External id": 11386,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11385 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664416, "dur": 5, + "args": { + "External id": 11387,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11386 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792664425, "dur": 22, + "args": { + "External id": 11388,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11387 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664426, "dur": 5, + "args": { + "External id": 11389,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11388 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792664432, "dur": 15, + "args": { + "External id": 11390,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11389 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792664434, "dur": 11, + "args": { + "External id": 11391,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11390 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664465, "dur": 5, + "args": { + "External id": 11392,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11391 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792664474, "dur": 21, + "args": { + "External id": 11393,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11392 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664475, "dur": 4, + "args": { + "External id": 11394,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11393 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792664481, "dur": 14, + "args": { + "External id": 11395,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11394 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792664482, "dur": 11, + "args": { + "External id": 11396,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11395 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664513, "dur": 5, + "args": { + "External id": 11397,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11396 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792664522, "dur": 21, + "args": { + "External id": 11398,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11397 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664523, "dur": 5, + "args": { + "External id": 11399,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11398 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792664529, "dur": 14, + "args": { + "External id": 11400,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11399 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792664530, "dur": 12, + "args": { + "External id": 11401,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11400 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664561, "dur": 5, + "args": { + "External id": 11402,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11401 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792664570, "dur": 21, + "args": { + "External id": 11403,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11402 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664571, "dur": 4, + "args": { + "External id": 11404,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11403 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792664577, "dur": 14, + "args": { + "External id": 11405,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11404 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792664578, "dur": 12, + "args": { + "External id": 11406,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11405 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664609, "dur": 5, + "args": { + "External id": 11407,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11406 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792664618, "dur": 21, + "args": { + "External id": 11408,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11407 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664619, "dur": 5, + "args": { + "External id": 11409,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11408 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792664625, "dur": 14, + "args": { + "External id": 11410,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11409 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792664627, "dur": 11, + "args": { + "External id": 11411,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11410 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664657, "dur": 5, + "args": { + "External id": 11412,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11411 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792664666, "dur": 21, + "args": { + "External id": 11413,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11412 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664667, "dur": 5, + "args": { + "External id": 11414,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11413 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792664673, "dur": 14, + "args": { + "External id": 11415,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11414 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792664674, "dur": 12, + "args": { + "External id": 11416,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11415 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664705, "dur": 5, + "args": { + "External id": 11417,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11416 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792664714, "dur": 22, + "args": { + "External id": 11418,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11417 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664715, "dur": 5, + "args": { + "External id": 11419,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11418 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792664721, "dur": 14, + "args": { + "External id": 11420,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11419 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792664723, "dur": 11, + "args": { + "External id": 11421,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11420 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664753, "dur": 6, + "args": { + "External id": 11422,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11421 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792664762, "dur": 22, + "args": { + "External id": 11423,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11422 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664764, "dur": 4, + "args": { + "External id": 11424,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11423 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792664769, "dur": 14, + "args": { + "External id": 11425,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11424 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792664771, "dur": 11, + "args": { + "External id": 11426,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11425 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664802, "dur": 5, + "args": { + "External id": 11427,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11426 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792664811, "dur": 21, + "args": { + "External id": 11428,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11427 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664812, "dur": 4, + "args": { + "External id": 11429,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11428 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792664817, "dur": 15, + "args": { + "External id": 11430,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11429 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792664819, "dur": 11, + "args": { + "External id": 11431,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11430 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664850, "dur": 5, + "args": { + "External id": 11432,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11431 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792664858, "dur": 21, + "args": { + "External id": 11433,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11432 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664860, "dur": 4, + "args": { + "External id": 11434,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11433 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792664865, "dur": 14, + "args": { + "External id": 11435,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11434 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792664867, "dur": 11, + "args": { + "External id": 11436,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11435 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664897, "dur": 6, + "args": { + "External id": 11437,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11436 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792664906, "dur": 23, + "args": { + "External id": 11438,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11437 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664908, "dur": 4, + "args": { + "External id": 11439,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11438 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792664913, "dur": 15, + "args": { + "External id": 11440,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11439 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792664915, "dur": 12, + "args": { + "External id": 11441,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11440 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664947, "dur": 5, + "args": { + "External id": 11442,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11441 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792664956, "dur": 21, + "args": { + "External id": 11443,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11442 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664957, "dur": 4, + "args": { + "External id": 11444,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11443 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792664962, "dur": 14, + "args": { + "External id": 11445,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11444 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792664964, "dur": 11, + "args": { + "External id": 11446,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11445 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792664995, "dur": 5, + "args": { + "External id": 11447,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11446 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792665004, "dur": 21, + "args": { + "External id": 11448,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11447 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665005, "dur": 4, + "args": { + "External id": 11449,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11448 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792665010, "dur": 14, + "args": { + "External id": 11450,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11449 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792665012, "dur": 11, + "args": { + "External id": 11451,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11450 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665042, "dur": 5, + "args": { + "External id": 11452,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11451 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792665051, "dur": 21, + "args": { + "External id": 11453,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11452 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665052, "dur": 5, + "args": { + "External id": 11454,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11453 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792665058, "dur": 14, + "args": { + "External id": 11455,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11454 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792665060, "dur": 11, + "args": { + "External id": 11456,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11455 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665090, "dur": 6, + "args": { + "External id": 11457,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11456 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792665100, "dur": 21, + "args": { + "External id": 11458,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11457 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665101, "dur": 4, + "args": { + "External id": 11459,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11458 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792665106, "dur": 15, + "args": { + "External id": 11460,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11459 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792665108, "dur": 11, + "args": { + "External id": 11461,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11460 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665138, "dur": 6, + "args": { + "External id": 11462,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11461 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792665148, "dur": 21, + "args": { + "External id": 11463,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11462 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665149, "dur": 4, + "args": { + "External id": 11464,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11463 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792665155, "dur": 13, + "args": { + "External id": 11465,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11464 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792665156, "dur": 11, + "args": { + "External id": 11466,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11465 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665186, "dur": 5, + "args": { + "External id": 11467,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11466 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792665195, "dur": 22, + "args": { + "External id": 11468,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11467 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665196, "dur": 5, + "args": { + "External id": 11469,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11468 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792665202, "dur": 14, + "args": { + "External id": 11470,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11469 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792665204, "dur": 11, + "args": { + "External id": 11471,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11470 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665235, "dur": 5, + "args": { + "External id": 11472,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11471 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792665244, "dur": 21, + "args": { + "External id": 11473,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11472 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665245, "dur": 5, + "args": { + "External id": 11474,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11473 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792665251, "dur": 14, + "args": { + "External id": 11475,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11474 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792665253, "dur": 11, + "args": { + "External id": 11476,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11475 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665283, "dur": 5, + "args": { + "External id": 11477,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11476 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792665292, "dur": 21, + "args": { + "External id": 11478,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11477 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665293, "dur": 5, + "args": { + "External id": 11479,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11478 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792665299, "dur": 14, + "args": { + "External id": 11480,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11479 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792665300, "dur": 11, + "args": { + "External id": 11481,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11480 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665331, "dur": 5, + "args": { + "External id": 11482,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11481 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792665340, "dur": 23, + "args": { + "External id": 11483,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11482 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665341, "dur": 5, + "args": { + "External id": 11484,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11483 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792665347, "dur": 15, + "args": { + "External id": 11485,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11484 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792665348, "dur": 13, + "args": { + "External id": 11486,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11485 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665383, "dur": 5, + "args": { + "External id": 11487,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11486 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792665392, "dur": 21, + "args": { + "External id": 11488,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11487 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665393, "dur": 5, + "args": { + "External id": 11489,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11488 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792665399, "dur": 14, + "args": { + "External id": 11490,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11489 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792665400, "dur": 11, + "args": { + "External id": 11491,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11490 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665431, "dur": 5, + "args": { + "External id": 11492,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11491 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792665440, "dur": 22, + "args": { + "External id": 11493,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11492 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665441, "dur": 5, + "args": { + "External id": 11494,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11493 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792665447, "dur": 15, + "args": { + "External id": 11495,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11494 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792665449, "dur": 11, + "args": { + "External id": 11496,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11495 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665480, "dur": 5, + "args": { + "External id": 11497,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11496 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792665489, "dur": 22, + "args": { + "External id": 11498,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11497 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665490, "dur": 5, + "args": { + "External id": 11499,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11498 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792665496, "dur": 14, + "args": { + "External id": 11500,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11499 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792665498, "dur": 11, + "args": { + "External id": 11501,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11500 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665529, "dur": 5, + "args": { + "External id": 11502,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11501 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792665537, "dur": 22, + "args": { + "External id": 11503,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11502 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665539, "dur": 4, + "args": { + "External id": 11504,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11503 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792665544, "dur": 14, + "args": { + "External id": 11505,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11504 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792665546, "dur": 11, + "args": { + "External id": 11506,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11505 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665576, "dur": 6, + "args": { + "External id": 11507,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11506 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792665586, "dur": 21, + "args": { + "External id": 11508,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11507 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665587, "dur": 4, + "args": { + "External id": 11509,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11508 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792665592, "dur": 15, + "args": { + "External id": 11510,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11509 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792665594, "dur": 11, + "args": { + "External id": 11511,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11510 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665625, "dur": 5, + "args": { + "External id": 11512,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11511 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792665634, "dur": 21, + "args": { + "External id": 11513,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11512 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665635, "dur": 4, + "args": { + "External id": 11514,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11513 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792665641, "dur": 14, + "args": { + "External id": 11515,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11514 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792665642, "dur": 11, + "args": { + "External id": 11516,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11515 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665673, "dur": 5, + "args": { + "External id": 11517,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11516 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792665682, "dur": 21, + "args": { + "External id": 11518,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11517 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665683, "dur": 4, + "args": { + "External id": 11519,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11518 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792665689, "dur": 14, + "args": { + "External id": 11520,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11519 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792665690, "dur": 11, + "args": { + "External id": 11521,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11520 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665721, "dur": 5, + "args": { + "External id": 11522,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11521 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792665730, "dur": 21, + "args": { + "External id": 11523,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11522 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665731, "dur": 4, + "args": { + "External id": 11524,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11523 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792665737, "dur": 13, + "args": { + "External id": 11525,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11524 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792665738, "dur": 11, + "args": { + "External id": 11526,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11525 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665768, "dur": 6, + "args": { + "External id": 11527,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11526 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792665778, "dur": 21, + "args": { + "External id": 11528,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11527 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665779, "dur": 4, + "args": { + "External id": 11529,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11528 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792665784, "dur": 14, + "args": { + "External id": 11530,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11529 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792665786, "dur": 11, + "args": { + "External id": 11531,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11530 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665817, "dur": 5, + "args": { + "External id": 11532,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11531 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792665826, "dur": 21, + "args": { + "External id": 11533,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11532 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665827, "dur": 5, + "args": { + "External id": 11534,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11533 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792665833, "dur": 14, + "args": { + "External id": 11535,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11534 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792665834, "dur": 11, + "args": { + "External id": 11536,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11535 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665865, "dur": 5, + "args": { + "External id": 11537,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11536 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792665874, "dur": 21, + "args": { + "External id": 11538,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11537 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665875, "dur": 5, + "args": { + "External id": 11539,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11538 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792665881, "dur": 13, + "args": { + "External id": 11540,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11539 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792665882, "dur": 11, + "args": { + "External id": 11541,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11540 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665913, "dur": 5, + "args": { + "External id": 11542,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11541 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792665922, "dur": 21, + "args": { + "External id": 11543,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11542 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665923, "dur": 5, + "args": { + "External id": 11544,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11543 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792665929, "dur": 14, + "args": { + "External id": 11545,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11544 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792665930, "dur": 11, + "args": { + "External id": 11546,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11545 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665961, "dur": 5, + "args": { + "External id": 11547,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11546 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792665970, "dur": 21, + "args": { + "External id": 11548,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11547 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792665971, "dur": 5, + "args": { + "External id": 11549,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11548 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792665977, "dur": 14, + "args": { + "External id": 11550,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11549 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792665979, "dur": 11, + "args": { + "External id": 11551,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11550 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666009, "dur": 5, + "args": { + "External id": 11552,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11551 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792666018, "dur": 21, + "args": { + "External id": 11553,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11552 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666020, "dur": 4, + "args": { + "External id": 11554,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11553 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792666025, "dur": 14, + "args": { + "External id": 11555,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11554 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792666027, "dur": 11, + "args": { + "External id": 11556,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11555 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666057, "dur": 5, + "args": { + "External id": 11557,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11556 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792666066, "dur": 21, + "args": { + "External id": 11558,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11557 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666067, "dur": 5, + "args": { + "External id": 11559,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11558 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792666073, "dur": 14, + "args": { + "External id": 11560,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11559 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792666075, "dur": 11, + "args": { + "External id": 11561,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11560 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666105, "dur": 5, + "args": { + "External id": 11562,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11561 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792666114, "dur": 21, + "args": { + "External id": 11563,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11562 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666115, "dur": 4, + "args": { + "External id": 11564,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11563 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792666120, "dur": 14, + "args": { + "External id": 11565,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11564 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792666122, "dur": 11, + "args": { + "External id": 11566,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11565 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666153, "dur": 5, + "args": { + "External id": 11567,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11566 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792666162, "dur": 21, + "args": { + "External id": 11568,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11567 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666163, "dur": 4, + "args": { + "External id": 11569,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11568 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792666169, "dur": 14, + "args": { + "External id": 11570,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11569 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792666170, "dur": 11, + "args": { + "External id": 11571,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11570 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666201, "dur": 5, + "args": { + "External id": 11572,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11571 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792666210, "dur": 21, + "args": { + "External id": 11573,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11572 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666211, "dur": 4, + "args": { + "External id": 11574,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11573 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792666216, "dur": 15, + "args": { + "External id": 11575,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11574 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792666218, "dur": 11, + "args": { + "External id": 11576,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11575 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666249, "dur": 5, + "args": { + "External id": 11577,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11576 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792666258, "dur": 21, + "args": { + "External id": 11578,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11577 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666259, "dur": 4, + "args": { + "External id": 11579,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11578 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792666265, "dur": 13, + "args": { + "External id": 11580,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11579 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792666266, "dur": 11, + "args": { + "External id": 11581,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11580 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666356, "dur": 6, + "args": { + "External id": 11582,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11581 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792666367, "dur": 25, + "args": { + "External id": 11583,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11582 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666368, "dur": 5, + "args": { + "External id": 11584,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11583 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792666374, "dur": 18, + "args": { + "External id": 11585,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11584 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792666376, "dur": 14, + "args": { + "External id": 11586,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11585 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666413, "dur": 5, + "args": { + "External id": 11587,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11586 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792666422, "dur": 22, + "args": { + "External id": 11588,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11587 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666423, "dur": 5, + "args": { + "External id": 11589,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11588 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792666429, "dur": 14, + "args": { + "External id": 11590,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11589 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792666431, "dur": 11, + "args": { + "External id": 11591,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11590 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666462, "dur": 5, + "args": { + "External id": 11592,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11591 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792666471, "dur": 21, + "args": { + "External id": 11593,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11592 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666472, "dur": 5, + "args": { + "External id": 11594,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11593 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792666478, "dur": 14, + "args": { + "External id": 11595,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11594 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792666480, "dur": 10, + "args": { + "External id": 11596,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11595 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666510, "dur": 5, + "args": { + "External id": 11597,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11596 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792666519, "dur": 22, + "args": { + "External id": 11598,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11597 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666520, "dur": 5, + "args": { + "External id": 11599,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11598 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792666526, "dur": 14, + "args": { + "External id": 11600,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11599 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792666528, "dur": 11, + "args": { + "External id": 11601,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11600 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666559, "dur": 5, + "args": { + "External id": 11602,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11601 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792666568, "dur": 21, + "args": { + "External id": 11603,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11602 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666569, "dur": 5, + "args": { + "External id": 11604,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11603 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792666575, "dur": 14, + "args": { + "External id": 11605,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11604 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792666576, "dur": 12, + "args": { + "External id": 11606,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11605 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666607, "dur": 5, + "args": { + "External id": 11607,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11606 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792666616, "dur": 21, + "args": { + "External id": 11608,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11607 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666617, "dur": 5, + "args": { + "External id": 11609,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11608 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792666623, "dur": 14, + "args": { + "External id": 11610,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11609 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792666624, "dur": 12, + "args": { + "External id": 11611,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11610 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666655, "dur": 5, + "args": { + "External id": 11612,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11611 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792666664, "dur": 22, + "args": { + "External id": 11613,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11612 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666666, "dur": 4, + "args": { + "External id": 11614,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11613 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792666671, "dur": 14, + "args": { + "External id": 11615,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11614 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792666673, "dur": 11, + "args": { + "External id": 11616,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11615 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666704, "dur": 5, + "args": { + "External id": 11617,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11616 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792666713, "dur": 22, + "args": { + "External id": 11618,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11617 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666714, "dur": 5, + "args": { + "External id": 11619,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11618 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792666720, "dur": 14, + "args": { + "External id": 11620,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11619 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792666722, "dur": 11, + "args": { + "External id": 11621,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11620 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666753, "dur": 5, + "args": { + "External id": 11622,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11621 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792666762, "dur": 21, + "args": { + "External id": 11623,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11622 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666763, "dur": 4, + "args": { + "External id": 11624,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11623 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792666769, "dur": 14, + "args": { + "External id": 11625,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11624 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792666770, "dur": 11, + "args": { + "External id": 11626,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11625 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666808, "dur": 6, + "args": { + "External id": 11627,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11626 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792666818, "dur": 22, + "args": { + "External id": 11628,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11627 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666819, "dur": 4, + "args": { + "External id": 11629,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11628 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792666825, "dur": 14, + "args": { + "External id": 11630,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11629 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792666826, "dur": 12, + "args": { + "External id": 11631,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11630 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666911, "dur": 6, + "args": { + "External id": 11632,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11631 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792666922, "dur": 24, + "args": { + "External id": 11633,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11632 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666923, "dur": 5, + "args": { + "External id": 11634,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11633 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792666929, "dur": 17, + "args": { + "External id": 11635,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11634 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792666931, "dur": 13, + "args": { + "External id": 11636,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11635 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666967, "dur": 5, + "args": { + "External id": 11637,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11636 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792666976, "dur": 22, + "args": { + "External id": 11638,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11637 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792666978, "dur": 4, + "args": { + "External id": 11639,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11638 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792666983, "dur": 15, + "args": { + "External id": 11640,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11639 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792666985, "dur": 11, + "args": { + "External id": 11641,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11640 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667017, "dur": 5, + "args": { + "External id": 11642,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11641 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792667026, "dur": 21, + "args": { + "External id": 11643,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11642 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667027, "dur": 5, + "args": { + "External id": 11644,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11643 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792667033, "dur": 14, + "args": { + "External id": 11645,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11644 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792667035, "dur": 11, + "args": { + "External id": 11646,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11645 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667065, "dur": 5, + "args": { + "External id": 11647,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11646 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792667074, "dur": 21, + "args": { + "External id": 11648,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11647 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667075, "dur": 5, + "args": { + "External id": 11649,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11648 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792667081, "dur": 14, + "args": { + "External id": 11650,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11649 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792667083, "dur": 11, + "args": { + "External id": 11651,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11650 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667113, "dur": 5, + "args": { + "External id": 11652,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11651 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792667122, "dur": 22, + "args": { + "External id": 11653,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11652 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667124, "dur": 4, + "args": { + "External id": 11654,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11653 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792667129, "dur": 14, + "args": { + "External id": 11655,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11654 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792667131, "dur": 11, + "args": { + "External id": 11656,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11655 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667161, "dur": 6, + "args": { + "External id": 11657,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11656 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792667171, "dur": 21, + "args": { + "External id": 11658,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11657 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667172, "dur": 4, + "args": { + "External id": 11659,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11658 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792667178, "dur": 14, + "args": { + "External id": 11660,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11659 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792667179, "dur": 11, + "args": { + "External id": 11661,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11660 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667210, "dur": 5, + "args": { + "External id": 11662,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11661 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792667219, "dur": 21, + "args": { + "External id": 11663,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11662 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667220, "dur": 5, + "args": { + "External id": 11664,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11663 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792667226, "dur": 14, + "args": { + "External id": 11665,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11664 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792667227, "dur": 12, + "args": { + "External id": 11666,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11665 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667258, "dur": 5, + "args": { + "External id": 11667,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11666 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792667267, "dur": 21, + "args": { + "External id": 11668,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11667 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667268, "dur": 5, + "args": { + "External id": 11669,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11668 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792667274, "dur": 14, + "args": { + "External id": 11670,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11669 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792667276, "dur": 11, + "args": { + "External id": 11671,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11670 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667306, "dur": 5, + "args": { + "External id": 11672,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11671 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792667315, "dur": 21, + "args": { + "External id": 11673,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11672 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667316, "dur": 5, + "args": { + "External id": 11674,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11673 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792667322, "dur": 14, + "args": { + "External id": 11675,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11674 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792667323, "dur": 11, + "args": { + "External id": 11676,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11675 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667354, "dur": 5, + "args": { + "External id": 11677,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11676 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792667364, "dur": 21, + "args": { + "External id": 11678,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11677 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667365, "dur": 4, + "args": { + "External id": 11679,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11678 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792667371, "dur": 14, + "args": { + "External id": 11680,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11679 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792667372, "dur": 11, + "args": { + "External id": 11681,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11680 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667403, "dur": 5, + "args": { + "External id": 11682,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11681 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792667412, "dur": 21, + "args": { + "External id": 11683,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11682 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667413, "dur": 5, + "args": { + "External id": 11684,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11683 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792667419, "dur": 14, + "args": { + "External id": 11685,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11684 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792667421, "dur": 11, + "args": { + "External id": 11686,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11685 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667451, "dur": 5, + "args": { + "External id": 11687,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11686 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792667460, "dur": 21, + "args": { + "External id": 11688,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11687 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667461, "dur": 5, + "args": { + "External id": 11689,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11688 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792667467, "dur": 14, + "args": { + "External id": 11690,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11689 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792667469, "dur": 10, + "args": { + "External id": 11691,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11690 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667499, "dur": 5, + "args": { + "External id": 11692,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11691 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792667508, "dur": 21, + "args": { + "External id": 11693,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11692 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667509, "dur": 5, + "args": { + "External id": 11694,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11693 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792667515, "dur": 14, + "args": { + "External id": 11695,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11694 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792667516, "dur": 12, + "args": { + "External id": 11696,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11695 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667548, "dur": 5, + "args": { + "External id": 11697,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11696 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792667557, "dur": 21, + "args": { + "External id": 11698,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11697 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667558, "dur": 4, + "args": { + "External id": 11699,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11698 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792667564, "dur": 14, + "args": { + "External id": 11700,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11699 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792667565, "dur": 11, + "args": { + "External id": 11701,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11700 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667596, "dur": 5, + "args": { + "External id": 11702,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11701 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792667605, "dur": 21, + "args": { + "External id": 11703,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11702 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667606, "dur": 5, + "args": { + "External id": 11704,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11703 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792667612, "dur": 14, + "args": { + "External id": 11705,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11704 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792667613, "dur": 11, + "args": { + "External id": 11706,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11705 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667644, "dur": 5, + "args": { + "External id": 11707,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11706 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792667653, "dur": 21, + "args": { + "External id": 11708,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11707 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667654, "dur": 4, + "args": { + "External id": 11709,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11708 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792667660, "dur": 14, + "args": { + "External id": 11710,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11709 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792667661, "dur": 11, + "args": { + "External id": 11711,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11710 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667691, "dur": 5, + "args": { + "External id": 11712,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11711 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792667700, "dur": 22, + "args": { + "External id": 11713,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11712 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667702, "dur": 4, + "args": { + "External id": 11714,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11713 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792667707, "dur": 14, + "args": { + "External id": 11715,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11714 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792667709, "dur": 11, + "args": { + "External id": 11716,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11715 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667739, "dur": 5, + "args": { + "External id": 11717,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11716 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792667748, "dur": 21, + "args": { + "External id": 11718,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11717 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667749, "dur": 5, + "args": { + "External id": 11719,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11718 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792667755, "dur": 14, + "args": { + "External id": 11720,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11719 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792667757, "dur": 11, + "args": { + "External id": 11721,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11720 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667787, "dur": 5, + "args": { + "External id": 11722,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11721 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792667796, "dur": 23, + "args": { + "External id": 11723,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11722 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667799, "dur": 4, + "args": { + "External id": 11724,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11723 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792667805, "dur": 14, + "args": { + "External id": 11725,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11724 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792667806, "dur": 11, + "args": { + "External id": 11726,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11725 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667837, "dur": 5, + "args": { + "External id": 11727,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11726 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792667846, "dur": 21, + "args": { + "External id": 11728,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11727 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667847, "dur": 4, + "args": { + "External id": 11729,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11728 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792667853, "dur": 14, + "args": { + "External id": 11730,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11729 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792667854, "dur": 11, + "args": { + "External id": 11731,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11730 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667885, "dur": 5, + "args": { + "External id": 11732,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11731 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792667894, "dur": 21, + "args": { + "External id": 11733,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11732 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667895, "dur": 5, + "args": { + "External id": 11734,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11733 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792667901, "dur": 14, + "args": { + "External id": 11735,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11734 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792667902, "dur": 12, + "args": { + "External id": 11736,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11735 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667933, "dur": 5, + "args": { + "External id": 11737,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11736 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792667942, "dur": 21, + "args": { + "External id": 11738,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11737 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667943, "dur": 4, + "args": { + "External id": 11739,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11738 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792667949, "dur": 13, + "args": { + "External id": 11740,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11739 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792667950, "dur": 11, + "args": { + "External id": 11741,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11740 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667981, "dur": 5, + "args": { + "External id": 11742,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11741 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792667990, "dur": 21, + "args": { + "External id": 11743,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11742 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792667991, "dur": 4, + "args": { + "External id": 11744,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11743 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792667996, "dur": 14, + "args": { + "External id": 11745,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11744 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792667998, "dur": 11, + "args": { + "External id": 11746,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11745 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668028, "dur": 6, + "args": { + "External id": 11747,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11746 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792668037, "dur": 21, + "args": { + "External id": 11748,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11747 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668039, "dur": 4, + "args": { + "External id": 11749,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11748 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792668044, "dur": 14, + "args": { + "External id": 11750,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11749 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792668046, "dur": 11, + "args": { + "External id": 11751,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11750 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668076, "dur": 5, + "args": { + "External id": 11752,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11751 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792668085, "dur": 21, + "args": { + "External id": 11753,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11752 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668086, "dur": 4, + "args": { + "External id": 11754,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11753 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792668091, "dur": 14, + "args": { + "External id": 11755,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11754 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792668093, "dur": 11, + "args": { + "External id": 11756,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11755 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668123, "dur": 5, + "args": { + "External id": 11757,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11756 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792668132, "dur": 21, + "args": { + "External id": 11758,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11757 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668133, "dur": 5, + "args": { + "External id": 11759,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11758 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792668139, "dur": 14, + "args": { + "External id": 11760,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11759 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792668140, "dur": 12, + "args": { + "External id": 11761,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11760 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668171, "dur": 5, + "args": { + "External id": 11762,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11761 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792668180, "dur": 21, + "args": { + "External id": 11763,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11762 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668181, "dur": 4, + "args": { + "External id": 11764,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11763 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792668187, "dur": 13, + "args": { + "External id": 11765,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11764 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792668188, "dur": 11, + "args": { + "External id": 11766,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11765 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668238, "dur": 6, + "args": { + "External id": 11767,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11766 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792668248, "dur": 23, + "args": { + "External id": 11768,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11767 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668249, "dur": 5, + "args": { + "External id": 11769,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11768 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792668255, "dur": 16, + "args": { + "External id": 11770,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11769 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792668257, "dur": 13, + "args": { + "External id": 11771,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11770 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668290, "dur": 5, + "args": { + "External id": 11772,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11771 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792668299, "dur": 21, + "args": { + "External id": 11773,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11772 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668300, "dur": 4, + "args": { + "External id": 11774,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11773 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792668306, "dur": 14, + "args": { + "External id": 11775,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11774 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792668307, "dur": 12, + "args": { + "External id": 11776,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11775 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668348, "dur": 5, + "args": { + "External id": 11777,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11776 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792668357, "dur": 22, + "args": { + "External id": 11778,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11777 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668359, "dur": 4, + "args": { + "External id": 11779,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11778 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792668364, "dur": 14, + "args": { + "External id": 11780,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11779 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792668366, "dur": 11, + "args": { + "External id": 11781,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11780 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668397, "dur": 5, + "args": { + "External id": 11782,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11781 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792668406, "dur": 21, + "args": { + "External id": 11783,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11782 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668407, "dur": 4, + "args": { + "External id": 11784,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11783 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792668413, "dur": 14, + "args": { + "External id": 11785,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11784 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792668414, "dur": 11, + "args": { + "External id": 11786,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11785 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668445, "dur": 5, + "args": { + "External id": 11787,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11786 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792668454, "dur": 21, + "args": { + "External id": 11788,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11787 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668455, "dur": 5, + "args": { + "External id": 11789,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11788 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792668461, "dur": 14, + "args": { + "External id": 11790,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11789 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792668462, "dur": 11, + "args": { + "External id": 11791,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11790 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668493, "dur": 5, + "args": { + "External id": 11792,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11791 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792668502, "dur": 21, + "args": { + "External id": 11793,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11792 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668503, "dur": 4, + "args": { + "External id": 11794,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11793 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792668509, "dur": 14, + "args": { + "External id": 11795,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11794 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792668510, "dur": 12, + "args": { + "External id": 11796,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11795 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668541, "dur": 5, + "args": { + "External id": 11797,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11796 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792668550, "dur": 21, + "args": { + "External id": 11798,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11797 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668551, "dur": 5, + "args": { + "External id": 11799,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11798 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792668557, "dur": 14, + "args": { + "External id": 11800,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11799 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792668559, "dur": 11, + "args": { + "External id": 11801,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11800 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668589, "dur": 5, + "args": { + "External id": 11802,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11801 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792668598, "dur": 21, + "args": { + "External id": 11803,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11802 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668599, "dur": 5, + "args": { + "External id": 11804,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11803 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792668605, "dur": 14, + "args": { + "External id": 11805,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11804 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792668607, "dur": 11, + "args": { + "External id": 11806,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11805 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668637, "dur": 5, + "args": { + "External id": 11807,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11806 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792668646, "dur": 21, + "args": { + "External id": 11808,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11807 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668647, "dur": 4, + "args": { + "External id": 11809,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11808 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792668653, "dur": 13, + "args": { + "External id": 11810,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11809 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792668654, "dur": 11, + "args": { + "External id": 11811,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11810 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668684, "dur": 6, + "args": { + "External id": 11812,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11811 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792668693, "dur": 23, + "args": { + "External id": 11813,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11812 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668695, "dur": 4, + "args": { + "External id": 11814,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11813 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792668700, "dur": 16, + "args": { + "External id": 11815,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11814 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792668702, "dur": 12, + "args": { + "External id": 11816,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11815 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668734, "dur": 5, + "args": { + "External id": 11817,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11816 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792668743, "dur": 21, + "args": { + "External id": 11818,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11817 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668744, "dur": 4, + "args": { + "External id": 11819,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11818 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792668750, "dur": 14, + "args": { + "External id": 11820,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11819 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792668751, "dur": 11, + "args": { + "External id": 11821,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11820 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668782, "dur": 5, + "args": { + "External id": 11822,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11821 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792668791, "dur": 21, + "args": { + "External id": 11823,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11822 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668792, "dur": 4, + "args": { + "External id": 11824,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11823 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792668797, "dur": 15, + "args": { + "External id": 11825,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11824 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792668799, "dur": 11, + "args": { + "External id": 11826,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11825 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668830, "dur": 5, + "args": { + "External id": 11827,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11826 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792668839, "dur": 21, + "args": { + "External id": 11828,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11827 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668840, "dur": 4, + "args": { + "External id": 11829,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11828 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792668845, "dur": 14, + "args": { + "External id": 11830,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11829 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792668847, "dur": 11, + "args": { + "External id": 11831,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11830 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668877, "dur": 5, + "args": { + "External id": 11832,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11831 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792668886, "dur": 22, + "args": { + "External id": 11833,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11832 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668887, "dur": 5, + "args": { + "External id": 11834,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11833 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792668893, "dur": 14, + "args": { + "External id": 11835,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11834 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792668895, "dur": 11, + "args": { + "External id": 11836,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11835 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668925, "dur": 5, + "args": { + "External id": 11837,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11836 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792668934, "dur": 21, + "args": { + "External id": 11838,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11837 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668936, "dur": 4, + "args": { + "External id": 11839,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11838 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792668941, "dur": 14, + "args": { + "External id": 11840,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11839 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792668943, "dur": 11, + "args": { + "External id": 11841,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11840 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668973, "dur": 5, + "args": { + "External id": 11842,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11841 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792668982, "dur": 21, + "args": { + "External id": 11843,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11842 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792668983, "dur": 4, + "args": { + "External id": 11844,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11843 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792668989, "dur": 14, + "args": { + "External id": 11845,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11844 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792668990, "dur": 12, + "args": { + "External id": 11846,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11845 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792669021, "dur": 5, + "args": { + "External id": 11847,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11846 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792669030, "dur": 21, + "args": { + "External id": 11848,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11847 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792669031, "dur": 4, + "args": { + "External id": 11849,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11848 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792669037, "dur": 13, + "args": { + "External id": 11850,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11849 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792669038, "dur": 11, + "args": { + "External id": 11851,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11850 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792669068, "dur": 5, + "args": { + "External id": 11852,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11851 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792669077, "dur": 21, + "args": { + "External id": 11853,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11852 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792669079, "dur": 4, + "args": { + "External id": 11854,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11853 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792669084, "dur": 14, + "args": { + "External id": 11855,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11854 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792669086, "dur": 11, + "args": { + "External id": 11856,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11855 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792669116, "dur": 5, + "args": { + "External id": 11857,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11856 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792669125, "dur": 21, + "args": { + "External id": 11858,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11857 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792669126, "dur": 5, + "args": { + "External id": 11859,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11858 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792669132, "dur": 14, + "args": { + "External id": 11860,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11859 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792669133, "dur": 11, + "args": { + "External id": 11861,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11860 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792669164, "dur": 5, + "args": { + "External id": 11862,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11861 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792669173, "dur": 21, + "args": { + "External id": 11863,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11862 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792669174, "dur": 5, + "args": { + "External id": 11864,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11863 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792669180, "dur": 14, + "args": { + "External id": 11865,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11864 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792669181, "dur": 11, + "args": { + "External id": 11866,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11865 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792669212, "dur": 5, + "args": { + "External id": 11867,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11866 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792669221, "dur": 21, + "args": { + "External id": 11868,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11867 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792669222, "dur": 4, + "args": { + "External id": 11869,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11868 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792669227, "dur": 14, + "args": { + "External id": 11870,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11869 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792669229, "dur": 11, + "args": { + "External id": 11871,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11870 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792669259, "dur": 6, + "args": { + "External id": 11872,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11871 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792669268, "dur": 21, + "args": { + "External id": 11873,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11872 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792669269, "dur": 5, + "args": { + "External id": 11874,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11873 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792669275, "dur": 14, + "args": { + "External id": 11875,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11874 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792669277, "dur": 11, + "args": { + "External id": 11876,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11875 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792669307, "dur": 5, + "args": { + "External id": 11877,"Record function id": 0, "Concrete Inputs": ["[4, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11876 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649792669316, "dur": 21, + "args": { + "External id": 11878,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11877 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792669317, "dur": 4, + "args": { + "External id": 11879,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11878 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649792669323, "dur": 14, + "args": { + "External id": 11880,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 11879 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649792669324, "dur": 11, + "args": { + "External id": 11881,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 11880 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::randn", "pid": 435, "tid": 435, + "ts": 1776649792669492, "dur": 53, + "args": { + "External id": 11882,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 11881 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792669497, "dur": 7, + "args": { + "External id": 11883,"Record function id": 0, "Concrete Inputs": ["[8, 4096]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11882 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::normal_", "pid": 435, "tid": 435, + "ts": 1776649792669509, "dur": 36, + "args": { + "External id": 11884,"Record function id": 0, "Concrete Inputs": ["", "0.", "1.", ""], "Input type": ["c10::Half", "Scalar", "Scalar", ""], "Input Dims": [[8, 4096], [], [], []], "Ev Idx": 11883 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::randn", "pid": 435, "tid": 435, + "ts": 1776649792669568, "dur": 29, + "args": { + "External id": 11885,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "", "5", "", "", "False"], "Input type": ["ScalarList", "", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11884 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792669571, "dur": 6, + "args": { + "External id": 11886,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11885 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::normal_", "pid": 435, "tid": 435, + "ts": 1776649792669579, "dur": 17, + "args": { + "External id": 11887,"Record function id": 0, "Concrete Inputs": ["", "0.", "1.", ""], "Input type": ["c10::Half", "Scalar", "Scalar", ""], "Input Dims": [[4096, 8192], [], [], []], "Ev Idx": 11886 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 435, "tid": 435, + "ts": 1776649792669625, "dur": 38, + "args": { + "External id": 11888,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "double"], "Input Dims": [[4096, 8192], []], "Ev Idx": 11887 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649792669677, "dur": 12, + "args": { + "External id": 11889,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 8192], []], "Ev Idx": 11888 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649792669680, "dur": 9, + "args": { + "External id": 11890,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 8192], []], "Ev Idx": 11889 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 435, "tid": 435, + "ts": 1776649792669694, "dur": 48, + "args": { + "External id": 11891,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[32, 128, 8192]], "Ev Idx": 11890 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792669699, "dur": 5, + "args": { + "External id": 11892,"Record function id": 0, "Concrete Inputs": ["[0]", "5", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11891 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 435, "tid": 435, + "ts": 1776649792669707, "dur": 34, + "args": { + "External id": 11893,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[32, 128, 8192], [0]], "Ev Idx": 11892 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 435, "tid": 435, + "ts": 1776649792669716, "dur": 8, + "args": { + "External id": 11894,"Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]", ""], "Input type": ["c10::Half", "ScalarList", ""], "Input Dims": [[0], [], []], "Ev Idx": 11893 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 435, "tid": 435, + "ts": 1776649792669750, "dur": 42, + "args": { + "External id": 11895,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1]", "False"], "Input type": ["c10::Half", "ScalarList", "Scalar"], "Input Dims": [[32, 128, 8192], [], []], "Ev Idx": 11894 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792669766, "dur": 1, + "args": { + "External id": 11896,"Record function id": 0, "Concrete Inputs": ["", "[32, 1, 8192]", "[8192, 0, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 11895 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 435, "tid": 435, + "ts": 1776649792669803, "dur": 27, + "args": { + "External id": 11897,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "double"], "Input Dims": [[32, 8192], []], "Ev Idx": 11896 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 435, "tid": 435, + "ts": 1776649792669840, "dur": 31, + "args": { + "External id": 11898,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "9.9999999999999995e-07", ""], "Input type": ["c10::Half", "Scalar", ""], "Input Dims": [[32, 8192], [], []], "Ev Idx": 11897 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792669845, "dur": 1, + "args": { + "External id": 11899,"Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 8192], [], [], [], []], "Ev Idx": 11898 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792669877, "dur": 1, + "args": { + "External id": 11900,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 8192], [], [], [], []], "Ev Idx": 11899 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649792669883, "dur": 7, + "args": { + "External id": 11901,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 8192], []], "Ev Idx": 11900 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792669888, "dur": 1, + "args": { + "External id": 11902,"Record function id": 0, "Concrete Inputs": ["", "[32, 1, 8192]", "[8192, 8192, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 11901 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 435, "tid": 435, + "ts": 1776649792669892, "dur": 25, + "args": { + "External id": 11903,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[32, 128, 8192], [32, 1, 8192]], "Ev Idx": 11902 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 435, "tid": 435, + "ts": 1776649792669922, "dur": 23, + "args": { + "External id": 11904,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[32, 128, 8192]], "Ev Idx": 11903 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 435, "tid": 435, + "ts": 1776649792669951, "dur": 16, + "args": { + "External id": 11905,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "-8", "7"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[32, 128, 8192], [], []], "Ev Idx": 11904 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792669952, "dur": 0, + "args": { + "External id": 11906,"Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 128, 8192], [], [], [], []], "Ev Idx": 11905 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792669972, "dur": 40, + "args": { + "External id": 11907,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 128, 8192], [], [], [], []], "Ev Idx": 11906 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649792669974, "dur": 38, + "args": { + "External id": 11908,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["c10::Half", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[32, 128, 8192], [], [], [], [], [], []], "Ev Idx": 11907 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649792669983, "dur": 7, + "args": { + "External id": 11909,"Record function id": 0, "Concrete Inputs": ["[32, 128, 8192]", "[1048576, 8192, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11908 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792669992, "dur": 19, + "args": { + "External id": 11910,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "c10::Half", "Scalar"], "Input Dims": [[32, 128, 8192], [32, 128, 8192], []], "Ev Idx": 11909 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 435, "tid": 435, + "ts": 1776649792670024, "dur": 31, + "args": { + "External id": 11911,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["signed char", "long int", "Scalar"], "Input Dims": [[32, 128, 8192], [], []], "Ev Idx": 11910 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792670059, "dur": 25, + "args": { + "External id": 11912,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "False", "False", ""], "Input type": ["signed char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[32, 128, 8192], [], [], [], []], "Ev Idx": 11911 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649792670060, "dur": 24, + "args": { + "External id": 11913,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "", "", "", "False", ""], "Input type": ["signed char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[32, 128, 8192], [], [], [], [], [], []], "Ev Idx": 11912 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649792670063, "dur": 5, + "args": { + "External id": 11914,"Record function id": 0, "Concrete Inputs": ["[32, 128, 8192]", "[1048576, 8192, 1]", "0", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11913 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792670069, "dur": 14, + "args": { + "External id": 11915,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["unsigned char", "signed char", "Scalar"], "Input Dims": [[32, 128, 8192], [32, 128, 8192], []], "Ev Idx": 11914 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649792670091, "dur": 5, + "args": { + "External id": 11916,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["unsigned char", "ScalarList"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 11915 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649792670093, "dur": 2, + "args": { + "External id": 11917,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["unsigned char", "ScalarList"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 11916 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649792670104, "dur": 7, + "args": { + "External id": 11918,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 11917 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792670108, "dur": 1, + "args": { + "External id": 11919,"Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]", "[8192, 1]", "0"], "Input type": ["unsigned char", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 8192], [], [], []], "Ev Idx": 11918 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649792670114, "dur": 4, + "args": { + "External id": 11920,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "2"], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 11919 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792670115, "dur": 2, + "args": { + "External id": 11921,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]", "[8192, 2]", "0"], "Input type": ["unsigned char", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 8192], [], [], []], "Ev Idx": 11920 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649792670124, "dur": 3, + "args": { + "External id": 11922,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 11921 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792670126, "dur": 0, + "args": { + "External id": 11923,"Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]", "[8192, 1]", "0"], "Input type": ["unsigned char", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 8192], [], [], []], "Ev Idx": 11922 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649792670128, "dur": 3, + "args": { + "External id": 11924,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "1", "9223372036854775807", "2"], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 11923 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792670130, "dur": 1, + "args": { + "External id": 11925,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096]", "[8192, 2]", "1"], "Input type": ["unsigned char", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 8192], [], [], []], "Ev Idx": 11924 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__lshift__", "pid": 435, "tid": 435, + "ts": 1776649792670139, "dur": 41, + "args": { + "External id": 11926,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 11925 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792670150, "dur": 6, + "args": { + "External id": 11927,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "0", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11926 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__or__", "pid": 435, "tid": 435, + "ts": 1776649792670187, "dur": 28, + "args": { + "External id": 11928,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["unsigned char", "unsigned char"], "Input Dims": [[4096, 4096], [4096, 4096]], "Ev Idx": 11927 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_or", "pid": 435, "tid": 435, + "ts": 1776649792670189, "dur": 26, + "args": { + "External id": 11929,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["unsigned char", "unsigned char"], "Input Dims": [[4096, 4096], [4096, 4096]], "Ev Idx": 11928 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792670221, "dur": 0, + "args": { + "External id": 11930,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 11929 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__and__", "pid": 435, "tid": 435, + "ts": 1776649792670237, "dur": 33, + "args": { + "External id": 11931,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "15"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 11930 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_and", "pid": 435, "tid": 435, + "ts": 1776649792670240, "dur": 29, + "args": { + "External id": 11932,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "15"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 11931 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_and", "pid": 435, "tid": 435, + "ts": 1776649792670244, "dur": 24, + "args": { + "External id": 11933,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["unsigned char", "long int"], "Input Dims": [[4096, 4096], []], "Ev Idx": 11932 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792670273, "dur": 26, + "args": { + "External id": 11934,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 11933 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649792670274, "dur": 25, + "args": { + "External id": 11935,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["unsigned char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], [], []], "Ev Idx": 11934 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649792670278, "dur": 5, + "args": { + "External id": 11936,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "[4096, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11935 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792670285, "dur": 13, + "args": { + "External id": 11937,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "unsigned char", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 11936 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__rshift__", "pid": 435, "tid": 435, + "ts": 1776649792670306, "dur": 28, + "args": { + "External id": 11938,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 11937 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792670312, "dur": 5, + "args": { + "External id": 11939,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "0", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11938 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792670338, "dur": 21, + "args": { + "External id": 11940,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 11939 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649792670339, "dur": 20, + "args": { + "External id": 11941,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["unsigned char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], [], []], "Ev Idx": 11940 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649792670341, "dur": 5, + "args": { + "External id": 11942,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "[4096, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11941 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792670347, "dur": 11, + "args": { + "External id": 11943,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "unsigned char", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 11942 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::stack", "pid": 435, "tid": 435, + "ts": 1776649792670372, "dur": 56, + "args": { + "External id": 11944,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["TensorList", "Scalar"], "Input Dims": [[[4096, 4096], [4096, 4096]], []], "Ev Idx": 11943 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649792670377, "dur": 3, + "args": { + "External id": 11945,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["signed char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 11944 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792670379, "dur": 1, + "args": { + "External id": 11946,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096, 1]", "[4096, 1, 1]", ""], "Input type": ["signed char", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 11945 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649792670381, "dur": 2, + "args": { + "External id": 11947,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["signed char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 11946 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792670383, "dur": 0, + "args": { + "External id": 11948,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096, 1]", "[4096, 1, 1]", ""], "Input type": ["signed char", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 11947 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 435, "tid": 435, + "ts": 1776649792670385, "dur": 40, + "args": { + "External id": 11949,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["TensorList", "Scalar"], "Input Dims": [[[4096, 4096, 1], [4096, 4096, 1]], []], "Ev Idx": 11948 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649792670434, "dur": 4, + "args": { + "External id": 11950,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["signed char", "ScalarList"], "Input Dims": [[4096, 4096, 2], []], "Ev Idx": 11949 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649792670435, "dur": 3, + "args": { + "External id": 11951,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["signed char", "ScalarList"], "Input Dims": [[4096, 4096, 2], []], "Ev Idx": 11950 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::repeat_interleave", "pid": 435, "tid": 435, + "ts": 1776649792670446, "dur": 66, + "args": { + "External id": 11952,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "128", "0", ""], "Input type": ["c10::Half", "Scalar", "Scalar", ""], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 11951 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649792670450, "dur": 3, + "args": { + "External id": 11953,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 8192], []], "Ev Idx": 11952 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792670452, "dur": 0, + "args": { + "External id": 11954,"Record function id": 0, "Concrete Inputs": ["", "[32, 1, 8192]", "[8192, 8192, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 11953 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 435, "tid": 435, + "ts": 1776649792670456, "dur": 7, + "args": { + "External id": 11955,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]", "False"], "Input type": ["c10::Half", "ScalarList", "Scalar"], "Input Dims": [[32, 1, 8192], [], []], "Ev Idx": 11954 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792670462, "dur": 0, + "args": { + "External id": 11956,"Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]", "[8192, 0, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 1, 8192], [], [], []], "Ev Idx": 11955 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649792670466, "dur": 38, + "args": { + "External id": 11957,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 11956 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649792670472, "dur": 10, + "args": { + "External id": 11958,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[32, 128, 8192], [], [], [], [], []], "Ev Idx": 11957 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792670475, "dur": 7, + "args": { + "External id": 11959,"Record function id": 0, "Concrete Inputs": ["[32, 128, 8192]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11958 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792670483, "dur": 20, + "args": { + "External id": 11960,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[32, 128, 8192], [32, 128, 8192], []], "Ev Idx": 11959 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::flatten", "pid": 435, "tid": 435, + "ts": 1776649792670507, "dur": 5, + "args": { + "External id": 11961,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[32, 128, 8192], [], []], "Ev Idx": 11960 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649792670510, "dur": 1, + "args": { + "External id": 11962,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 11961 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 435, "tid": 435, + "ts": 1776649792670520, "dur": 22, + "args": { + "External id": 11963,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["signed char", "long int", "Scalar"], "Input Dims": [[4096, 8192], [], []], "Ev Idx": 11962 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792670547, "dur": 26, + "args": { + "External id": 11964,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["signed char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 11963 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649792670549, "dur": 24, + "args": { + "External id": 11965,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "", "", "", "False", ""], "Input type": ["signed char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 8192], [], [], [], [], [], []], "Ev Idx": 11964 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649792670552, "dur": 5, + "args": { + "External id": 11966,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11965 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792670558, "dur": 14, + "args": { + "External id": 11967,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "signed char", "Scalar"], "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 11966 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 435, "tid": 435, + "ts": 1776649792670579, "dur": 19, + "args": { + "External id": 11968,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4096, 8192], [4096, 8192]], "Ev Idx": 11967 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792670655, "dur": 161, + "args": { + "External id": 11969,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11968 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792670661, "dur": 154, + "args": { + "External id": 11970,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11969 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792670823, "dur": 31, + "args": { + "External id": 11971,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11970 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792670826, "dur": 27, + "args": { + "External id": 11972,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11971 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792670859, "dur": 26, + "args": { + "External id": 11973,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11972 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792670861, "dur": 24, + "args": { + "External id": 11974,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11973 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792670890, "dur": 26, + "args": { + "External id": 11975,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11974 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792670892, "dur": 24, + "args": { + "External id": 11976,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11975 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792670921, "dur": 25, + "args": { + "External id": 11977,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11976 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792670923, "dur": 23, + "args": { + "External id": 11978,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11977 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792670951, "dur": 25, + "args": { + "External id": 11979,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11978 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792670953, "dur": 23, + "args": { + "External id": 11980,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11979 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792670981, "dur": 25, + "args": { + "External id": 11981,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11980 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792670983, "dur": 23, + "args": { + "External id": 11982,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11981 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792671011, "dur": 25, + "args": { + "External id": 11983,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11982 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792671013, "dur": 23, + "args": { + "External id": 11984,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11983 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792671041, "dur": 25, + "args": { + "External id": 11985,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11984 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792671043, "dur": 22, + "args": { + "External id": 11986,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11985 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792671071, "dur": 25, + "args": { + "External id": 11987,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11986 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792671073, "dur": 22, + "args": { + "External id": 11988,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11987 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792672270, "dur": 38, + "args": { + "External id": 11989,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11988 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792672274, "dur": 33, + "args": { + "External id": 11990,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11989 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792672314, "dur": 28, + "args": { + "External id": 11991,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11990 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792672317, "dur": 24, + "args": { + "External id": 11992,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11991 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792672346, "dur": 26, + "args": { + "External id": 11993,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11992 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792672349, "dur": 22, + "args": { + "External id": 11994,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11993 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792672377, "dur": 26, + "args": { + "External id": 11995,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11994 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792672379, "dur": 23, + "args": { + "External id": 11996,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11995 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792672407, "dur": 25, + "args": { + "External id": 11997,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11996 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792672409, "dur": 22, + "args": { + "External id": 11998,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11997 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792672437, "dur": 25, + "args": { + "External id": 11999,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11998 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792672439, "dur": 23, + "args": { + "External id": 12000,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 11999 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792672467, "dur": 25, + "args": { + "External id": 12001,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12000 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792672469, "dur": 23, + "args": { + "External id": 12002,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12001 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792672497, "dur": 26, + "args": { + "External id": 12003,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12002 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792672499, "dur": 23, + "args": { + "External id": 12004,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12003 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792672527, "dur": 26, + "args": { + "External id": 12005,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12004 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792672529, "dur": 23, + "args": { + "External id": 12006,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12005 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792672558, "dur": 26, + "args": { + "External id": 12007,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12006 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792672560, "dur": 23, + "args": { + "External id": 12008,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12007 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792672588, "dur": 26, + "args": { + "External id": 12009,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12008 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792672590, "dur": 23, + "args": { + "External id": 12010,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12009 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792672618, "dur": 26, + "args": { + "External id": 12011,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12010 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792672620, "dur": 23, + "args": { + "External id": 12012,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12011 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792672648, "dur": 25, + "args": { + "External id": 12013,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12012 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792672650, "dur": 22, + "args": { + "External id": 12014,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12013 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792672678, "dur": 25, + "args": { + "External id": 12015,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12014 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792672680, "dur": 22, + "args": { + "External id": 12016,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12015 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792672707, "dur": 26, + "args": { + "External id": 12017,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12016 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792672710, "dur": 22, + "args": { + "External id": 12018,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12017 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792672737, "dur": 25, + "args": { + "External id": 12019,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12018 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792672739, "dur": 23, + "args": { + "External id": 12020,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12019 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792672767, "dur": 25, + "args": { + "External id": 12021,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12020 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792672769, "dur": 23, + "args": { + "External id": 12022,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12021 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792672797, "dur": 25, + "args": { + "External id": 12023,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12022 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792672799, "dur": 23, + "args": { + "External id": 12024,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12023 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792672827, "dur": 25, + "args": { + "External id": 12025,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12024 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792672829, "dur": 22, + "args": { + "External id": 12026,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12025 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792672857, "dur": 25, + "args": { + "External id": 12027,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12026 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792672859, "dur": 22, + "args": { + "External id": 12028,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12027 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792672886, "dur": 26, + "args": { + "External id": 12029,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12028 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792672889, "dur": 22, + "args": { + "External id": 12030,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12029 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792672916, "dur": 25, + "args": { + "External id": 12031,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12030 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792672918, "dur": 23, + "args": { + "External id": 12032,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12031 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792672946, "dur": 25, + "args": { + "External id": 12033,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12032 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792672948, "dur": 22, + "args": { + "External id": 12034,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12033 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792672975, "dur": 26, + "args": { + "External id": 12035,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12034 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792672978, "dur": 22, + "args": { + "External id": 12036,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12035 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792673005, "dur": 25, + "args": { + "External id": 12037,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12036 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792673007, "dur": 23, + "args": { + "External id": 12038,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12037 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792673035, "dur": 25, + "args": { + "External id": 12039,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12038 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792673037, "dur": 22, + "args": { + "External id": 12040,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12039 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792673064, "dur": 26, + "args": { + "External id": 12041,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12040 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792673066, "dur": 23, + "args": { + "External id": 12042,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12041 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792673094, "dur": 25, + "args": { + "External id": 12043,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12042 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792673096, "dur": 23, + "args": { + "External id": 12044,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12043 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792673124, "dur": 25, + "args": { + "External id": 12045,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12044 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792673126, "dur": 22, + "args": { + "External id": 12046,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12045 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792673153, "dur": 25, + "args": { + "External id": 12047,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12046 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792673155, "dur": 23, + "args": { + "External id": 12048,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12047 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792673183, "dur": 25, + "args": { + "External id": 12049,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12048 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792673185, "dur": 22, + "args": { + "External id": 12050,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12049 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792673213, "dur": 25, + "args": { + "External id": 12051,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12050 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792673215, "dur": 22, + "args": { + "External id": 12052,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12051 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792673242, "dur": 25, + "args": { + "External id": 12053,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12052 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792673244, "dur": 23, + "args": { + "External id": 12054,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12053 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792673272, "dur": 25, + "args": { + "External id": 12055,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12054 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792673274, "dur": 23, + "args": { + "External id": 12056,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12055 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792673301, "dur": 26, + "args": { + "External id": 12057,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12056 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792673304, "dur": 22, + "args": { + "External id": 12058,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12057 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792673331, "dur": 25, + "args": { + "External id": 12059,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12058 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792673333, "dur": 23, + "args": { + "External id": 12060,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12059 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792673361, "dur": 25, + "args": { + "External id": 12061,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12060 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792673363, "dur": 22, + "args": { + "External id": 12062,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12061 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792673391, "dur": 25, + "args": { + "External id": 12063,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12062 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792673393, "dur": 22, + "args": { + "External id": 12064,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12063 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792673420, "dur": 25, + "args": { + "External id": 12065,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12064 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792673422, "dur": 23, + "args": { + "External id": 12066,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12065 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792673450, "dur": 25, + "args": { + "External id": 12067,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12066 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792673452, "dur": 22, + "args": { + "External id": 12068,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12067 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792673479, "dur": 26, + "args": { + "External id": 12069,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12068 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792673482, "dur": 22, + "args": { + "External id": 12070,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12069 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792673509, "dur": 26, + "args": { + "External id": 12071,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12070 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792673511, "dur": 23, + "args": { + "External id": 12072,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12071 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792673539, "dur": 26, + "args": { + "External id": 12073,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12072 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792673541, "dur": 23, + "args": { + "External id": 12074,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12073 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792673569, "dur": 25, + "args": { + "External id": 12075,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12074 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792673571, "dur": 22, + "args": { + "External id": 12076,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12075 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792673598, "dur": 26, + "args": { + "External id": 12077,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12076 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792673601, "dur": 22, + "args": { + "External id": 12078,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12077 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792673628, "dur": 26, + "args": { + "External id": 12079,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12078 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792673630, "dur": 23, + "args": { + "External id": 12080,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12079 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792673658, "dur": 25, + "args": { + "External id": 12081,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12080 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792673660, "dur": 23, + "args": { + "External id": 12082,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12081 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792673688, "dur": 25, + "args": { + "External id": 12083,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12082 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792673690, "dur": 22, + "args": { + "External id": 12084,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12083 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792673717, "dur": 26, + "args": { + "External id": 12085,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12084 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792673720, "dur": 22, + "args": { + "External id": 12086,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12085 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792673747, "dur": 26, + "args": { + "External id": 12087,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12086 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792673749, "dur": 23, + "args": { + "External id": 12088,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12087 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792674756, "dur": 35, + "args": { + "External id": 12089,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12088 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792674760, "dur": 30, + "args": { + "External id": 12090,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12089 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792674797, "dur": 27, + "args": { + "External id": 12091,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12090 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792674800, "dur": 23, + "args": { + "External id": 12092,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12091 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792674829, "dur": 26, + "args": { + "External id": 12093,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12092 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792674831, "dur": 23, + "args": { + "External id": 12094,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12093 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792674859, "dur": 26, + "args": { + "External id": 12095,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12094 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792674861, "dur": 23, + "args": { + "External id": 12096,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12095 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792674889, "dur": 26, + "args": { + "External id": 12097,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12096 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792674892, "dur": 22, + "args": { + "External id": 12098,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12097 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792674919, "dur": 25, + "args": { + "External id": 12099,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12098 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792674921, "dur": 23, + "args": { + "External id": 12100,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12099 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792674949, "dur": 25, + "args": { + "External id": 12101,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12100 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792674951, "dur": 22, + "args": { + "External id": 12102,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12101 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792674979, "dur": 25, + "args": { + "External id": 12103,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12102 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792674981, "dur": 22, + "args": { + "External id": 12104,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12103 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792675008, "dur": 26, + "args": { + "External id": 12105,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12104 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792675010, "dur": 23, + "args": { + "External id": 12106,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12105 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792675038, "dur": 25, + "args": { + "External id": 12107,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12106 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792675040, "dur": 22, + "args": { + "External id": 12108,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12107 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792675309, "dur": 33, + "args": { + "External id": 12109,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12108 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792675312, "dur": 29, + "args": { + "External id": 12110,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12109 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792675348, "dur": 26, + "args": { + "External id": 12111,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12110 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792675350, "dur": 24, + "args": { + "External id": 12112,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12111 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792675379, "dur": 30, + "args": { + "External id": 12113,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12112 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792675381, "dur": 27, + "args": { + "External id": 12114,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12113 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792675414, "dur": 26, + "args": { + "External id": 12115,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12114 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792675416, "dur": 23, + "args": { + "External id": 12116,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12115 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792675444, "dur": 26, + "args": { + "External id": 12117,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12116 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792675446, "dur": 23, + "args": { + "External id": 12118,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12117 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792675474, "dur": 26, + "args": { + "External id": 12119,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12118 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792675476, "dur": 23, + "args": { + "External id": 12120,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12119 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792675504, "dur": 25, + "args": { + "External id": 12121,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12120 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792675506, "dur": 22, + "args": { + "External id": 12122,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12121 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792675533, "dur": 26, + "args": { + "External id": 12123,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12122 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792675536, "dur": 22, + "args": { + "External id": 12124,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12123 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792675563, "dur": 26, + "args": { + "External id": 12125,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12124 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792675565, "dur": 23, + "args": { + "External id": 12126,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12125 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792675593, "dur": 25, + "args": { + "External id": 12127,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12126 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792675595, "dur": 22, + "args": { + "External id": 12128,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12127 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792675623, "dur": 25, + "args": { + "External id": 12129,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12128 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792675625, "dur": 22, + "args": { + "External id": 12130,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12129 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792675652, "dur": 25, + "args": { + "External id": 12131,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12130 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792675654, "dur": 22, + "args": { + "External id": 12132,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12131 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792675682, "dur": 25, + "args": { + "External id": 12133,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12132 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792675684, "dur": 22, + "args": { + "External id": 12134,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12133 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792675712, "dur": 25, + "args": { + "External id": 12135,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12134 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792675714, "dur": 23, + "args": { + "External id": 12136,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12135 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792675742, "dur": 25, + "args": { + "External id": 12137,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12136 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792675744, "dur": 22, + "args": { + "External id": 12138,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12137 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792675771, "dur": 26, + "args": { + "External id": 12139,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12138 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792675774, "dur": 22, + "args": { + "External id": 12140,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12139 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792675801, "dur": 25, + "args": { + "External id": 12141,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12140 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792675803, "dur": 23, + "args": { + "External id": 12142,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12141 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792675831, "dur": 25, + "args": { + "External id": 12143,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12142 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792675833, "dur": 22, + "args": { + "External id": 12144,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12143 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792675860, "dur": 26, + "args": { + "External id": 12145,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12144 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792675862, "dur": 23, + "args": { + "External id": 12146,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12145 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792675890, "dur": 25, + "args": { + "External id": 12147,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12146 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792675892, "dur": 22, + "args": { + "External id": 12148,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12147 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792675919, "dur": 26, + "args": { + "External id": 12149,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12148 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792675922, "dur": 22, + "args": { + "External id": 12150,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12149 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792675949, "dur": 25, + "args": { + "External id": 12151,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12150 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792675951, "dur": 22, + "args": { + "External id": 12152,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12151 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792675979, "dur": 25, + "args": { + "External id": 12153,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12152 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792675981, "dur": 22, + "args": { + "External id": 12154,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12153 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676008, "dur": 25, + "args": { + "External id": 12155,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12154 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676010, "dur": 23, + "args": { + "External id": 12156,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12155 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676038, "dur": 25, + "args": { + "External id": 12157,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12156 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676040, "dur": 22, + "args": { + "External id": 12158,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12157 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676067, "dur": 26, + "args": { + "External id": 12159,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12158 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676069, "dur": 23, + "args": { + "External id": 12160,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12159 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676097, "dur": 25, + "args": { + "External id": 12161,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12160 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676100, "dur": 22, + "args": { + "External id": 12162,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12161 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676127, "dur": 25, + "args": { + "External id": 12163,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12162 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676129, "dur": 22, + "args": { + "External id": 12164,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12163 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676156, "dur": 26, + "args": { + "External id": 12165,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12164 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676159, "dur": 22, + "args": { + "External id": 12166,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12165 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676186, "dur": 25, + "args": { + "External id": 12167,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12166 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676188, "dur": 22, + "args": { + "External id": 12168,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12167 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676270, "dur": 33, + "args": { + "External id": 12169,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12168 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676274, "dur": 28, + "args": { + "External id": 12170,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12169 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676308, "dur": 27, + "args": { + "External id": 12171,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12170 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676311, "dur": 23, + "args": { + "External id": 12172,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12171 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676339, "dur": 26, + "args": { + "External id": 12173,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12172 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676341, "dur": 23, + "args": { + "External id": 12174,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12173 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676369, "dur": 26, + "args": { + "External id": 12175,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12174 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676371, "dur": 23, + "args": { + "External id": 12176,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12175 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676399, "dur": 25, + "args": { + "External id": 12177,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12176 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676401, "dur": 22, + "args": { + "External id": 12178,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12177 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676428, "dur": 26, + "args": { + "External id": 12179,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12178 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676430, "dur": 23, + "args": { + "External id": 12180,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12179 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676458, "dur": 26, + "args": { + "External id": 12181,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12180 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676461, "dur": 22, + "args": { + "External id": 12182,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12181 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676489, "dur": 25, + "args": { + "External id": 12183,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12182 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676491, "dur": 22, + "args": { + "External id": 12184,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12183 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676519, "dur": 25, + "args": { + "External id": 12185,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12184 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676521, "dur": 22, + "args": { + "External id": 12186,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12185 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676549, "dur": 25, + "args": { + "External id": 12187,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12186 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676551, "dur": 22, + "args": { + "External id": 12188,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12187 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676578, "dur": 26, + "args": { + "External id": 12189,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12188 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676581, "dur": 22, + "args": { + "External id": 12190,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12189 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676608, "dur": 25, + "args": { + "External id": 12191,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12190 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676610, "dur": 22, + "args": { + "External id": 12192,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12191 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676637, "dur": 25, + "args": { + "External id": 12193,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12192 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676640, "dur": 22, + "args": { + "External id": 12194,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12193 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676667, "dur": 25, + "args": { + "External id": 12195,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12194 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676669, "dur": 22, + "args": { + "External id": 12196,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12195 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676696, "dur": 26, + "args": { + "External id": 12197,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12196 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676699, "dur": 22, + "args": { + "External id": 12198,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12197 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676726, "dur": 26, + "args": { + "External id": 12199,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12198 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676729, "dur": 22, + "args": { + "External id": 12200,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12199 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676756, "dur": 25, + "args": { + "External id": 12201,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12200 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676758, "dur": 23, + "args": { + "External id": 12202,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12201 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676786, "dur": 31, + "args": { + "External id": 12203,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12202 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676788, "dur": 28, + "args": { + "External id": 12204,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12203 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676822, "dur": 26, + "args": { + "External id": 12205,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12204 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676824, "dur": 23, + "args": { + "External id": 12206,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12205 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792676853, "dur": 25, + "args": { + "External id": 12207,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12206 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792676855, "dur": 22, + "args": { + "External id": 12208,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12207 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792677813, "dur": 35, + "args": { + "External id": 12209,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12208 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792677817, "dur": 30, + "args": { + "External id": 12210,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12209 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792677854, "dur": 27, + "args": { + "External id": 12211,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12210 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792677857, "dur": 23, + "args": { + "External id": 12212,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12211 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792677886, "dur": 26, + "args": { + "External id": 12213,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12212 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792677888, "dur": 23, + "args": { + "External id": 12214,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12213 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792677916, "dur": 26, + "args": { + "External id": 12215,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12214 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792677918, "dur": 23, + "args": { + "External id": 12216,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12215 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792677946, "dur": 25, + "args": { + "External id": 12217,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12216 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792677948, "dur": 23, + "args": { + "External id": 12218,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12217 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792677976, "dur": 25, + "args": { + "External id": 12219,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12218 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792677978, "dur": 22, + "args": { + "External id": 12220,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12219 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792678005, "dur": 25, + "args": { + "External id": 12221,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12220 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792678008, "dur": 22, + "args": { + "External id": 12222,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12221 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792678035, "dur": 25, + "args": { + "External id": 12223,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12222 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792678037, "dur": 22, + "args": { + "External id": 12224,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12223 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792678064, "dur": 26, + "args": { + "External id": 12225,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12224 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792678067, "dur": 22, + "args": { + "External id": 12226,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12225 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792678094, "dur": 25, + "args": { + "External id": 12227,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12226 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792678097, "dur": 22, + "args": { + "External id": 12228,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12227 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792678364, "dur": 34, + "args": { + "External id": 12229,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12228 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792678368, "dur": 29, + "args": { + "External id": 12230,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12229 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792678403, "dur": 27, + "args": { + "External id": 12231,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12230 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792678406, "dur": 23, + "args": { + "External id": 12232,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12231 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792678434, "dur": 25, + "args": { + "External id": 12233,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12232 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792678436, "dur": 22, + "args": { + "External id": 12234,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12233 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792678464, "dur": 25, + "args": { + "External id": 12235,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12234 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792678466, "dur": 22, + "args": { + "External id": 12236,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12235 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792678493, "dur": 26, + "args": { + "External id": 12237,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12236 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792678496, "dur": 22, + "args": { + "External id": 12238,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12237 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792678523, "dur": 25, + "args": { + "External id": 12239,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12238 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792678525, "dur": 23, + "args": { + "External id": 12240,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12239 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792678552, "dur": 26, + "args": { + "External id": 12241,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12240 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792678555, "dur": 22, + "args": { + "External id": 12242,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12241 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792678582, "dur": 25, + "args": { + "External id": 12243,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12242 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792678584, "dur": 23, + "args": { + "External id": 12244,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12243 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792678612, "dur": 25, + "args": { + "External id": 12245,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12244 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792678614, "dur": 22, + "args": { + "External id": 12246,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12245 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792678641, "dur": 25, + "args": { + "External id": 12247,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12246 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792678643, "dur": 23, + "args": { + "External id": 12248,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12247 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792678671, "dur": 25, + "args": { + "External id": 12249,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12248 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792678673, "dur": 22, + "args": { + "External id": 12250,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12249 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792678701, "dur": 25, + "args": { + "External id": 12251,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12250 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792678703, "dur": 22, + "args": { + "External id": 12252,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12251 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792678730, "dur": 25, + "args": { + "External id": 12253,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12252 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792678732, "dur": 22, + "args": { + "External id": 12254,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12253 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792678759, "dur": 25, + "args": { + "External id": 12255,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12254 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792678762, "dur": 22, + "args": { + "External id": 12256,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12255 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792678789, "dur": 25, + "args": { + "External id": 12257,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12256 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792678791, "dur": 22, + "args": { + "External id": 12258,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12257 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792678818, "dur": 26, + "args": { + "External id": 12259,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12258 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792678821, "dur": 22, + "args": { + "External id": 12260,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12259 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792678848, "dur": 26, + "args": { + "External id": 12261,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12260 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792678851, "dur": 22, + "args": { + "External id": 12262,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12261 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792678879, "dur": 26, + "args": { + "External id": 12263,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12262 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792678882, "dur": 22, + "args": { + "External id": 12264,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12263 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792678909, "dur": 26, + "args": { + "External id": 12265,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12264 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792678911, "dur": 23, + "args": { + "External id": 12266,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12265 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792678939, "dur": 25, + "args": { + "External id": 12267,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12266 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792678941, "dur": 23, + "args": { + "External id": 12268,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12267 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792678969, "dur": 25, + "args": { + "External id": 12269,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12268 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792678971, "dur": 22, + "args": { + "External id": 12270,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12269 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792678999, "dur": 25, + "args": { + "External id": 12271,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12270 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679001, "dur": 22, + "args": { + "External id": 12272,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12271 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679028, "dur": 25, + "args": { + "External id": 12273,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12272 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679030, "dur": 23, + "args": { + "External id": 12274,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12273 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679058, "dur": 25, + "args": { + "External id": 12275,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12274 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679060, "dur": 23, + "args": { + "External id": 12276,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12275 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679088, "dur": 25, + "args": { + "External id": 12277,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12276 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679090, "dur": 23, + "args": { + "External id": 12278,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12277 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679118, "dur": 25, + "args": { + "External id": 12279,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12278 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679120, "dur": 22, + "args": { + "External id": 12280,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12279 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679147, "dur": 26, + "args": { + "External id": 12281,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12280 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679149, "dur": 23, + "args": { + "External id": 12282,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12281 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679177, "dur": 25, + "args": { + "External id": 12283,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12282 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679179, "dur": 22, + "args": { + "External id": 12284,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12283 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679207, "dur": 25, + "args": { + "External id": 12285,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12284 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679209, "dur": 22, + "args": { + "External id": 12286,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12285 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679236, "dur": 26, + "args": { + "External id": 12287,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12286 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679239, "dur": 22, + "args": { + "External id": 12288,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12287 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679276, "dur": 26, + "args": { + "External id": 12289,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12288 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679278, "dur": 23, + "args": { + "External id": 12290,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12289 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679307, "dur": 26, + "args": { + "External id": 12291,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12290 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679309, "dur": 23, + "args": { + "External id": 12292,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12291 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679337, "dur": 25, + "args": { + "External id": 12293,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12292 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679339, "dur": 23, + "args": { + "External id": 12294,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12293 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679367, "dur": 25, + "args": { + "External id": 12295,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12294 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679369, "dur": 22, + "args": { + "External id": 12296,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12295 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679397, "dur": 29, + "args": { + "External id": 12297,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12296 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679400, "dur": 26, + "args": { + "External id": 12298,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12297 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679430, "dur": 26, + "args": { + "External id": 12299,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12298 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679433, "dur": 22, + "args": { + "External id": 12300,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12299 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679460, "dur": 26, + "args": { + "External id": 12301,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12300 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679462, "dur": 23, + "args": { + "External id": 12302,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12301 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679490, "dur": 25, + "args": { + "External id": 12303,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12302 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679492, "dur": 22, + "args": { + "External id": 12304,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12303 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679519, "dur": 26, + "args": { + "External id": 12305,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12304 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679521, "dur": 23, + "args": { + "External id": 12306,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12305 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679549, "dur": 25, + "args": { + "External id": 12307,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12306 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679551, "dur": 23, + "args": { + "External id": 12308,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12307 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679579, "dur": 25, + "args": { + "External id": 12309,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12308 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679581, "dur": 22, + "args": { + "External id": 12310,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12309 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679608, "dur": 26, + "args": { + "External id": 12311,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12310 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679610, "dur": 23, + "args": { + "External id": 12312,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12311 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679638, "dur": 25, + "args": { + "External id": 12313,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12312 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679640, "dur": 22, + "args": { + "External id": 12314,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12313 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679667, "dur": 26, + "args": { + "External id": 12315,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12314 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679670, "dur": 22, + "args": { + "External id": 12316,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12315 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679697, "dur": 25, + "args": { + "External id": 12317,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12316 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679699, "dur": 23, + "args": { + "External id": 12318,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12317 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679727, "dur": 25, + "args": { + "External id": 12319,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12318 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679729, "dur": 22, + "args": { + "External id": 12320,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12319 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679756, "dur": 26, + "args": { + "External id": 12321,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12320 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679759, "dur": 22, + "args": { + "External id": 12322,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12321 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679786, "dur": 25, + "args": { + "External id": 12323,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12322 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679788, "dur": 23, + "args": { + "External id": 12324,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12323 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679816, "dur": 25, + "args": { + "External id": 12325,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12324 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679818, "dur": 22, + "args": { + "External id": 12326,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12325 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792679845, "dur": 25, + "args": { + "External id": 12327,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12326 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792679848, "dur": 22, + "args": { + "External id": 12328,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12327 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792680872, "dur": 37, + "args": { + "External id": 12329,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12328 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792680875, "dur": 33, + "args": { + "External id": 12330,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12329 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792680915, "dur": 27, + "args": { + "External id": 12331,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12330 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792680918, "dur": 24, + "args": { + "External id": 12332,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12331 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792680947, "dur": 29, + "args": { + "External id": 12333,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12332 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792680952, "dur": 23, + "args": { + "External id": 12334,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12333 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792680981, "dur": 26, + "args": { + "External id": 12335,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12334 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792680983, "dur": 23, + "args": { + "External id": 12336,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12335 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792681011, "dur": 26, + "args": { + "External id": 12337,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12336 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792681014, "dur": 22, + "args": { + "External id": 12338,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12337 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792681041, "dur": 26, + "args": { + "External id": 12339,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12338 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792681043, "dur": 23, + "args": { + "External id": 12340,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12339 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792681071, "dur": 25, + "args": { + "External id": 12341,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12340 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792681073, "dur": 22, + "args": { + "External id": 12342,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12341 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792681100, "dur": 26, + "args": { + "External id": 12343,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12342 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792681103, "dur": 22, + "args": { + "External id": 12344,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12343 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792681130, "dur": 25, + "args": { + "External id": 12345,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12344 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792681132, "dur": 23, + "args": { + "External id": 12346,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12345 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792681160, "dur": 25, + "args": { + "External id": 12347,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12346 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792681162, "dur": 22, + "args": { + "External id": 12348,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12347 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792681417, "dur": 34, + "args": { + "External id": 12349,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12348 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792681420, "dur": 30, + "args": { + "External id": 12350,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12349 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792681457, "dur": 27, + "args": { + "External id": 12351,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12350 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792681459, "dur": 24, + "args": { + "External id": 12352,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12351 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792681488, "dur": 26, + "args": { + "External id": 12353,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12352 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792681490, "dur": 23, + "args": { + "External id": 12354,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12353 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792681518, "dur": 26, + "args": { + "External id": 12355,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12354 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792681520, "dur": 23, + "args": { + "External id": 12356,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12355 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792681549, "dur": 25, + "args": { + "External id": 12357,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12356 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792681551, "dur": 22, + "args": { + "External id": 12358,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12357 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792681578, "dur": 26, + "args": { + "External id": 12359,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12358 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792681581, "dur": 22, + "args": { + "External id": 12360,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12359 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792681608, "dur": 26, + "args": { + "External id": 12361,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12360 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792681610, "dur": 23, + "args": { + "External id": 12362,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12361 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792681638, "dur": 26, + "args": { + "External id": 12363,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12362 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792681640, "dur": 23, + "args": { + "External id": 12364,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12363 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792681668, "dur": 25, + "args": { + "External id": 12365,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12364 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792681670, "dur": 23, + "args": { + "External id": 12366,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12365 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792681698, "dur": 25, + "args": { + "External id": 12367,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12366 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792681700, "dur": 23, + "args": { + "External id": 12368,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12367 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792681728, "dur": 25, + "args": { + "External id": 12369,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12368 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792681730, "dur": 22, + "args": { + "External id": 12370,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12369 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792681757, "dur": 25, + "args": { + "External id": 12371,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12370 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792681759, "dur": 23, + "args": { + "External id": 12372,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12371 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792681787, "dur": 25, + "args": { + "External id": 12373,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12372 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792681789, "dur": 23, + "args": { + "External id": 12374,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12373 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792681817, "dur": 25, + "args": { + "External id": 12375,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12374 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792681819, "dur": 23, + "args": { + "External id": 12376,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12375 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792681847, "dur": 25, + "args": { + "External id": 12377,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12376 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792681849, "dur": 22, + "args": { + "External id": 12378,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12377 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792681876, "dur": 25, + "args": { + "External id": 12379,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12378 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792681878, "dur": 23, + "args": { + "External id": 12380,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12379 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792681906, "dur": 25, + "args": { + "External id": 12381,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12380 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792681908, "dur": 23, + "args": { + "External id": 12382,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12381 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792681936, "dur": 25, + "args": { + "External id": 12383,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12382 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792681938, "dur": 22, + "args": { + "External id": 12384,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12383 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792681965, "dur": 26, + "args": { + "External id": 12385,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12384 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792681968, "dur": 22, + "args": { + "External id": 12386,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12385 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792681995, "dur": 25, + "args": { + "External id": 12387,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12386 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792681997, "dur": 22, + "args": { + "External id": 12388,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12387 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682025, "dur": 25, + "args": { + "External id": 12389,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12388 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682027, "dur": 22, + "args": { + "External id": 12390,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12389 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682055, "dur": 26, + "args": { + "External id": 12391,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12390 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682057, "dur": 23, + "args": { + "External id": 12392,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12391 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682085, "dur": 25, + "args": { + "External id": 12393,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12392 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682087, "dur": 22, + "args": { + "External id": 12394,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12393 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682114, "dur": 26, + "args": { + "External id": 12395,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12394 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682117, "dur": 22, + "args": { + "External id": 12396,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12395 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682144, "dur": 26, + "args": { + "External id": 12397,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12396 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682146, "dur": 23, + "args": { + "External id": 12398,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12397 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682174, "dur": 26, + "args": { + "External id": 12399,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12398 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682176, "dur": 23, + "args": { + "External id": 12400,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12399 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682204, "dur": 25, + "args": { + "External id": 12401,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12400 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682206, "dur": 23, + "args": { + "External id": 12402,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12401 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682234, "dur": 25, + "args": { + "External id": 12403,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12402 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682236, "dur": 22, + "args": { + "External id": 12404,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12403 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682263, "dur": 26, + "args": { + "External id": 12405,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12404 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682265, "dur": 23, + "args": { + "External id": 12406,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12405 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682293, "dur": 25, + "args": { + "External id": 12407,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12406 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682295, "dur": 23, + "args": { + "External id": 12408,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12407 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682323, "dur": 25, + "args": { + "External id": 12409,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12408 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682325, "dur": 23, + "args": { + "External id": 12410,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12409 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682353, "dur": 25, + "args": { + "External id": 12411,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12410 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682355, "dur": 22, + "args": { + "External id": 12412,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12411 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682382, "dur": 26, + "args": { + "External id": 12413,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12412 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682384, "dur": 23, + "args": { + "External id": 12414,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12413 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682412, "dur": 25, + "args": { + "External id": 12415,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12414 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682414, "dur": 23, + "args": { + "External id": 12416,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12415 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682442, "dur": 26, + "args": { + "External id": 12417,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12416 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682444, "dur": 23, + "args": { + "External id": 12418,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12417 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682472, "dur": 25, + "args": { + "External id": 12419,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12418 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682474, "dur": 22, + "args": { + "External id": 12420,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12419 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682502, "dur": 25, + "args": { + "External id": 12421,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12420 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682504, "dur": 23, + "args": { + "External id": 12422,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12421 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682532, "dur": 29, + "args": { + "External id": 12423,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12422 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682534, "dur": 26, + "args": { + "External id": 12424,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12423 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682566, "dur": 25, + "args": { + "External id": 12425,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12424 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682568, "dur": 23, + "args": { + "External id": 12426,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12425 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682596, "dur": 25, + "args": { + "External id": 12427,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12426 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682598, "dur": 22, + "args": { + "External id": 12428,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12427 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682625, "dur": 26, + "args": { + "External id": 12429,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12428 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682628, "dur": 22, + "args": { + "External id": 12430,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12429 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682655, "dur": 25, + "args": { + "External id": 12431,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12430 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682657, "dur": 23, + "args": { + "External id": 12432,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12431 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682685, "dur": 25, + "args": { + "External id": 12433,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12432 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682687, "dur": 22, + "args": { + "External id": 12434,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12433 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682714, "dur": 26, + "args": { + "External id": 12435,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12434 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682717, "dur": 22, + "args": { + "External id": 12436,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12435 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682744, "dur": 25, + "args": { + "External id": 12437,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12436 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682746, "dur": 23, + "args": { + "External id": 12438,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12437 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682774, "dur": 25, + "args": { + "External id": 12439,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12438 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682776, "dur": 22, + "args": { + "External id": 12440,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12439 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682803, "dur": 26, + "args": { + "External id": 12441,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12440 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682805, "dur": 23, + "args": { + "External id": 12442,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12441 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682833, "dur": 26, + "args": { + "External id": 12443,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12442 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682835, "dur": 23, + "args": { + "External id": 12444,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12443 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682863, "dur": 26, + "args": { + "External id": 12445,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12444 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682865, "dur": 23, + "args": { + "External id": 12446,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12445 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792682893, "dur": 25, + "args": { + "External id": 12447,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12446 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792682895, "dur": 22, + "args": { + "External id": 12448,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12447 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792683913, "dur": 34, + "args": { + "External id": 12449,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12448 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792683917, "dur": 29, + "args": { + "External id": 12450,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12449 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792683952, "dur": 28, + "args": { + "External id": 12451,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12450 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792683955, "dur": 24, + "args": { + "External id": 12452,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12451 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792683984, "dur": 26, + "args": { + "External id": 12453,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12452 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792683986, "dur": 24, + "args": { + "External id": 12454,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12453 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792684015, "dur": 26, + "args": { + "External id": 12455,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12454 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792684017, "dur": 23, + "args": { + "External id": 12456,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12455 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792684045, "dur": 26, + "args": { + "External id": 12457,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12456 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792684047, "dur": 23, + "args": { + "External id": 12458,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12457 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792684076, "dur": 25, + "args": { + "External id": 12459,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12458 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792684078, "dur": 23, + "args": { + "External id": 12460,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12459 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792684106, "dur": 25, + "args": { + "External id": 12461,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12460 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792684108, "dur": 22, + "args": { + "External id": 12462,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12461 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792684135, "dur": 26, + "args": { + "External id": 12463,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12462 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792684138, "dur": 23, + "args": { + "External id": 12464,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12463 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792684166, "dur": 25, + "args": { + "External id": 12465,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12464 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792684168, "dur": 22, + "args": { + "External id": 12466,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12465 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792684196, "dur": 45, + "args": { + "External id": 12467,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12466 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792684198, "dur": 42, + "args": { + "External id": 12468,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12467 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792684463, "dur": 35, + "args": { + "External id": 12469,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12468 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792684467, "dur": 31, + "args": { + "External id": 12470,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12469 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792684504, "dur": 27, + "args": { + "External id": 12471,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12470 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792684506, "dur": 24, + "args": { + "External id": 12472,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12471 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792684536, "dur": 26, + "args": { + "External id": 12473,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12472 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792684538, "dur": 23, + "args": { + "External id": 12474,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12473 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792684566, "dur": 26, + "args": { + "External id": 12475,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12474 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792684569, "dur": 22, + "args": { + "External id": 12476,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12475 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792684596, "dur": 26, + "args": { + "External id": 12477,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12476 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792684599, "dur": 22, + "args": { + "External id": 12478,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12477 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792684626, "dur": 26, + "args": { + "External id": 12479,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12478 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792684629, "dur": 22, + "args": { + "External id": 12480,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12479 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792684656, "dur": 26, + "args": { + "External id": 12481,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12480 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792684659, "dur": 22, + "args": { + "External id": 12482,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12481 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792684686, "dur": 26, + "args": { + "External id": 12483,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12482 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792684689, "dur": 22, + "args": { + "External id": 12484,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12483 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792684716, "dur": 26, + "args": { + "External id": 12485,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12484 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792684719, "dur": 22, + "args": { + "External id": 12486,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12485 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792684746, "dur": 26, + "args": { + "External id": 12487,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12486 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792684749, "dur": 22, + "args": { + "External id": 12488,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12487 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792684776, "dur": 26, + "args": { + "External id": 12489,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12488 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792684779, "dur": 23, + "args": { + "External id": 12490,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12489 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792684807, "dur": 26, + "args": { + "External id": 12491,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12490 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792684809, "dur": 23, + "args": { + "External id": 12492,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12491 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792684837, "dur": 26, + "args": { + "External id": 12493,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12492 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792684839, "dur": 23, + "args": { + "External id": 12494,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12493 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792684867, "dur": 26, + "args": { + "External id": 12495,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12494 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792684870, "dur": 22, + "args": { + "External id": 12496,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12495 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792684897, "dur": 26, + "args": { + "External id": 12497,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12496 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792684900, "dur": 22, + "args": { + "External id": 12498,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12497 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792684928, "dur": 25, + "args": { + "External id": 12499,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12498 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792684930, "dur": 22, + "args": { + "External id": 12500,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12499 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792684957, "dur": 26, + "args": { + "External id": 12501,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12500 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792684960, "dur": 22, + "args": { + "External id": 12502,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12501 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792684987, "dur": 26, + "args": { + "External id": 12503,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12502 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792684989, "dur": 23, + "args": { + "External id": 12504,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12503 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685017, "dur": 25, + "args": { + "External id": 12505,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12504 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685019, "dur": 23, + "args": { + "External id": 12506,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12505 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685047, "dur": 25, + "args": { + "External id": 12507,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12506 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685049, "dur": 23, + "args": { + "External id": 12508,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12507 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685077, "dur": 26, + "args": { + "External id": 12509,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12508 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685080, "dur": 23, + "args": { + "External id": 12510,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12509 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685108, "dur": 26, + "args": { + "External id": 12511,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12510 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685110, "dur": 23, + "args": { + "External id": 12512,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12511 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685138, "dur": 25, + "args": { + "External id": 12513,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12512 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685140, "dur": 23, + "args": { + "External id": 12514,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12513 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685168, "dur": 26, + "args": { + "External id": 12515,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12514 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685170, "dur": 23, + "args": { + "External id": 12516,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12515 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685198, "dur": 26, + "args": { + "External id": 12517,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12516 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685200, "dur": 23, + "args": { + "External id": 12518,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12517 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685229, "dur": 25, + "args": { + "External id": 12519,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12518 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685231, "dur": 23, + "args": { + "External id": 12520,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12519 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685259, "dur": 25, + "args": { + "External id": 12521,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12520 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685261, "dur": 22, + "args": { + "External id": 12522,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12521 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685289, "dur": 25, + "args": { + "External id": 12523,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12522 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685291, "dur": 22, + "args": { + "External id": 12524,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12523 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685319, "dur": 25, + "args": { + "External id": 12525,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12524 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685321, "dur": 22, + "args": { + "External id": 12526,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12525 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685348, "dur": 26, + "args": { + "External id": 12527,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12526 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685351, "dur": 22, + "args": { + "External id": 12528,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12527 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685378, "dur": 26, + "args": { + "External id": 12529,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12528 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685381, "dur": 22, + "args": { + "External id": 12530,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12529 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685408, "dur": 26, + "args": { + "External id": 12531,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12530 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685410, "dur": 23, + "args": { + "External id": 12532,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12531 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685438, "dur": 26, + "args": { + "External id": 12533,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12532 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685441, "dur": 22, + "args": { + "External id": 12534,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12533 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685468, "dur": 26, + "args": { + "External id": 12535,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12534 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685470, "dur": 23, + "args": { + "External id": 12536,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12535 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685498, "dur": 25, + "args": { + "External id": 12537,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12536 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685500, "dur": 23, + "args": { + "External id": 12538,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12537 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685528, "dur": 25, + "args": { + "External id": 12539,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12538 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685530, "dur": 22, + "args": { + "External id": 12540,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12539 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685558, "dur": 25, + "args": { + "External id": 12541,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12540 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685560, "dur": 23, + "args": { + "External id": 12542,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12541 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685588, "dur": 25, + "args": { + "External id": 12543,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12542 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685590, "dur": 23, + "args": { + "External id": 12544,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12543 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685618, "dur": 26, + "args": { + "External id": 12545,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12544 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685620, "dur": 23, + "args": { + "External id": 12546,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12545 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685648, "dur": 26, + "args": { + "External id": 12547,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12546 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685650, "dur": 23, + "args": { + "External id": 12548,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12547 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685678, "dur": 26, + "args": { + "External id": 12549,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12548 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685680, "dur": 23, + "args": { + "External id": 12550,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12549 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685708, "dur": 25, + "args": { + "External id": 12551,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12550 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685710, "dur": 23, + "args": { + "External id": 12552,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12551 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685738, "dur": 25, + "args": { + "External id": 12553,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12552 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685740, "dur": 23, + "args": { + "External id": 12554,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12553 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685768, "dur": 25, + "args": { + "External id": 12555,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12554 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685770, "dur": 22, + "args": { + "External id": 12556,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12555 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685798, "dur": 25, + "args": { + "External id": 12557,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12556 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685800, "dur": 22, + "args": { + "External id": 12558,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12557 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685827, "dur": 25, + "args": { + "External id": 12559,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12558 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685829, "dur": 23, + "args": { + "External id": 12560,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12559 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685857, "dur": 25, + "args": { + "External id": 12561,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12560 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685859, "dur": 23, + "args": { + "External id": 12562,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12561 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685887, "dur": 25, + "args": { + "External id": 12563,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12562 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685889, "dur": 23, + "args": { + "External id": 12564,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12563 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685917, "dur": 25, + "args": { + "External id": 12565,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12564 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685919, "dur": 22, + "args": { + "External id": 12566,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12565 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 435, "tid": 435, + "ts": 1776649792685947, "dur": 25, + "args": { + "External id": 12567,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12566 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 435, "tid": 435, + "ts": 1776649792685949, "dur": 22, + "args": { + "External id": 12568,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8, 4096], [4096, 8192]], "Ev Idx": 12567 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__and__", "pid": 435, "tid": 435, + "ts": 1776649792686996, "dur": 36, + "args": { + "External id": 12569,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "15"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 12568 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_and", "pid": 435, "tid": 435, + "ts": 1776649792686998, "dur": 34, + "args": { + "External id": 12570,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "15"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 12569 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::bitwise_and", "pid": 435, "tid": 435, + "ts": 1776649792687003, "dur": 27, + "args": { + "External id": 12571,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["unsigned char", "long int"], "Input Dims": [[4096, 4096], []], "Ev Idx": 12570 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792687038, "dur": 30, + "args": { + "External id": 12572,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 12571 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649792687039, "dur": 28, + "args": { + "External id": 12573,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["unsigned char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], [], []], "Ev Idx": 12572 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649792687044, "dur": 6, + "args": { + "External id": 12574,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "[4096, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12573 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792687051, "dur": 16, + "args": { + "External id": 12575,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "unsigned char", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 12574 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::__rshift__", "pid": 435, "tid": 435, + "ts": 1776649792687074, "dur": 25, + "args": { + "External id": 12576,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4"], "Input type": ["unsigned char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 12575 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792687081, "dur": 5, + "args": { + "External id": 12577,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "0", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12576 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792687103, "dur": 22, + "args": { + "External id": 12578,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "False", "False", ""], "Input type": ["unsigned char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], []], "Ev Idx": 12577 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649792687104, "dur": 21, + "args": { + "External id": 12579,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "", "", "", "False", ""], "Input type": ["unsigned char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 4096], [], [], [], [], [], []], "Ev Idx": 12578 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649792687106, "dur": 5, + "args": { + "External id": 12580,"Record function id": 0, "Concrete Inputs": ["[4096, 4096]", "[4096, 1]", "1", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12579 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792687112, "dur": 12, + "args": { + "External id": 12581,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["signed char", "unsigned char", "Scalar"], "Input Dims": [[4096, 4096], [4096, 4096], []], "Ev Idx": 12580 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::stack", "pid": 435, "tid": 435, + "ts": 1776649792687134, "dur": 34, + "args": { + "External id": 12582,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["TensorList", "Scalar"], "Input Dims": [[[4096, 4096], [4096, 4096]], []], "Ev Idx": 12581 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649792687136, "dur": 4, + "args": { + "External id": 12583,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["signed char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 12582 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792687139, "dur": 1, + "args": { + "External id": 12584,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096, 1]", "[4096, 1, 1]", ""], "Input type": ["signed char", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 12583 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649792687141, "dur": 2, + "args": { + "External id": 12585,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["signed char", "Scalar"], "Input Dims": [[4096, 4096], []], "Ev Idx": 12584 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792687143, "dur": 0, + "args": { + "External id": 12586,"Record function id": 0, "Concrete Inputs": ["", "[4096, 4096, 1]", "[4096, 1, 1]", ""], "Input type": ["signed char", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 4096], [], [], []], "Ev Idx": 12585 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 435, "tid": 435, + "ts": 1776649792687145, "dur": 21, + "args": { + "External id": 12587,"Record function id": 0, "Concrete Inputs": ["", "-1"], "Input type": ["TensorList", "Scalar"], "Input Dims": [[[4096, 4096, 1], [4096, 4096, 1]], []], "Ev Idx": 12586 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649792687174, "dur": 5, + "args": { + "External id": 12588,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["signed char", "ScalarList"], "Input Dims": [[4096, 4096, 2], []], "Ev Idx": 12587 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649792687176, "dur": 3, + "args": { + "External id": 12589,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["signed char", "ScalarList"], "Input Dims": [[4096, 4096, 2], []], "Ev Idx": 12588 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::repeat_interleave", "pid": 435, "tid": 435, + "ts": 1776649792687184, "dur": 47, + "args": { + "External id": 12590,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "128", "0", ""], "Input type": ["c10::Half", "Scalar", "Scalar", ""], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 12589 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 435, "tid": 435, + "ts": 1776649792687185, "dur": 4, + "args": { + "External id": 12591,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 8192], []], "Ev Idx": 12590 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792687187, "dur": 1, + "args": { + "External id": 12592,"Record function id": 0, "Concrete Inputs": ["", "[32, 1, 8192]", "[8192, 8192, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 12591 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 435, "tid": 435, + "ts": 1776649792687190, "dur": 5, + "args": { + "External id": 12593,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]", "False"], "Input type": ["c10::Half", "ScalarList", "Scalar"], "Input Dims": [[32, 1, 8192], [], []], "Ev Idx": 12592 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792687193, "dur": 1, + "args": { + "External id": 12594,"Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]", "[8192, 0, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 1, 8192], [], [], []], "Ev Idx": 12593 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649792687196, "dur": 29, + "args": { + "External id": 12595,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 12594 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649792687198, "dur": 9, + "args": { + "External id": 12596,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[32, 128, 8192], [], [], [], [], []], "Ev Idx": 12595 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792687200, "dur": 6, + "args": { + "External id": 12597,"Record function id": 0, "Concrete Inputs": ["[32, 128, 8192]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12596 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792687208, "dur": 16, + "args": { + "External id": 12598,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[32, 128, 8192], [32, 128, 8192], []], "Ev Idx": 12597 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::flatten", "pid": 435, "tid": 435, + "ts": 1776649792687227, "dur": 3, + "args": { + "External id": 12599,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[32, 128, 8192], [], []], "Ev Idx": 12598 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649792687228, "dur": 2, + "args": { + "External id": 12600,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 12599 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 435, "tid": 435, + "ts": 1776649792687238, "dur": 19, + "args": { + "External id": 12601,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["signed char", "long int", "Scalar"], "Input Dims": [[4096, 8192], [], []], "Ev Idx": 12600 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792687262, "dur": 24, + "args": { + "External id": 12602,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["signed char", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[4096, 8192], [], [], [], []], "Ev Idx": 12601 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649792687263, "dur": 23, + "args": { + "External id": 12603,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "", "", "", "False", ""], "Input type": ["signed char", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[4096, 8192], [], [], [], [], [], []], "Ev Idx": 12602 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649792687266, "dur": 6, + "args": { + "External id": 12604,"Record function id": 0, "Concrete Inputs": ["[4096, 8192]", "[8192, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12603 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792687273, "dur": 12, + "args": { + "External id": 12605,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "signed char", "Scalar"], "Input Dims": [[4096, 8192], [4096, 8192], []], "Ev Idx": 12604 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 435, "tid": 435, + "ts": 1776649792687292, "dur": 16, + "args": { + "External id": 12606,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[4096, 8192], [4096, 8192]], "Ev Idx": 12605 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 435, "tid": 435, + "ts": 1776649792687320, "dur": 11, + "args": { + "External id": 12607,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[4096, 8192]], "Ev Idx": 12606 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 435, "tid": 435, + "ts": 1776649792687324, "dur": 6, + "args": { + "External id": 12608,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[4096, 8192], [], []], "Ev Idx": 12607 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792687328, "dur": 1, + "args": { + "External id": 12609,"Record function id": 0, "Concrete Inputs": ["", "[8192, 4096]", "[1, 8192]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 8192], [], [], []], "Ev Idx": 12608 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 435, "tid": 435, + "ts": 1776649792687337, "dur": 27, + "args": { + "External id": 12610,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[8192, 4096], []], "Ev Idx": 12609 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649792687339, "dur": 25, + "args": { + "External id": 12611,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[8192, 4096], []], "Ev Idx": 12610 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649792687340, "dur": 8, + "args": { + "External id": 12612,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[8192, 4096], [], [], [], [], []], "Ev Idx": 12611 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792687342, "dur": 5, + "args": { + "External id": 12613,"Record function id": 0, "Concrete Inputs": ["[8192, 4096]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12612 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792687349, "dur": 14, + "args": { + "External id": 12614,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[8192, 4096], [8192, 4096], []], "Ev Idx": 12613 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649792687407, "dur": 63, + "args": { + "External id": 12615,"Record function id": 0, "Concrete Inputs": ["[8192, 4096]", "", "", "", "False", ""], "Input type": ["ScalarList", "", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12614 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 435, "tid": 435, + "ts": 1776649792687480, "dur": 7, + "args": { + "External id": 12616,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Dims": [[8192, 4096]], "Ev Idx": 12615 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 435, "tid": 435, + "ts": 1776649792687483, "dur": 4, + "args": { + "External id": 12617,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Dims": [[8192, 4096]], "Ev Idx": 12616 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::uniform_", "pid": 435, "tid": 435, + "ts": 1776649792687550, "dur": 213210, + "args": { + "External id": 12618,"Record function id": 0, "Concrete Inputs": ["", "-0.015624999999999998", "0.015624999999999998", ""], "Input type": ["float", "Scalar", "Scalar", ""], "Input Dims": [[8192, 4096], [], [], []], "Ev Idx": 12617 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792900920, "dur": 189763, + "args": { + "External id": 12619,"Record function id": 0, "Concrete Inputs": ["", "", "5", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[8192, 4096], [], [], [], [], []], "Ev Idx": 12618 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649792900926, "dur": 189754, + "args": { + "External id": 12620,"Record function id": 0, "Concrete Inputs": ["", "5", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[8192, 4096], [], [], [], [], [], []], "Ev Idx": 12619 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649792900952, "dur": 56, + "args": { + "External id": 12621,"Record function id": 0, "Concrete Inputs": ["[8192, 4096]", "[4096, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12620 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792901014, "dur": 189657, + "args": { + "External id": 12622,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "float", "Scalar"], "Input Dims": [[8192, 4096], [8192, 4096], []], "Ev Idx": 12621 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649792901034, "dur": 75215, + "args": { + "External id": 12623,"Record function id": 0, "Concrete Inputs": ["", "5", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[8192, 4096], [], [], [], []], "Ev Idx": 12622 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649792901036, "dur": 75211, + "args": { + "External id": 12624,"Record function id": 0, "Concrete Inputs": ["", "5", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[8192, 4096], [], [], [], [], [], []], "Ev Idx": 12623 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649792901038, "dur": 46, + "args": { + "External id": 12625,"Record function id": 0, "Concrete Inputs": ["[8192, 4096]", "[4096, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12624 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792901085, "dur": 75153, + "args": { + "External id": 12626,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "float", "Scalar"], "Input Dims": [[8192, 4096], [8192, 4096], []], "Ev Idx": 12625 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::expand_as", "pid": 435, "tid": 435, + "ts": 1776649792976270, "dur": 27, + "args": { + "External id": 12627,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[8192, 4096], [8192, 4096]], "Ev Idx": 12626 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 435, "tid": 435, + "ts": 1776649792976282, "dur": 15, + "args": { + "External id": 12628,"Record function id": 0, "Concrete Inputs": ["", "[8192, 4096]", "False"], "Input type": ["c10::Half", "ScalarList", "Scalar"], "Input Dims": [[8192, 4096], [], []], "Ev Idx": 12627 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649792976290, "dur": 6, + "args": { + "External id": 12629,"Record function id": 0, "Concrete Inputs": ["", "[8192, 4096]", "[4096, 1]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[8192, 4096], [], [], []], "Ev Idx": 12628 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649792976302, "dur": 103365, + "args": { + "External id": 12630,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[8192, 4096], [8192, 4096], []], "Ev Idx": 12629 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_has_compatible_shallow_copy_type", "pid": 435, "tid": 435, + "ts": 1776649793090778, "dur": 2, + "args": { + "External id": 12631,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "c10::Half"], "Input Dims": [[8192, 4096], [8192, 4096]], "Ev Idx": 12630 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_has_compatible_shallow_copy_type", "pid": 435, "tid": 435, + "ts": 1776649793090803, "dur": 0, + "args": { + "External id": 12632,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "c10::Half"], "Input Dims": [[8192, 4096], [8192, 4096]], "Ev Idx": 12631 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649793106325, "dur": 93, + "args": { + "External id": 12633,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[8192, 4096], [8192, 4096], []], "Ev Idx": 12632 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793106507, "dur": 15, + "args": { + "External id": 12634,"Record function id": 0, "Concrete Inputs": ["[256, 16384]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12633 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793106541, "dur": 1, + "args": { + "External id": 12635,"Record function id": 0, "Concrete Inputs": ["[32, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12634 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793106556, "dur": 10, + "args": { + "External id": 12636,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12635 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793106559, "dur": 1, + "args": { + "External id": 12637,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12636 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793106563, "dur": 2, + "args": { + "External id": 12638,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12637 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649793106600, "dur": 1394, + "args": { + "External id": 12639,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 16384], [], [], [], [], [], [], []], "Ev Idx": 12638 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649793106603, "dur": 1389, + "args": { + "External id": 12640,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[256, 16384], [], [], [], [], [], []], "Ev Idx": 12639 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649793106614, "dur": 23, + "args": { + "External id": 12641,"Record function id": 0, "Concrete Inputs": ["[256, 16384]", "[16384, 1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12640 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649793106639, "dur": 1349, + "args": { + "External id": 12642,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 16384], [256, 16384], []], "Ev Idx": 12641 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649793108036, "dur": 138, + "args": { + "External id": 12643,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[32, 8192], [], [], [], [], [], [], []], "Ev Idx": 12642 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649793108039, "dur": 135, + "args": { + "External id": 12644,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[32, 8192], [], [], [], [], [], []], "Ev Idx": 12643 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649793108050, "dur": 19, + "args": { + "External id": 12645,"Record function id": 0, "Concrete Inputs": ["[32, 8192]", "[8192, 1]", "5", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12644 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649793108071, "dur": 102, + "args": { + "External id": 12646,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[32, 8192], [32, 8192], []], "Ev Idx": 12645 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649793108182, "dur": 86, + "args": { + "External id": 12647,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[1024], [], [], [], [], [], [], []], "Ev Idx": 12646 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649793108183, "dur": 84, + "args": { + "External id": 12648,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[1024], [], [], [], [], [], []], "Ev Idx": 12647 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649793108186, "dur": 6, + "args": { + "External id": 12649,"Record function id": 0, "Concrete Inputs": ["[1024]", "[1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12648 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649793108194, "dur": 72, + "args": { + "External id": 12650,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[1024], [1024], []], "Ev Idx": 12649 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 435, "tid": 435, + "ts": 1776649793108299, "dur": 30, + "args": { + "External id": 12651,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[32, 8192]], "Ev Idx": 12650 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 435, "tid": 435, + "ts": 1776649793108306, "dur": 18, + "args": { + "External id": 12652,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[32, 8192], [], []], "Ev Idx": 12651 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649793108317, "dur": 6, + "args": { + "External id": 12653,"Record function id": 0, "Concrete Inputs": ["", "[8192, 32]", "[1, 8192]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 12652 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 435, "tid": 435, + "ts": 1776649793108337, "dur": 3, + "args": { + "External id": 12654,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[8192, 4096]], "Ev Idx": 12653 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 435, "tid": 435, + "ts": 1776649793108337, "dur": 2, + "args": { + "External id": 12655,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::Half", "Scalar", "Scalar"], "Input Dims": [[8192, 4096], [], []], "Ev Idx": 12654 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649793108339, "dur": 0, + "args": { + "External id": 12656,"Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]", "[1, 4096]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[8192, 4096], [], [], []], "Ev Idx": 12655 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649793108353, "dur": 28, + "args": { + "External id": 12657,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 128, 8192]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 8192], []], "Ev Idx": 12656 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 435, "tid": 435, + "ts": 1776649793108371, "dur": 9, + "args": { + "External id": 12658,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]", "[128, 1, 4096]"], "Input type": ["c10::Half", "ScalarList", "ScalarList"], "Input Dims": [[4096, 8192], [], []], "Ev Idx": 12657 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 435, "tid": 435, + "ts": 1776649793108390, "dur": 7, + "args": { + "External id": 12659,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1, 0, 2]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 12658 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649793108396, "dur": 0, + "args": { + "External id": 12660,"Record function id": 0, "Concrete Inputs": ["", "[128, 32, 8192]", "[1, 128, 4096]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[32, 128, 8192], [], [], []], "Ev Idx": 12659 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649793108400, "dur": 107, + "args": { + "External id": 12661,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[128, -1]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[128, 32, 8192], []], "Ev Idx": 12660 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649793108404, "dur": 93, + "args": { + "External id": 12662,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[128, 32, 8192], []], "Ev Idx": 12661 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649793108410, "dur": 21, + "args": { + "External id": 12663,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[128, 32, 8192], [], [], [], [], []], "Ev Idx": 12662 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793108415, "dur": 15, + "args": { + "External id": 12664,"Record function id": 0, "Concrete Inputs": ["[128, 32, 8192]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12663 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649793108433, "dur": 63, + "args": { + "External id": 12665,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[128, 32, 8192], [128, 32, 8192], []], "Ev Idx": 12664 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 435, "tid": 435, + "ts": 1776649793108501, "dur": 5, + "args": { + "External id": 12666,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[128, 262144]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[128, 32, 8192], []], "Ev Idx": 12665 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649793108515, "dur": 32, + "args": { + "External id": 12667,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1, -1]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[8192, 32], []], "Ev Idx": 12666 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649793108517, "dur": 27, + "args": { + "External id": 12668,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::Half", "Scalar"], "Input Dims": [[8192, 32], []], "Ev Idx": 12667 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649793108519, "dur": 8, + "args": { + "External id": 12669,"Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "0"], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[8192, 32], [], [], [], [], []], "Ev Idx": 12668 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793108520, "dur": 6, + "args": { + "External id": 12670,"Record function id": 0, "Concrete Inputs": ["[8192, 32]", "5", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12669 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649793108528, "dur": 15, + "args": { + "External id": 12671,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[8192, 32], [8192, 32], []], "Ev Idx": 12670 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 435, "tid": 435, + "ts": 1776649793108545, "dur": 2, + "args": { + "External id": 12672,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1, 262144]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[8192, 32], []], "Ev Idx": 12671 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 435, "tid": 435, + "ts": 1776649793108558, "dur": 36, + "args": { + "External id": 12673,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", "c10::Half"], "Input Dims": [[128, 262144], [1, 262144]], "Ev Idx": 12672 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::round", "pid": 435, "tid": 435, + "ts": 1776649793108600, "dur": 25, + "args": { + "External id": 12674,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::Half"], "Input Dims": [[128, 262144]], "Ev Idx": 12673 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649793108633, "dur": 31, + "args": { + "External id": 12675,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[128, 262144], [], [], [], []], "Ev Idx": 12674 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649793108635, "dur": 29, + "args": { + "External id": 12676,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "", "", "", "False", ""], "Input type": ["c10::Half", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[128, 262144], [], [], [], [], [], []], "Ev Idx": 12675 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649793108639, "dur": 5, + "args": { + "External id": 12677,"Record function id": 0, "Concrete Inputs": ["[128, 262144]", "[262144, 1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12676 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649793108646, "dur": 17, + "args": { + "External id": 12678,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "c10::Half", "Scalar"], "Input Dims": [[128, 262144], [128, 262144], []], "Ev Idx": 12677 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 435, "tid": 435, + "ts": 1776649793108685, "dur": 31, + "args": { + "External id": 12679,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["int", "long int", "Scalar"], "Input Dims": [[128, 262144], [], []], "Ev Idx": 12678 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 435, "tid": 435, + "ts": 1776649793108726, "dur": 33, + "args": { + "External id": 12680,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "15"], "Input type": ["int", "Scalar", "Scalar"], "Input Dims": [[128, 262144], [], []], "Ev Idx": 12679 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649793108732, "dur": 0, + "args": { + "External id": 12681,"Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[128, 262144], [], [], [], []], "Ev Idx": 12680 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649793108768, "dur": 8, + "args": { + "External id": 12682,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[128, -1, 8192]"], "Input type": ["int", "ScalarList"], "Input Dims": [[128, 262144], []], "Ev Idx": 12681 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649793108770, "dur": 6, + "args": { + "External id": 12683,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[128, -1, 8192]"], "Input type": ["int", "ScalarList"], "Input Dims": [[128, 262144], []], "Ev Idx": 12682 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 435, "tid": 435, + "ts": 1776649793108779, "dur": 5, + "args": { + "External id": 12684,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[1, 0, 2]"], "Input type": ["int", "ScalarList"], "Input Dims": [[128, 32, 8192], []], "Ev Idx": 12683 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649793108782, "dur": 1, + "args": { + "External id": 12685,"Record function id": 0, "Concrete Inputs": ["", "[32, 128, 8192]", "[8192, 262144, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Dims": [[128, 32, 8192], [], [], []], "Ev Idx": 12684 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649793108787, "dur": 37, + "args": { + "External id": 12686,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["int", "ScalarList"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 12685 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649793108790, "dur": 30, + "args": { + "External id": 12687,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 12686 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649793108791, "dur": 9, + "args": { + "External id": 12688,"Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "0"], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[32, 128, 8192], [], [], [], [], []], "Ev Idx": 12687 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793108793, "dur": 7, + "args": { + "External id": 12689,"Record function id": 0, "Concrete Inputs": ["[32, 128, 8192]", "3", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12688 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649793108801, "dur": 18, + "args": { + "External id": 12690,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[32, 128, 8192], [32, 128, 8192], []], "Ev Idx": 12689 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 435, "tid": 435, + "ts": 1776649793108822, "dur": 1, + "args": { + "External id": 12691,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[4096, 8192]"], "Input type": ["int", "ScalarList"], "Input Dims": [[32, 128, 8192], []], "Ev Idx": 12690 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649793108833, "dur": 3, + "args": { + "External id": 12692,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 64]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[1, 262144], []], "Ev Idx": 12691 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649793108834, "dur": 2, + "args": { + "External id": 12693,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 64]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[1, 262144], []], "Ev Idx": 12692 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649793108851, "dur": 12, + "args": { + "External id": 12694,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[4096, 64], [], [], [], []], "Ev Idx": 12693 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649793108859, "dur": 2, + "args": { + "External id": 12695,"Record function id": 0, "Concrete Inputs": ["", "[4096, 64]", "[64, 1]", "0"], "Input type": ["c10::Half", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[4096, 64], [], [], []], "Ev Idx": 12694 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793108933, "dur": 3, + "args": { + "External id": 12696,"Record function id": 0, "Concrete Inputs": ["[64]", "4", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12695 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649793108945, "dur": 1639, + "args": { + "External id": 12697,"Record function id": 0, "Concrete Inputs": ["", "", "4", "False", "False", ""], "Input type": ["long int", "", "Scalar", "Scalar", "Scalar", ""], "Input Dims": [[64], [], [], [], [], []], "Ev Idx": 12696 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649793108946, "dur": 1638, + "args": { + "External id": 12698,"Record function id": 0, "Concrete Inputs": ["", "4", "", "", "", "False", ""], "Input type": ["long int", "Scalar", "", "", "", "Scalar", ""], "Input Dims": [[64], [], [], [], [], [], []], "Ev Idx": 12697 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649793108948, "dur": 9, + "args": { + "External id": 12699,"Record function id": 0, "Concrete Inputs": ["[64]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12698 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649793108958, "dur": 1626, + "args": { + "External id": 12700,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Dims": [[64], [64], []], "Ev Idx": 12699 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 435, "tid": 435, + "ts": 1776649793110588, "dur": 1, + "args": { + "External id": 12701,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Dims": [[64]], "Ev Idx": 12700 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::index", "pid": 435, "tid": 435, + "ts": 1776649793110596, "dur": 52, + "args": { + "External id": 12702,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::Half", ""], "Input Dims": [[4096, 64], []], "Ev Idx": 12701 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649793110610, "dur": 1, + "args": { + "External id": 12703,"Record function id": 0, "Concrete Inputs": ["", "[4096, 64]", "[64, 0]", ""], "Input type": ["c10::Half", "ScalarList", "ScalarList", ""], "Input Dims": [[4096, 64], [], [], []], "Ev Idx": 12702 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649793110613, "dur": 2, + "args": { + "External id": 12704,"Record function id": 0, "Concrete Inputs": ["", "[1, 64]"], "Input type": ["long int", "ScalarList"], "Input Dims": [[64], []], "Ev Idx": 12703 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649793110614, "dur": 1, + "args": { + "External id": 12705,"Record function id": 0, "Concrete Inputs": ["", "[1, 64]"], "Input type": ["long int", "ScalarList"], "Input Dims": [[64], []], "Ev Idx": 12704 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649793110662, "dur": 3, + "args": { + "External id": 12706,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 8192]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 64], []], "Ev Idx": 12705 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649793110663, "dur": 2, + "args": { + "External id": 12707,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 8192]"], "Input type": ["c10::Half", "ScalarList"], "Input Dims": [[4096, 64], []], "Ev Idx": 12706 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649793110669, "dur": 3, + "args": { + "External id": 12708,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 16, 512, 16]"], "Input type": ["int", "ScalarList"], "Input Dims": [[4096, 8192], []], "Ev Idx": 12707 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649793110670, "dur": 2, + "args": { + "External id": 12709,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 16, 512, 16]"], "Input type": ["int", "ScalarList"], "Input Dims": [[4096, 8192], []], "Ev Idx": 12708 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 435, "tid": 435, + "ts": 1776649793110675, "dur": 5, + "args": { + "External id": 12710,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[0, 2, 1, 3]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 16, 512, 16], []], "Ev Idx": 12709 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649793110678, "dur": 1, + "args": { + "External id": 12711,"Record function id": 0, "Concrete Inputs": ["", "[256, 512, 16, 16]", "[131072, 16, 8192, 1]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Dims": [[256, 16, 512, 16], [], [], []], "Ev Idx": 12710 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649793110685, "dur": 35, + "args": { + "External id": 12712,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 131072]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 512, 16, 16], []], "Ev Idx": 12711 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 435, "tid": 435, + "ts": 1776649793110687, "dur": 30, + "args": { + "External id": 12713,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[256, 512, 16, 16], []], "Ev Idx": 12712 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 435, "tid": 435, + "ts": 1776649793110689, "dur": 9, + "args": { + "External id": 12714,"Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "0"], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[256, 512, 16, 16], [], [], [], [], []], "Ev Idx": 12713 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793110691, "dur": 7, + "args": { + "External id": 12715,"Record function id": 0, "Concrete Inputs": ["[256, 512, 16, 16]", "3", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12714 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649793110700, "dur": 16, + "args": { + "External id": 12716,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 512, 16, 16], [256, 512, 16, 16], []], "Ev Idx": 12715 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 435, "tid": 435, + "ts": 1776649793110719, "dur": 1, + "args": { + "External id": 12717,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 131072]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 512, 16, 16], []], "Ev Idx": 12716 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649793110730, "dur": 3, + "args": { + "External id": 12718,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 131072], []], "Ev Idx": 12717 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649793110731, "dur": 2, + "args": { + "External id": 12719,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 1024]"], "Input type": ["int", "ScalarList"], "Input Dims": [[256, 131072], []], "Ev Idx": 12718 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649793110738, "dur": 5, + "args": { + "External id": 12720,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[32768, 1024], [], [], [], []], "Ev Idx": 12719 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649793110741, "dur": 1, + "args": { + "External id": 12721,"Record function id": 0, "Concrete Inputs": ["", "[32768, 1024]", "[1024, 1]", "0"], "Input type": ["int", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[32768, 1024], [], [], []], "Ev Idx": 12720 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::index", "pid": 435, "tid": 435, + "ts": 1776649793110746, "dur": 231, + "args": { + "External id": 12722,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["int", ""], "Input Dims": [[32768, 1024], []], "Ev Idx": 12721 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649793110749, "dur": 196, + "args": { + "External id": 12723,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "False", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Dims": [[1024], [], [], [], [], [], [], []], "Ev Idx": 12722 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649793110752, "dur": 193, + "args": { + "External id": 12724,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Dims": [[1024], [], [], [], [], [], []], "Ev Idx": 12723 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649793110754, "dur": 6, + "args": { + "External id": 12725,"Record function id": 0, "Concrete Inputs": ["[1024]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12724 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649793110762, "dur": 182, + "args": { + "External id": 12726,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Dims": [[1024], [1024], []], "Ev Idx": 12725 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649793110948, "dur": 1, + "args": { + "External id": 12727,"Record function id": 0, "Concrete Inputs": ["", "[32768, 1024]", "[1024, 0]", ""], "Input type": ["int", "ScalarList", "ScalarList", ""], "Input Dims": [[32768, 1024], [], [], []], "Ev Idx": 12726 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649793110950, "dur": 2, + "args": { + "External id": 12728,"Record function id": 0, "Concrete Inputs": ["", "[1, 1024]"], "Input type": ["long int", "ScalarList"], "Input Dims": [[1024], []], "Ev Idx": 12727 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649793110951, "dur": 1, + "args": { + "External id": 12729,"Record function id": 0, "Concrete Inputs": ["", "[1, 1024]"], "Input type": ["long int", "ScalarList"], "Input Dims": [[1024], []], "Ev Idx": 12728 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 435, "tid": 435, + "ts": 1776649793110988, "dur": 3, + "args": { + "External id": 12730,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 131072]"], "Input type": ["int", "ScalarList"], "Input Dims": [[32768, 1024], []], "Ev Idx": 12729 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 435, "tid": 435, + "ts": 1776649793110989, "dur": 2, + "args": { + "External id": 12731,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[256, 131072]"], "Input type": ["int", "ScalarList"], "Input Dims": [[32768, 1024], []], "Ev Idx": 12730 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649793111950, "dur": 98108, + "args": { + "External id": 12732,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 131072], [], [], [], [], [], [], []], "Ev Idx": 12731 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649793111952, "dur": 98105, + "args": { + "External id": 12733,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[256, 131072], [], [], [], [], [], []], "Ev Idx": 12732 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649793111959, "dur": 79, + "args": { + "External id": 12734,"Record function id": 0, "Concrete Inputs": ["[256, 131072]", "[131072, 1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12733 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649793112041, "dur": 98013, + "args": { + "External id": 12735,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 131072], [256, 131072], []], "Ev Idx": 12734 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 435, "tid": 435, + "ts": 1776649793210085, "dur": 18, + "args": { + "External id": 12736,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 131072]], "Ev Idx": 12735 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 435, "tid": 435, + "ts": 1776649793210090, "dur": 13, + "args": { + "External id": 12737,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 131072]], "Ev Idx": 12736 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649793210105, "dur": 0, + "args": { + "External id": 12738,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 131072], [], [], [], [], [], [], []], "Ev Idx": 12737 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::resolve_conj", "pid": 435, "tid": 435, + "ts": 1776649793210108, "dur": 1, + "args": { + "External id": 12739,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 131072]], "Ev Idx": 12738 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::resolve_neg", "pid": 435, "tid": 435, + "ts": 1776649793210110, "dur": 1, + "args": { + "External id": 12740,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 131072]], "Ev Idx": 12739 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 435, "tid": 435, + "ts": 1776649793326639, "dur": 10, + "args": { + "External id": 12741,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[256, 16384]], "Ev Idx": 12740 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649793326680, "dur": 1507, + "args": { + "External id": 12742,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 16384], [], [], [], [], [], [], []], "Ev Idx": 12741 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 435, "tid": 435, + "ts": 1776649793326685, "dur": 1500, + "args": { + "External id": 12743,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Dims": [[256, 16384], [], [], [], [], [], []], "Ev Idx": 12742 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 435, "tid": 435, + "ts": 1776649793326707, "dur": 62, + "args": { + "External id": 12744,"Record function id": 0, "Concrete Inputs": ["[256, 16384]", "[16384, 1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12743 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649793326773, "dur": 1407, + "args": { + "External id": 12745,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 16384], [256, 16384], []], "Ev Idx": 12744 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649793328294, "dur": 1, + "args": { + "External id": 12746,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "0", "", "", "False", "False", ""], "Input type": ["int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[256, 16384], [], [], [], [], [], [], []], "Ev Idx": 12745 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649793328327, "dur": 32, + "args": { + "External id": 12747,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[256, 16384], [], [], [], []], "Ev Idx": 12746 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649793328346, "dur": 5, + "args": { + "External id": 12748,"Record function id": 0, "Concrete Inputs": ["", "[256, 16384]", "[16384, 1]", "0"], "Input type": ["int", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[256, 16384], [], [], []], "Ev Idx": 12747 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649793328361, "dur": 7, + "args": { + "External id": 12749,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[256, 16384], [], [], [], []], "Ev Idx": 12748 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649793328365, "dur": 1, + "args": { + "External id": 12750,"Record function id": 0, "Concrete Inputs": ["", "[256, 16384]", "[16384, 1]", "0"], "Input type": ["int", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[256, 16384], [], [], []], "Ev Idx": 12749 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649793328373, "dur": 76, + "args": { + "External id": 12751,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "int", "Scalar"], "Input Dims": [[256, 16384], [256, 16384], []], "Ev Idx": 12750 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 435, "tid": 435, + "ts": 1776649793328461, "dur": 1, + "args": { + "External id": 12752,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "5", "0", "", "", "False", "False", ""], "Input type": ["c10::Half", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Dims": [[32, 8192], [], [], [], [], [], [], []], "Ev Idx": 12751 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649793328467, "dur": 6, + "args": { + "External id": 12753,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[32, 8192], [], [], [], []], "Ev Idx": 12752 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649793328470, "dur": 1, + "args": { + "External id": 12754,"Record function id": 0, "Concrete Inputs": ["", "[32, 8192]", "[8192, 1]", "0"], "Input type": ["c10::Half", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 12753 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 435, "tid": 435, + "ts": 1776649793328474, "dur": 3, + "args": { + "External id": 12755,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::Half", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Dims": [[32, 8192], [], [], [], []], "Ev Idx": 12754 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 435, "tid": 435, + "ts": 1776649793328476, "dur": 0, + "args": { + "External id": 12756,"Record function id": 0, "Concrete Inputs": ["", "[32, 8192]", "[8192, 1]", "0"], "Input type": ["c10::Half", "ScalarList", "ScalarList", "Scalar"], "Input Dims": [[32, 8192], [], [], []], "Ev Idx": 12755 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 435, "tid": 435, + "ts": 1776649793328479, "dur": 22, + "args": { + "External id": 12757,"Sequence number": 0, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::Half", "c10::Half", "Scalar"], "Input Dims": [[32, 8192], [32, 8192], []], "Ev Idx": 12756 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793329634, "dur": 26, + "args": { + "External id": 12758,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12757 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793329672, "dur": 61, + "args": { + "External id": 12759,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12758 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793329676, "dur": 5, + "args": { + "External id": 12760,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12759 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793329685, "dur": 47, + "args": { + "External id": 12761,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12760 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793329691, "dur": 40, + "args": { + "External id": 12762,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12761 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793329776, "dur": 6, + "args": { + "External id": 12763,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12762 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793329786, "dur": 23, + "args": { + "External id": 12764,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12763 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793329787, "dur": 5, + "args": { + "External id": 12765,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12764 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793329793, "dur": 15, + "args": { + "External id": 12766,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12765 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793329795, "dur": 12, + "args": { + "External id": 12767,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12766 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793329828, "dur": 6, + "args": { + "External id": 12768,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12767 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793329838, "dur": 22, + "args": { + "External id": 12769,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12768 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793329839, "dur": 5, + "args": { + "External id": 12770,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12769 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793329845, "dur": 14, + "args": { + "External id": 12771,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12770 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793329847, "dur": 11, + "args": { + "External id": 12772,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12771 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793329879, "dur": 5, + "args": { + "External id": 12773,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12772 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793329888, "dur": 21, + "args": { + "External id": 12774,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12773 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793329889, "dur": 5, + "args": { + "External id": 12775,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12774 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793329895, "dur": 14, + "args": { + "External id": 12776,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12775 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793329897, "dur": 11, + "args": { + "External id": 12777,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12776 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793329927, "dur": 5, + "args": { + "External id": 12778,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12777 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793329937, "dur": 21, + "args": { + "External id": 12779,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12778 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793329938, "dur": 4, + "args": { + "External id": 12780,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12779 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793329944, "dur": 13, + "args": { + "External id": 12781,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12780 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793329945, "dur": 11, + "args": { + "External id": 12782,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12781 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793329976, "dur": 5, + "args": { + "External id": 12783,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12782 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793329985, "dur": 21, + "args": { + "External id": 12784,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12783 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793329987, "dur": 4, + "args": { + "External id": 12785,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12784 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793329992, "dur": 14, + "args": { + "External id": 12786,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12785 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793329994, "dur": 11, + "args": { + "External id": 12787,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12786 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330024, "dur": 5, + "args": { + "External id": 12788,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12787 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793330033, "dur": 22, + "args": { + "External id": 12789,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12788 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330035, "dur": 4, + "args": { + "External id": 12790,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12789 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793330040, "dur": 15, + "args": { + "External id": 12791,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12790 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793330042, "dur": 11, + "args": { + "External id": 12792,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12791 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330073, "dur": 6, + "args": { + "External id": 12793,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12792 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793330082, "dur": 21, + "args": { + "External id": 12794,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12793 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330084, "dur": 4, + "args": { + "External id": 12795,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12794 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793330089, "dur": 14, + "args": { + "External id": 12796,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12795 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793330091, "dur": 11, + "args": { + "External id": 12797,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12796 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330121, "dur": 5, + "args": { + "External id": 12798,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12797 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793330130, "dur": 34, + "args": { + "External id": 12799,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12798 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330132, "dur": 4, + "args": { + "External id": 12800,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12799 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793330148, "dur": 15, + "args": { + "External id": 12801,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12800 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793330150, "dur": 12, + "args": { + "External id": 12802,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12801 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330182, "dur": 5, + "args": { + "External id": 12803,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12802 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793330191, "dur": 21, + "args": { + "External id": 12804,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12803 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330193, "dur": 4, + "args": { + "External id": 12805,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12804 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793330198, "dur": 14, + "args": { + "External id": 12806,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12805 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793330200, "dur": 11, + "args": { + "External id": 12807,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12806 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330335, "dur": 7, + "args": { + "External id": 12808,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12807 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793330347, "dur": 25, + "args": { + "External id": 12809,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12808 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330348, "dur": 5, + "args": { + "External id": 12810,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12809 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793330354, "dur": 17, + "args": { + "External id": 12811,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12810 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793330356, "dur": 14, + "args": { + "External id": 12812,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12811 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330392, "dur": 5, + "args": { + "External id": 12813,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12812 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793330402, "dur": 21, + "args": { + "External id": 12814,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12813 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330403, "dur": 4, + "args": { + "External id": 12815,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12814 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793330409, "dur": 14, + "args": { + "External id": 12816,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12815 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793330410, "dur": 12, + "args": { + "External id": 12817,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12816 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330442, "dur": 5, + "args": { + "External id": 12818,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12817 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793330451, "dur": 21, + "args": { + "External id": 12819,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12818 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330452, "dur": 4, + "args": { + "External id": 12820,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12819 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793330458, "dur": 14, + "args": { + "External id": 12821,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12820 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793330459, "dur": 11, + "args": { + "External id": 12822,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12821 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330489, "dur": 6, + "args": { + "External id": 12823,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12822 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793330499, "dur": 21, + "args": { + "External id": 12824,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12823 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330500, "dur": 4, + "args": { + "External id": 12825,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12824 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793330505, "dur": 14, + "args": { + "External id": 12826,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12825 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793330507, "dur": 11, + "args": { + "External id": 12827,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12826 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330537, "dur": 5, + "args": { + "External id": 12828,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12827 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793330546, "dur": 21, + "args": { + "External id": 12829,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12828 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330548, "dur": 4, + "args": { + "External id": 12830,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12829 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793330553, "dur": 14, + "args": { + "External id": 12831,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12830 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793330555, "dur": 11, + "args": { + "External id": 12832,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12831 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330585, "dur": 5, + "args": { + "External id": 12833,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12832 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793330594, "dur": 22, + "args": { + "External id": 12834,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12833 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330595, "dur": 5, + "args": { + "External id": 12835,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12834 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793330601, "dur": 15, + "args": { + "External id": 12836,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12835 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793330603, "dur": 11, + "args": { + "External id": 12837,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12836 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330636, "dur": 6, + "args": { + "External id": 12838,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12837 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793330646, "dur": 23, + "args": { + "External id": 12839,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12838 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330649, "dur": 4, + "args": { + "External id": 12840,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12839 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793330655, "dur": 13, + "args": { + "External id": 12841,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12840 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793330656, "dur": 11, + "args": { + "External id": 12842,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12841 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330686, "dur": 5, + "args": { + "External id": 12843,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12842 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793330695, "dur": 21, + "args": { + "External id": 12844,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12843 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330696, "dur": 5, + "args": { + "External id": 12845,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12844 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793330702, "dur": 13, + "args": { + "External id": 12846,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12845 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793330703, "dur": 11, + "args": { + "External id": 12847,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12846 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330733, "dur": 6, + "args": { + "External id": 12848,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12847 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793330742, "dur": 21, + "args": { + "External id": 12849,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12848 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330744, "dur": 4, + "args": { + "External id": 12850,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12849 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793330749, "dur": 14, + "args": { + "External id": 12851,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12850 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793330751, "dur": 11, + "args": { + "External id": 12852,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12851 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330781, "dur": 5, + "args": { + "External id": 12853,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12852 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793330790, "dur": 20, + "args": { + "External id": 12854,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12853 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330791, "dur": 4, + "args": { + "External id": 12855,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12854 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793330796, "dur": 14, + "args": { + "External id": 12856,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12855 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793330798, "dur": 11, + "args": { + "External id": 12857,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12856 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330828, "dur": 5, + "args": { + "External id": 12858,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12857 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793330837, "dur": 21, + "args": { + "External id": 12859,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12858 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330838, "dur": 5, + "args": { + "External id": 12860,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12859 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793330844, "dur": 13, + "args": { + "External id": 12861,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12860 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793330846, "dur": 10, + "args": { + "External id": 12862,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12861 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330876, "dur": 5, + "args": { + "External id": 12863,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12862 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793330885, "dur": 21, + "args": { + "External id": 12864,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12863 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330887, "dur": 4, + "args": { + "External id": 12865,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12864 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793330892, "dur": 14, + "args": { + "External id": 12866,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12865 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793330894, "dur": 11, + "args": { + "External id": 12867,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12866 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330924, "dur": 5, + "args": { + "External id": 12868,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12867 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793330933, "dur": 21, + "args": { + "External id": 12869,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12868 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330934, "dur": 5, + "args": { + "External id": 12870,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12869 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793330940, "dur": 14, + "args": { + "External id": 12871,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12870 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793330942, "dur": 11, + "args": { + "External id": 12872,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12871 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330972, "dur": 5, + "args": { + "External id": 12873,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12872 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793330981, "dur": 21, + "args": { + "External id": 12874,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12873 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793330982, "dur": 4, + "args": { + "External id": 12875,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12874 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793330988, "dur": 13, + "args": { + "External id": 12876,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12875 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793330989, "dur": 11, + "args": { + "External id": 12877,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12876 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331020, "dur": 5, + "args": { + "External id": 12878,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12877 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793331029, "dur": 20, + "args": { + "External id": 12879,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12878 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331030, "dur": 4, + "args": { + "External id": 12880,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12879 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793331035, "dur": 14, + "args": { + "External id": 12881,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12880 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793331037, "dur": 11, + "args": { + "External id": 12882,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12881 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331067, "dur": 5, + "args": { + "External id": 12883,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12882 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793331076, "dur": 21, + "args": { + "External id": 12884,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12883 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331077, "dur": 5, + "args": { + "External id": 12885,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12884 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793331083, "dur": 14, + "args": { + "External id": 12886,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12885 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793331085, "dur": 10, + "args": { + "External id": 12887,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12886 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331115, "dur": 5, + "args": { + "External id": 12888,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12887 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793331124, "dur": 21, + "args": { + "External id": 12889,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12888 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331125, "dur": 5, + "args": { + "External id": 12890,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12889 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793331131, "dur": 13, + "args": { + "External id": 12891,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12890 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793331132, "dur": 11, + "args": { + "External id": 12892,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12891 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331162, "dur": 5, + "args": { + "External id": 12893,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12892 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793331171, "dur": 21, + "args": { + "External id": 12894,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12893 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331172, "dur": 5, + "args": { + "External id": 12895,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12894 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793331178, "dur": 13, + "args": { + "External id": 12896,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12895 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793331179, "dur": 11, + "args": { + "External id": 12897,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12896 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331209, "dur": 5, + "args": { + "External id": 12898,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12897 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793331218, "dur": 21, + "args": { + "External id": 12899,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12898 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331219, "dur": 5, + "args": { + "External id": 12900,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12899 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793331225, "dur": 13, + "args": { + "External id": 12901,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12900 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793331226, "dur": 11, + "args": { + "External id": 12902,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12901 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331256, "dur": 6, + "args": { + "External id": 12903,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12902 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793331265, "dur": 21, + "args": { + "External id": 12904,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12903 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331266, "dur": 5, + "args": { + "External id": 12905,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12904 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793331272, "dur": 14, + "args": { + "External id": 12906,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12905 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793331274, "dur": 11, + "args": { + "External id": 12907,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12906 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331304, "dur": 5, + "args": { + "External id": 12908,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12907 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793331313, "dur": 21, + "args": { + "External id": 12909,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12908 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331314, "dur": 4, + "args": { + "External id": 12910,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12909 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793331320, "dur": 13, + "args": { + "External id": 12911,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12910 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793331321, "dur": 11, + "args": { + "External id": 12912,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12911 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331351, "dur": 5, + "args": { + "External id": 12913,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12912 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793331360, "dur": 21, + "args": { + "External id": 12914,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12913 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331361, "dur": 4, + "args": { + "External id": 12915,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12914 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793331367, "dur": 13, + "args": { + "External id": 12916,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12915 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793331368, "dur": 11, + "args": { + "External id": 12917,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12916 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331398, "dur": 6, + "args": { + "External id": 12918,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12917 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793331407, "dur": 21, + "args": { + "External id": 12919,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12918 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331409, "dur": 4, + "args": { + "External id": 12920,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12919 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793331414, "dur": 14, + "args": { + "External id": 12921,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12920 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793331416, "dur": 11, + "args": { + "External id": 12922,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12921 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331446, "dur": 5, + "args": { + "External id": 12923,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12922 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793331455, "dur": 21, + "args": { + "External id": 12924,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12923 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331456, "dur": 5, + "args": { + "External id": 12925,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12924 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793331462, "dur": 14, + "args": { + "External id": 12926,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12925 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793331464, "dur": 10, + "args": { + "External id": 12927,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12926 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331494, "dur": 5, + "args": { + "External id": 12928,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12927 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793331503, "dur": 20, + "args": { + "External id": 12929,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12928 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331504, "dur": 4, + "args": { + "External id": 12930,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12929 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793331509, "dur": 14, + "args": { + "External id": 12931,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12930 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793331511, "dur": 11, + "args": { + "External id": 12932,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12931 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331541, "dur": 5, + "args": { + "External id": 12933,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12932 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793331550, "dur": 21, + "args": { + "External id": 12934,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12933 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331551, "dur": 5, + "args": { + "External id": 12935,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12934 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793331557, "dur": 13, + "args": { + "External id": 12936,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12935 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793331559, "dur": 10, + "args": { + "External id": 12937,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12936 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331588, "dur": 5, + "args": { + "External id": 12938,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12937 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793331597, "dur": 21, + "args": { + "External id": 12939,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12938 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331598, "dur": 5, + "args": { + "External id": 12940,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12939 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793331604, "dur": 14, + "args": { + "External id": 12941,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12940 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793331606, "dur": 10, + "args": { + "External id": 12942,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12941 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331638, "dur": 6, + "args": { + "External id": 12943,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12942 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793331648, "dur": 23, + "args": { + "External id": 12944,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12943 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331650, "dur": 5, + "args": { + "External id": 12945,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12944 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793331656, "dur": 15, + "args": { + "External id": 12946,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12945 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793331659, "dur": 10, + "args": { + "External id": 12947,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12946 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331688, "dur": 6, + "args": { + "External id": 12948,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12947 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793331697, "dur": 21, + "args": { + "External id": 12949,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12948 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331699, "dur": 4, + "args": { + "External id": 12950,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12949 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793331704, "dur": 14, + "args": { + "External id": 12951,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12950 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793331706, "dur": 11, + "args": { + "External id": 12952,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12951 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331736, "dur": 5, + "args": { + "External id": 12953,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12952 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793331745, "dur": 21, + "args": { + "External id": 12954,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12953 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331746, "dur": 4, + "args": { + "External id": 12955,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12954 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793331752, "dur": 13, + "args": { + "External id": 12956,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12955 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793331753, "dur": 11, + "args": { + "External id": 12957,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12956 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331783, "dur": 5, + "args": { + "External id": 12958,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12957 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793331792, "dur": 21, + "args": { + "External id": 12959,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12958 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331793, "dur": 5, + "args": { + "External id": 12960,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12959 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793331799, "dur": 14, + "args": { + "External id": 12961,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12960 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793331800, "dur": 11, + "args": { + "External id": 12962,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12961 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331830, "dur": 5, + "args": { + "External id": 12963,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12962 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793331839, "dur": 21, + "args": { + "External id": 12964,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12963 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331840, "dur": 5, + "args": { + "External id": 12965,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12964 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793331846, "dur": 13, + "args": { + "External id": 12966,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12965 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793331847, "dur": 11, + "args": { + "External id": 12967,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12966 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331877, "dur": 5, + "args": { + "External id": 12968,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12967 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793331886, "dur": 21, + "args": { + "External id": 12969,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12968 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331887, "dur": 5, + "args": { + "External id": 12970,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12969 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793331893, "dur": 14, + "args": { + "External id": 12971,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12970 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793331895, "dur": 10, + "args": { + "External id": 12972,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12971 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331925, "dur": 5, + "args": { + "External id": 12973,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12972 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793331934, "dur": 21, + "args": { + "External id": 12974,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12973 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331935, "dur": 5, + "args": { + "External id": 12975,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12974 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793331941, "dur": 13, + "args": { + "External id": 12976,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12975 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793331943, "dur": 10, + "args": { + "External id": 12977,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12976 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331973, "dur": 5, + "args": { + "External id": 12978,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12977 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793331982, "dur": 20, + "args": { + "External id": 12979,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12978 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793331983, "dur": 4, + "args": { + "External id": 12980,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12979 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793331988, "dur": 14, + "args": { + "External id": 12981,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12980 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793331990, "dur": 11, + "args": { + "External id": 12982,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12981 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332020, "dur": 5, + "args": { + "External id": 12983,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12982 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793332029, "dur": 21, + "args": { + "External id": 12984,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12983 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332030, "dur": 4, + "args": { + "External id": 12985,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12984 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793332036, "dur": 13, + "args": { + "External id": 12986,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12985 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793332037, "dur": 11, + "args": { + "External id": 12987,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12986 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332067, "dur": 5, + "args": { + "External id": 12988,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12987 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793332076, "dur": 21, + "args": { + "External id": 12989,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12988 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332077, "dur": 5, + "args": { + "External id": 12990,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12989 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793332083, "dur": 14, + "args": { + "External id": 12991,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12990 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793332085, "dur": 11, + "args": { + "External id": 12992,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12991 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332115, "dur": 5, + "args": { + "External id": 12993,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12992 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793332124, "dur": 20, + "args": { + "External id": 12994,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12993 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332125, "dur": 4, + "args": { + "External id": 12995,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12994 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793332130, "dur": 14, + "args": { + "External id": 12996,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 12995 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793332132, "dur": 11, + "args": { + "External id": 12997,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 12996 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332162, "dur": 5, + "args": { + "External id": 12998,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12997 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793332171, "dur": 21, + "args": { + "External id": 12999,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 12998 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332172, "dur": 4, + "args": { + "External id": 13000,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12999 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793332178, "dur": 13, + "args": { + "External id": 13001,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13000 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793332179, "dur": 11, + "args": { + "External id": 13002,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13001 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332210, "dur": 5, + "args": { + "External id": 13003,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13002 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793332240, "dur": 24, + "args": { + "External id": 13004,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13003 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332241, "dur": 5, + "args": { + "External id": 13005,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13004 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793332248, "dur": 15, + "args": { + "External id": 13006,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13005 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793332250, "dur": 12, + "args": { + "External id": 13007,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13006 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332283, "dur": 6, + "args": { + "External id": 13008,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13007 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793332293, "dur": 21, + "args": { + "External id": 13009,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13008 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332294, "dur": 4, + "args": { + "External id": 13010,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13009 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793332299, "dur": 15, + "args": { + "External id": 13011,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13010 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793332301, "dur": 11, + "args": { + "External id": 13012,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13011 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332332, "dur": 5, + "args": { + "External id": 13013,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13012 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793332341, "dur": 21, + "args": { + "External id": 13014,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13013 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332342, "dur": 5, + "args": { + "External id": 13015,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13014 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793332348, "dur": 14, + "args": { + "External id": 13016,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13015 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793332349, "dur": 11, + "args": { + "External id": 13017,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13016 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332380, "dur": 5, + "args": { + "External id": 13018,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13017 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793332389, "dur": 21, + "args": { + "External id": 13019,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13018 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332390, "dur": 4, + "args": { + "External id": 13020,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13019 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793332396, "dur": 13, + "args": { + "External id": 13021,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13020 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793332397, "dur": 11, + "args": { + "External id": 13022,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13021 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332427, "dur": 5, + "args": { + "External id": 13023,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13022 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793332436, "dur": 21, + "args": { + "External id": 13024,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13023 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332437, "dur": 5, + "args": { + "External id": 13025,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13024 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793332443, "dur": 13, + "args": { + "External id": 13026,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13025 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793332445, "dur": 10, + "args": { + "External id": 13027,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13026 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332474, "dur": 5, + "args": { + "External id": 13028,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13027 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793332483, "dur": 27, + "args": { + "External id": 13029,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13028 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332485, "dur": 4, + "args": { + "External id": 13030,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13029 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793332490, "dur": 20, + "args": { + "External id": 13031,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13030 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793332492, "dur": 17, + "args": { + "External id": 13032,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13031 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332530, "dur": 5, + "args": { + "External id": 13033,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13032 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793332539, "dur": 21, + "args": { + "External id": 13034,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13033 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332540, "dur": 5, + "args": { + "External id": 13035,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13034 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793332546, "dur": 14, + "args": { + "External id": 13036,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13035 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793332548, "dur": 11, + "args": { + "External id": 13037,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13036 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332578, "dur": 5, + "args": { + "External id": 13038,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13037 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793332587, "dur": 22, + "args": { + "External id": 13039,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13038 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332588, "dur": 5, + "args": { + "External id": 13040,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13039 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793332594, "dur": 14, + "args": { + "External id": 13041,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13040 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793332596, "dur": 11, + "args": { + "External id": 13042,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13041 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332626, "dur": 5, + "args": { + "External id": 13043,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13042 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793332635, "dur": 21, + "args": { + "External id": 13044,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13043 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332636, "dur": 5, + "args": { + "External id": 13045,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13044 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793332642, "dur": 14, + "args": { + "External id": 13046,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13045 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793332644, "dur": 11, + "args": { + "External id": 13047,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13046 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332674, "dur": 5, + "args": { + "External id": 13048,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13047 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793332683, "dur": 21, + "args": { + "External id": 13049,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13048 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332684, "dur": 5, + "args": { + "External id": 13050,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13049 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793332690, "dur": 14, + "args": { + "External id": 13051,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13050 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793332692, "dur": 11, + "args": { + "External id": 13052,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13051 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332722, "dur": 5, + "args": { + "External id": 13053,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13052 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793332731, "dur": 21, + "args": { + "External id": 13054,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13053 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332732, "dur": 4, + "args": { + "External id": 13055,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13054 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793332738, "dur": 13, + "args": { + "External id": 13056,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13055 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793332739, "dur": 11, + "args": { + "External id": 13057,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13056 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332836, "dur": 7, + "args": { + "External id": 13058,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13057 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793332848, "dur": 25, + "args": { + "External id": 13059,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13058 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332849, "dur": 5, + "args": { + "External id": 13060,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13059 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793332855, "dur": 17, + "args": { + "External id": 13061,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13060 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793332857, "dur": 14, + "args": { + "External id": 13062,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13061 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332894, "dur": 5, + "args": { + "External id": 13063,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13062 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793332903, "dur": 21, + "args": { + "External id": 13064,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13063 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332904, "dur": 5, + "args": { + "External id": 13065,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13064 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793332910, "dur": 14, + "args": { + "External id": 13066,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13065 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793332912, "dur": 11, + "args": { + "External id": 13067,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13066 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332942, "dur": 6, + "args": { + "External id": 13068,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13067 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793332952, "dur": 21, + "args": { + "External id": 13069,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13068 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332953, "dur": 4, + "args": { + "External id": 13070,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13069 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793332958, "dur": 15, + "args": { + "External id": 13071,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13070 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793332960, "dur": 11, + "args": { + "External id": 13072,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13071 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793332991, "dur": 5, + "args": { + "External id": 13073,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13072 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793333000, "dur": 21, + "args": { + "External id": 13074,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13073 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333001, "dur": 5, + "args": { + "External id": 13075,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13074 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793333007, "dur": 14, + "args": { + "External id": 13076,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13075 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793333008, "dur": 11, + "args": { + "External id": 13077,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13076 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333039, "dur": 5, + "args": { + "External id": 13078,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13077 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793333048, "dur": 20, + "args": { + "External id": 13079,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13078 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333049, "dur": 4, + "args": { + "External id": 13080,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13079 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793333054, "dur": 14, + "args": { + "External id": 13081,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13080 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793333056, "dur": 11, + "args": { + "External id": 13082,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13081 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333086, "dur": 5, + "args": { + "External id": 13083,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13082 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793333095, "dur": 21, + "args": { + "External id": 13084,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13083 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333096, "dur": 5, + "args": { + "External id": 13085,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13084 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793333102, "dur": 14, + "args": { + "External id": 13086,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13085 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793333104, "dur": 10, + "args": { + "External id": 13087,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13086 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333134, "dur": 5, + "args": { + "External id": 13088,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13087 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793333143, "dur": 20, + "args": { + "External id": 13089,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13088 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333144, "dur": 4, + "args": { + "External id": 13090,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13089 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793333149, "dur": 14, + "args": { + "External id": 13091,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13090 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793333151, "dur": 11, + "args": { + "External id": 13092,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13091 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333181, "dur": 5, + "args": { + "External id": 13093,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13092 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793333190, "dur": 20, + "args": { + "External id": 13094,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13093 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333191, "dur": 4, + "args": { + "External id": 13095,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13094 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793333197, "dur": 13, + "args": { + "External id": 13096,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13095 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793333198, "dur": 11, + "args": { + "External id": 13097,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13096 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333228, "dur": 5, + "args": { + "External id": 13098,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13097 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793333237, "dur": 21, + "args": { + "External id": 13099,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13098 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333238, "dur": 4, + "args": { + "External id": 13100,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13099 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793333243, "dur": 14, + "args": { + "External id": 13101,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13100 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793333245, "dur": 11, + "args": { + "External id": 13102,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13101 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333275, "dur": 5, + "args": { + "External id": 13103,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13102 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793333284, "dur": 21, + "args": { + "External id": 13104,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13103 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333285, "dur": 5, + "args": { + "External id": 13105,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13104 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793333291, "dur": 14, + "args": { + "External id": 13106,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13105 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793333293, "dur": 10, + "args": { + "External id": 13107,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13106 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333375, "dur": 6, + "args": { + "External id": 13108,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13107 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793333386, "dur": 25, + "args": { + "External id": 13109,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13108 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333387, "dur": 5, + "args": { + "External id": 13110,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13109 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793333393, "dur": 17, + "args": { + "External id": 13111,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13110 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793333395, "dur": 14, + "args": { + "External id": 13112,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13111 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333431, "dur": 5, + "args": { + "External id": 13113,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13112 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793333440, "dur": 22, + "args": { + "External id": 13114,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13113 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333442, "dur": 4, + "args": { + "External id": 13115,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13114 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793333447, "dur": 14, + "args": { + "External id": 13116,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13115 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793333449, "dur": 11, + "args": { + "External id": 13117,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13116 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333480, "dur": 5, + "args": { + "External id": 13118,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13117 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793333489, "dur": 21, + "args": { + "External id": 13119,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13118 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333490, "dur": 4, + "args": { + "External id": 13120,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13119 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793333496, "dur": 13, + "args": { + "External id": 13121,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13120 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793333497, "dur": 11, + "args": { + "External id": 13122,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13121 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333527, "dur": 5, + "args": { + "External id": 13123,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13122 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793333536, "dur": 21, + "args": { + "External id": 13124,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13123 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333537, "dur": 5, + "args": { + "External id": 13125,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13124 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793333543, "dur": 14, + "args": { + "External id": 13126,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13125 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793333545, "dur": 11, + "args": { + "External id": 13127,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13126 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333574, "dur": 6, + "args": { + "External id": 13128,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13127 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793333584, "dur": 20, + "args": { + "External id": 13129,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13128 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333585, "dur": 4, + "args": { + "External id": 13130,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13129 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793333590, "dur": 14, + "args": { + "External id": 13131,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13130 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793333592, "dur": 11, + "args": { + "External id": 13132,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13131 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333622, "dur": 5, + "args": { + "External id": 13133,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13132 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793333631, "dur": 21, + "args": { + "External id": 13134,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13133 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333632, "dur": 5, + "args": { + "External id": 13135,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13134 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793333638, "dur": 14, + "args": { + "External id": 13136,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13135 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793333640, "dur": 10, + "args": { + "External id": 13137,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13136 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333669, "dur": 5, + "args": { + "External id": 13138,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13137 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793333678, "dur": 22, + "args": { + "External id": 13139,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13138 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333680, "dur": 4, + "args": { + "External id": 13140,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13139 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793333685, "dur": 14, + "args": { + "External id": 13141,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13140 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793333687, "dur": 11, + "args": { + "External id": 13142,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13141 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333717, "dur": 5, + "args": { + "External id": 13143,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13142 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793333726, "dur": 21, + "args": { + "External id": 13144,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13143 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333727, "dur": 5, + "args": { + "External id": 13145,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13144 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793333733, "dur": 13, + "args": { + "External id": 13146,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13145 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793333735, "dur": 10, + "args": { + "External id": 13147,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13146 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333764, "dur": 5, + "args": { + "External id": 13148,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13147 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793333773, "dur": 22, + "args": { + "External id": 13149,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13148 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333774, "dur": 5, + "args": { + "External id": 13150,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13149 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793333780, "dur": 14, + "args": { + "External id": 13151,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13150 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793333782, "dur": 11, + "args": { + "External id": 13152,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13151 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333813, "dur": 5, + "args": { + "External id": 13153,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13152 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793333822, "dur": 20, + "args": { + "External id": 13154,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13153 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333823, "dur": 4, + "args": { + "External id": 13155,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13154 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793333829, "dur": 13, + "args": { + "External id": 13156,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13155 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793333830, "dur": 11, + "args": { + "External id": 13157,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13156 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333860, "dur": 5, + "args": { + "External id": 13158,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13157 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793333869, "dur": 21, + "args": { + "External id": 13159,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13158 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333870, "dur": 4, + "args": { + "External id": 13160,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13159 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793333876, "dur": 13, + "args": { + "External id": 13161,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13160 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793333877, "dur": 11, + "args": { + "External id": 13162,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13161 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333907, "dur": 6, + "args": { + "External id": 13163,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13162 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793333916, "dur": 21, + "args": { + "External id": 13164,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13163 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333918, "dur": 4, + "args": { + "External id": 13165,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13164 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793333923, "dur": 14, + "args": { + "External id": 13166,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13165 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793333925, "dur": 10, + "args": { + "External id": 13167,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13166 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333955, "dur": 5, + "args": { + "External id": 13168,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13167 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793333964, "dur": 21, + "args": { + "External id": 13169,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13168 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793333965, "dur": 4, + "args": { + "External id": 13170,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13169 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793333971, "dur": 13, + "args": { + "External id": 13171,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13170 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793333972, "dur": 11, + "args": { + "External id": 13172,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13171 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334002, "dur": 5, + "args": { + "External id": 13173,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13172 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793334011, "dur": 21, + "args": { + "External id": 13174,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13173 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334013, "dur": 4, + "args": { + "External id": 13175,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13174 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793334018, "dur": 14, + "args": { + "External id": 13176,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13175 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793334020, "dur": 10, + "args": { + "External id": 13177,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13176 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334049, "dur": 6, + "args": { + "External id": 13178,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13177 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793334058, "dur": 21, + "args": { + "External id": 13179,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13178 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334060, "dur": 4, + "args": { + "External id": 13180,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13179 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793334065, "dur": 14, + "args": { + "External id": 13181,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13180 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793334067, "dur": 11, + "args": { + "External id": 13182,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13181 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334097, "dur": 5, + "args": { + "External id": 13183,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13182 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793334106, "dur": 22, + "args": { + "External id": 13184,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13183 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334107, "dur": 5, + "args": { + "External id": 13185,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13184 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793334113, "dur": 14, + "args": { + "External id": 13186,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13185 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793334115, "dur": 11, + "args": { + "External id": 13187,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13186 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334145, "dur": 5, + "args": { + "External id": 13188,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13187 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793334154, "dur": 21, + "args": { + "External id": 13189,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13188 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334155, "dur": 5, + "args": { + "External id": 13190,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13189 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793334161, "dur": 13, + "args": { + "External id": 13191,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13190 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793334163, "dur": 10, + "args": { + "External id": 13192,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13191 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334192, "dur": 5, + "args": { + "External id": 13193,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13192 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793334201, "dur": 21, + "args": { + "External id": 13194,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13193 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334202, "dur": 5, + "args": { + "External id": 13195,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13194 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793334208, "dur": 14, + "args": { + "External id": 13196,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13195 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793334210, "dur": 10, + "args": { + "External id": 13197,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13196 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334239, "dur": 6, + "args": { + "External id": 13198,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13197 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793334248, "dur": 21, + "args": { + "External id": 13199,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13198 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334250, "dur": 4, + "args": { + "External id": 13200,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13199 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793334255, "dur": 14, + "args": { + "External id": 13201,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13200 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793334257, "dur": 10, + "args": { + "External id": 13202,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13201 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334286, "dur": 6, + "args": { + "External id": 13203,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13202 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793334295, "dur": 21, + "args": { + "External id": 13204,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13203 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334297, "dur": 4, + "args": { + "External id": 13205,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13204 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793334302, "dur": 14, + "args": { + "External id": 13206,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13205 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793334304, "dur": 11, + "args": { + "External id": 13207,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13206 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334334, "dur": 5, + "args": { + "External id": 13208,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13207 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793334343, "dur": 20, + "args": { + "External id": 13209,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13208 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334344, "dur": 4, + "args": { + "External id": 13210,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13209 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793334349, "dur": 14, + "args": { + "External id": 13211,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13210 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793334351, "dur": 11, + "args": { + "External id": 13212,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13211 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334381, "dur": 5, + "args": { + "External id": 13213,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13212 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793334390, "dur": 21, + "args": { + "External id": 13214,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13213 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334391, "dur": 4, + "args": { + "External id": 13215,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13214 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793334397, "dur": 13, + "args": { + "External id": 13216,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13215 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793334398, "dur": 11, + "args": { + "External id": 13217,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13216 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334428, "dur": 5, + "args": { + "External id": 13218,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13217 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793334437, "dur": 21, + "args": { + "External id": 13219,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13218 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334438, "dur": 5, + "args": { + "External id": 13220,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13219 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793334444, "dur": 13, + "args": { + "External id": 13221,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13220 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793334445, "dur": 11, + "args": { + "External id": 13222,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13221 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334475, "dur": 6, + "args": { + "External id": 13223,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13222 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793334484, "dur": 21, + "args": { + "External id": 13224,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13223 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334486, "dur": 4, + "args": { + "External id": 13225,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13224 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793334491, "dur": 14, + "args": { + "External id": 13226,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13225 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793334493, "dur": 11, + "args": { + "External id": 13227,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13226 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334523, "dur": 5, + "args": { + "External id": 13228,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13227 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793334532, "dur": 21, + "args": { + "External id": 13229,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13228 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334533, "dur": 5, + "args": { + "External id": 13230,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13229 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793334539, "dur": 14, + "args": { + "External id": 13231,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13230 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793334541, "dur": 10, + "args": { + "External id": 13232,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13231 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334570, "dur": 5, + "args": { + "External id": 13233,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13232 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793334579, "dur": 21, + "args": { + "External id": 13234,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13233 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334580, "dur": 5, + "args": { + "External id": 13235,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13234 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793334586, "dur": 14, + "args": { + "External id": 13236,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13235 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793334588, "dur": 10, + "args": { + "External id": 13237,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13236 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334617, "dur": 5, + "args": { + "External id": 13238,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13237 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793334626, "dur": 24, + "args": { + "External id": 13239,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13238 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334628, "dur": 6, + "args": { + "External id": 13240,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13239 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793334635, "dur": 15, + "args": { + "External id": 13241,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13240 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793334637, "dur": 11, + "args": { + "External id": 13242,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13241 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334669, "dur": 5, + "args": { + "External id": 13243,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13242 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793334678, "dur": 21, + "args": { + "External id": 13244,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13243 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334679, "dur": 4, + "args": { + "External id": 13245,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13244 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793334685, "dur": 13, + "args": { + "External id": 13246,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13245 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793334686, "dur": 11, + "args": { + "External id": 13247,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13246 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334716, "dur": 6, + "args": { + "External id": 13248,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13247 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793334725, "dur": 22, + "args": { + "External id": 13249,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13248 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334727, "dur": 4, + "args": { + "External id": 13250,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13249 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793334732, "dur": 14, + "args": { + "External id": 13251,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13250 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793334734, "dur": 11, + "args": { + "External id": 13252,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13251 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334764, "dur": 5, + "args": { + "External id": 13253,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13252 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793334773, "dur": 21, + "args": { + "External id": 13254,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13253 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334775, "dur": 4, + "args": { + "External id": 13255,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13254 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793334780, "dur": 14, + "args": { + "External id": 13256,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13255 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793334782, "dur": 11, + "args": { + "External id": 13257,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13256 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334812, "dur": 5, + "args": { + "External id": 13258,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13257 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793334821, "dur": 21, + "args": { + "External id": 13259,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13258 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334822, "dur": 5, + "args": { + "External id": 13260,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13259 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793334828, "dur": 14, + "args": { + "External id": 13261,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13260 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793334830, "dur": 11, + "args": { + "External id": 13262,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13261 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334860, "dur": 5, + "args": { + "External id": 13263,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13262 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793334869, "dur": 21, + "args": { + "External id": 13264,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13263 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334870, "dur": 5, + "args": { + "External id": 13265,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13264 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793334876, "dur": 14, + "args": { + "External id": 13266,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13265 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793334878, "dur": 11, + "args": { + "External id": 13267,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13266 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334908, "dur": 5, + "args": { + "External id": 13268,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13267 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793334917, "dur": 21, + "args": { + "External id": 13269,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13268 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334918, "dur": 5, + "args": { + "External id": 13270,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13269 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793334924, "dur": 14, + "args": { + "External id": 13271,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13270 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793334926, "dur": 11, + "args": { + "External id": 13272,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13271 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334957, "dur": 5, + "args": { + "External id": 13273,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13272 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793334966, "dur": 21, + "args": { + "External id": 13274,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13273 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793334967, "dur": 4, + "args": { + "External id": 13275,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13274 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793334972, "dur": 14, + "args": { + "External id": 13276,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13275 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793334974, "dur": 11, + "args": { + "External id": 13277,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13276 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335005, "dur": 5, + "args": { + "External id": 13278,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13277 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793335014, "dur": 21, + "args": { + "External id": 13279,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13278 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335015, "dur": 4, + "args": { + "External id": 13280,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13279 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793335020, "dur": 14, + "args": { + "External id": 13281,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13280 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793335022, "dur": 11, + "args": { + "External id": 13282,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13281 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335053, "dur": 5, + "args": { + "External id": 13283,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13282 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793335062, "dur": 21, + "args": { + "External id": 13284,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13283 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335063, "dur": 4, + "args": { + "External id": 13285,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13284 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793335068, "dur": 14, + "args": { + "External id": 13286,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13285 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793335070, "dur": 11, + "args": { + "External id": 13287,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13286 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335100, "dur": 5, + "args": { + "External id": 13288,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13287 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793335109, "dur": 21, + "args": { + "External id": 13289,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13288 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335110, "dur": 5, + "args": { + "External id": 13290,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13289 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793335116, "dur": 14, + "args": { + "External id": 13291,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13290 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793335118, "dur": 11, + "args": { + "External id": 13292,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13291 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335148, "dur": 5, + "args": { + "External id": 13293,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13292 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793335157, "dur": 21, + "args": { + "External id": 13294,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13293 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335158, "dur": 5, + "args": { + "External id": 13295,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13294 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793335164, "dur": 14, + "args": { + "External id": 13296,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13295 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793335166, "dur": 11, + "args": { + "External id": 13297,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13296 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335196, "dur": 5, + "args": { + "External id": 13298,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13297 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793335205, "dur": 21, + "args": { + "External id": 13299,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13298 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335207, "dur": 4, + "args": { + "External id": 13300,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13299 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793335212, "dur": 14, + "args": { + "External id": 13301,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13300 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793335214, "dur": 11, + "args": { + "External id": 13302,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13301 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335244, "dur": 5, + "args": { + "External id": 13303,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13302 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793335253, "dur": 21, + "args": { + "External id": 13304,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13303 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335254, "dur": 5, + "args": { + "External id": 13305,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13304 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793335260, "dur": 14, + "args": { + "External id": 13306,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13305 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793335262, "dur": 11, + "args": { + "External id": 13307,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13306 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335292, "dur": 5, + "args": { + "External id": 13308,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13307 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793335300, "dur": 21, + "args": { + "External id": 13309,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13308 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335302, "dur": 4, + "args": { + "External id": 13310,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13309 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793335307, "dur": 14, + "args": { + "External id": 13311,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13310 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793335309, "dur": 11, + "args": { + "External id": 13312,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13311 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335348, "dur": 5, + "args": { + "External id": 13313,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13312 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793335358, "dur": 21, + "args": { + "External id": 13314,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13313 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335359, "dur": 4, + "args": { + "External id": 13315,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13314 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793335365, "dur": 14, + "args": { + "External id": 13316,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13315 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793335366, "dur": 11, + "args": { + "External id": 13317,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13316 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335397, "dur": 5, + "args": { + "External id": 13318,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13317 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793335406, "dur": 21, + "args": { + "External id": 13319,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13318 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335407, "dur": 5, + "args": { + "External id": 13320,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13319 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793335413, "dur": 14, + "args": { + "External id": 13321,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13320 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793335415, "dur": 11, + "args": { + "External id": 13322,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13321 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335445, "dur": 5, + "args": { + "External id": 13323,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13322 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793335454, "dur": 21, + "args": { + "External id": 13324,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13323 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335456, "dur": 4, + "args": { + "External id": 13325,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13324 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793335461, "dur": 14, + "args": { + "External id": 13326,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13325 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793335463, "dur": 11, + "args": { + "External id": 13327,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13326 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335493, "dur": 5, + "args": { + "External id": 13328,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13327 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793335502, "dur": 22, + "args": { + "External id": 13329,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13328 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335503, "dur": 5, + "args": { + "External id": 13330,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13329 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793335509, "dur": 14, + "args": { + "External id": 13331,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13330 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793335511, "dur": 11, + "args": { + "External id": 13332,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13331 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335541, "dur": 5, + "args": { + "External id": 13333,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13332 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793335550, "dur": 21, + "args": { + "External id": 13334,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13333 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335551, "dur": 5, + "args": { + "External id": 13335,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13334 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793335557, "dur": 14, + "args": { + "External id": 13336,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13335 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793335559, "dur": 11, + "args": { + "External id": 13337,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13336 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335589, "dur": 5, + "args": { + "External id": 13338,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13337 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793335598, "dur": 21, + "args": { + "External id": 13339,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13338 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335599, "dur": 4, + "args": { + "External id": 13340,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13339 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793335605, "dur": 14, + "args": { + "External id": 13341,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13340 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793335606, "dur": 11, + "args": { + "External id": 13342,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13341 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335639, "dur": 5, + "args": { + "External id": 13343,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13342 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793335648, "dur": 23, + "args": { + "External id": 13344,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13343 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335650, "dur": 5, + "args": { + "External id": 13345,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13344 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793335657, "dur": 14, + "args": { + "External id": 13346,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13345 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793335659, "dur": 11, + "args": { + "External id": 13347,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13346 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335689, "dur": 5, + "args": { + "External id": 13348,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13347 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793335698, "dur": 21, + "args": { + "External id": 13349,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13348 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335699, "dur": 5, + "args": { + "External id": 13350,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13349 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793335705, "dur": 14, + "args": { + "External id": 13351,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13350 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793335706, "dur": 11, + "args": { + "External id": 13352,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13351 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335738, "dur": 6, + "args": { + "External id": 13353,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13352 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793335748, "dur": 21, + "args": { + "External id": 13354,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13353 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335749, "dur": 4, + "args": { + "External id": 13355,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13354 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793335755, "dur": 14, + "args": { + "External id": 13356,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13355 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793335757, "dur": 11, + "args": { + "External id": 13357,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13356 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335846, "dur": 7, + "args": { + "External id": 13358,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13357 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793335858, "dur": 25, + "args": { + "External id": 13359,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13358 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335859, "dur": 5, + "args": { + "External id": 13360,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13359 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793335865, "dur": 17, + "args": { + "External id": 13361,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13360 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793335867, "dur": 14, + "args": { + "External id": 13362,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13361 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335903, "dur": 6, + "args": { + "External id": 13363,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13362 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793335913, "dur": 22, + "args": { + "External id": 13364,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13363 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335914, "dur": 4, + "args": { + "External id": 13365,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13364 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793335920, "dur": 14, + "args": { + "External id": 13366,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13365 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793335921, "dur": 12, + "args": { + "External id": 13367,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13366 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335953, "dur": 5, + "args": { + "External id": 13368,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13367 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793335962, "dur": 22, + "args": { + "External id": 13369,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13368 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793335963, "dur": 5, + "args": { + "External id": 13370,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13369 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793335969, "dur": 14, + "args": { + "External id": 13371,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13370 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793335971, "dur": 11, + "args": { + "External id": 13372,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13371 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336002, "dur": 5, + "args": { + "External id": 13373,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13372 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793336011, "dur": 21, + "args": { + "External id": 13374,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13373 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336012, "dur": 4, + "args": { + "External id": 13375,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13374 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793336018, "dur": 14, + "args": { + "External id": 13376,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13375 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793336019, "dur": 11, + "args": { + "External id": 13377,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13376 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336050, "dur": 5, + "args": { + "External id": 13378,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13377 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793336059, "dur": 22, + "args": { + "External id": 13379,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13378 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336060, "dur": 5, + "args": { + "External id": 13380,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13379 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793336066, "dur": 14, + "args": { + "External id": 13381,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13380 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793336068, "dur": 11, + "args": { + "External id": 13382,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13381 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336099, "dur": 5, + "args": { + "External id": 13383,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13382 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793336108, "dur": 21, + "args": { + "External id": 13384,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13383 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336109, "dur": 4, + "args": { + "External id": 13385,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13384 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793336115, "dur": 13, + "args": { + "External id": 13386,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13385 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793336116, "dur": 11, + "args": { + "External id": 13387,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13386 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336147, "dur": 5, + "args": { + "External id": 13388,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13387 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793336156, "dur": 21, + "args": { + "External id": 13389,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13388 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336157, "dur": 4, + "args": { + "External id": 13390,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13389 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793336163, "dur": 13, + "args": { + "External id": 13391,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13390 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793336164, "dur": 11, + "args": { + "External id": 13392,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13391 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336195, "dur": 5, + "args": { + "External id": 13393,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13392 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793336204, "dur": 40, + "args": { + "External id": 13394,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13393 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336205, "dur": 4, + "args": { + "External id": 13395,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13394 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793336210, "dur": 34, + "args": { + "External id": 13396,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13395 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793336212, "dur": 30, + "args": { + "External id": 13397,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13396 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336265, "dur": 5, + "args": { + "External id": 13398,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13397 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793336274, "dur": 22, + "args": { + "External id": 13399,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13398 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336276, "dur": 4, + "args": { + "External id": 13400,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13399 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793336282, "dur": 14, + "args": { + "External id": 13401,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13400 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793336283, "dur": 12, + "args": { + "External id": 13402,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13401 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336314, "dur": 6, + "args": { + "External id": 13403,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13402 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793336324, "dur": 21, + "args": { + "External id": 13404,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13403 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336325, "dur": 4, + "args": { + "External id": 13405,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13404 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793336330, "dur": 14, + "args": { + "External id": 13406,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13405 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793336332, "dur": 11, + "args": { + "External id": 13407,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13406 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336416, "dur": 6, + "args": { + "External id": 13408,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13407 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793336427, "dur": 25, + "args": { + "External id": 13409,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13408 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336429, "dur": 4, + "args": { + "External id": 13410,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13409 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793336435, "dur": 17, + "args": { + "External id": 13411,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13410 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793336436, "dur": 14, + "args": { + "External id": 13412,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13411 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336472, "dur": 6, + "args": { + "External id": 13413,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13412 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793336482, "dur": 22, + "args": { + "External id": 13414,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13413 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336483, "dur": 4, + "args": { + "External id": 13415,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13414 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793336489, "dur": 14, + "args": { + "External id": 13416,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13415 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793336490, "dur": 12, + "args": { + "External id": 13417,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13416 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336522, "dur": 5, + "args": { + "External id": 13418,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13417 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793336531, "dur": 21, + "args": { + "External id": 13419,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13418 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336532, "dur": 5, + "args": { + "External id": 13420,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13419 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793336538, "dur": 14, + "args": { + "External id": 13421,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13420 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793336540, "dur": 11, + "args": { + "External id": 13422,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13421 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336571, "dur": 5, + "args": { + "External id": 13423,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13422 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793336580, "dur": 21, + "args": { + "External id": 13424,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13423 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336581, "dur": 4, + "args": { + "External id": 13425,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13424 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793336587, "dur": 14, + "args": { + "External id": 13426,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13425 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793336588, "dur": 11, + "args": { + "External id": 13427,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13426 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336619, "dur": 6, + "args": { + "External id": 13428,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13427 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793336629, "dur": 21, + "args": { + "External id": 13429,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13428 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336630, "dur": 4, + "args": { + "External id": 13430,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13429 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793336635, "dur": 14, + "args": { + "External id": 13431,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13430 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793336637, "dur": 11, + "args": { + "External id": 13432,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13431 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336667, "dur": 6, + "args": { + "External id": 13433,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13432 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793336677, "dur": 21, + "args": { + "External id": 13434,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13433 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336678, "dur": 4, + "args": { + "External id": 13435,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13434 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793336683, "dur": 14, + "args": { + "External id": 13436,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13435 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793336685, "dur": 11, + "args": { + "External id": 13437,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13436 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336716, "dur": 5, + "args": { + "External id": 13438,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13437 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793336725, "dur": 21, + "args": { + "External id": 13439,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13438 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336726, "dur": 4, + "args": { + "External id": 13440,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13439 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793336732, "dur": 14, + "args": { + "External id": 13441,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13440 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793336733, "dur": 11, + "args": { + "External id": 13442,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13441 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336763, "dur": 6, + "args": { + "External id": 13443,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13442 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793336772, "dur": 22, + "args": { + "External id": 13444,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13443 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336774, "dur": 4, + "args": { + "External id": 13445,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13444 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793336779, "dur": 14, + "args": { + "External id": 13446,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13445 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793336781, "dur": 11, + "args": { + "External id": 13447,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13446 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336812, "dur": 5, + "args": { + "External id": 13448,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13447 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793336821, "dur": 21, + "args": { + "External id": 13449,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13448 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336822, "dur": 5, + "args": { + "External id": 13450,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13449 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793336828, "dur": 14, + "args": { + "External id": 13451,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13450 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793336829, "dur": 12, + "args": { + "External id": 13452,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13451 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336860, "dur": 5, + "args": { + "External id": 13453,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13452 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793336869, "dur": 21, + "args": { + "External id": 13454,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13453 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336870, "dur": 5, + "args": { + "External id": 13455,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13454 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793336876, "dur": 14, + "args": { + "External id": 13456,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13455 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793336878, "dur": 10, + "args": { + "External id": 13457,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13456 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336908, "dur": 5, + "args": { + "External id": 13458,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13457 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793336917, "dur": 21, + "args": { + "External id": 13459,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13458 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336918, "dur": 4, + "args": { + "External id": 13460,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13459 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793336924, "dur": 14, + "args": { + "External id": 13461,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13460 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793336925, "dur": 12, + "args": { + "External id": 13462,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13461 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336956, "dur": 5, + "args": { + "External id": 13463,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13462 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793336965, "dur": 21, + "args": { + "External id": 13464,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13463 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793336966, "dur": 4, + "args": { + "External id": 13465,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13464 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793336972, "dur": 14, + "args": { + "External id": 13466,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13465 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793336973, "dur": 11, + "args": { + "External id": 13467,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13466 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337004, "dur": 5, + "args": { + "External id": 13468,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13467 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793337013, "dur": 21, + "args": { + "External id": 13469,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13468 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337014, "dur": 5, + "args": { + "External id": 13470,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13469 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793337020, "dur": 14, + "args": { + "External id": 13471,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13470 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793337022, "dur": 11, + "args": { + "External id": 13472,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13471 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337053, "dur": 5, + "args": { + "External id": 13473,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13472 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793337062, "dur": 21, + "args": { + "External id": 13474,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13473 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337063, "dur": 4, + "args": { + "External id": 13475,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13474 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793337069, "dur": 14, + "args": { + "External id": 13476,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13475 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793337070, "dur": 11, + "args": { + "External id": 13477,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13476 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337101, "dur": 5, + "args": { + "External id": 13478,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13477 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793337110, "dur": 21, + "args": { + "External id": 13479,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13478 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337111, "dur": 4, + "args": { + "External id": 13480,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13479 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793337117, "dur": 14, + "args": { + "External id": 13481,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13480 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793337118, "dur": 11, + "args": { + "External id": 13482,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13481 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337149, "dur": 5, + "args": { + "External id": 13483,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13482 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793337158, "dur": 21, + "args": { + "External id": 13484,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13483 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337159, "dur": 5, + "args": { + "External id": 13485,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13484 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793337165, "dur": 14, + "args": { + "External id": 13486,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13485 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793337167, "dur": 10, + "args": { + "External id": 13487,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13486 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337197, "dur": 5, + "args": { + "External id": 13488,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13487 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793337206, "dur": 22, + "args": { + "External id": 13489,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13488 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337207, "dur": 5, + "args": { + "External id": 13490,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13489 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793337213, "dur": 14, + "args": { + "External id": 13491,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13490 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793337215, "dur": 11, + "args": { + "External id": 13492,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13491 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337245, "dur": 5, + "args": { + "External id": 13493,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13492 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793337255, "dur": 21, + "args": { + "External id": 13494,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13493 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337256, "dur": 4, + "args": { + "External id": 13495,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13494 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793337261, "dur": 15, + "args": { + "External id": 13496,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13495 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793337263, "dur": 11, + "args": { + "External id": 13497,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13496 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337294, "dur": 5, + "args": { + "External id": 13498,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13497 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793337303, "dur": 21, + "args": { + "External id": 13499,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13498 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337304, "dur": 5, + "args": { + "External id": 13500,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13499 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793337310, "dur": 14, + "args": { + "External id": 13501,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13500 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793337312, "dur": 11, + "args": { + "External id": 13502,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13501 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337342, "dur": 5, + "args": { + "External id": 13503,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13502 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793337351, "dur": 21, + "args": { + "External id": 13504,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13503 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337352, "dur": 5, + "args": { + "External id": 13505,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13504 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793337358, "dur": 14, + "args": { + "External id": 13506,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13505 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793337359, "dur": 11, + "args": { + "External id": 13507,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13506 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337390, "dur": 5, + "args": { + "External id": 13508,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13507 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793337399, "dur": 21, + "args": { + "External id": 13509,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13508 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337400, "dur": 5, + "args": { + "External id": 13510,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13509 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793337406, "dur": 14, + "args": { + "External id": 13511,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13510 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793337408, "dur": 11, + "args": { + "External id": 13512,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13511 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337438, "dur": 5, + "args": { + "External id": 13513,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13512 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793337447, "dur": 21, + "args": { + "External id": 13514,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13513 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337448, "dur": 5, + "args": { + "External id": 13515,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13514 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793337454, "dur": 14, + "args": { + "External id": 13516,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13515 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793337455, "dur": 12, + "args": { + "External id": 13517,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13516 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337486, "dur": 5, + "args": { + "External id": 13518,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13517 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793337495, "dur": 21, + "args": { + "External id": 13519,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13518 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337496, "dur": 5, + "args": { + "External id": 13520,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13519 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793337502, "dur": 14, + "args": { + "External id": 13521,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13520 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793337504, "dur": 11, + "args": { + "External id": 13522,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13521 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337534, "dur": 5, + "args": { + "External id": 13523,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13522 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793337543, "dur": 22, + "args": { + "External id": 13524,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13523 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337544, "dur": 5, + "args": { + "External id": 13525,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13524 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793337550, "dur": 14, + "args": { + "External id": 13526,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13525 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793337552, "dur": 11, + "args": { + "External id": 13527,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13526 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337583, "dur": 5, + "args": { + "External id": 13528,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13527 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793337592, "dur": 21, + "args": { + "External id": 13529,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13528 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337593, "dur": 4, + "args": { + "External id": 13530,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13529 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793337599, "dur": 14, + "args": { + "External id": 13531,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13530 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793337600, "dur": 11, + "args": { + "External id": 13532,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13531 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337630, "dur": 7, + "args": { + "External id": 13533,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13532 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793337642, "dur": 23, + "args": { + "External id": 13534,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13533 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337643, "dur": 5, + "args": { + "External id": 13535,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13534 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793337649, "dur": 16, + "args": { + "External id": 13536,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13535 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793337651, "dur": 13, + "args": { + "External id": 13537,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13536 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337683, "dur": 5, + "args": { + "External id": 13538,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13537 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793337692, "dur": 21, + "args": { + "External id": 13539,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13538 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337693, "dur": 4, + "args": { + "External id": 13540,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13539 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793337699, "dur": 14, + "args": { + "External id": 13541,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13540 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793337700, "dur": 11, + "args": { + "External id": 13542,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13541 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337731, "dur": 5, + "args": { + "External id": 13543,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13542 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793337740, "dur": 21, + "args": { + "External id": 13544,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13543 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337741, "dur": 5, + "args": { + "External id": 13545,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13544 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793337747, "dur": 14, + "args": { + "External id": 13546,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13545 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793337748, "dur": 11, + "args": { + "External id": 13547,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13546 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337778, "dur": 6, + "args": { + "External id": 13548,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13547 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793337788, "dur": 21, + "args": { + "External id": 13549,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13548 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337789, "dur": 4, + "args": { + "External id": 13550,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13549 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793337794, "dur": 14, + "args": { + "External id": 13551,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13550 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793337796, "dur": 11, + "args": { + "External id": 13552,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13551 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337827, "dur": 5, + "args": { + "External id": 13553,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13552 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793337836, "dur": 21, + "args": { + "External id": 13554,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13553 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337837, "dur": 4, + "args": { + "External id": 13555,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13554 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793337843, "dur": 14, + "args": { + "External id": 13556,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13555 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793337844, "dur": 11, + "args": { + "External id": 13557,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13556 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337875, "dur": 5, + "args": { + "External id": 13558,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13557 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793337884, "dur": 22, + "args": { + "External id": 13559,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13558 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337885, "dur": 4, + "args": { + "External id": 13560,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13559 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793337891, "dur": 14, + "args": { + "External id": 13561,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13560 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793337893, "dur": 11, + "args": { + "External id": 13562,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13561 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337924, "dur": 5, + "args": { + "External id": 13563,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13562 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793337933, "dur": 21, + "args": { + "External id": 13564,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13563 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337934, "dur": 4, + "args": { + "External id": 13565,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13564 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793337940, "dur": 14, + "args": { + "External id": 13566,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13565 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793337941, "dur": 11, + "args": { + "External id": 13567,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13566 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337972, "dur": 5, + "args": { + "External id": 13568,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13567 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793337981, "dur": 21, + "args": { + "External id": 13569,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13568 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793337982, "dur": 4, + "args": { + "External id": 13570,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13569 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793337988, "dur": 14, + "args": { + "External id": 13571,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13570 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793337989, "dur": 11, + "args": { + "External id": 13572,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13571 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338020, "dur": 5, + "args": { + "External id": 13573,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13572 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793338029, "dur": 21, + "args": { + "External id": 13574,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13573 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338030, "dur": 4, + "args": { + "External id": 13575,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13574 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793338035, "dur": 15, + "args": { + "External id": 13576,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13575 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793338037, "dur": 11, + "args": { + "External id": 13577,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13576 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338068, "dur": 5, + "args": { + "External id": 13578,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13577 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793338077, "dur": 21, + "args": { + "External id": 13579,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13578 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338078, "dur": 4, + "args": { + "External id": 13580,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13579 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793338083, "dur": 14, + "args": { + "External id": 13581,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13580 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793338085, "dur": 11, + "args": { + "External id": 13582,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13581 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338115, "dur": 5, + "args": { + "External id": 13583,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13582 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793338124, "dur": 22, + "args": { + "External id": 13584,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13583 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338126, "dur": 4, + "args": { + "External id": 13585,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13584 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793338131, "dur": 14, + "args": { + "External id": 13586,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13585 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793338133, "dur": 11, + "args": { + "External id": 13587,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13586 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338163, "dur": 6, + "args": { + "External id": 13588,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13587 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793338172, "dur": 22, + "args": { + "External id": 13589,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13588 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338174, "dur": 4, + "args": { + "External id": 13590,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13589 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793338179, "dur": 14, + "args": { + "External id": 13591,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13590 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793338181, "dur": 11, + "args": { + "External id": 13592,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13591 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338211, "dur": 5, + "args": { + "External id": 13593,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13592 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793338220, "dur": 22, + "args": { + "External id": 13594,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13593 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338222, "dur": 4, + "args": { + "External id": 13595,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13594 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793338227, "dur": 14, + "args": { + "External id": 13596,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13595 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793338229, "dur": 11, + "args": { + "External id": 13597,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13596 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338260, "dur": 5, + "args": { + "External id": 13598,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13597 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793338269, "dur": 21, + "args": { + "External id": 13599,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13598 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338270, "dur": 4, + "args": { + "External id": 13600,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13599 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793338276, "dur": 14, + "args": { + "External id": 13601,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13600 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793338277, "dur": 11, + "args": { + "External id": 13602,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13601 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338308, "dur": 5, + "args": { + "External id": 13603,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13602 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793338317, "dur": 21, + "args": { + "External id": 13604,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13603 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338318, "dur": 4, + "args": { + "External id": 13605,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13604 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793338324, "dur": 14, + "args": { + "External id": 13606,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13605 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793338325, "dur": 11, + "args": { + "External id": 13607,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13606 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338356, "dur": 5, + "args": { + "External id": 13608,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13607 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793338365, "dur": 21, + "args": { + "External id": 13609,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13608 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338366, "dur": 5, + "args": { + "External id": 13610,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13609 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793338372, "dur": 14, + "args": { + "External id": 13611,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13610 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793338374, "dur": 11, + "args": { + "External id": 13612,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13611 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338404, "dur": 5, + "args": { + "External id": 13613,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13612 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793338413, "dur": 22, + "args": { + "External id": 13614,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13613 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338414, "dur": 5, + "args": { + "External id": 13615,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13614 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793338420, "dur": 14, + "args": { + "External id": 13616,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13615 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793338422, "dur": 11, + "args": { + "External id": 13617,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13616 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338452, "dur": 5, + "args": { + "External id": 13618,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13617 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793338461, "dur": 22, + "args": { + "External id": 13619,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13618 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338463, "dur": 4, + "args": { + "External id": 13620,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13619 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793338468, "dur": 15, + "args": { + "External id": 13621,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13620 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793338470, "dur": 11, + "args": { + "External id": 13622,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13621 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338501, "dur": 5, + "args": { + "External id": 13623,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13622 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793338510, "dur": 21, + "args": { + "External id": 13624,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13623 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338511, "dur": 4, + "args": { + "External id": 13625,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13624 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793338516, "dur": 14, + "args": { + "External id": 13626,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13625 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793338518, "dur": 11, + "args": { + "External id": 13627,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13626 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338549, "dur": 5, + "args": { + "External id": 13628,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13627 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793338558, "dur": 21, + "args": { + "External id": 13629,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13628 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338559, "dur": 5, + "args": { + "External id": 13630,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13629 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793338565, "dur": 14, + "args": { + "External id": 13631,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13630 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793338567, "dur": 11, + "args": { + "External id": 13632,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13631 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338597, "dur": 5, + "args": { + "External id": 13633,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13632 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793338606, "dur": 21, + "args": { + "External id": 13634,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13633 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338608, "dur": 4, + "args": { + "External id": 13635,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13634 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793338613, "dur": 14, + "args": { + "External id": 13636,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13635 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793338615, "dur": 11, + "args": { + "External id": 13637,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13636 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338645, "dur": 5, + "args": { + "External id": 13638,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13637 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793338654, "dur": 21, + "args": { + "External id": 13639,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13638 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338655, "dur": 5, + "args": { + "External id": 13640,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13639 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793338661, "dur": 14, + "args": { + "External id": 13641,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13640 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793338663, "dur": 11, + "args": { + "External id": 13642,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13641 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338693, "dur": 5, + "args": { + "External id": 13643,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13642 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793338702, "dur": 22, + "args": { + "External id": 13644,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13643 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338703, "dur": 5, + "args": { + "External id": 13645,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13644 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793338709, "dur": 15, + "args": { + "External id": 13646,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13645 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793338711, "dur": 11, + "args": { + "External id": 13647,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13646 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338742, "dur": 5, + "args": { + "External id": 13648,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13647 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793338750, "dur": 22, + "args": { + "External id": 13649,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13648 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338752, "dur": 4, + "args": { + "External id": 13650,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13649 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793338757, "dur": 14, + "args": { + "External id": 13651,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13650 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793338759, "dur": 11, + "args": { + "External id": 13652,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13651 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338789, "dur": 5, + "args": { + "External id": 13653,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13652 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793338798, "dur": 25, + "args": { + "External id": 13654,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13653 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338800, "dur": 4, + "args": { + "External id": 13655,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13654 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793338805, "dur": 18, + "args": { + "External id": 13656,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13655 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793338810, "dur": 11, + "args": { + "External id": 13657,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13656 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338900, "dur": 6, + "args": { + "External id": 13658,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13657 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793338911, "dur": 25, + "args": { + "External id": 13659,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13658 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338912, "dur": 5, + "args": { + "External id": 13660,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13659 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793338918, "dur": 18, + "args": { + "External id": 13661,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13660 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793338920, "dur": 14, + "args": { + "External id": 13662,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13661 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338957, "dur": 5, + "args": { + "External id": 13663,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13662 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793338966, "dur": 22, + "args": { + "External id": 13664,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13663 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793338967, "dur": 4, + "args": { + "External id": 13665,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13664 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793338973, "dur": 14, + "args": { + "External id": 13666,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13665 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793338974, "dur": 12, + "args": { + "External id": 13667,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13666 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339006, "dur": 5, + "args": { + "External id": 13668,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13667 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793339015, "dur": 22, + "args": { + "External id": 13669,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13668 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339016, "dur": 5, + "args": { + "External id": 13670,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13669 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793339022, "dur": 14, + "args": { + "External id": 13671,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13670 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793339024, "dur": 11, + "args": { + "External id": 13672,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13671 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339054, "dur": 5, + "args": { + "External id": 13673,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13672 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793339063, "dur": 21, + "args": { + "External id": 13674,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13673 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339064, "dur": 5, + "args": { + "External id": 13675,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13674 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793339070, "dur": 14, + "args": { + "External id": 13676,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13675 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793339072, "dur": 11, + "args": { + "External id": 13677,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13676 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339102, "dur": 6, + "args": { + "External id": 13678,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13677 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793339111, "dur": 22, + "args": { + "External id": 13679,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13678 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339113, "dur": 4, + "args": { + "External id": 13680,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13679 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793339118, "dur": 14, + "args": { + "External id": 13681,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13680 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793339120, "dur": 11, + "args": { + "External id": 13682,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13681 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339151, "dur": 5, + "args": { + "External id": 13683,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13682 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793339159, "dur": 22, + "args": { + "External id": 13684,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13683 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339161, "dur": 4, + "args": { + "External id": 13685,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13684 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793339166, "dur": 14, + "args": { + "External id": 13686,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13685 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793339168, "dur": 11, + "args": { + "External id": 13687,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13686 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339199, "dur": 5, + "args": { + "External id": 13688,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13687 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793339208, "dur": 21, + "args": { + "External id": 13689,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13688 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339209, "dur": 4, + "args": { + "External id": 13690,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13689 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793339214, "dur": 14, + "args": { + "External id": 13691,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13690 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793339216, "dur": 11, + "args": { + "External id": 13692,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13691 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339247, "dur": 5, + "args": { + "External id": 13693,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13692 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793339256, "dur": 21, + "args": { + "External id": 13694,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13693 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339257, "dur": 4, + "args": { + "External id": 13695,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13694 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793339263, "dur": 13, + "args": { + "External id": 13696,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13695 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793339264, "dur": 11, + "args": { + "External id": 13697,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13696 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339295, "dur": 5, + "args": { + "External id": 13698,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13697 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793339304, "dur": 21, + "args": { + "External id": 13699,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13698 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339305, "dur": 4, + "args": { + "External id": 13700,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13699 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793339310, "dur": 14, + "args": { + "External id": 13701,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13700 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793339312, "dur": 11, + "args": { + "External id": 13702,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13701 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339343, "dur": 5, + "args": { + "External id": 13703,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13702 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793339352, "dur": 21, + "args": { + "External id": 13704,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13703 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339353, "dur": 4, + "args": { + "External id": 13705,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13704 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793339359, "dur": 14, + "args": { + "External id": 13706,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13705 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793339360, "dur": 11, + "args": { + "External id": 13707,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13706 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339444, "dur": 6, + "args": { + "External id": 13708,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13707 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793339455, "dur": 25, + "args": { + "External id": 13709,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13708 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339457, "dur": 4, + "args": { + "External id": 13710,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13709 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793339463, "dur": 17, + "args": { + "External id": 13711,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13710 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793339464, "dur": 15, + "args": { + "External id": 13712,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13711 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339500, "dur": 6, + "args": { + "External id": 13713,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13712 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793339510, "dur": 21, + "args": { + "External id": 13714,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13713 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339511, "dur": 4, + "args": { + "External id": 13715,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13714 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793339517, "dur": 14, + "args": { + "External id": 13716,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13715 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793339518, "dur": 12, + "args": { + "External id": 13717,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13716 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339549, "dur": 5, + "args": { + "External id": 13718,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13717 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793339558, "dur": 22, + "args": { + "External id": 13719,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13718 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339560, "dur": 4, + "args": { + "External id": 13720,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13719 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793339565, "dur": 14, + "args": { + "External id": 13721,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13720 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793339567, "dur": 11, + "args": { + "External id": 13722,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13721 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339597, "dur": 5, + "args": { + "External id": 13723,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13722 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793339606, "dur": 22, + "args": { + "External id": 13724,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13723 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339608, "dur": 4, + "args": { + "External id": 13725,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13724 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793339613, "dur": 15, + "args": { + "External id": 13726,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13725 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793339615, "dur": 11, + "args": { + "External id": 13727,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13726 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339648, "dur": 7, + "args": { + "External id": 13728,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13727 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793339659, "dur": 21, + "args": { + "External id": 13729,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13728 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339660, "dur": 5, + "args": { + "External id": 13730,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13729 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793339666, "dur": 14, + "args": { + "External id": 13731,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13730 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793339668, "dur": 11, + "args": { + "External id": 13732,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13731 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339698, "dur": 6, + "args": { + "External id": 13733,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13732 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793339707, "dur": 22, + "args": { + "External id": 13734,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13733 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339709, "dur": 4, + "args": { + "External id": 13735,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13734 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793339714, "dur": 14, + "args": { + "External id": 13736,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13735 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793339716, "dur": 11, + "args": { + "External id": 13737,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13736 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339747, "dur": 5, + "args": { + "External id": 13738,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13737 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793339756, "dur": 21, + "args": { + "External id": 13739,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13738 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339757, "dur": 4, + "args": { + "External id": 13740,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13739 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793339763, "dur": 14, + "args": { + "External id": 13741,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13740 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793339764, "dur": 11, + "args": { + "External id": 13742,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13741 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339795, "dur": 5, + "args": { + "External id": 13743,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13742 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793339804, "dur": 21, + "args": { + "External id": 13744,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13743 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339805, "dur": 4, + "args": { + "External id": 13745,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13744 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793339811, "dur": 14, + "args": { + "External id": 13746,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13745 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793339812, "dur": 11, + "args": { + "External id": 13747,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13746 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339843, "dur": 5, + "args": { + "External id": 13748,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13747 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793339852, "dur": 21, + "args": { + "External id": 13749,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13748 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339853, "dur": 4, + "args": { + "External id": 13750,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13749 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793339859, "dur": 14, + "args": { + "External id": 13751,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13750 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793339860, "dur": 11, + "args": { + "External id": 13752,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13751 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339891, "dur": 5, + "args": { + "External id": 13753,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13752 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793339900, "dur": 21, + "args": { + "External id": 13754,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13753 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339901, "dur": 5, + "args": { + "External id": 13755,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13754 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793339907, "dur": 14, + "args": { + "External id": 13756,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13755 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793339909, "dur": 11, + "args": { + "External id": 13757,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13756 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339939, "dur": 5, + "args": { + "External id": 13758,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13757 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793339948, "dur": 21, + "args": { + "External id": 13759,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13758 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339949, "dur": 5, + "args": { + "External id": 13760,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13759 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793339955, "dur": 14, + "args": { + "External id": 13761,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13760 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793339957, "dur": 11, + "args": { + "External id": 13762,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13761 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339987, "dur": 6, + "args": { + "External id": 13763,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13762 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793339997, "dur": 21, + "args": { + "External id": 13764,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13763 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793339998, "dur": 5, + "args": { + "External id": 13765,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13764 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793340004, "dur": 14, + "args": { + "External id": 13766,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13765 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793340006, "dur": 10, + "args": { + "External id": 13767,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13766 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340036, "dur": 5, + "args": { + "External id": 13768,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13767 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793340045, "dur": 21, + "args": { + "External id": 13769,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13768 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340046, "dur": 5, + "args": { + "External id": 13770,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13769 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793340052, "dur": 13, + "args": { + "External id": 13771,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13770 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793340053, "dur": 11, + "args": { + "External id": 13772,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13771 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340084, "dur": 5, + "args": { + "External id": 13773,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13772 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793340093, "dur": 21, + "args": { + "External id": 13774,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13773 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340094, "dur": 4, + "args": { + "External id": 13775,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13774 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793340099, "dur": 14, + "args": { + "External id": 13776,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13775 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793340101, "dur": 11, + "args": { + "External id": 13777,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13776 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340131, "dur": 5, + "args": { + "External id": 13778,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13777 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793340140, "dur": 21, + "args": { + "External id": 13779,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13778 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340142, "dur": 4, + "args": { + "External id": 13780,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13779 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793340147, "dur": 14, + "args": { + "External id": 13781,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13780 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793340149, "dur": 11, + "args": { + "External id": 13782,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13781 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340179, "dur": 5, + "args": { + "External id": 13783,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13782 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793340188, "dur": 21, + "args": { + "External id": 13784,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13783 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340189, "dur": 5, + "args": { + "External id": 13785,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13784 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793340195, "dur": 14, + "args": { + "External id": 13786,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13785 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793340197, "dur": 11, + "args": { + "External id": 13787,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13786 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340248, "dur": 6, + "args": { + "External id": 13788,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13787 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793340258, "dur": 23, + "args": { + "External id": 13789,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13788 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340259, "dur": 4, + "args": { + "External id": 13790,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13789 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793340265, "dur": 16, + "args": { + "External id": 13791,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13790 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793340267, "dur": 12, + "args": { + "External id": 13792,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13791 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340300, "dur": 5, + "args": { + "External id": 13793,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13792 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793340309, "dur": 21, + "args": { + "External id": 13794,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13793 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340310, "dur": 5, + "args": { + "External id": 13795,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13794 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793340316, "dur": 14, + "args": { + "External id": 13796,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13795 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793340318, "dur": 11, + "args": { + "External id": 13797,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13796 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340349, "dur": 5, + "args": { + "External id": 13798,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13797 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793340358, "dur": 21, + "args": { + "External id": 13799,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13798 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340359, "dur": 5, + "args": { + "External id": 13800,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13799 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793340365, "dur": 14, + "args": { + "External id": 13801,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13800 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793340367, "dur": 11, + "args": { + "External id": 13802,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13801 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340397, "dur": 5, + "args": { + "External id": 13803,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13802 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793340406, "dur": 21, + "args": { + "External id": 13804,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13803 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340407, "dur": 5, + "args": { + "External id": 13805,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13804 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793340413, "dur": 14, + "args": { + "External id": 13806,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13805 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793340415, "dur": 11, + "args": { + "External id": 13807,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13806 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340445, "dur": 5, + "args": { + "External id": 13808,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13807 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793340454, "dur": 21, + "args": { + "External id": 13809,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13808 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340455, "dur": 5, + "args": { + "External id": 13810,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13809 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793340461, "dur": 14, + "args": { + "External id": 13811,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13810 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793340463, "dur": 11, + "args": { + "External id": 13812,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13811 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340493, "dur": 5, + "args": { + "External id": 13813,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13812 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793340502, "dur": 21, + "args": { + "External id": 13814,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13813 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340503, "dur": 4, + "args": { + "External id": 13815,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13814 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793340509, "dur": 13, + "args": { + "External id": 13816,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13815 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793340510, "dur": 11, + "args": { + "External id": 13817,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13816 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340540, "dur": 6, + "args": { + "External id": 13818,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13817 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793340549, "dur": 21, + "args": { + "External id": 13819,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13818 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340551, "dur": 4, + "args": { + "External id": 13820,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13819 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793340556, "dur": 14, + "args": { + "External id": 13821,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13820 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793340558, "dur": 11, + "args": { + "External id": 13822,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13821 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340588, "dur": 5, + "args": { + "External id": 13823,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13822 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793340597, "dur": 32, + "args": { + "External id": 13824,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13823 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340609, "dur": 4, + "args": { + "External id": 13825,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13824 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793340614, "dur": 14, + "args": { + "External id": 13826,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13825 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793340616, "dur": 11, + "args": { + "External id": 13827,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13826 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340647, "dur": 5, + "args": { + "External id": 13828,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13827 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793340656, "dur": 20, + "args": { + "External id": 13829,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13828 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340657, "dur": 4, + "args": { + "External id": 13830,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13829 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793340663, "dur": 13, + "args": { + "External id": 13831,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13830 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793340664, "dur": 11, + "args": { + "External id": 13832,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13831 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340694, "dur": 5, + "args": { + "External id": 13833,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13832 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793340703, "dur": 22, + "args": { + "External id": 13834,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13833 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340704, "dur": 5, + "args": { + "External id": 13835,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13834 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793340710, "dur": 14, + "args": { + "External id": 13836,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13835 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793340712, "dur": 11, + "args": { + "External id": 13837,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13836 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340742, "dur": 5, + "args": { + "External id": 13838,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13837 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793340751, "dur": 21, + "args": { + "External id": 13839,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13838 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340752, "dur": 5, + "args": { + "External id": 13840,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13839 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793340758, "dur": 14, + "args": { + "External id": 13841,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13840 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793340760, "dur": 10, + "args": { + "External id": 13842,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13841 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340790, "dur": 5, + "args": { + "External id": 13843,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13842 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793340799, "dur": 21, + "args": { + "External id": 13844,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13843 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340800, "dur": 4, + "args": { + "External id": 13845,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13844 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793340806, "dur": 14, + "args": { + "External id": 13846,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13845 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793340807, "dur": 11, + "args": { + "External id": 13847,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13846 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340838, "dur": 5, + "args": { + "External id": 13848,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13847 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793340847, "dur": 21, + "args": { + "External id": 13849,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13848 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340848, "dur": 4, + "args": { + "External id": 13850,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13849 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793340853, "dur": 14, + "args": { + "External id": 13851,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13850 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793340855, "dur": 11, + "args": { + "External id": 13852,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13851 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340885, "dur": 5, + "args": { + "External id": 13853,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13852 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793340894, "dur": 21, + "args": { + "External id": 13854,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13853 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340895, "dur": 5, + "args": { + "External id": 13855,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13854 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793340901, "dur": 14, + "args": { + "External id": 13856,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13855 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793340903, "dur": 10, + "args": { + "External id": 13857,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13856 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340933, "dur": 5, + "args": { + "External id": 13858,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13857 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793340942, "dur": 20, + "args": { + "External id": 13859,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13858 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340943, "dur": 4, + "args": { + "External id": 13860,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13859 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793340949, "dur": 13, + "args": { + "External id": 13861,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13860 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793340950, "dur": 11, + "args": { + "External id": 13862,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13861 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340980, "dur": 5, + "args": { + "External id": 13863,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13862 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793340989, "dur": 23, + "args": { + "External id": 13864,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13863 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793340993, "dur": 4, + "args": { + "External id": 13865,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13864 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793340998, "dur": 14, + "args": { + "External id": 13866,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13865 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793341000, "dur": 11, + "args": { + "External id": 13867,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13866 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341030, "dur": 5, + "args": { + "External id": 13868,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13867 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793341039, "dur": 21, + "args": { + "External id": 13869,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13868 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341040, "dur": 5, + "args": { + "External id": 13870,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13869 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793341046, "dur": 14, + "args": { + "External id": 13871,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13870 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793341048, "dur": 10, + "args": { + "External id": 13872,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13871 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341078, "dur": 5, + "args": { + "External id": 13873,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13872 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793341087, "dur": 21, + "args": { + "External id": 13874,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13873 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341088, "dur": 4, + "args": { + "External id": 13875,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13874 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793341093, "dur": 14, + "args": { + "External id": 13876,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13875 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793341095, "dur": 11, + "args": { + "External id": 13877,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13876 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341125, "dur": 5, + "args": { + "External id": 13878,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13877 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793341134, "dur": 21, + "args": { + "External id": 13879,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13878 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341135, "dur": 5, + "args": { + "External id": 13880,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13879 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793341141, "dur": 13, + "args": { + "External id": 13881,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13880 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793341142, "dur": 11, + "args": { + "External id": 13882,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13881 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341172, "dur": 6, + "args": { + "External id": 13883,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13882 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793341182, "dur": 20, + "args": { + "External id": 13884,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13883 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341183, "dur": 4, + "args": { + "External id": 13885,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13884 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793341188, "dur": 14, + "args": { + "External id": 13886,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13885 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793341190, "dur": 11, + "args": { + "External id": 13887,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13886 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341220, "dur": 5, + "args": { + "External id": 13888,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13887 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793341229, "dur": 21, + "args": { + "External id": 13889,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13888 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341230, "dur": 5, + "args": { + "External id": 13890,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13889 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793341236, "dur": 13, + "args": { + "External id": 13891,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13890 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793341237, "dur": 11, + "args": { + "External id": 13892,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13891 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341268, "dur": 5, + "args": { + "External id": 13893,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13892 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793341277, "dur": 20, + "args": { + "External id": 13894,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13893 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341278, "dur": 4, + "args": { + "External id": 13895,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13894 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793341283, "dur": 14, + "args": { + "External id": 13896,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13895 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793341285, "dur": 11, + "args": { + "External id": 13897,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13896 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341315, "dur": 5, + "args": { + "External id": 13898,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13897 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793341324, "dur": 21, + "args": { + "External id": 13899,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13898 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341325, "dur": 5, + "args": { + "External id": 13900,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13899 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793341331, "dur": 14, + "args": { + "External id": 13901,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13900 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793341333, "dur": 10, + "args": { + "External id": 13902,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13901 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341363, "dur": 5, + "args": { + "External id": 13903,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13902 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793341372, "dur": 20, + "args": { + "External id": 13904,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13903 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341373, "dur": 4, + "args": { + "External id": 13905,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13904 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793341378, "dur": 14, + "args": { + "External id": 13906,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13905 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793341380, "dur": 11, + "args": { + "External id": 13907,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13906 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341410, "dur": 5, + "args": { + "External id": 13908,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13907 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793341419, "dur": 21, + "args": { + "External id": 13909,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13908 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341420, "dur": 5, + "args": { + "External id": 13910,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13909 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793341426, "dur": 14, + "args": { + "External id": 13911,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13910 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793341427, "dur": 11, + "args": { + "External id": 13912,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13911 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341458, "dur": 5, + "args": { + "External id": 13913,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13912 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793341467, "dur": 20, + "args": { + "External id": 13914,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13913 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341468, "dur": 4, + "args": { + "External id": 13915,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13914 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793341473, "dur": 14, + "args": { + "External id": 13916,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13915 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793341475, "dur": 11, + "args": { + "External id": 13917,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13916 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341505, "dur": 5, + "args": { + "External id": 13918,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13917 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793341514, "dur": 21, + "args": { + "External id": 13919,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13918 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341515, "dur": 5, + "args": { + "External id": 13920,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13919 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793341521, "dur": 14, + "args": { + "External id": 13921,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13920 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793341523, "dur": 11, + "args": { + "External id": 13922,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13921 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341553, "dur": 5, + "args": { + "External id": 13923,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13922 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793341562, "dur": 21, + "args": { + "External id": 13924,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13923 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341563, "dur": 5, + "args": { + "External id": 13925,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13924 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793341569, "dur": 14, + "args": { + "External id": 13926,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13925 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793341570, "dur": 11, + "args": { + "External id": 13927,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13926 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341600, "dur": 6, + "args": { + "External id": 13928,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13927 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793341609, "dur": 21, + "args": { + "External id": 13929,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13928 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341611, "dur": 4, + "args": { + "External id": 13930,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13929 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793341616, "dur": 14, + "args": { + "External id": 13931,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13930 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793341618, "dur": 11, + "args": { + "External id": 13932,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13931 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341648, "dur": 5, + "args": { + "External id": 13933,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13932 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793341657, "dur": 21, + "args": { + "External id": 13934,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13933 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341658, "dur": 5, + "args": { + "External id": 13935,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13934 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793341664, "dur": 14, + "args": { + "External id": 13936,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13935 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793341666, "dur": 10, + "args": { + "External id": 13937,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13936 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341696, "dur": 5, + "args": { + "External id": 13938,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13937 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793341705, "dur": 21, + "args": { + "External id": 13939,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13938 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341706, "dur": 4, + "args": { + "External id": 13940,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13939 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793341711, "dur": 14, + "args": { + "External id": 13941,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13940 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793341713, "dur": 11, + "args": { + "External id": 13942,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13941 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341744, "dur": 5, + "args": { + "External id": 13943,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13942 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793341753, "dur": 20, + "args": { + "External id": 13944,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13943 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341754, "dur": 4, + "args": { + "External id": 13945,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13944 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793341759, "dur": 14, + "args": { + "External id": 13946,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13945 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793341761, "dur": 11, + "args": { + "External id": 13947,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13946 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341791, "dur": 5, + "args": { + "External id": 13948,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13947 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793341800, "dur": 21, + "args": { + "External id": 13949,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13948 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341801, "dur": 5, + "args": { + "External id": 13950,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13949 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793341807, "dur": 13, + "args": { + "External id": 13951,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13950 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793341808, "dur": 11, + "args": { + "External id": 13952,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13951 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341838, "dur": 5, + "args": { + "External id": 13953,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13952 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793341847, "dur": 21, + "args": { + "External id": 13954,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13953 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341848, "dur": 5, + "args": { + "External id": 13955,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13954 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793341854, "dur": 14, + "args": { + "External id": 13956,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13955 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793341856, "dur": 10, + "args": { + "External id": 13957,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13956 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341946, "dur": 6, + "args": { + "External id": 13958,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13957 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793341957, "dur": 25, + "args": { + "External id": 13959,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13958 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793341959, "dur": 4, + "args": { + "External id": 13960,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13959 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793341965, "dur": 17, + "args": { + "External id": 13961,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13960 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793341967, "dur": 14, + "args": { + "External id": 13962,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13961 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342003, "dur": 5, + "args": { + "External id": 13963,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13962 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793342012, "dur": 22, + "args": { + "External id": 13964,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13963 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342013, "dur": 5, + "args": { + "External id": 13965,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13964 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793342019, "dur": 14, + "args": { + "External id": 13966,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13965 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793342021, "dur": 11, + "args": { + "External id": 13967,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13966 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342052, "dur": 5, + "args": { + "External id": 13968,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13967 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793342061, "dur": 22, + "args": { + "External id": 13969,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13968 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342063, "dur": 5, + "args": { + "External id": 13970,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13969 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793342069, "dur": 14, + "args": { + "External id": 13971,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13970 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793342070, "dur": 12, + "args": { + "External id": 13972,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13971 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342101, "dur": 5, + "args": { + "External id": 13973,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13972 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793342110, "dur": 21, + "args": { + "External id": 13974,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13973 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342111, "dur": 5, + "args": { + "External id": 13975,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13974 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793342117, "dur": 14, + "args": { + "External id": 13976,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13975 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793342119, "dur": 10, + "args": { + "External id": 13977,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13976 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342149, "dur": 5, + "args": { + "External id": 13978,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13977 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793342158, "dur": 21, + "args": { + "External id": 13979,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13978 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342159, "dur": 5, + "args": { + "External id": 13980,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13979 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793342165, "dur": 14, + "args": { + "External id": 13981,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13980 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793342167, "dur": 11, + "args": { + "External id": 13982,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13981 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342198, "dur": 5, + "args": { + "External id": 13983,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13982 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793342207, "dur": 21, + "args": { + "External id": 13984,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13983 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342208, "dur": 4, + "args": { + "External id": 13985,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13984 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793342213, "dur": 14, + "args": { + "External id": 13986,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13985 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793342215, "dur": 11, + "args": { + "External id": 13987,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13986 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342246, "dur": 5, + "args": { + "External id": 13988,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13987 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793342255, "dur": 25, + "args": { + "External id": 13989,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13988 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342257, "dur": 5, + "args": { + "External id": 13990,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13989 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793342264, "dur": 15, + "args": { + "External id": 13991,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13990 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793342266, "dur": 12, + "args": { + "External id": 13992,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13991 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342299, "dur": 5, + "args": { + "External id": 13993,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13992 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793342308, "dur": 21, + "args": { + "External id": 13994,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13993 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342309, "dur": 4, + "args": { + "External id": 13995,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13994 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793342315, "dur": 13, + "args": { + "External id": 13996,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 13995 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793342316, "dur": 11, + "args": { + "External id": 13997,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 13996 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342347, "dur": 5, + "args": { + "External id": 13998,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13997 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793342356, "dur": 21, + "args": { + "External id": 13999,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 13998 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342357, "dur": 4, + "args": { + "External id": 14000,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13999 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793342363, "dur": 13, + "args": { + "External id": 14001,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14000 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793342364, "dur": 11, + "args": { + "External id": 14002,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14001 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342395, "dur": 5, + "args": { + "External id": 14003,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14002 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793342403, "dur": 22, + "args": { + "External id": 14004,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14003 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342405, "dur": 4, + "args": { + "External id": 14005,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14004 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793342410, "dur": 14, + "args": { + "External id": 14006,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14005 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793342412, "dur": 11, + "args": { + "External id": 14007,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14006 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342495, "dur": 6, + "args": { + "External id": 14008,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14007 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793342511, "dur": 25, + "args": { + "External id": 14009,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14008 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342512, "dur": 5, + "args": { + "External id": 14010,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14009 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793342519, "dur": 17, + "args": { + "External id": 14011,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14010 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793342520, "dur": 15, + "args": { + "External id": 14012,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14011 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342558, "dur": 5, + "args": { + "External id": 14013,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14012 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793342567, "dur": 21, + "args": { + "External id": 14014,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14013 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342568, "dur": 5, + "args": { + "External id": 14015,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14014 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793342574, "dur": 14, + "args": { + "External id": 14016,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14015 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793342576, "dur": 11, + "args": { + "External id": 14017,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14016 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342607, "dur": 5, + "args": { + "External id": 14018,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14017 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793342616, "dur": 21, + "args": { + "External id": 14019,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14018 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342617, "dur": 4, + "args": { + "External id": 14020,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14019 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793342623, "dur": 14, + "args": { + "External id": 14021,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14020 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793342624, "dur": 12, + "args": { + "External id": 14022,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14021 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342656, "dur": 5, + "args": { + "External id": 14023,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14022 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793342665, "dur": 21, + "args": { + "External id": 14024,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14023 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342666, "dur": 4, + "args": { + "External id": 14025,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14024 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793342672, "dur": 13, + "args": { + "External id": 14026,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14025 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793342673, "dur": 11, + "args": { + "External id": 14027,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14026 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342704, "dur": 5, + "args": { + "External id": 14028,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14027 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793342713, "dur": 21, + "args": { + "External id": 14029,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14028 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342714, "dur": 4, + "args": { + "External id": 14030,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14029 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793342720, "dur": 13, + "args": { + "External id": 14031,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14030 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793342721, "dur": 11, + "args": { + "External id": 14032,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14031 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342751, "dur": 6, + "args": { + "External id": 14033,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14032 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793342760, "dur": 21, + "args": { + "External id": 14034,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14033 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342761, "dur": 5, + "args": { + "External id": 14035,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14034 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793342767, "dur": 14, + "args": { + "External id": 14036,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14035 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793342769, "dur": 11, + "args": { + "External id": 14037,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14036 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342799, "dur": 5, + "args": { + "External id": 14038,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14037 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793342808, "dur": 21, + "args": { + "External id": 14039,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14038 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342810, "dur": 4, + "args": { + "External id": 14040,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14039 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793342815, "dur": 14, + "args": { + "External id": 14041,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14040 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793342817, "dur": 11, + "args": { + "External id": 14042,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14041 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342847, "dur": 5, + "args": { + "External id": 14043,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14042 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793342856, "dur": 23, + "args": { + "External id": 14044,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14043 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342857, "dur": 5, + "args": { + "External id": 14045,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14044 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793342863, "dur": 16, + "args": { + "External id": 14046,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14045 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793342865, "dur": 12, + "args": { + "External id": 14047,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14046 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342899, "dur": 5, + "args": { + "External id": 14048,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14047 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793342908, "dur": 21, + "args": { + "External id": 14049,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14048 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342909, "dur": 5, + "args": { + "External id": 14050,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14049 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793342915, "dur": 14, + "args": { + "External id": 14051,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14050 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793342917, "dur": 10, + "args": { + "External id": 14052,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14051 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342947, "dur": 5, + "args": { + "External id": 14053,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14052 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793342956, "dur": 21, + "args": { + "External id": 14054,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14053 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342957, "dur": 5, + "args": { + "External id": 14055,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14054 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793342963, "dur": 13, + "args": { + "External id": 14056,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14055 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793342964, "dur": 11, + "args": { + "External id": 14057,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14056 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793342994, "dur": 6, + "args": { + "External id": 14058,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14057 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793343004, "dur": 20, + "args": { + "External id": 14059,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14058 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343005, "dur": 4, + "args": { + "External id": 14060,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14059 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793343010, "dur": 14, + "args": { + "External id": 14061,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14060 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793343012, "dur": 11, + "args": { + "External id": 14062,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14061 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343043, "dur": 5, + "args": { + "External id": 14063,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14062 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793343052, "dur": 21, + "args": { + "External id": 14064,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14063 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343053, "dur": 4, + "args": { + "External id": 14065,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14064 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793343058, "dur": 14, + "args": { + "External id": 14066,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14065 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793343060, "dur": 11, + "args": { + "External id": 14067,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14066 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343091, "dur": 5, + "args": { + "External id": 14068,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14067 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793343100, "dur": 20, + "args": { + "External id": 14069,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14068 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343101, "dur": 4, + "args": { + "External id": 14070,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14069 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793343106, "dur": 14, + "args": { + "External id": 14071,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14070 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793343108, "dur": 11, + "args": { + "External id": 14072,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14071 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343138, "dur": 5, + "args": { + "External id": 14073,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14072 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793343147, "dur": 21, + "args": { + "External id": 14074,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14073 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343148, "dur": 4, + "args": { + "External id": 14075,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14074 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793343154, "dur": 14, + "args": { + "External id": 14076,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14075 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793343155, "dur": 11, + "args": { + "External id": 14077,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14076 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343186, "dur": 5, + "args": { + "External id": 14078,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14077 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793343195, "dur": 21, + "args": { + "External id": 14079,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14078 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343196, "dur": 4, + "args": { + "External id": 14080,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14079 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793343202, "dur": 13, + "args": { + "External id": 14081,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14080 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793343203, "dur": 11, + "args": { + "External id": 14082,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14081 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343234, "dur": 5, + "args": { + "External id": 14083,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14082 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793343242, "dur": 21, + "args": { + "External id": 14084,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14083 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343244, "dur": 4, + "args": { + "External id": 14085,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14084 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793343249, "dur": 14, + "args": { + "External id": 14086,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14085 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793343251, "dur": 11, + "args": { + "External id": 14087,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14086 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343281, "dur": 6, + "args": { + "External id": 14088,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14087 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793343290, "dur": 22, + "args": { + "External id": 14089,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14088 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343292, "dur": 4, + "args": { + "External id": 14090,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14089 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793343297, "dur": 14, + "args": { + "External id": 14091,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14090 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793343299, "dur": 11, + "args": { + "External id": 14092,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14091 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343329, "dur": 5, + "args": { + "External id": 14093,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14092 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793343338, "dur": 21, + "args": { + "External id": 14094,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14093 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343340, "dur": 4, + "args": { + "External id": 14095,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14094 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793343345, "dur": 14, + "args": { + "External id": 14096,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14095 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793343347, "dur": 11, + "args": { + "External id": 14097,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14096 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343377, "dur": 6, + "args": { + "External id": 14098,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14097 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793343387, "dur": 20, + "args": { + "External id": 14099,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14098 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343388, "dur": 4, + "args": { + "External id": 14100,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14099 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793343393, "dur": 14, + "args": { + "External id": 14101,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14100 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793343395, "dur": 11, + "args": { + "External id": 14102,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14101 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343425, "dur": 6, + "args": { + "External id": 14103,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14102 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793343434, "dur": 21, + "args": { + "External id": 14104,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14103 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343436, "dur": 4, + "args": { + "External id": 14105,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14104 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793343441, "dur": 14, + "args": { + "External id": 14106,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14105 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793343443, "dur": 11, + "args": { + "External id": 14107,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14106 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343473, "dur": 5, + "args": { + "External id": 14108,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14107 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793343482, "dur": 21, + "args": { + "External id": 14109,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14108 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343483, "dur": 5, + "args": { + "External id": 14110,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14109 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793343489, "dur": 13, + "args": { + "External id": 14111,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14110 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793343490, "dur": 11, + "args": { + "External id": 14112,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14111 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343521, "dur": 5, + "args": { + "External id": 14113,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14112 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793343530, "dur": 20, + "args": { + "External id": 14114,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14113 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343531, "dur": 4, + "args": { + "External id": 14115,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14114 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793343536, "dur": 14, + "args": { + "External id": 14116,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14115 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793343538, "dur": 11, + "args": { + "External id": 14117,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14116 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343568, "dur": 5, + "args": { + "External id": 14118,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14117 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793343577, "dur": 21, + "args": { + "External id": 14119,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14118 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343579, "dur": 4, + "args": { + "External id": 14120,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14119 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793343584, "dur": 14, + "args": { + "External id": 14121,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14120 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793343586, "dur": 11, + "args": { + "External id": 14122,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14121 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343616, "dur": 5, + "args": { + "External id": 14123,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14122 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793343625, "dur": 21, + "args": { + "External id": 14124,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14123 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343626, "dur": 5, + "args": { + "External id": 14125,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14124 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793343632, "dur": 14, + "args": { + "External id": 14126,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14125 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793343634, "dur": 11, + "args": { + "External id": 14127,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14126 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343664, "dur": 5, + "args": { + "External id": 14128,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14127 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793343673, "dur": 21, + "args": { + "External id": 14129,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14128 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343674, "dur": 5, + "args": { + "External id": 14130,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14129 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793343680, "dur": 14, + "args": { + "External id": 14131,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14130 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793343682, "dur": 11, + "args": { + "External id": 14132,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14131 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343712, "dur": 5, + "args": { + "External id": 14133,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14132 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793343721, "dur": 21, + "args": { + "External id": 14134,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14133 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343722, "dur": 5, + "args": { + "External id": 14135,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14134 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793343728, "dur": 14, + "args": { + "External id": 14136,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14135 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793343730, "dur": 10, + "args": { + "External id": 14137,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14136 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343759, "dur": 6, + "args": { + "External id": 14138,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14137 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793343769, "dur": 21, + "args": { + "External id": 14139,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14138 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343770, "dur": 4, + "args": { + "External id": 14140,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14139 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793343775, "dur": 15, + "args": { + "External id": 14141,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14140 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793343777, "dur": 11, + "args": { + "External id": 14142,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14141 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343808, "dur": 5, + "args": { + "External id": 14143,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14142 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793343817, "dur": 21, + "args": { + "External id": 14144,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14143 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343818, "dur": 4, + "args": { + "External id": 14145,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14144 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793343824, "dur": 14, + "args": { + "External id": 14146,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14145 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793343825, "dur": 11, + "args": { + "External id": 14147,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14146 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343856, "dur": 5, + "args": { + "External id": 14148,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14147 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793343865, "dur": 20, + "args": { + "External id": 14149,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14148 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343866, "dur": 4, + "args": { + "External id": 14150,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14149 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793343872, "dur": 13, + "args": { + "External id": 14151,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14150 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793343873, "dur": 11, + "args": { + "External id": 14152,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14151 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343903, "dur": 5, + "args": { + "External id": 14153,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14152 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793343912, "dur": 21, + "args": { + "External id": 14154,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14153 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343913, "dur": 5, + "args": { + "External id": 14155,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14154 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793343919, "dur": 14, + "args": { + "External id": 14156,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14155 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793343921, "dur": 11, + "args": { + "External id": 14157,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14156 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343951, "dur": 5, + "args": { + "External id": 14158,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14157 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793343960, "dur": 22, + "args": { + "External id": 14159,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14158 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343962, "dur": 4, + "args": { + "External id": 14160,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14159 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793343967, "dur": 14, + "args": { + "External id": 14161,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14160 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793343969, "dur": 11, + "args": { + "External id": 14162,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14161 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793343999, "dur": 5, + "args": { + "External id": 14163,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14162 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793344008, "dur": 22, + "args": { + "External id": 14164,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14163 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344010, "dur": 4, + "args": { + "External id": 14165,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14164 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793344015, "dur": 15, + "args": { + "External id": 14166,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14165 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793344017, "dur": 11, + "args": { + "External id": 14167,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14166 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344048, "dur": 5, + "args": { + "External id": 14168,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14167 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793344057, "dur": 21, + "args": { + "External id": 14169,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14168 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344058, "dur": 5, + "args": { + "External id": 14170,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14169 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793344064, "dur": 14, + "args": { + "External id": 14171,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14170 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793344065, "dur": 11, + "args": { + "External id": 14172,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14171 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344096, "dur": 5, + "args": { + "External id": 14173,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14172 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793344105, "dur": 21, + "args": { + "External id": 14174,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14173 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344107, "dur": 4, + "args": { + "External id": 14175,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14174 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793344112, "dur": 14, + "args": { + "External id": 14176,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14175 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793344114, "dur": 11, + "args": { + "External id": 14177,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14176 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344144, "dur": 5, + "args": { + "External id": 14178,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14177 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793344153, "dur": 21, + "args": { + "External id": 14179,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14178 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344154, "dur": 5, + "args": { + "External id": 14180,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14179 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793344160, "dur": 14, + "args": { + "External id": 14181,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14180 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793344161, "dur": 11, + "args": { + "External id": 14182,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14181 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344192, "dur": 5, + "args": { + "External id": 14183,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14182 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793344201, "dur": 39, + "args": { + "External id": 14184,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14183 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344202, "dur": 4, + "args": { + "External id": 14185,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14184 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793344207, "dur": 33, + "args": { + "External id": 14186,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14185 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793344209, "dur": 29, + "args": { + "External id": 14187,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14186 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344262, "dur": 6, + "args": { + "External id": 14188,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14187 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793344272, "dur": 22, + "args": { + "External id": 14189,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14188 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344273, "dur": 4, + "args": { + "External id": 14190,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14189 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793344279, "dur": 14, + "args": { + "External id": 14191,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14190 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793344280, "dur": 12, + "args": { + "External id": 14192,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14191 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344311, "dur": 5, + "args": { + "External id": 14193,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14192 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793344320, "dur": 22, + "args": { + "External id": 14194,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14193 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344321, "dur": 5, + "args": { + "External id": 14195,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14194 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793344327, "dur": 14, + "args": { + "External id": 14196,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14195 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793344329, "dur": 11, + "args": { + "External id": 14197,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14196 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344360, "dur": 5, + "args": { + "External id": 14198,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14197 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793344369, "dur": 22, + "args": { + "External id": 14199,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14198 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344370, "dur": 5, + "args": { + "External id": 14200,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14199 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793344376, "dur": 14, + "args": { + "External id": 14201,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14200 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793344378, "dur": 11, + "args": { + "External id": 14202,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14201 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344409, "dur": 5, + "args": { + "External id": 14203,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14202 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793344418, "dur": 21, + "args": { + "External id": 14204,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14203 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344419, "dur": 4, + "args": { + "External id": 14205,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14204 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793344424, "dur": 14, + "args": { + "External id": 14206,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14205 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793344426, "dur": 11, + "args": { + "External id": 14207,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14206 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344457, "dur": 5, + "args": { + "External id": 14208,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14207 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793344466, "dur": 21, + "args": { + "External id": 14209,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14208 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344467, "dur": 4, + "args": { + "External id": 14210,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14209 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793344472, "dur": 14, + "args": { + "External id": 14211,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14210 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793344474, "dur": 11, + "args": { + "External id": 14212,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14211 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344505, "dur": 5, + "args": { + "External id": 14213,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14212 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793344514, "dur": 21, + "args": { + "External id": 14214,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14213 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344515, "dur": 4, + "args": { + "External id": 14215,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14214 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793344520, "dur": 14, + "args": { + "External id": 14216,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14215 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793344522, "dur": 11, + "args": { + "External id": 14217,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14216 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344553, "dur": 5, + "args": { + "External id": 14218,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14217 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793344561, "dur": 22, + "args": { + "External id": 14219,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14218 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344563, "dur": 4, + "args": { + "External id": 14220,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14219 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793344568, "dur": 14, + "args": { + "External id": 14221,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14220 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793344570, "dur": 11, + "args": { + "External id": 14222,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14221 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344600, "dur": 6, + "args": { + "External id": 14223,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14222 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793344610, "dur": 21, + "args": { + "External id": 14224,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14223 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344611, "dur": 4, + "args": { + "External id": 14225,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14224 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793344616, "dur": 14, + "args": { + "External id": 14226,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14225 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793344618, "dur": 11, + "args": { + "External id": 14227,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14226 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344648, "dur": 6, + "args": { + "External id": 14228,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14227 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793344657, "dur": 22, + "args": { + "External id": 14229,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14228 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344659, "dur": 4, + "args": { + "External id": 14230,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14229 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793344664, "dur": 14, + "args": { + "External id": 14231,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14230 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793344666, "dur": 11, + "args": { + "External id": 14232,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14231 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344697, "dur": 5, + "args": { + "External id": 14233,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14232 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793344706, "dur": 21, + "args": { + "External id": 14234,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14233 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344707, "dur": 4, + "args": { + "External id": 14235,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14234 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793344713, "dur": 13, + "args": { + "External id": 14236,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14235 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793344714, "dur": 11, + "args": { + "External id": 14237,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14236 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344745, "dur": 5, + "args": { + "External id": 14238,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14237 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793344754, "dur": 21, + "args": { + "External id": 14239,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14238 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344755, "dur": 4, + "args": { + "External id": 14240,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14239 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793344761, "dur": 13, + "args": { + "External id": 14241,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14240 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793344762, "dur": 11, + "args": { + "External id": 14242,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14241 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344792, "dur": 5, + "args": { + "External id": 14243,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14242 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793344801, "dur": 23, + "args": { + "External id": 14244,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14243 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344803, "dur": 4, + "args": { + "External id": 14245,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14244 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793344808, "dur": 16, + "args": { + "External id": 14246,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14245 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793344810, "dur": 13, + "args": { + "External id": 14247,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14246 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344844, "dur": 5, + "args": { + "External id": 14248,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14247 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793344853, "dur": 21, + "args": { + "External id": 14249,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14248 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344855, "dur": 4, + "args": { + "External id": 14250,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14249 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793344860, "dur": 14, + "args": { + "External id": 14251,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14250 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793344862, "dur": 11, + "args": { + "External id": 14252,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14251 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344892, "dur": 5, + "args": { + "External id": 14253,"Record function id": 0, "Concrete Inputs": ["[8, 8192]", "5", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14252 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 435, "tid": 435, + "ts": 1776649793344901, "dur": 21, + "args": { + "External id": 14254,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Dims": [[], [], [], [], []], "Ev Idx": 14253 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 435, "tid": 435, + "ts": 1776649793344902, "dur": 4, + "args": { + "External id": 14255,"Record function id": 0, "Concrete Inputs": ["[1024]", "3", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 14254 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 435, "tid": 435, + "ts": 1776649793344908, "dur": 13, + "args": { + "External id": 14256,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["int"], "Input Dims": [[1024]], "Ev Idx": 14255 + } + }, + { + "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 435, "tid": 435, + "ts": 1776649793344909, "dur": 11, + "args": { + "External id": 14257,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["int", "Scalar"], "Input Dims": [[1024], []], "Ev Idx": 14256 + } + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791331726, "dur": 1, + "args": { + "External id": 7480, + "cbid": 210, "correlation": 57317 + } + }, + { + "ph": "f", "id": 57317, "pid": 435, "tid": 435, "ts": 1776649791331726, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791331738, "dur": 41, + "args": { + "External id": 7480, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57318, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57318, "pid": 0, "tid": 7, "ts": 1776649791331738, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791331728, "dur": 7, + "args": { + "External id": 7480, + "cbid": 211, "correlation": 57318 + } + }, + { + "ph": "s", "id": 57318, "pid": 435, "tid": 435, "ts": 1776649791331728, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791331766, "dur": 108, + "args": { + "External id": 7482, + "cbid": 210, "correlation": 57345 + } + }, + { + "ph": "f", "id": 57345, "pid": 435, "tid": 435, "ts": 1776649791331766, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791331890, "dur": 41, + "args": { + "External id": 7482, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57346, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57346, "pid": 0, "tid": 7, "ts": 1776649791331890, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791331877, "dur": 10, + "args": { + "External id": 7482, + "cbid": 211, "correlation": 57346 + } + }, + { + "ph": "s", "id": 57346, "pid": 435, "tid": 435, "ts": 1776649791331877, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791331924, "dur": 0, + "args": { + "External id": 7484, + "cbid": 210, "correlation": 57373 + } + }, + { + "ph": "f", "id": 57373, "pid": 435, "tid": 435, "ts": 1776649791331924, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791331937, "dur": 40, + "args": { + "External id": 7484, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57374, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57374, "pid": 0, "tid": 7, "ts": 1776649791331937, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791331925, "dur": 6, + "args": { + "External id": 7484, + "cbid": 211, "correlation": 57374 + } + }, + { + "ph": "s", "id": 57374, "pid": 435, "tid": 435, "ts": 1776649791331925, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791331956, "dur": 0, + "args": { + "External id": 7486, + "cbid": 210, "correlation": 57401 + } + }, + { + "ph": "f", "id": 57401, "pid": 435, "tid": 435, "ts": 1776649791331956, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791331979, "dur": 41, + "args": { + "External id": 7486, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57402, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57402, "pid": 0, "tid": 7, "ts": 1776649791331979, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791331957, "dur": 6, + "args": { + "External id": 7486, + "cbid": 211, "correlation": 57402 + } + }, + { + "ph": "s", "id": 57402, "pid": 435, "tid": 435, "ts": 1776649791331957, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791331987, "dur": 0, + "args": { + "External id": 7488, + "cbid": 210, "correlation": 57429 + } + }, + { + "ph": "f", "id": 57429, "pid": 435, "tid": 435, "ts": 1776649791331987, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791332021, "dur": 41, + "args": { + "External id": 7488, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57430, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57430, "pid": 0, "tid": 7, "ts": 1776649791332021, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791331988, "dur": 8, + "args": { + "External id": 7488, + "cbid": 211, "correlation": 57430 + } + }, + { + "ph": "s", "id": 57430, "pid": 435, "tid": 435, "ts": 1776649791331988, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791332020, "dur": 31, + "args": { + "External id": 7490, + "cbid": 210, "correlation": 57457 + } + }, + { + "ph": "f", "id": 57457, "pid": 435, "tid": 435, "ts": 1776649791332020, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791332066, "dur": 41, + "args": { + "External id": 7490, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57458, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57458, "pid": 0, "tid": 7, "ts": 1776649791332066, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791332052, "dur": 9, + "args": { + "External id": 7490, + "cbid": 211, "correlation": 57458 + } + }, + { + "ph": "s", "id": 57458, "pid": 435, "tid": 435, "ts": 1776649791332052, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791332091, "dur": 0, + "args": { + "External id": 7492, + "cbid": 210, "correlation": 57485 + } + }, + { + "ph": "f", "id": 57485, "pid": 435, "tid": 435, "ts": 1776649791332091, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791332108, "dur": 40, + "args": { + "External id": 7492, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57486, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57486, "pid": 0, "tid": 7, "ts": 1776649791332108, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791332092, "dur": 6, + "args": { + "External id": 7492, + "cbid": 211, "correlation": 57486 + } + }, + { + "ph": "s", "id": 57486, "pid": 435, "tid": 435, "ts": 1776649791332092, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791332124, "dur": 0, + "args": { + "External id": 7494, + "cbid": 210, "correlation": 57513 + } + }, + { + "ph": "f", "id": 57513, "pid": 435, "tid": 435, "ts": 1776649791332124, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791332150, "dur": 41, + "args": { + "External id": 7494, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57514, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57514, "pid": 0, "tid": 7, "ts": 1776649791332150, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791332126, "dur": 6, + "args": { + "External id": 7494, + "cbid": 211, "correlation": 57514 + } + }, + { + "ph": "s", "id": 57514, "pid": 435, "tid": 435, "ts": 1776649791332126, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791332155, "dur": 0, + "args": { + "External id": 7496, + "cbid": 210, "correlation": 57541 + } + }, + { + "ph": "f", "id": 57541, "pid": 435, "tid": 435, "ts": 1776649791332155, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791332193, "dur": 46, + "args": { + "External id": 7496, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57542, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57542, "pid": 0, "tid": 7, "ts": 1776649791332193, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791332157, "dur": 6, + "args": { + "External id": 7496, + "cbid": 211, "correlation": 57542 + } + }, + { + "ph": "s", "id": 57542, "pid": 435, "tid": 435, "ts": 1776649791332157, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791332186, "dur": 0, + "args": { + "External id": 7498, + "cbid": 210, "correlation": 57569 + } + }, + { + "ph": "f", "id": 57569, "pid": 435, "tid": 435, "ts": 1776649791332186, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791332241, "dur": 49, + "args": { + "External id": 7498, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57570, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57570, "pid": 0, "tid": 7, "ts": 1776649791332241, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791332188, "dur": 7, + "args": { + "External id": 7498, + "cbid": 211, "correlation": 57570 + } + }, + { + "ph": "s", "id": 57570, "pid": 435, "tid": 435, "ts": 1776649791332188, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791332238, "dur": 1, + "args": { + "External id": 7500, + "cbid": 210, "correlation": 57597 + } + }, + { + "ph": "f", "id": 57597, "pid": 435, "tid": 435, "ts": 1776649791332238, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791332291, "dur": 49, + "args": { + "External id": 7500, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57598, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57598, "pid": 0, "tid": 7, "ts": 1776649791332291, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791332240, "dur": 7, + "args": { + "External id": 7500, + "cbid": 211, "correlation": 57598 + } + }, + { + "ph": "s", "id": 57598, "pid": 435, "tid": 435, "ts": 1776649791332240, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791332274, "dur": 0, + "args": { + "External id": 7502, + "cbid": 210, "correlation": 57625 + } + }, + { + "ph": "f", "id": 57625, "pid": 435, "tid": 435, "ts": 1776649791332274, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791332342, "dur": 52, + "args": { + "External id": 7502, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57626, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57626, "pid": 0, "tid": 7, "ts": 1776649791332342, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791332276, "dur": 6, + "args": { + "External id": 7502, + "cbid": 211, "correlation": 57626 + } + }, + { + "ph": "s", "id": 57626, "pid": 435, "tid": 435, "ts": 1776649791332276, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791332307, "dur": 0, + "args": { + "External id": 7504, + "cbid": 210, "correlation": 57653 + } + }, + { + "ph": "f", "id": 57653, "pid": 435, "tid": 435, "ts": 1776649791332307, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791332395, "dur": 49, + "args": { + "External id": 7504, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57654, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57654, "pid": 0, "tid": 7, "ts": 1776649791332395, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791332308, "dur": 6, + "args": { + "External id": 7504, + "cbid": 211, "correlation": 57654 + } + }, + { + "ph": "s", "id": 57654, "pid": 435, "tid": 435, "ts": 1776649791332308, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791332337, "dur": 0, + "args": { + "External id": 7506, + "cbid": 210, "correlation": 57681 + } + }, + { + "ph": "f", "id": 57681, "pid": 435, "tid": 435, "ts": 1776649791332337, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791332446, "dur": 49, + "args": { + "External id": 7506, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57682, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57682, "pid": 0, "tid": 7, "ts": 1776649791332446, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791332339, "dur": 6, + "args": { + "External id": 7506, + "cbid": 211, "correlation": 57682 + } + }, + { + "ph": "s", "id": 57682, "pid": 435, "tid": 435, "ts": 1776649791332339, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791332368, "dur": 0, + "args": { + "External id": 7508, + "cbid": 210, "correlation": 57709 + } + }, + { + "ph": "f", "id": 57709, "pid": 435, "tid": 435, "ts": 1776649791332368, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791332497, "dur": 49, + "args": { + "External id": 7508, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57710, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57710, "pid": 0, "tid": 7, "ts": 1776649791332497, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791332369, "dur": 6, + "args": { + "External id": 7508, + "cbid": 211, "correlation": 57710 + } + }, + { + "ph": "s", "id": 57710, "pid": 435, "tid": 435, "ts": 1776649791332369, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791332400, "dur": 0, + "args": { + "External id": 7510, + "cbid": 210, "correlation": 57737 + } + }, + { + "ph": "f", "id": 57737, "pid": 435, "tid": 435, "ts": 1776649791332400, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791332548, "dur": 49, + "args": { + "External id": 7510, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57738, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57738, "pid": 0, "tid": 7, "ts": 1776649791332548, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791332401, "dur": 6, + "args": { + "External id": 7510, + "cbid": 211, "correlation": 57738 + } + }, + { + "ph": "s", "id": 57738, "pid": 435, "tid": 435, "ts": 1776649791332401, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791332431, "dur": 0, + "args": { + "External id": 7512, + "cbid": 210, "correlation": 57765 + } + }, + { + "ph": "f", "id": 57765, "pid": 435, "tid": 435, "ts": 1776649791332431, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791332599, "dur": 46, + "args": { + "External id": 7512, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57766, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57766, "pid": 0, "tid": 7, "ts": 1776649791332599, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791332432, "dur": 6, + "args": { + "External id": 7512, + "cbid": 211, "correlation": 57766 + } + }, + { + "ph": "s", "id": 57766, "pid": 435, "tid": 435, "ts": 1776649791332432, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791332461, "dur": 0, + "args": { + "External id": 7514, + "cbid": 210, "correlation": 57793 + } + }, + { + "ph": "f", "id": 57793, "pid": 435, "tid": 435, "ts": 1776649791332461, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791332646, "dur": 47, + "args": { + "External id": 7514, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57794, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57794, "pid": 0, "tid": 7, "ts": 1776649791332646, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791332462, "dur": 7, + "args": { + "External id": 7514, + "cbid": 211, "correlation": 57794 + } + }, + { + "ph": "s", "id": 57794, "pid": 435, "tid": 435, "ts": 1776649791332462, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791332494, "dur": 0, + "args": { + "External id": 7516, + "cbid": 210, "correlation": 57821 + } + }, + { + "ph": "f", "id": 57821, "pid": 435, "tid": 435, "ts": 1776649791332494, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791332695, "dur": 46, + "args": { + "External id": 7516, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57822, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57822, "pid": 0, "tid": 7, "ts": 1776649791332695, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791332495, "dur": 5, + "args": { + "External id": 7516, + "cbid": 211, "correlation": 57822 + } + }, + { + "ph": "s", "id": 57822, "pid": 435, "tid": 435, "ts": 1776649791332495, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791332524, "dur": 0, + "args": { + "External id": 7518, + "cbid": 210, "correlation": 57849 + } + }, + { + "ph": "f", "id": 57849, "pid": 435, "tid": 435, "ts": 1776649791332524, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791332743, "dur": 43, + "args": { + "External id": 7518, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57850, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57850, "pid": 0, "tid": 7, "ts": 1776649791332743, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791332526, "dur": 6, + "args": { + "External id": 7518, + "cbid": 211, "correlation": 57850 + } + }, + { + "ph": "s", "id": 57850, "pid": 435, "tid": 435, "ts": 1776649791332526, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791332556, "dur": 0, + "args": { + "External id": 7520, + "cbid": 210, "correlation": 57877 + } + }, + { + "ph": "f", "id": 57877, "pid": 435, "tid": 435, "ts": 1776649791332556, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791332788, "dur": 46, + "args": { + "External id": 7520, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57878, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57878, "pid": 0, "tid": 7, "ts": 1776649791332788, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791332558, "dur": 6, + "args": { + "External id": 7520, + "cbid": 211, "correlation": 57878 + } + }, + { + "ph": "s", "id": 57878, "pid": 435, "tid": 435, "ts": 1776649791332558, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791332587, "dur": 0, + "args": { + "External id": 7522, + "cbid": 210, "correlation": 57905 + } + }, + { + "ph": "f", "id": 57905, "pid": 435, "tid": 435, "ts": 1776649791332587, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791332836, "dur": 44, + "args": { + "External id": 7522, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57906, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57906, "pid": 0, "tid": 7, "ts": 1776649791332836, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791332588, "dur": 5, + "args": { + "External id": 7522, + "cbid": 211, "correlation": 57906 + } + }, + { + "ph": "s", "id": 57906, "pid": 435, "tid": 435, "ts": 1776649791332588, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791332617, "dur": 0, + "args": { + "External id": 7524, + "cbid": 210, "correlation": 57933 + } + }, + { + "ph": "f", "id": 57933, "pid": 435, "tid": 435, "ts": 1776649791332617, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791332882, "dur": 46, + "args": { + "External id": 7524, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57934, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57934, "pid": 0, "tid": 7, "ts": 1776649791332882, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791332618, "dur": 5, + "args": { + "External id": 7524, + "cbid": 211, "correlation": 57934 + } + }, + { + "ph": "s", "id": 57934, "pid": 435, "tid": 435, "ts": 1776649791332618, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791332795, "dur": 0, + "args": { + "External id": 7526, + "cbid": 210, "correlation": 57961 + } + }, + { + "ph": "f", "id": 57961, "pid": 435, "tid": 435, "ts": 1776649791332795, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791332930, "dur": 46, + "args": { + "External id": 7526, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57962, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57962, "pid": 0, "tid": 7, "ts": 1776649791332930, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791332796, "dur": 6, + "args": { + "External id": 7526, + "cbid": 211, "correlation": 57962 + } + }, + { + "ph": "s", "id": 57962, "pid": 435, "tid": 435, "ts": 1776649791332796, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791332827, "dur": 0, + "args": { + "External id": 7528, + "cbid": 210, "correlation": 57989 + } + }, + { + "ph": "f", "id": 57989, "pid": 435, "tid": 435, "ts": 1776649791332827, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791332978, "dur": 47, + "args": { + "External id": 7528, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57990, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57990, "pid": 0, "tid": 7, "ts": 1776649791332978, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791332828, "dur": 6, + "args": { + "External id": 7528, + "cbid": 211, "correlation": 57990 + } + }, + { + "ph": "s", "id": 57990, "pid": 435, "tid": 435, "ts": 1776649791332828, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791332862, "dur": 0, + "args": { + "External id": 7530, + "cbid": 210, "correlation": 58017 + } + }, + { + "ph": "f", "id": 58017, "pid": 435, "tid": 435, "ts": 1776649791332862, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791333026, "dur": 49, + "args": { + "External id": 7530, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58018, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58018, "pid": 0, "tid": 7, "ts": 1776649791333026, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791332863, "dur": 7, + "args": { + "External id": 7530, + "cbid": 211, "correlation": 58018 + } + }, + { + "ph": "s", "id": 58018, "pid": 435, "tid": 435, "ts": 1776649791332863, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791332894, "dur": 0, + "args": { + "External id": 7532, + "cbid": 210, "correlation": 58045 + } + }, + { + "ph": "f", "id": 58045, "pid": 435, "tid": 435, "ts": 1776649791332894, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791333077, "dur": 47, + "args": { + "External id": 7532, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58046, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58046, "pid": 0, "tid": 7, "ts": 1776649791333077, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791332896, "dur": 6, + "args": { + "External id": 7532, + "cbid": 211, "correlation": 58046 + } + }, + { + "ph": "s", "id": 58046, "pid": 435, "tid": 435, "ts": 1776649791332896, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791332925, "dur": 0, + "args": { + "External id": 7534, + "cbid": 210, "correlation": 58073 + } + }, + { + "ph": "f", "id": 58073, "pid": 435, "tid": 435, "ts": 1776649791332925, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791333126, "dur": 49, + "args": { + "External id": 7534, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58074, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58074, "pid": 0, "tid": 7, "ts": 1776649791333126, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791332926, "dur": 6, + "args": { + "External id": 7534, + "cbid": 211, "correlation": 58074 + } + }, + { + "ph": "s", "id": 58074, "pid": 435, "tid": 435, "ts": 1776649791332926, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791332956, "dur": 0, + "args": { + "External id": 7536, + "cbid": 210, "correlation": 58101 + } + }, + { + "ph": "f", "id": 58101, "pid": 435, "tid": 435, "ts": 1776649791332956, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791333176, "dur": 50, + "args": { + "External id": 7536, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58102, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58102, "pid": 0, "tid": 7, "ts": 1776649791333176, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791332957, "dur": 6, + "args": { + "External id": 7536, + "cbid": 211, "correlation": 58102 + } + }, + { + "ph": "s", "id": 58102, "pid": 435, "tid": 435, "ts": 1776649791332957, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791332986, "dur": 0, + "args": { + "External id": 7538, + "cbid": 210, "correlation": 58129 + } + }, + { + "ph": "f", "id": 58129, "pid": 435, "tid": 435, "ts": 1776649791332986, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791333227, "dur": 47, + "args": { + "External id": 7538, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58130, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58130, "pid": 0, "tid": 7, "ts": 1776649791333227, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791332987, "dur": 6, + "args": { + "External id": 7538, + "cbid": 211, "correlation": 58130 + } + }, + { + "ph": "s", "id": 58130, "pid": 435, "tid": 435, "ts": 1776649791332987, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791333016, "dur": 0, + "args": { + "External id": 7540, + "cbid": 210, "correlation": 58157 + } + }, + { + "ph": "f", "id": 58157, "pid": 435, "tid": 435, "ts": 1776649791333016, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791333276, "dur": 48, + "args": { + "External id": 7540, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58158, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58158, "pid": 0, "tid": 7, "ts": 1776649791333276, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791333017, "dur": 5, + "args": { + "External id": 7540, + "cbid": 211, "correlation": 58158 + } + }, + { + "ph": "s", "id": 58158, "pid": 435, "tid": 435, "ts": 1776649791333017, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791333046, "dur": 0, + "args": { + "External id": 7542, + "cbid": 210, "correlation": 58185 + } + }, + { + "ph": "f", "id": 58185, "pid": 435, "tid": 435, "ts": 1776649791333046, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791333326, "dur": 46, + "args": { + "External id": 7542, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58186, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58186, "pid": 0, "tid": 7, "ts": 1776649791333326, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791333047, "dur": 6, + "args": { + "External id": 7542, + "cbid": 211, "correlation": 58186 + } + }, + { + "ph": "s", "id": 58186, "pid": 435, "tid": 435, "ts": 1776649791333047, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791333076, "dur": 0, + "args": { + "External id": 7544, + "cbid": 210, "correlation": 58213 + } + }, + { + "ph": "f", "id": 58213, "pid": 435, "tid": 435, "ts": 1776649791333076, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791333374, "dur": 49, + "args": { + "External id": 7544, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58214, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58214, "pid": 0, "tid": 7, "ts": 1776649791333374, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791333078, "dur": 6, + "args": { + "External id": 7544, + "cbid": 211, "correlation": 58214 + } + }, + { + "ph": "s", "id": 58214, "pid": 435, "tid": 435, "ts": 1776649791333078, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791333106, "dur": 0, + "args": { + "External id": 7546, + "cbid": 210, "correlation": 58241 + } + }, + { + "ph": "f", "id": 58241, "pid": 435, "tid": 435, "ts": 1776649791333106, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791333425, "dur": 49, + "args": { + "External id": 7546, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58242, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58242, "pid": 0, "tid": 7, "ts": 1776649791333425, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791333108, "dur": 6, + "args": { + "External id": 7546, + "cbid": 211, "correlation": 58242 + } + }, + { + "ph": "s", "id": 58242, "pid": 435, "tid": 435, "ts": 1776649791333108, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791333137, "dur": 0, + "args": { + "External id": 7548, + "cbid": 210, "correlation": 58269 + } + }, + { + "ph": "f", "id": 58269, "pid": 435, "tid": 435, "ts": 1776649791333137, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791333476, "dur": 48, + "args": { + "External id": 7548, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58270, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58270, "pid": 0, "tid": 7, "ts": 1776649791333476, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791333138, "dur": 6, + "args": { + "External id": 7548, + "cbid": 211, "correlation": 58270 + } + }, + { + "ph": "s", "id": 58270, "pid": 435, "tid": 435, "ts": 1776649791333138, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791333167, "dur": 0, + "args": { + "External id": 7550, + "cbid": 210, "correlation": 58297 + } + }, + { + "ph": "f", "id": 58297, "pid": 435, "tid": 435, "ts": 1776649791333167, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791333527, "dur": 47, + "args": { + "External id": 7550, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58298, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58298, "pid": 0, "tid": 7, "ts": 1776649791333527, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791333168, "dur": 6, + "args": { + "External id": 7550, + "cbid": 211, "correlation": 58298 + } + }, + { + "ph": "s", "id": 58298, "pid": 435, "tid": 435, "ts": 1776649791333168, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791333197, "dur": 0, + "args": { + "External id": 7552, + "cbid": 210, "correlation": 58325 + } + }, + { + "ph": "f", "id": 58325, "pid": 435, "tid": 435, "ts": 1776649791333197, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791333575, "dur": 44, + "args": { + "External id": 7552, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58326, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58326, "pid": 0, "tid": 7, "ts": 1776649791333575, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791333198, "dur": 6, + "args": { + "External id": 7552, + "cbid": 211, "correlation": 58326 + } + }, + { + "ph": "s", "id": 58326, "pid": 435, "tid": 435, "ts": 1776649791333198, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791333227, "dur": 0, + "args": { + "External id": 7554, + "cbid": 210, "correlation": 58353 + } + }, + { + "ph": "f", "id": 58353, "pid": 435, "tid": 435, "ts": 1776649791333227, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791333621, "dur": 49, + "args": { + "External id": 7554, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58354, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58354, "pid": 0, "tid": 7, "ts": 1776649791333621, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791333229, "dur": 6, + "args": { + "External id": 7554, + "cbid": 211, "correlation": 58354 + } + }, + { + "ph": "s", "id": 58354, "pid": 435, "tid": 435, "ts": 1776649791333229, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791333257, "dur": 0, + "args": { + "External id": 7556, + "cbid": 210, "correlation": 58381 + } + }, + { + "ph": "f", "id": 58381, "pid": 435, "tid": 435, "ts": 1776649791333257, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791333671, "dur": 44, + "args": { + "External id": 7556, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58382, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58382, "pid": 0, "tid": 7, "ts": 1776649791333671, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791333258, "dur": 5, + "args": { + "External id": 7556, + "cbid": 211, "correlation": 58382 + } + }, + { + "ph": "s", "id": 58382, "pid": 435, "tid": 435, "ts": 1776649791333258, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791333287, "dur": 0, + "args": { + "External id": 7558, + "cbid": 210, "correlation": 58409 + } + }, + { + "ph": "f", "id": 58409, "pid": 435, "tid": 435, "ts": 1776649791333287, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791333717, "dur": 46, + "args": { + "External id": 7558, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58410, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58410, "pid": 0, "tid": 7, "ts": 1776649791333717, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791333289, "dur": 6, + "args": { + "External id": 7558, + "cbid": 211, "correlation": 58410 + } + }, + { + "ph": "s", "id": 58410, "pid": 435, "tid": 435, "ts": 1776649791333289, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791333317, "dur": 0, + "args": { + "External id": 7560, + "cbid": 210, "correlation": 58437 + } + }, + { + "ph": "f", "id": 58437, "pid": 435, "tid": 435, "ts": 1776649791333317, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791333765, "dur": 44, + "args": { + "External id": 7560, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58438, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58438, "pid": 0, "tid": 7, "ts": 1776649791333765, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791333319, "dur": 5, + "args": { + "External id": 7560, + "cbid": 211, "correlation": 58438 + } + }, + { + "ph": "s", "id": 58438, "pid": 435, "tid": 435, "ts": 1776649791333319, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791333347, "dur": 0, + "args": { + "External id": 7562, + "cbid": 210, "correlation": 58465 + } + }, + { + "ph": "f", "id": 58465, "pid": 435, "tid": 435, "ts": 1776649791333347, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791333811, "dur": 46, + "args": { + "External id": 7562, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58466, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58466, "pid": 0, "tid": 7, "ts": 1776649791333811, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791333349, "dur": 6, + "args": { + "External id": 7562, + "cbid": 211, "correlation": 58466 + } + }, + { + "ph": "s", "id": 58466, "pid": 435, "tid": 435, "ts": 1776649791333349, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791333378, "dur": 0, + "args": { + "External id": 7564, + "cbid": 210, "correlation": 58493 + } + }, + { + "ph": "f", "id": 58493, "pid": 435, "tid": 435, "ts": 1776649791333378, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791333859, "dur": 51, + "args": { + "External id": 7564, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58494, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58494, "pid": 0, "tid": 7, "ts": 1776649791333859, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791333379, "dur": 5, + "args": { + "External id": 7564, + "cbid": 211, "correlation": 58494 + } + }, + { + "ph": "s", "id": 58494, "pid": 435, "tid": 435, "ts": 1776649791333379, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791333408, "dur": 0, + "args": { + "External id": 7566, + "cbid": 210, "correlation": 58521 + } + }, + { + "ph": "f", "id": 58521, "pid": 435, "tid": 435, "ts": 1776649791333408, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791333912, "dur": 46, + "args": { + "External id": 7566, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58522, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58522, "pid": 0, "tid": 7, "ts": 1776649791333912, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791333409, "dur": 6, + "args": { + "External id": 7566, + "cbid": 211, "correlation": 58522 + } + }, + { + "ph": "s", "id": 58522, "pid": 435, "tid": 435, "ts": 1776649791333409, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791333438, "dur": 0, + "args": { + "External id": 7568, + "cbid": 210, "correlation": 58549 + } + }, + { + "ph": "f", "id": 58549, "pid": 435, "tid": 435, "ts": 1776649791333438, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791333960, "dur": 47, + "args": { + "External id": 7568, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58550, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58550, "pid": 0, "tid": 7, "ts": 1776649791333960, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791333440, "dur": 5, + "args": { + "External id": 7568, + "cbid": 211, "correlation": 58550 + } + }, + { + "ph": "s", "id": 58550, "pid": 435, "tid": 435, "ts": 1776649791333440, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791333468, "dur": 0, + "args": { + "External id": 7570, + "cbid": 210, "correlation": 58577 + } + }, + { + "ph": "f", "id": 58577, "pid": 435, "tid": 435, "ts": 1776649791333468, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791334008, "dur": 47, + "args": { + "External id": 7570, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58578, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58578, "pid": 0, "tid": 7, "ts": 1776649791334008, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791333469, "dur": 6, + "args": { + "External id": 7570, + "cbid": 211, "correlation": 58578 + } + }, + { + "ph": "s", "id": 58578, "pid": 435, "tid": 435, "ts": 1776649791333469, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791333498, "dur": 0, + "args": { + "External id": 7572, + "cbid": 210, "correlation": 58605 + } + }, + { + "ph": "f", "id": 58605, "pid": 435, "tid": 435, "ts": 1776649791333498, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791334057, "dur": 46, + "args": { + "External id": 7572, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58606, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58606, "pid": 0, "tid": 7, "ts": 1776649791334057, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791333500, "dur": 6, + "args": { + "External id": 7572, + "cbid": 211, "correlation": 58606 + } + }, + { + "ph": "s", "id": 58606, "pid": 435, "tid": 435, "ts": 1776649791333500, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791333529, "dur": 0, + "args": { + "External id": 7574, + "cbid": 210, "correlation": 58633 + } + }, + { + "ph": "f", "id": 58633, "pid": 435, "tid": 435, "ts": 1776649791333529, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791334104, "dur": 46, + "args": { + "External id": 7574, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58634, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58634, "pid": 0, "tid": 7, "ts": 1776649791334104, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791333530, "dur": 5, + "args": { + "External id": 7574, + "cbid": 211, "correlation": 58634 + } + }, + { + "ph": "s", "id": 58634, "pid": 435, "tid": 435, "ts": 1776649791333530, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791333559, "dur": 0, + "args": { + "External id": 7576, + "cbid": 210, "correlation": 58661 + } + }, + { + "ph": "f", "id": 58661, "pid": 435, "tid": 435, "ts": 1776649791333559, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791334153, "dur": 48, + "args": { + "External id": 7576, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58662, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58662, "pid": 0, "tid": 7, "ts": 1776649791334153, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791333560, "dur": 6, + "args": { + "External id": 7576, + "cbid": 211, "correlation": 58662 + } + }, + { + "ph": "s", "id": 58662, "pid": 435, "tid": 435, "ts": 1776649791333560, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791333603, "dur": 4, + "args": { + "External id": 58673, + "cbid": 135, "correlation": 58673 + } + }, + { + "ph": "f", "id": 58673, "pid": 435, "tid": 435, "ts": 1776649791333603, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791333619, "dur": 592, + "args": { + "External id": 58678, + "cbid": 165, "correlation": 58678 + } + }, + { + "ph": "s", "id": 58678, "pid": 435, "tid": 435, "ts": 1776649791333619, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649791334214, "dur": 5, + "args": { + "External id": 58681, + "cbid": 139, "correlation": 58681 + } + }, + { + "ph": "f", "id": 58681, "pid": 435, "tid": 435, "ts": 1776649791334214, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791334254, "dur": 1, + "args": { + "External id": 7578, + "cbid": 210, "correlation": 58715 + } + }, + { + "ph": "f", "id": 58715, "pid": 435, "tid": 435, "ts": 1776649791334254, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791334267, "dur": 40, + "args": { + "External id": 7578, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58716, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58716, "pid": 0, "tid": 7, "ts": 1776649791334267, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791334256, "dur": 9, + "args": { + "External id": 7578, + "cbid": 211, "correlation": 58716 + } + }, + { + "ph": "s", "id": 58716, "pid": 435, "tid": 435, "ts": 1776649791334256, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791334291, "dur": 0, + "args": { + "External id": 7580, + "cbid": 210, "correlation": 58743 + } + }, + { + "ph": "f", "id": 58743, "pid": 435, "tid": 435, "ts": 1776649791334291, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791334309, "dur": 43, + "args": { + "External id": 7580, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58744, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58744, "pid": 0, "tid": 7, "ts": 1776649791334309, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791334292, "dur": 6, + "args": { + "External id": 7580, + "cbid": 211, "correlation": 58744 + } + }, + { + "ph": "s", "id": 58744, "pid": 435, "tid": 435, "ts": 1776649791334292, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791334322, "dur": 0, + "args": { + "External id": 7582, + "cbid": 210, "correlation": 58771 + } + }, + { + "ph": "f", "id": 58771, "pid": 435, "tid": 435, "ts": 1776649791334322, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791334354, "dur": 46, + "args": { + "External id": 7582, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58772, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58772, "pid": 0, "tid": 7, "ts": 1776649791334354, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791334323, "dur": 6, + "args": { + "External id": 7582, + "cbid": 211, "correlation": 58772 + } + }, + { + "ph": "s", "id": 58772, "pid": 435, "tid": 435, "ts": 1776649791334323, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791334353, "dur": 0, + "args": { + "External id": 7584, + "cbid": 210, "correlation": 58799 + } + }, + { + "ph": "f", "id": 58799, "pid": 435, "tid": 435, "ts": 1776649791334353, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791334402, "dur": 46, + "args": { + "External id": 7584, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58800, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58800, "pid": 0, "tid": 7, "ts": 1776649791334402, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791334354, "dur": 6, + "args": { + "External id": 7584, + "cbid": 211, "correlation": 58800 + } + }, + { + "ph": "s", "id": 58800, "pid": 435, "tid": 435, "ts": 1776649791334354, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791334383, "dur": 0, + "args": { + "External id": 7586, + "cbid": 210, "correlation": 58827 + } + }, + { + "ph": "f", "id": 58827, "pid": 435, "tid": 435, "ts": 1776649791334383, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791334450, "dur": 47, + "args": { + "External id": 7586, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58828, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58828, "pid": 0, "tid": 7, "ts": 1776649791334450, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791334385, "dur": 5, + "args": { + "External id": 7586, + "cbid": 211, "correlation": 58828 + } + }, + { + "ph": "s", "id": 58828, "pid": 435, "tid": 435, "ts": 1776649791334385, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791334413, "dur": 0, + "args": { + "External id": 7588, + "cbid": 210, "correlation": 58855 + } + }, + { + "ph": "f", "id": 58855, "pid": 435, "tid": 435, "ts": 1776649791334413, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791334499, "dur": 49, + "args": { + "External id": 7588, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58856, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58856, "pid": 0, "tid": 7, "ts": 1776649791334499, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791334415, "dur": 6, + "args": { + "External id": 7588, + "cbid": 211, "correlation": 58856 + } + }, + { + "ph": "s", "id": 58856, "pid": 435, "tid": 435, "ts": 1776649791334415, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791334444, "dur": 0, + "args": { + "External id": 7590, + "cbid": 210, "correlation": 58883 + } + }, + { + "ph": "f", "id": 58883, "pid": 435, "tid": 435, "ts": 1776649791334444, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791334549, "dur": 49, + "args": { + "External id": 7590, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58884, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58884, "pid": 0, "tid": 7, "ts": 1776649791334549, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791334445, "dur": 5, + "args": { + "External id": 7590, + "cbid": 211, "correlation": 58884 + } + }, + { + "ph": "s", "id": 58884, "pid": 435, "tid": 435, "ts": 1776649791334445, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791334473, "dur": 0, + "args": { + "External id": 7592, + "cbid": 210, "correlation": 58911 + } + }, + { + "ph": "f", "id": 58911, "pid": 435, "tid": 435, "ts": 1776649791334473, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791334600, "dur": 44, + "args": { + "External id": 7592, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58912, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58912, "pid": 0, "tid": 7, "ts": 1776649791334600, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791334474, "dur": 6, + "args": { + "External id": 7592, + "cbid": 211, "correlation": 58912 + } + }, + { + "ph": "s", "id": 58912, "pid": 435, "tid": 435, "ts": 1776649791334474, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791334503, "dur": 0, + "args": { + "External id": 7594, + "cbid": 210, "correlation": 58939 + } + }, + { + "ph": "f", "id": 58939, "pid": 435, "tid": 435, "ts": 1776649791334503, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791334646, "dur": 41, + "args": { + "External id": 7594, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58940, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58940, "pid": 0, "tid": 7, "ts": 1776649791334646, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791334505, "dur": 6, + "args": { + "External id": 7594, + "cbid": 211, "correlation": 58940 + } + }, + { + "ph": "s", "id": 58940, "pid": 435, "tid": 435, "ts": 1776649791334505, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791334535, "dur": 0, + "args": { + "External id": 7596, + "cbid": 210, "correlation": 58967 + } + }, + { + "ph": "f", "id": 58967, "pid": 435, "tid": 435, "ts": 1776649791334535, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791334689, "dur": 48, + "args": { + "External id": 7596, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 58968, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 58968, "pid": 0, "tid": 7, "ts": 1776649791334689, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791334536, "dur": 6, + "args": { + "External id": 7596, + "cbid": 211, "correlation": 58968 + } + }, + { + "ph": "s", "id": 58968, "pid": 435, "tid": 435, "ts": 1776649791334536, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791334561, "dur": 180, + "args": { + "External id": 58974, + "cbid": 165, "correlation": 58974 + } + }, + { + "ph": "s", "id": 58974, "pid": 435, "tid": 435, "ts": 1776649791334561, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791334762, "dur": 3, + "args": { + "External id": 58983, + "cbid": 135, "correlation": 58983 + } + }, + { + "ph": "f", "id": 58983, "pid": 435, "tid": 435, "ts": 1776649791334762, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791334792, "dur": 1, + "args": { + "External id": 7598, + "cbid": 210, "correlation": 59009 + } + }, + { + "ph": "f", "id": 59009, "pid": 435, "tid": 435, "ts": 1776649791334792, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791334804, "dur": 40, + "args": { + "External id": 7598, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59010, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59010, "pid": 0, "tid": 7, "ts": 1776649791334804, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791334794, "dur": 7, + "args": { + "External id": 7598, + "cbid": 211, "correlation": 59010 + } + }, + { + "ph": "s", "id": 59010, "pid": 435, "tid": 435, "ts": 1776649791334794, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791334827, "dur": 0, + "args": { + "External id": 7600, + "cbid": 210, "correlation": 59037 + } + }, + { + "ph": "f", "id": 59037, "pid": 435, "tid": 435, "ts": 1776649791334827, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791334846, "dur": 43, + "args": { + "External id": 7600, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59038, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59038, "pid": 0, "tid": 7, "ts": 1776649791334846, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791334829, "dur": 6, + "args": { + "External id": 7600, + "cbid": 211, "correlation": 59038 + } + }, + { + "ph": "s", "id": 59038, "pid": 435, "tid": 435, "ts": 1776649791334829, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791334858, "dur": 0, + "args": { + "External id": 7602, + "cbid": 210, "correlation": 59065 + } + }, + { + "ph": "f", "id": 59065, "pid": 435, "tid": 435, "ts": 1776649791334858, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791334891, "dur": 49, + "args": { + "External id": 7602, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59066, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59066, "pid": 0, "tid": 7, "ts": 1776649791334891, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791334859, "dur": 6, + "args": { + "External id": 7602, + "cbid": 211, "correlation": 59066 + } + }, + { + "ph": "s", "id": 59066, "pid": 435, "tid": 435, "ts": 1776649791334859, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791334889, "dur": 0, + "args": { + "External id": 7604, + "cbid": 210, "correlation": 59093 + } + }, + { + "ph": "f", "id": 59093, "pid": 435, "tid": 435, "ts": 1776649791334889, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791334942, "dur": 46, + "args": { + "External id": 7604, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59094, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59094, "pid": 0, "tid": 7, "ts": 1776649791334942, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791334890, "dur": 6, + "args": { + "External id": 7604, + "cbid": 211, "correlation": 59094 + } + }, + { + "ph": "s", "id": 59094, "pid": 435, "tid": 435, "ts": 1776649791334890, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791334919, "dur": 0, + "args": { + "External id": 7606, + "cbid": 210, "correlation": 59121 + } + }, + { + "ph": "f", "id": 59121, "pid": 435, "tid": 435, "ts": 1776649791334919, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791334990, "dur": 49, + "args": { + "External id": 7606, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59122, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59122, "pid": 0, "tid": 7, "ts": 1776649791334990, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791334920, "dur": 5, + "args": { + "External id": 7606, + "cbid": 211, "correlation": 59122 + } + }, + { + "ph": "s", "id": 59122, "pid": 435, "tid": 435, "ts": 1776649791334920, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791334948, "dur": 0, + "args": { + "External id": 7608, + "cbid": 210, "correlation": 59149 + } + }, + { + "ph": "f", "id": 59149, "pid": 435, "tid": 435, "ts": 1776649791334948, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791335040, "dur": 46, + "args": { + "External id": 7608, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59150, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59150, "pid": 0, "tid": 7, "ts": 1776649791335040, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791334950, "dur": 11, + "args": { + "External id": 7608, + "cbid": 211, "correlation": 59150 + } + }, + { + "ph": "s", "id": 59150, "pid": 435, "tid": 435, "ts": 1776649791334950, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791334986, "dur": 0, + "args": { + "External id": 7610, + "cbid": 210, "correlation": 59177 + } + }, + { + "ph": "f", "id": 59177, "pid": 435, "tid": 435, "ts": 1776649791334986, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791335088, "dur": 50, + "args": { + "External id": 7610, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59178, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59178, "pid": 0, "tid": 7, "ts": 1776649791335088, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791334988, "dur": 6, + "args": { + "External id": 7610, + "cbid": 211, "correlation": 59178 + } + }, + { + "ph": "s", "id": 59178, "pid": 435, "tid": 435, "ts": 1776649791334988, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335017, "dur": 0, + "args": { + "External id": 7612, + "cbid": 210, "correlation": 59205 + } + }, + { + "ph": "f", "id": 59205, "pid": 435, "tid": 435, "ts": 1776649791335017, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791335140, "dur": 47, + "args": { + "External id": 7612, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59206, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59206, "pid": 0, "tid": 7, "ts": 1776649791335140, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335018, "dur": 6, + "args": { + "External id": 7612, + "cbid": 211, "correlation": 59206 + } + }, + { + "ph": "s", "id": 59206, "pid": 435, "tid": 435, "ts": 1776649791335018, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335047, "dur": 0, + "args": { + "External id": 7614, + "cbid": 210, "correlation": 59233 + } + }, + { + "ph": "f", "id": 59233, "pid": 435, "tid": 435, "ts": 1776649791335047, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791335189, "dur": 47, + "args": { + "External id": 7614, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59234, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59234, "pid": 0, "tid": 7, "ts": 1776649791335189, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335049, "dur": 5, + "args": { + "External id": 7614, + "cbid": 211, "correlation": 59234 + } + }, + { + "ph": "s", "id": 59234, "pid": 435, "tid": 435, "ts": 1776649791335049, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335077, "dur": 0, + "args": { + "External id": 7616, + "cbid": 210, "correlation": 59261 + } + }, + { + "ph": "f", "id": 59261, "pid": 435, "tid": 435, "ts": 1776649791335077, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791335237, "dur": 44, + "args": { + "External id": 7616, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59262, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59262, "pid": 0, "tid": 7, "ts": 1776649791335237, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335079, "dur": 5, + "args": { + "External id": 7616, + "cbid": 211, "correlation": 59262 + } + }, + { + "ph": "s", "id": 59262, "pid": 435, "tid": 435, "ts": 1776649791335079, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335107, "dur": 0, + "args": { + "External id": 7618, + "cbid": 210, "correlation": 59289 + } + }, + { + "ph": "f", "id": 59289, "pid": 435, "tid": 435, "ts": 1776649791335107, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791335283, "dur": 48, + "args": { + "External id": 7618, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59290, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59290, "pid": 0, "tid": 7, "ts": 1776649791335283, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335109, "dur": 6, + "args": { + "External id": 7618, + "cbid": 211, "correlation": 59290 + } + }, + { + "ph": "s", "id": 59290, "pid": 435, "tid": 435, "ts": 1776649791335109, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335138, "dur": 0, + "args": { + "External id": 7620, + "cbid": 210, "correlation": 59317 + } + }, + { + "ph": "f", "id": 59317, "pid": 435, "tid": 435, "ts": 1776649791335138, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791335333, "dur": 46, + "args": { + "External id": 7620, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59318, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59318, "pid": 0, "tid": 7, "ts": 1776649791335333, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335139, "dur": 6, + "args": { + "External id": 7620, + "cbid": 211, "correlation": 59318 + } + }, + { + "ph": "s", "id": 59318, "pid": 435, "tid": 435, "ts": 1776649791335139, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335219, "dur": 0, + "args": { + "External id": 7622, + "cbid": 210, "correlation": 59345 + } + }, + { + "ph": "f", "id": 59345, "pid": 435, "tid": 435, "ts": 1776649791335219, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791335380, "dur": 46, + "args": { + "External id": 7622, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59346, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59346, "pid": 0, "tid": 7, "ts": 1776649791335380, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335221, "dur": 6, + "args": { + "External id": 7622, + "cbid": 211, "correlation": 59346 + } + }, + { + "ph": "s", "id": 59346, "pid": 435, "tid": 435, "ts": 1776649791335221, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335250, "dur": 0, + "args": { + "External id": 7624, + "cbid": 210, "correlation": 59373 + } + }, + { + "ph": "f", "id": 59373, "pid": 435, "tid": 435, "ts": 1776649791335250, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791335428, "dur": 44, + "args": { + "External id": 7624, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59374, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59374, "pid": 0, "tid": 7, "ts": 1776649791335428, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335252, "dur": 6, + "args": { + "External id": 7624, + "cbid": 211, "correlation": 59374 + } + }, + { + "ph": "s", "id": 59374, "pid": 435, "tid": 435, "ts": 1776649791335252, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335281, "dur": 0, + "args": { + "External id": 7626, + "cbid": 210, "correlation": 59401 + } + }, + { + "ph": "f", "id": 59401, "pid": 435, "tid": 435, "ts": 1776649791335281, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791335474, "dur": 46, + "args": { + "External id": 7626, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59402, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59402, "pid": 0, "tid": 7, "ts": 1776649791335474, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335282, "dur": 6, + "args": { + "External id": 7626, + "cbid": 211, "correlation": 59402 + } + }, + { + "ph": "s", "id": 59402, "pid": 435, "tid": 435, "ts": 1776649791335282, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335311, "dur": 0, + "args": { + "External id": 7628, + "cbid": 210, "correlation": 59429 + } + }, + { + "ph": "f", "id": 59429, "pid": 435, "tid": 435, "ts": 1776649791335311, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791335521, "dur": 49, + "args": { + "External id": 7628, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59430, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59430, "pid": 0, "tid": 7, "ts": 1776649791335521, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335312, "dur": 5, + "args": { + "External id": 7628, + "cbid": 211, "correlation": 59430 + } + }, + { + "ph": "s", "id": 59430, "pid": 435, "tid": 435, "ts": 1776649791335312, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335340, "dur": 0, + "args": { + "External id": 7630, + "cbid": 210, "correlation": 59457 + } + }, + { + "ph": "f", "id": 59457, "pid": 435, "tid": 435, "ts": 1776649791335340, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791335572, "dur": 49, + "args": { + "External id": 7630, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59458, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59458, "pid": 0, "tid": 7, "ts": 1776649791335572, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335341, "dur": 6, + "args": { + "External id": 7630, + "cbid": 211, "correlation": 59458 + } + }, + { + "ph": "s", "id": 59458, "pid": 435, "tid": 435, "ts": 1776649791335341, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335370, "dur": 0, + "args": { + "External id": 7632, + "cbid": 210, "correlation": 59485 + } + }, + { + "ph": "f", "id": 59485, "pid": 435, "tid": 435, "ts": 1776649791335370, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791335622, "dur": 46, + "args": { + "External id": 7632, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59486, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59486, "pid": 0, "tid": 7, "ts": 1776649791335622, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335371, "dur": 6, + "args": { + "External id": 7632, + "cbid": 211, "correlation": 59486 + } + }, + { + "ph": "s", "id": 59486, "pid": 435, "tid": 435, "ts": 1776649791335371, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335400, "dur": 0, + "args": { + "External id": 7634, + "cbid": 210, "correlation": 59513 + } + }, + { + "ph": "f", "id": 59513, "pid": 435, "tid": 435, "ts": 1776649791335400, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791335671, "dur": 46, + "args": { + "External id": 7634, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59514, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59514, "pid": 0, "tid": 7, "ts": 1776649791335671, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335401, "dur": 6, + "args": { + "External id": 7634, + "cbid": 211, "correlation": 59514 + } + }, + { + "ph": "s", "id": 59514, "pid": 435, "tid": 435, "ts": 1776649791335401, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335430, "dur": 0, + "args": { + "External id": 7636, + "cbid": 210, "correlation": 59541 + } + }, + { + "ph": "f", "id": 59541, "pid": 435, "tid": 435, "ts": 1776649791335430, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791335719, "dur": 44, + "args": { + "External id": 7636, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59542, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59542, "pid": 0, "tid": 7, "ts": 1776649791335719, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335431, "dur": 6, + "args": { + "External id": 7636, + "cbid": 211, "correlation": 59542 + } + }, + { + "ph": "s", "id": 59542, "pid": 435, "tid": 435, "ts": 1776649791335431, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335460, "dur": 0, + "args": { + "External id": 7638, + "cbid": 210, "correlation": 59569 + } + }, + { + "ph": "f", "id": 59569, "pid": 435, "tid": 435, "ts": 1776649791335460, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791335764, "dur": 49, + "args": { + "External id": 7638, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59570, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59570, "pid": 0, "tid": 7, "ts": 1776649791335764, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335461, "dur": 6, + "args": { + "External id": 7638, + "cbid": 211, "correlation": 59570 + } + }, + { + "ph": "s", "id": 59570, "pid": 435, "tid": 435, "ts": 1776649791335461, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335490, "dur": 0, + "args": { + "External id": 7640, + "cbid": 210, "correlation": 59597 + } + }, + { + "ph": "f", "id": 59597, "pid": 435, "tid": 435, "ts": 1776649791335490, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791335816, "dur": 48, + "args": { + "External id": 7640, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59598, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59598, "pid": 0, "tid": 7, "ts": 1776649791335816, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335491, "dur": 6, + "args": { + "External id": 7640, + "cbid": 211, "correlation": 59598 + } + }, + { + "ph": "s", "id": 59598, "pid": 435, "tid": 435, "ts": 1776649791335491, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335520, "dur": 0, + "args": { + "External id": 7642, + "cbid": 210, "correlation": 59625 + } + }, + { + "ph": "f", "id": 59625, "pid": 435, "tid": 435, "ts": 1776649791335520, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791335866, "dur": 49, + "args": { + "External id": 7642, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59626, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59626, "pid": 0, "tid": 7, "ts": 1776649791335866, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335522, "dur": 6, + "args": { + "External id": 7642, + "cbid": 211, "correlation": 59626 + } + }, + { + "ph": "s", "id": 59626, "pid": 435, "tid": 435, "ts": 1776649791335522, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335551, "dur": 0, + "args": { + "External id": 7644, + "cbid": 210, "correlation": 59653 + } + }, + { + "ph": "f", "id": 59653, "pid": 435, "tid": 435, "ts": 1776649791335551, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791335917, "dur": 49, + "args": { + "External id": 7644, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59654, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59654, "pid": 0, "tid": 7, "ts": 1776649791335917, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335552, "dur": 5, + "args": { + "External id": 7644, + "cbid": 211, "correlation": 59654 + } + }, + { + "ph": "s", "id": 59654, "pid": 435, "tid": 435, "ts": 1776649791335552, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335580, "dur": 0, + "args": { + "External id": 7646, + "cbid": 210, "correlation": 59681 + } + }, + { + "ph": "f", "id": 59681, "pid": 435, "tid": 435, "ts": 1776649791335580, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791335968, "dur": 46, + "args": { + "External id": 7646, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59682, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59682, "pid": 0, "tid": 7, "ts": 1776649791335968, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335582, "dur": 6, + "args": { + "External id": 7646, + "cbid": 211, "correlation": 59682 + } + }, + { + "ph": "s", "id": 59682, "pid": 435, "tid": 435, "ts": 1776649791335582, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335610, "dur": 0, + "args": { + "External id": 7648, + "cbid": 210, "correlation": 59709 + } + }, + { + "ph": "f", "id": 59709, "pid": 435, "tid": 435, "ts": 1776649791335610, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791336016, "dur": 47, + "args": { + "External id": 7648, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59710, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59710, "pid": 0, "tid": 7, "ts": 1776649791336016, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335612, "dur": 5, + "args": { + "External id": 7648, + "cbid": 211, "correlation": 59710 + } + }, + { + "ph": "s", "id": 59710, "pid": 435, "tid": 435, "ts": 1776649791335612, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335640, "dur": 0, + "args": { + "External id": 7650, + "cbid": 210, "correlation": 59737 + } + }, + { + "ph": "f", "id": 59737, "pid": 435, "tid": 435, "ts": 1776649791335640, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791336064, "dur": 44, + "args": { + "External id": 7650, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59738, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59738, "pid": 0, "tid": 7, "ts": 1776649791336064, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335641, "dur": 5, + "args": { + "External id": 7650, + "cbid": 211, "correlation": 59738 + } + }, + { + "ph": "s", "id": 59738, "pid": 435, "tid": 435, "ts": 1776649791335641, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335669, "dur": 0, + "args": { + "External id": 7652, + "cbid": 210, "correlation": 59765 + } + }, + { + "ph": "f", "id": 59765, "pid": 435, "tid": 435, "ts": 1776649791335669, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791336110, "dur": 46, + "args": { + "External id": 7652, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59766, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59766, "pid": 0, "tid": 7, "ts": 1776649791336110, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335671, "dur": 6, + "args": { + "External id": 7652, + "cbid": 211, "correlation": 59766 + } + }, + { + "ph": "s", "id": 59766, "pid": 435, "tid": 435, "ts": 1776649791335671, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335701, "dur": 0, + "args": { + "External id": 7654, + "cbid": 210, "correlation": 59793 + } + }, + { + "ph": "f", "id": 59793, "pid": 435, "tid": 435, "ts": 1776649791335701, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791336158, "dur": 49, + "args": { + "External id": 7654, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59794, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59794, "pid": 0, "tid": 7, "ts": 1776649791336158, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335702, "dur": 5, + "args": { + "External id": 7654, + "cbid": 211, "correlation": 59794 + } + }, + { + "ph": "s", "id": 59794, "pid": 435, "tid": 435, "ts": 1776649791335702, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335731, "dur": 0, + "args": { + "External id": 7656, + "cbid": 210, "correlation": 59821 + } + }, + { + "ph": "f", "id": 59821, "pid": 435, "tid": 435, "ts": 1776649791335731, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791336208, "dur": 46, + "args": { + "External id": 7656, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59822, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59822, "pid": 0, "tid": 7, "ts": 1776649791336208, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335732, "dur": 5, + "args": { + "External id": 7656, + "cbid": 211, "correlation": 59822 + } + }, + { + "ph": "s", "id": 59822, "pid": 435, "tid": 435, "ts": 1776649791335732, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335760, "dur": 0, + "args": { + "External id": 7658, + "cbid": 210, "correlation": 59849 + } + }, + { + "ph": "f", "id": 59849, "pid": 435, "tid": 435, "ts": 1776649791335760, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791336256, "dur": 49, + "args": { + "External id": 7658, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59850, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59850, "pid": 0, "tid": 7, "ts": 1776649791336256, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335761, "dur": 6, + "args": { + "External id": 7658, + "cbid": 211, "correlation": 59850 + } + }, + { + "ph": "s", "id": 59850, "pid": 435, "tid": 435, "ts": 1776649791335761, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335790, "dur": 0, + "args": { + "External id": 7660, + "cbid": 210, "correlation": 59877 + } + }, + { + "ph": "f", "id": 59877, "pid": 435, "tid": 435, "ts": 1776649791335790, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791336307, "dur": 49, + "args": { + "External id": 7660, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59878, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59878, "pid": 0, "tid": 7, "ts": 1776649791336307, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335792, "dur": 6, + "args": { + "External id": 7660, + "cbid": 211, "correlation": 59878 + } + }, + { + "ph": "s", "id": 59878, "pid": 435, "tid": 435, "ts": 1776649791335792, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335820, "dur": 0, + "args": { + "External id": 7662, + "cbid": 210, "correlation": 59905 + } + }, + { + "ph": "f", "id": 59905, "pid": 435, "tid": 435, "ts": 1776649791335820, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791336357, "dur": 46, + "args": { + "External id": 7662, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59906, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59906, "pid": 0, "tid": 7, "ts": 1776649791336357, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335821, "dur": 5, + "args": { + "External id": 7662, + "cbid": 211, "correlation": 59906 + } + }, + { + "ph": "s", "id": 59906, "pid": 435, "tid": 435, "ts": 1776649791335821, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335850, "dur": 0, + "args": { + "External id": 7664, + "cbid": 210, "correlation": 59933 + } + }, + { + "ph": "f", "id": 59933, "pid": 435, "tid": 435, "ts": 1776649791335850, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791336406, "dur": 46, + "args": { + "External id": 7664, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59934, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59934, "pid": 0, "tid": 7, "ts": 1776649791336406, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335851, "dur": 5, + "args": { + "External id": 7664, + "cbid": 211, "correlation": 59934 + } + }, + { + "ph": "s", "id": 59934, "pid": 435, "tid": 435, "ts": 1776649791335851, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335880, "dur": 0, + "args": { + "External id": 7666, + "cbid": 210, "correlation": 59961 + } + }, + { + "ph": "f", "id": 59961, "pid": 435, "tid": 435, "ts": 1776649791335880, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791336453, "dur": 50, + "args": { + "External id": 7666, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59962, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59962, "pid": 0, "tid": 7, "ts": 1776649791336453, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335881, "dur": 6, + "args": { + "External id": 7666, + "cbid": 211, "correlation": 59962 + } + }, + { + "ph": "s", "id": 59962, "pid": 435, "tid": 435, "ts": 1776649791335881, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335910, "dur": 0, + "args": { + "External id": 7668, + "cbid": 210, "correlation": 59989 + } + }, + { + "ph": "f", "id": 59989, "pid": 435, "tid": 435, "ts": 1776649791335910, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791336505, "dur": 43, + "args": { + "External id": 7668, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 59990, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 59990, "pid": 0, "tid": 7, "ts": 1776649791336505, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335911, "dur": 5, + "args": { + "External id": 7668, + "cbid": 211, "correlation": 59990 + } + }, + { + "ph": "s", "id": 59990, "pid": 435, "tid": 435, "ts": 1776649791335911, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335939, "dur": 0, + "args": { + "External id": 7670, + "cbid": 210, "correlation": 60017 + } + }, + { + "ph": "f", "id": 60017, "pid": 435, "tid": 435, "ts": 1776649791335939, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791336550, "dur": 46, + "args": { + "External id": 7670, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60018, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60018, "pid": 0, "tid": 7, "ts": 1776649791336550, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335940, "dur": 6, + "args": { + "External id": 7670, + "cbid": 211, "correlation": 60018 + } + }, + { + "ph": "s", "id": 60018, "pid": 435, "tid": 435, "ts": 1776649791335940, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335969, "dur": 0, + "args": { + "External id": 7672, + "cbid": 210, "correlation": 60045 + } + }, + { + "ph": "f", "id": 60045, "pid": 435, "tid": 435, "ts": 1776649791335969, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791336598, "dur": 44, + "args": { + "External id": 7672, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60046, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60046, "pid": 0, "tid": 7, "ts": 1776649791336598, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335970, "dur": 5, + "args": { + "External id": 7672, + "cbid": 211, "correlation": 60046 + } + }, + { + "ph": "s", "id": 60046, "pid": 435, "tid": 435, "ts": 1776649791335970, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791335998, "dur": 0, + "args": { + "External id": 7674, + "cbid": 210, "correlation": 60073 + } + }, + { + "ph": "f", "id": 60073, "pid": 435, "tid": 435, "ts": 1776649791335998, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791336644, "dur": 47, + "args": { + "External id": 7674, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60074, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60074, "pid": 0, "tid": 7, "ts": 1776649791336644, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791335999, "dur": 6, + "args": { + "External id": 7674, + "cbid": 211, "correlation": 60074 + } + }, + { + "ph": "s", "id": 60074, "pid": 435, "tid": 435, "ts": 1776649791335999, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791336028, "dur": 0, + "args": { + "External id": 7676, + "cbid": 210, "correlation": 60101 + } + }, + { + "ph": "f", "id": 60101, "pid": 435, "tid": 435, "ts": 1776649791336028, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791336692, "dur": 48, + "args": { + "External id": 7676, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60102, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60102, "pid": 0, "tid": 7, "ts": 1776649791336692, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791336029, "dur": 6, + "args": { + "External id": 7676, + "cbid": 211, "correlation": 60102 + } + }, + { + "ph": "s", "id": 60102, "pid": 435, "tid": 435, "ts": 1776649791336029, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791336058, "dur": 0, + "args": { + "External id": 7678, + "cbid": 210, "correlation": 60129 + } + }, + { + "ph": "f", "id": 60129, "pid": 435, "tid": 435, "ts": 1776649791336058, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791336743, "dur": 46, + "args": { + "External id": 7678, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60130, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60130, "pid": 0, "tid": 7, "ts": 1776649791336743, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791336059, "dur": 5, + "args": { + "External id": 7678, + "cbid": 211, "correlation": 60130 + } + }, + { + "ph": "s", "id": 60130, "pid": 435, "tid": 435, "ts": 1776649791336059, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791336087, "dur": 0, + "args": { + "External id": 7680, + "cbid": 210, "correlation": 60157 + } + }, + { + "ph": "f", "id": 60157, "pid": 435, "tid": 435, "ts": 1776649791336087, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791336791, "dur": 46, + "args": { + "External id": 7680, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60158, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60158, "pid": 0, "tid": 7, "ts": 1776649791336791, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791336089, "dur": 6, + "args": { + "External id": 7680, + "cbid": 211, "correlation": 60158 + } + }, + { + "ph": "s", "id": 60158, "pid": 435, "tid": 435, "ts": 1776649791336089, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791336128, "dur": 0, + "args": { + "External id": 7682, + "cbid": 210, "correlation": 60185 + } + }, + { + "ph": "f", "id": 60185, "pid": 435, "tid": 435, "ts": 1776649791336128, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791336839, "dur": 47, + "args": { + "External id": 7682, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60186, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60186, "pid": 0, "tid": 7, "ts": 1776649791336839, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791336129, "dur": 6, + "args": { + "External id": 7682, + "cbid": 211, "correlation": 60186 + } + }, + { + "ph": "s", "id": 60186, "pid": 435, "tid": 435, "ts": 1776649791336129, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791336158, "dur": 0, + "args": { + "External id": 7684, + "cbid": 210, "correlation": 60213 + } + }, + { + "ph": "f", "id": 60213, "pid": 435, "tid": 435, "ts": 1776649791336158, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791336887, "dur": 48, + "args": { + "External id": 7684, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60214, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60214, "pid": 0, "tid": 7, "ts": 1776649791336887, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791336159, "dur": 6, + "args": { + "External id": 7684, + "cbid": 211, "correlation": 60214 + } + }, + { + "ph": "s", "id": 60214, "pid": 435, "tid": 435, "ts": 1776649791336159, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791336188, "dur": 0, + "args": { + "External id": 7686, + "cbid": 210, "correlation": 60241 + } + }, + { + "ph": "f", "id": 60241, "pid": 435, "tid": 435, "ts": 1776649791336188, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791336938, "dur": 44, + "args": { + "External id": 7686, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60242, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60242, "pid": 0, "tid": 7, "ts": 1776649791336938, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791336189, "dur": 5, + "args": { + "External id": 7686, + "cbid": 211, "correlation": 60242 + } + }, + { + "ph": "s", "id": 60242, "pid": 435, "tid": 435, "ts": 1776649791336189, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791336235, "dur": 1, + "args": { + "External id": 7688, + "cbid": 210, "correlation": 60269 + } + }, + { + "ph": "f", "id": 60269, "pid": 435, "tid": 435, "ts": 1776649791336235, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791336983, "dur": 48, + "args": { + "External id": 7688, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60270, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60270, "pid": 0, "tid": 7, "ts": 1776649791336983, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791336237, "dur": 6, + "args": { + "External id": 7688, + "cbid": 211, "correlation": 60270 + } + }, + { + "ph": "s", "id": 60270, "pid": 435, "tid": 435, "ts": 1776649791336237, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791336271, "dur": 0, + "args": { + "External id": 7690, + "cbid": 210, "correlation": 60297 + } + }, + { + "ph": "f", "id": 60297, "pid": 435, "tid": 435, "ts": 1776649791336271, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791337033, "dur": 51, + "args": { + "External id": 7690, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60298, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60298, "pid": 0, "tid": 7, "ts": 1776649791337033, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791336273, "dur": 6, + "args": { + "External id": 7690, + "cbid": 211, "correlation": 60298 + } + }, + { + "ph": "s", "id": 60298, "pid": 435, "tid": 435, "ts": 1776649791336273, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791336302, "dur": 0, + "args": { + "External id": 7692, + "cbid": 210, "correlation": 60325 + } + }, + { + "ph": "f", "id": 60325, "pid": 435, "tid": 435, "ts": 1776649791336302, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791337087, "dur": 43, + "args": { + "External id": 7692, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60326, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60326, "pid": 0, "tid": 7, "ts": 1776649791337087, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791336303, "dur": 6, + "args": { + "External id": 7692, + "cbid": 211, "correlation": 60326 + } + }, + { + "ph": "s", "id": 60326, "pid": 435, "tid": 435, "ts": 1776649791336303, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791336332, "dur": 0, + "args": { + "External id": 7694, + "cbid": 210, "correlation": 60353 + } + }, + { + "ph": "f", "id": 60353, "pid": 435, "tid": 435, "ts": 1776649791336332, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791337132, "dur": 49, + "args": { + "External id": 7694, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60354, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60354, "pid": 0, "tid": 7, "ts": 1776649791337132, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791336334, "dur": 6, + "args": { + "External id": 7694, + "cbid": 211, "correlation": 60354 + } + }, + { + "ph": "s", "id": 60354, "pid": 435, "tid": 435, "ts": 1776649791336334, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791336363, "dur": 0, + "args": { + "External id": 7696, + "cbid": 210, "correlation": 60381 + } + }, + { + "ph": "f", "id": 60381, "pid": 435, "tid": 435, "ts": 1776649791336363, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791337183, "dur": 48, + "args": { + "External id": 7696, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60382, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60382, "pid": 0, "tid": 7, "ts": 1776649791337183, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791336364, "dur": 5, + "args": { + "External id": 7696, + "cbid": 211, "correlation": 60382 + } + }, + { + "ph": "s", "id": 60382, "pid": 435, "tid": 435, "ts": 1776649791336364, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791336398, "dur": 4, + "args": { + "External id": 60393, + "cbid": 135, "correlation": 60393 + } + }, + { + "ph": "f", "id": 60393, "pid": 435, "tid": 435, "ts": 1776649791336398, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791336413, "dur": 828, + "args": { + "External id": 60398, + "cbid": 165, "correlation": 60398 + } + }, + { + "ph": "s", "id": 60398, "pid": 435, "tid": 435, "ts": 1776649791336413, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649791337244, "dur": 4, + "args": { + "External id": 60401, + "cbid": 139, "correlation": 60401 + } + }, + { + "ph": "f", "id": 60401, "pid": 435, "tid": 435, "ts": 1776649791337244, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791337282, "dur": 1, + "args": { + "External id": 7698, + "cbid": 210, "correlation": 60435 + } + }, + { + "ph": "f", "id": 60435, "pid": 435, "tid": 435, "ts": 1776649791337282, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791337294, "dur": 40, + "args": { + "External id": 7698, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60436, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60436, "pid": 0, "tid": 7, "ts": 1776649791337294, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791337284, "dur": 8, + "args": { + "External id": 7698, + "cbid": 211, "correlation": 60436 + } + }, + { + "ph": "s", "id": 60436, "pid": 435, "tid": 435, "ts": 1776649791337284, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791337317, "dur": 0, + "args": { + "External id": 7700, + "cbid": 210, "correlation": 60463 + } + }, + { + "ph": "f", "id": 60463, "pid": 435, "tid": 435, "ts": 1776649791337317, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791337336, "dur": 43, + "args": { + "External id": 7700, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60464, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60464, "pid": 0, "tid": 7, "ts": 1776649791337336, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791337318, "dur": 6, + "args": { + "External id": 7700, + "cbid": 211, "correlation": 60464 + } + }, + { + "ph": "s", "id": 60464, "pid": 435, "tid": 435, "ts": 1776649791337318, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791337347, "dur": 0, + "args": { + "External id": 7702, + "cbid": 210, "correlation": 60491 + } + }, + { + "ph": "f", "id": 60491, "pid": 435, "tid": 435, "ts": 1776649791337347, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791337381, "dur": 47, + "args": { + "External id": 7702, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60492, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60492, "pid": 0, "tid": 7, "ts": 1776649791337381, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791337349, "dur": 6, + "args": { + "External id": 7702, + "cbid": 211, "correlation": 60492 + } + }, + { + "ph": "s", "id": 60492, "pid": 435, "tid": 435, "ts": 1776649791337349, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791337378, "dur": 0, + "args": { + "External id": 7704, + "cbid": 210, "correlation": 60519 + } + }, + { + "ph": "f", "id": 60519, "pid": 435, "tid": 435, "ts": 1776649791337378, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791337430, "dur": 45, + "args": { + "External id": 7704, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60520, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60520, "pid": 0, "tid": 7, "ts": 1776649791337430, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791337379, "dur": 6, + "args": { + "External id": 7704, + "cbid": 211, "correlation": 60520 + } + }, + { + "ph": "s", "id": 60520, "pid": 435, "tid": 435, "ts": 1776649791337379, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791337408, "dur": 0, + "args": { + "External id": 7706, + "cbid": 210, "correlation": 60547 + } + }, + { + "ph": "f", "id": 60547, "pid": 435, "tid": 435, "ts": 1776649791337408, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791337477, "dur": 46, + "args": { + "External id": 7706, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60548, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60548, "pid": 0, "tid": 7, "ts": 1776649791337477, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791337409, "dur": 5, + "args": { + "External id": 7706, + "cbid": 211, "correlation": 60548 + } + }, + { + "ph": "s", "id": 60548, "pid": 435, "tid": 435, "ts": 1776649791337409, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791337438, "dur": 0, + "args": { + "External id": 7708, + "cbid": 210, "correlation": 60575 + } + }, + { + "ph": "f", "id": 60575, "pid": 435, "tid": 435, "ts": 1776649791337438, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791337525, "dur": 46, + "args": { + "External id": 7708, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60576, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60576, "pid": 0, "tid": 7, "ts": 1776649791337525, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791337439, "dur": 5, + "args": { + "External id": 7708, + "cbid": 211, "correlation": 60576 + } + }, + { + "ph": "s", "id": 60576, "pid": 435, "tid": 435, "ts": 1776649791337439, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791337467, "dur": 0, + "args": { + "External id": 7710, + "cbid": 210, "correlation": 60603 + } + }, + { + "ph": "f", "id": 60603, "pid": 435, "tid": 435, "ts": 1776649791337467, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791337573, "dur": 46, + "args": { + "External id": 7710, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60604, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60604, "pid": 0, "tid": 7, "ts": 1776649791337573, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791337468, "dur": 5, + "args": { + "External id": 7710, + "cbid": 211, "correlation": 60604 + } + }, + { + "ph": "s", "id": 60604, "pid": 435, "tid": 435, "ts": 1776649791337468, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791337497, "dur": 0, + "args": { + "External id": 7712, + "cbid": 210, "correlation": 60631 + } + }, + { + "ph": "f", "id": 60631, "pid": 435, "tid": 435, "ts": 1776649791337497, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791337621, "dur": 46, + "args": { + "External id": 7712, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60632, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60632, "pid": 0, "tid": 7, "ts": 1776649791337621, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791337498, "dur": 6, + "args": { + "External id": 7712, + "cbid": 211, "correlation": 60632 + } + }, + { + "ph": "s", "id": 60632, "pid": 435, "tid": 435, "ts": 1776649791337498, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791337527, "dur": 0, + "args": { + "External id": 7714, + "cbid": 210, "correlation": 60659 + } + }, + { + "ph": "f", "id": 60659, "pid": 435, "tid": 435, "ts": 1776649791337527, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791337669, "dur": 52, + "args": { + "External id": 7714, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60660, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60660, "pid": 0, "tid": 7, "ts": 1776649791337669, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791337528, "dur": 5, + "args": { + "External id": 7714, + "cbid": 211, "correlation": 60660 + } + }, + { + "ph": "s", "id": 60660, "pid": 435, "tid": 435, "ts": 1776649791337528, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791337556, "dur": 0, + "args": { + "External id": 7716, + "cbid": 210, "correlation": 60687 + } + }, + { + "ph": "f", "id": 60687, "pid": 435, "tid": 435, "ts": 1776649791337556, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791337723, "dur": 47, + "args": { + "External id": 7716, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60688, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60688, "pid": 0, "tid": 7, "ts": 1776649791337723, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791337557, "dur": 5, + "args": { + "External id": 7716, + "cbid": 211, "correlation": 60688 + } + }, + { + "ph": "s", "id": 60688, "pid": 435, "tid": 435, "ts": 1776649791337557, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791337582, "dur": 192, + "args": { + "External id": 60694, + "cbid": 165, "correlation": 60694 + } + }, + { + "ph": "s", "id": 60694, "pid": 435, "tid": 435, "ts": 1776649791337582, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791337795, "dur": 3, + "args": { + "External id": 60703, + "cbid": 135, "correlation": 60703 + } + }, + { + "ph": "f", "id": 60703, "pid": 435, "tid": 435, "ts": 1776649791337795, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791337826, "dur": 1, + "args": { + "External id": 7718, + "cbid": 210, "correlation": 60729 + } + }, + { + "ph": "f", "id": 60729, "pid": 435, "tid": 435, "ts": 1776649791337826, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791337838, "dur": 41, + "args": { + "External id": 7718, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60730, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60730, "pid": 0, "tid": 7, "ts": 1776649791337838, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791337828, "dur": 7, + "args": { + "External id": 7718, + "cbid": 211, "correlation": 60730 + } + }, + { + "ph": "s", "id": 60730, "pid": 435, "tid": 435, "ts": 1776649791337828, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791337861, "dur": 0, + "args": { + "External id": 7720, + "cbid": 210, "correlation": 60757 + } + }, + { + "ph": "f", "id": 60757, "pid": 435, "tid": 435, "ts": 1776649791337861, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791337880, "dur": 43, + "args": { + "External id": 7720, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60758, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60758, "pid": 0, "tid": 7, "ts": 1776649791337880, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791337862, "dur": 6, + "args": { + "External id": 7720, + "cbid": 211, "correlation": 60758 + } + }, + { + "ph": "s", "id": 60758, "pid": 435, "tid": 435, "ts": 1776649791337862, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791337891, "dur": 0, + "args": { + "External id": 7722, + "cbid": 210, "correlation": 60785 + } + }, + { + "ph": "f", "id": 60785, "pid": 435, "tid": 435, "ts": 1776649791337891, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791337925, "dur": 40, + "args": { + "External id": 7722, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60786, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60786, "pid": 0, "tid": 7, "ts": 1776649791337925, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791337893, "dur": 6, + "args": { + "External id": 7722, + "cbid": 211, "correlation": 60786 + } + }, + { + "ph": "s", "id": 60786, "pid": 435, "tid": 435, "ts": 1776649791337893, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791337922, "dur": 0, + "args": { + "External id": 7724, + "cbid": 210, "correlation": 60813 + } + }, + { + "ph": "f", "id": 60813, "pid": 435, "tid": 435, "ts": 1776649791337922, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791337968, "dur": 45, + "args": { + "External id": 7724, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60814, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60814, "pid": 0, "tid": 7, "ts": 1776649791337968, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791337923, "dur": 6, + "args": { + "External id": 7724, + "cbid": 211, "correlation": 60814 + } + }, + { + "ph": "s", "id": 60814, "pid": 435, "tid": 435, "ts": 1776649791337923, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791337953, "dur": 0, + "args": { + "External id": 7726, + "cbid": 210, "correlation": 60841 + } + }, + { + "ph": "f", "id": 60841, "pid": 435, "tid": 435, "ts": 1776649791337953, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791338015, "dur": 46, + "args": { + "External id": 7726, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60842, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60842, "pid": 0, "tid": 7, "ts": 1776649791338015, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791337954, "dur": 6, + "args": { + "External id": 7726, + "cbid": 211, "correlation": 60842 + } + }, + { + "ph": "s", "id": 60842, "pid": 435, "tid": 435, "ts": 1776649791337954, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791337983, "dur": 0, + "args": { + "External id": 7728, + "cbid": 210, "correlation": 60869 + } + }, + { + "ph": "f", "id": 60869, "pid": 435, "tid": 435, "ts": 1776649791337983, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791338063, "dur": 52, + "args": { + "External id": 7728, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60870, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60870, "pid": 0, "tid": 7, "ts": 1776649791338063, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791337984, "dur": 6, + "args": { + "External id": 7728, + "cbid": 211, "correlation": 60870 + } + }, + { + "ph": "s", "id": 60870, "pid": 435, "tid": 435, "ts": 1776649791337984, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338013, "dur": 0, + "args": { + "External id": 7730, + "cbid": 210, "correlation": 60897 + } + }, + { + "ph": "f", "id": 60897, "pid": 435, "tid": 435, "ts": 1776649791338013, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791338117, "dur": 46, + "args": { + "External id": 7730, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60898, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60898, "pid": 0, "tid": 7, "ts": 1776649791338117, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338015, "dur": 5, + "args": { + "External id": 7730, + "cbid": 211, "correlation": 60898 + } + }, + { + "ph": "s", "id": 60898, "pid": 435, "tid": 435, "ts": 1776649791338015, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338043, "dur": 0, + "args": { + "External id": 7732, + "cbid": 210, "correlation": 60925 + } + }, + { + "ph": "f", "id": 60925, "pid": 435, "tid": 435, "ts": 1776649791338043, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791338165, "dur": 44, + "args": { + "External id": 7732, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60926, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60926, "pid": 0, "tid": 7, "ts": 1776649791338165, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338044, "dur": 5, + "args": { + "External id": 7732, + "cbid": 211, "correlation": 60926 + } + }, + { + "ph": "s", "id": 60926, "pid": 435, "tid": 435, "ts": 1776649791338044, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338073, "dur": 0, + "args": { + "External id": 7734, + "cbid": 210, "correlation": 60953 + } + }, + { + "ph": "f", "id": 60953, "pid": 435, "tid": 435, "ts": 1776649791338073, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791338210, "dur": 49, + "args": { + "External id": 7734, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60954, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60954, "pid": 0, "tid": 7, "ts": 1776649791338210, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338074, "dur": 6, + "args": { + "External id": 7734, + "cbid": 211, "correlation": 60954 + } + }, + { + "ph": "s", "id": 60954, "pid": 435, "tid": 435, "ts": 1776649791338074, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338103, "dur": 0, + "args": { + "External id": 7736, + "cbid": 210, "correlation": 60981 + } + }, + { + "ph": "f", "id": 60981, "pid": 435, "tid": 435, "ts": 1776649791338103, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791338261, "dur": 48, + "args": { + "External id": 7736, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 60982, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 60982, "pid": 0, "tid": 7, "ts": 1776649791338261, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338104, "dur": 5, + "args": { + "External id": 7736, + "cbid": 211, "correlation": 60982 + } + }, + { + "ph": "s", "id": 60982, "pid": 435, "tid": 435, "ts": 1776649791338104, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338133, "dur": 0, + "args": { + "External id": 7738, + "cbid": 210, "correlation": 61009 + } + }, + { + "ph": "f", "id": 61009, "pid": 435, "tid": 435, "ts": 1776649791338133, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791338311, "dur": 49, + "args": { + "External id": 7738, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61010, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61010, "pid": 0, "tid": 7, "ts": 1776649791338311, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338135, "dur": 6, + "args": { + "External id": 7738, + "cbid": 211, "correlation": 61010 + } + }, + { + "ph": "s", "id": 61010, "pid": 435, "tid": 435, "ts": 1776649791338135, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338163, "dur": 0, + "args": { + "External id": 7740, + "cbid": 210, "correlation": 61037 + } + }, + { + "ph": "f", "id": 61037, "pid": 435, "tid": 435, "ts": 1776649791338163, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791338361, "dur": 47, + "args": { + "External id": 7740, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61038, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61038, "pid": 0, "tid": 7, "ts": 1776649791338361, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338165, "dur": 6, + "args": { + "External id": 7740, + "cbid": 211, "correlation": 61038 + } + }, + { + "ph": "s", "id": 61038, "pid": 435, "tid": 435, "ts": 1776649791338165, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338193, "dur": 0, + "args": { + "External id": 7742, + "cbid": 210, "correlation": 61065 + } + }, + { + "ph": "f", "id": 61065, "pid": 435, "tid": 435, "ts": 1776649791338193, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791338410, "dur": 46, + "args": { + "External id": 7742, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61066, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61066, "pid": 0, "tid": 7, "ts": 1776649791338410, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338195, "dur": 5, + "args": { + "External id": 7742, + "cbid": 211, "correlation": 61066 + } + }, + { + "ph": "s", "id": 61066, "pid": 435, "tid": 435, "ts": 1776649791338195, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338223, "dur": 0, + "args": { + "External id": 7744, + "cbid": 210, "correlation": 61093 + } + }, + { + "ph": "f", "id": 61093, "pid": 435, "tid": 435, "ts": 1776649791338223, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791338457, "dur": 49, + "args": { + "External id": 7744, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61094, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61094, "pid": 0, "tid": 7, "ts": 1776649791338457, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338224, "dur": 6, + "args": { + "External id": 7744, + "cbid": 211, "correlation": 61094 + } + }, + { + "ph": "s", "id": 61094, "pid": 435, "tid": 435, "ts": 1776649791338224, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338253, "dur": 0, + "args": { + "External id": 7746, + "cbid": 210, "correlation": 61121 + } + }, + { + "ph": "f", "id": 61121, "pid": 435, "tid": 435, "ts": 1776649791338253, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791338508, "dur": 41, + "args": { + "External id": 7746, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61122, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61122, "pid": 0, "tid": 7, "ts": 1776649791338508, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338254, "dur": 5, + "args": { + "External id": 7746, + "cbid": 211, "correlation": 61122 + } + }, + { + "ph": "s", "id": 61122, "pid": 435, "tid": 435, "ts": 1776649791338254, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338283, "dur": 0, + "args": { + "External id": 7748, + "cbid": 210, "correlation": 61149 + } + }, + { + "ph": "f", "id": 61149, "pid": 435, "tid": 435, "ts": 1776649791338283, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791338550, "dur": 47, + "args": { + "External id": 7748, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61150, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61150, "pid": 0, "tid": 7, "ts": 1776649791338550, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338284, "dur": 6, + "args": { + "External id": 7748, + "cbid": 211, "correlation": 61150 + } + }, + { + "ph": "s", "id": 61150, "pid": 435, "tid": 435, "ts": 1776649791338284, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338313, "dur": 0, + "args": { + "External id": 7750, + "cbid": 210, "correlation": 61177 + } + }, + { + "ph": "f", "id": 61177, "pid": 435, "tid": 435, "ts": 1776649791338313, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791338600, "dur": 46, + "args": { + "External id": 7750, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61178, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61178, "pid": 0, "tid": 7, "ts": 1776649791338600, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338314, "dur": 5, + "args": { + "External id": 7750, + "cbid": 211, "correlation": 61178 + } + }, + { + "ph": "s", "id": 61178, "pid": 435, "tid": 435, "ts": 1776649791338314, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338342, "dur": 0, + "args": { + "External id": 7752, + "cbid": 210, "correlation": 61205 + } + }, + { + "ph": "f", "id": 61205, "pid": 435, "tid": 435, "ts": 1776649791338342, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791338648, "dur": 46, + "args": { + "External id": 7752, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61206, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61206, "pid": 0, "tid": 7, "ts": 1776649791338648, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338344, "dur": 5, + "args": { + "External id": 7752, + "cbid": 211, "correlation": 61206 + } + }, + { + "ph": "s", "id": 61206, "pid": 435, "tid": 435, "ts": 1776649791338344, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338372, "dur": 0, + "args": { + "External id": 7754, + "cbid": 210, "correlation": 61233 + } + }, + { + "ph": "f", "id": 61233, "pid": 435, "tid": 435, "ts": 1776649791338372, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791338696, "dur": 47, + "args": { + "External id": 7754, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61234, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61234, "pid": 0, "tid": 7, "ts": 1776649791338696, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338373, "dur": 6, + "args": { + "External id": 7754, + "cbid": 211, "correlation": 61234 + } + }, + { + "ph": "s", "id": 61234, "pid": 435, "tid": 435, "ts": 1776649791338373, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338402, "dur": 0, + "args": { + "External id": 7756, + "cbid": 210, "correlation": 61261 + } + }, + { + "ph": "f", "id": 61261, "pid": 435, "tid": 435, "ts": 1776649791338402, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791338744, "dur": 44, + "args": { + "External id": 7756, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61262, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61262, "pid": 0, "tid": 7, "ts": 1776649791338744, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338403, "dur": 5, + "args": { + "External id": 7756, + "cbid": 211, "correlation": 61262 + } + }, + { + "ph": "s", "id": 61262, "pid": 435, "tid": 435, "ts": 1776649791338403, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338431, "dur": 0, + "args": { + "External id": 7758, + "cbid": 210, "correlation": 61289 + } + }, + { + "ph": "f", "id": 61289, "pid": 435, "tid": 435, "ts": 1776649791338431, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791338790, "dur": 49, + "args": { + "External id": 7758, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61290, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61290, "pid": 0, "tid": 7, "ts": 1776649791338790, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338433, "dur": 5, + "args": { + "External id": 7758, + "cbid": 211, "correlation": 61290 + } + }, + { + "ph": "s", "id": 61290, "pid": 435, "tid": 435, "ts": 1776649791338433, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338461, "dur": 0, + "args": { + "External id": 7760, + "cbid": 210, "correlation": 61317 + } + }, + { + "ph": "f", "id": 61317, "pid": 435, "tid": 435, "ts": 1776649791338461, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791338841, "dur": 47, + "args": { + "External id": 7760, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61318, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61318, "pid": 0, "tid": 7, "ts": 1776649791338841, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338462, "dur": 5, + "args": { + "External id": 7760, + "cbid": 211, "correlation": 61318 + } + }, + { + "ph": "s", "id": 61318, "pid": 435, "tid": 435, "ts": 1776649791338462, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338491, "dur": 0, + "args": { + "External id": 7762, + "cbid": 210, "correlation": 61345 + } + }, + { + "ph": "f", "id": 61345, "pid": 435, "tid": 435, "ts": 1776649791338491, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791338890, "dur": 49, + "args": { + "External id": 7762, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61346, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61346, "pid": 0, "tid": 7, "ts": 1776649791338890, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338492, "dur": 5, + "args": { + "External id": 7762, + "cbid": 211, "correlation": 61346 + } + }, + { + "ph": "s", "id": 61346, "pid": 435, "tid": 435, "ts": 1776649791338492, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338520, "dur": 0, + "args": { + "External id": 7764, + "cbid": 210, "correlation": 61373 + } + }, + { + "ph": "f", "id": 61373, "pid": 435, "tid": 435, "ts": 1776649791338520, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791338941, "dur": 49, + "args": { + "External id": 7764, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61374, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61374, "pid": 0, "tid": 7, "ts": 1776649791338941, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338522, "dur": 5, + "args": { + "External id": 7764, + "cbid": 211, "correlation": 61374 + } + }, + { + "ph": "s", "id": 61374, "pid": 435, "tid": 435, "ts": 1776649791338522, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338550, "dur": 0, + "args": { + "External id": 7766, + "cbid": 210, "correlation": 61401 + } + }, + { + "ph": "f", "id": 61401, "pid": 435, "tid": 435, "ts": 1776649791338550, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791338992, "dur": 49, + "args": { + "External id": 7766, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61402, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61402, "pid": 0, "tid": 7, "ts": 1776649791338992, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338552, "dur": 5, + "args": { + "External id": 7766, + "cbid": 211, "correlation": 61402 + } + }, + { + "ph": "s", "id": 61402, "pid": 435, "tid": 435, "ts": 1776649791338552, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338580, "dur": 0, + "args": { + "External id": 7768, + "cbid": 210, "correlation": 61429 + } + }, + { + "ph": "f", "id": 61429, "pid": 435, "tid": 435, "ts": 1776649791338580, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791339043, "dur": 46, + "args": { + "External id": 7768, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61430, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61430, "pid": 0, "tid": 7, "ts": 1776649791339043, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338581, "dur": 5, + "args": { + "External id": 7768, + "cbid": 211, "correlation": 61430 + } + }, + { + "ph": "s", "id": 61430, "pid": 435, "tid": 435, "ts": 1776649791338581, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338610, "dur": 0, + "args": { + "External id": 7770, + "cbid": 210, "correlation": 61457 + } + }, + { + "ph": "f", "id": 61457, "pid": 435, "tid": 435, "ts": 1776649791338610, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791339091, "dur": 46, + "args": { + "External id": 7770, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61458, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61458, "pid": 0, "tid": 7, "ts": 1776649791339091, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338611, "dur": 6, + "args": { + "External id": 7770, + "cbid": 211, "correlation": 61458 + } + }, + { + "ph": "s", "id": 61458, "pid": 435, "tid": 435, "ts": 1776649791338611, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338640, "dur": 0, + "args": { + "External id": 7772, + "cbid": 210, "correlation": 61485 + } + }, + { + "ph": "f", "id": 61485, "pid": 435, "tid": 435, "ts": 1776649791338640, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791339139, "dur": 44, + "args": { + "External id": 7772, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61486, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61486, "pid": 0, "tid": 7, "ts": 1776649791339139, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338641, "dur": 5, + "args": { + "External id": 7772, + "cbid": 211, "correlation": 61486 + } + }, + { + "ph": "s", "id": 61486, "pid": 435, "tid": 435, "ts": 1776649791338641, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338669, "dur": 0, + "args": { + "External id": 7774, + "cbid": 210, "correlation": 61513 + } + }, + { + "ph": "f", "id": 61513, "pid": 435, "tid": 435, "ts": 1776649791338669, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791339184, "dur": 44, + "args": { + "External id": 7774, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61514, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61514, "pid": 0, "tid": 7, "ts": 1776649791339184, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338671, "dur": 5, + "args": { + "External id": 7774, + "cbid": 211, "correlation": 61514 + } + }, + { + "ph": "s", "id": 61514, "pid": 435, "tid": 435, "ts": 1776649791338671, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338699, "dur": 0, + "args": { + "External id": 7776, + "cbid": 210, "correlation": 61541 + } + }, + { + "ph": "f", "id": 61541, "pid": 435, "tid": 435, "ts": 1776649791338699, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791339229, "dur": 44, + "args": { + "External id": 7776, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61542, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61542, "pid": 0, "tid": 7, "ts": 1776649791339229, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338700, "dur": 6, + "args": { + "External id": 7776, + "cbid": 211, "correlation": 61542 + } + }, + { + "ph": "s", "id": 61542, "pid": 435, "tid": 435, "ts": 1776649791338700, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338729, "dur": 0, + "args": { + "External id": 7778, + "cbid": 210, "correlation": 61569 + } + }, + { + "ph": "f", "id": 61569, "pid": 435, "tid": 435, "ts": 1776649791338729, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791339275, "dur": 47, + "args": { + "External id": 7778, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61570, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61570, "pid": 0, "tid": 7, "ts": 1776649791339275, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338730, "dur": 5, + "args": { + "External id": 7778, + "cbid": 211, "correlation": 61570 + } + }, + { + "ph": "s", "id": 61570, "pid": 435, "tid": 435, "ts": 1776649791338730, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338759, "dur": 0, + "args": { + "External id": 7780, + "cbid": 210, "correlation": 61597 + } + }, + { + "ph": "f", "id": 61597, "pid": 435, "tid": 435, "ts": 1776649791338759, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791339324, "dur": 47, + "args": { + "External id": 7780, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61598, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61598, "pid": 0, "tid": 7, "ts": 1776649791339324, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338760, "dur": 6, + "args": { + "External id": 7780, + "cbid": 211, "correlation": 61598 + } + }, + { + "ph": "s", "id": 61598, "pid": 435, "tid": 435, "ts": 1776649791338760, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338791, "dur": 0, + "args": { + "External id": 7782, + "cbid": 210, "correlation": 61625 + } + }, + { + "ph": "f", "id": 61625, "pid": 435, "tid": 435, "ts": 1776649791338791, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791339372, "dur": 49, + "args": { + "External id": 7782, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61626, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61626, "pid": 0, "tid": 7, "ts": 1776649791339372, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338792, "dur": 5, + "args": { + "External id": 7782, + "cbid": 211, "correlation": 61626 + } + }, + { + "ph": "s", "id": 61626, "pid": 435, "tid": 435, "ts": 1776649791338792, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338821, "dur": 0, + "args": { + "External id": 7784, + "cbid": 210, "correlation": 61653 + } + }, + { + "ph": "f", "id": 61653, "pid": 435, "tid": 435, "ts": 1776649791338821, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791339423, "dur": 49, + "args": { + "External id": 7784, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61654, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61654, "pid": 0, "tid": 7, "ts": 1776649791339423, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338822, "dur": 6, + "args": { + "External id": 7784, + "cbid": 211, "correlation": 61654 + } + }, + { + "ph": "s", "id": 61654, "pid": 435, "tid": 435, "ts": 1776649791338822, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338851, "dur": 0, + "args": { + "External id": 7786, + "cbid": 210, "correlation": 61681 + } + }, + { + "ph": "f", "id": 61681, "pid": 435, "tid": 435, "ts": 1776649791338851, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791339473, "dur": 46, + "args": { + "External id": 7786, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61682, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61682, "pid": 0, "tid": 7, "ts": 1776649791339473, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338852, "dur": 6, + "args": { + "External id": 7786, + "cbid": 211, "correlation": 61682 + } + }, + { + "ph": "s", "id": 61682, "pid": 435, "tid": 435, "ts": 1776649791338852, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338881, "dur": 0, + "args": { + "External id": 7788, + "cbid": 210, "correlation": 61709 + } + }, + { + "ph": "f", "id": 61709, "pid": 435, "tid": 435, "ts": 1776649791338881, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791339521, "dur": 47, + "args": { + "External id": 7788, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61710, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61710, "pid": 0, "tid": 7, "ts": 1776649791339521, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338882, "dur": 5, + "args": { + "External id": 7788, + "cbid": 211, "correlation": 61710 + } + }, + { + "ph": "s", "id": 61710, "pid": 435, "tid": 435, "ts": 1776649791338882, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338911, "dur": 0, + "args": { + "External id": 7790, + "cbid": 210, "correlation": 61737 + } + }, + { + "ph": "f", "id": 61737, "pid": 435, "tid": 435, "ts": 1776649791338911, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791339570, "dur": 49, + "args": { + "External id": 7790, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61738, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61738, "pid": 0, "tid": 7, "ts": 1776649791339570, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338912, "dur": 5, + "args": { + "External id": 7790, + "cbid": 211, "correlation": 61738 + } + }, + { + "ph": "s", "id": 61738, "pid": 435, "tid": 435, "ts": 1776649791338912, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338941, "dur": 0, + "args": { + "External id": 7792, + "cbid": 210, "correlation": 61765 + } + }, + { + "ph": "f", "id": 61765, "pid": 435, "tid": 435, "ts": 1776649791338941, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791339621, "dur": 49, + "args": { + "External id": 7792, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61766, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61766, "pid": 0, "tid": 7, "ts": 1776649791339621, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338942, "dur": 6, + "args": { + "External id": 7792, + "cbid": 211, "correlation": 61766 + } + }, + { + "ph": "s", "id": 61766, "pid": 435, "tid": 435, "ts": 1776649791338942, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791338970, "dur": 0, + "args": { + "External id": 7794, + "cbid": 210, "correlation": 61793 + } + }, + { + "ph": "f", "id": 61793, "pid": 435, "tid": 435, "ts": 1776649791338970, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791339671, "dur": 48, + "args": { + "External id": 7794, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61794, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61794, "pid": 0, "tid": 7, "ts": 1776649791339671, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791338972, "dur": 5, + "args": { + "External id": 7794, + "cbid": 211, "correlation": 61794 + } + }, + { + "ph": "s", "id": 61794, "pid": 435, "tid": 435, "ts": 1776649791338972, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791339000, "dur": 0, + "args": { + "External id": 7796, + "cbid": 210, "correlation": 61821 + } + }, + { + "ph": "f", "id": 61821, "pid": 435, "tid": 435, "ts": 1776649791339000, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791339722, "dur": 47, + "args": { + "External id": 7796, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61822, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61822, "pid": 0, "tid": 7, "ts": 1776649791339722, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791339001, "dur": 5, + "args": { + "External id": 7796, + "cbid": 211, "correlation": 61822 + } + }, + { + "ph": "s", "id": 61822, "pid": 435, "tid": 435, "ts": 1776649791339001, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791339030, "dur": 0, + "args": { + "External id": 7798, + "cbid": 210, "correlation": 61849 + } + }, + { + "ph": "f", "id": 61849, "pid": 435, "tid": 435, "ts": 1776649791339030, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791339770, "dur": 44, + "args": { + "External id": 7798, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61850, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61850, "pid": 0, "tid": 7, "ts": 1776649791339770, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791339031, "dur": 5, + "args": { + "External id": 7798, + "cbid": 211, "correlation": 61850 + } + }, + { + "ph": "s", "id": 61850, "pid": 435, "tid": 435, "ts": 1776649791339031, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791339059, "dur": 0, + "args": { + "External id": 7800, + "cbid": 210, "correlation": 61877 + } + }, + { + "ph": "f", "id": 61877, "pid": 435, "tid": 435, "ts": 1776649791339059, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791339816, "dur": 46, + "args": { + "External id": 7800, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61878, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61878, "pid": 0, "tid": 7, "ts": 1776649791339816, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791339061, "dur": 6, + "args": { + "External id": 7800, + "cbid": 211, "correlation": 61878 + } + }, + { + "ph": "s", "id": 61878, "pid": 435, "tid": 435, "ts": 1776649791339061, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791339089, "dur": 0, + "args": { + "External id": 7802, + "cbid": 210, "correlation": 61905 + } + }, + { + "ph": "f", "id": 61905, "pid": 435, "tid": 435, "ts": 1776649791339089, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791339863, "dur": 48, + "args": { + "External id": 7802, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61906, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61906, "pid": 0, "tid": 7, "ts": 1776649791339863, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791339091, "dur": 6, + "args": { + "External id": 7802, + "cbid": 211, "correlation": 61906 + } + }, + { + "ph": "s", "id": 61906, "pid": 435, "tid": 435, "ts": 1776649791339091, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791339119, "dur": 0, + "args": { + "External id": 7804, + "cbid": 210, "correlation": 61933 + } + }, + { + "ph": "f", "id": 61933, "pid": 435, "tid": 435, "ts": 1776649791339119, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791339913, "dur": 43, + "args": { + "External id": 7804, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61934, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61934, "pid": 0, "tid": 7, "ts": 1776649791339913, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791339121, "dur": 5, + "args": { + "External id": 7804, + "cbid": 211, "correlation": 61934 + } + }, + { + "ph": "s", "id": 61934, "pid": 435, "tid": 435, "ts": 1776649791339121, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791339149, "dur": 0, + "args": { + "External id": 7806, + "cbid": 210, "correlation": 61961 + } + }, + { + "ph": "f", "id": 61961, "pid": 435, "tid": 435, "ts": 1776649791339149, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791339958, "dur": 44, + "args": { + "External id": 7806, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61962, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61962, "pid": 0, "tid": 7, "ts": 1776649791339958, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791339151, "dur": 6, + "args": { + "External id": 7806, + "cbid": 211, "correlation": 61962 + } + }, + { + "ph": "s", "id": 61962, "pid": 435, "tid": 435, "ts": 1776649791339151, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791339179, "dur": 0, + "args": { + "External id": 7808, + "cbid": 210, "correlation": 61989 + } + }, + { + "ph": "f", "id": 61989, "pid": 435, "tid": 435, "ts": 1776649791339179, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791340003, "dur": 49, + "args": { + "External id": 7808, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 61990, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 61990, "pid": 0, "tid": 7, "ts": 1776649791340003, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791339181, "dur": 6, + "args": { + "External id": 7808, + "cbid": 211, "correlation": 61990 + } + }, + { + "ph": "s", "id": 61990, "pid": 435, "tid": 435, "ts": 1776649791339181, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791339209, "dur": 0, + "args": { + "External id": 7810, + "cbid": 210, "correlation": 62017 + } + }, + { + "ph": "f", "id": 62017, "pid": 435, "tid": 435, "ts": 1776649791339209, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791340054, "dur": 49, + "args": { + "External id": 7810, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62018, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62018, "pid": 0, "tid": 7, "ts": 1776649791340054, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791339211, "dur": 6, + "args": { + "External id": 7810, + "cbid": 211, "correlation": 62018 + } + }, + { + "ph": "s", "id": 62018, "pid": 435, "tid": 435, "ts": 1776649791339211, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791339239, "dur": 0, + "args": { + "External id": 7812, + "cbid": 210, "correlation": 62045 + } + }, + { + "ph": "f", "id": 62045, "pid": 435, "tid": 435, "ts": 1776649791339239, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791340105, "dur": 47, + "args": { + "External id": 7812, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62046, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62046, "pid": 0, "tid": 7, "ts": 1776649791340105, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791339241, "dur": 6, + "args": { + "External id": 7812, + "cbid": 211, "correlation": 62046 + } + }, + { + "ph": "s", "id": 62046, "pid": 435, "tid": 435, "ts": 1776649791339241, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791339269, "dur": 0, + "args": { + "External id": 7814, + "cbid": 210, "correlation": 62073 + } + }, + { + "ph": "f", "id": 62073, "pid": 435, "tid": 435, "ts": 1776649791339269, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791340154, "dur": 49, + "args": { + "External id": 7814, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62074, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62074, "pid": 0, "tid": 7, "ts": 1776649791340154, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791339271, "dur": 5, + "args": { + "External id": 7814, + "cbid": 211, "correlation": 62074 + } + }, + { + "ph": "s", "id": 62074, "pid": 435, "tid": 435, "ts": 1776649791339271, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791339299, "dur": 0, + "args": { + "External id": 7816, + "cbid": 210, "correlation": 62101 + } + }, + { + "ph": "f", "id": 62101, "pid": 435, "tid": 435, "ts": 1776649791339299, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791340205, "dur": 50, + "args": { + "External id": 7816, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62102, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62102, "pid": 0, "tid": 7, "ts": 1776649791340205, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791339301, "dur": 6, + "args": { + "External id": 7816, + "cbid": 211, "correlation": 62102 + } + }, + { + "ph": "s", "id": 62102, "pid": 435, "tid": 435, "ts": 1776649791339301, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791339333, "dur": 4, + "args": { + "External id": 62113, + "cbid": 135, "correlation": 62113 + } + }, + { + "ph": "f", "id": 62113, "pid": 435, "tid": 435, "ts": 1776649791339333, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791339348, "dur": 916, + "args": { + "External id": 62118, + "cbid": 165, "correlation": 62118 + } + }, + { + "ph": "s", "id": 62118, "pid": 435, "tid": 435, "ts": 1776649791339348, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649791340268, "dur": 5, + "args": { + "External id": 62121, + "cbid": 139, "correlation": 62121 + } + }, + { + "ph": "f", "id": 62121, "pid": 435, "tid": 435, "ts": 1776649791340268, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor >, at::detail::Array >(int, at::native::AUnaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791340398, "dur": 33, + "args": { + "External id": 7819, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62138, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 62138, "pid": 0, "tid": 7, "ts": 1776649791340398, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791340384, "dur": 13, + "args": { + "External id": 7819, + "cbid": 211, "correlation": 62138 + } + }, + { + "ph": "s", "id": 62138, "pid": 435, "tid": 435, "ts": 1776649791340384, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649791340458, "dur": 57, + "args": { + "External id": 7823, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62151, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 62151, "pid": 0, "tid": 7, "ts": 1776649791340458, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791340447, "dur": 9, + "args": { + "External id": 7823, + "cbid": 211, "correlation": 62151 + } + }, + { + "ph": "s", "id": 62151, "pid": 435, "tid": 435, "ts": 1776649791340447, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BUnaryFunctor, at::detail::Array >(int, at::native::BUnaryFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791340517, "dur": 33, + "args": { + "External id": 7824, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62165, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 62165, "pid": 0, "tid": 7, "ts": 1776649791340517, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791340490, "dur": 8, + "args": { + "External id": 7824, + "cbid": 211, "correlation": 62165 + } + }, + { + "ph": "s", "id": 62165, "pid": 435, "tid": 435, "ts": 1776649791340490, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649791340551, "dur": 57, + "args": { + "External id": 7829, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62178, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 62178, "pid": 0, "tid": 7, "ts": 1776649791340551, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791340518, "dur": 6, + "args": { + "External id": 7829, + "cbid": 211, "correlation": 62178 + } + }, + { + "ph": "s", "id": 62178, "pid": 435, "tid": 435, "ts": 1776649791340518, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy_contig, unsigned int, 3, 128, 1>(at::native::(anonymous namespace)::OpaqueType<1u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata, unsigned int, 128, 1>, at::native::(anonymous namespace)::TensorSizeStride, int, unsigned int)", "pid": 0, "tid": 7, + "ts": 1776649791340610, "dur": 126, + "args": { + "External id": 7835, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62190, + "registers per thread": 21, + "shared memory": 0, + "blocks per SM": 4.000000, + "warps per SM": 64.000000, + "grid": [216, 2, 1], + "block": [512, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 62190, "pid": 0, "tid": 7, "ts": 1776649791340610, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791340581, "dur": 9, + "args": { + "External id": 7835, + "cbid": 211, "correlation": 62190 + } + }, + { + "ph": "s", "id": 62190, "pid": 435, "tid": 435, "ts": 1776649791340581, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649791340738, "dur": 136, + "args": { + "External id": 7846, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62203, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 62203, "pid": 0, "tid": 7, "ts": 1776649791340738, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791340661, "dur": 8, + "args": { + "External id": 7846, + "cbid": 211, "correlation": 62203 + } + }, + { + "ph": "s", "id": 62203, "pid": 435, "tid": 435, "ts": 1776649791340661, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, at::detail::Array >(int, at::native::CUDAFunctorOnSelf_add, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791340875, "dur": 63, + "args": { + "External id": 7849, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62214, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 62214, "pid": 0, "tid": 7, "ts": 1776649791340875, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791340706, "dur": 8, + "args": { + "External id": 7849, + "cbid": 211, "correlation": 62214 + } + }, + { + "ph": "s", "id": 62214, "pid": 435, "tid": 435, "ts": 1776649791340706, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649791340939, "dur": 179, + "args": { + "External id": 7853, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62227, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 62227, "pid": 0, "tid": 7, "ts": 1776649791340939, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791340737, "dur": 7, + "args": { + "External id": 7853, + "cbid": 211, "correlation": 62227 + } + }, + { + "ph": "s", "id": 62227, "pid": 435, "tid": 435, "ts": 1776649791340737, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649791340767, "dur": 1, + "args": { + "External id": 7854, + "cbid": 317, "correlation": 62237 + } + }, + { + "ph": "f", "id": 62237, "pid": 435, "tid": 435, "ts": 1776649791340767, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 435, "tid": 435, + "ts": 1776649791340769, "dur": 423, + "args": { + "External id": 7854, + "cbid": 20, "correlation": 62238 + } + }, + { + "ph": "f", "id": 62238, "pid": 435, "tid": 435, "ts": 1776649791340769, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor >, at::detail::Array >(int, at::native::BinaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791341220, "dur": 114, + "args": { + "External id": 7854, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62242, + "registers per thread": 22, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 62242, "pid": 0, "tid": 7, "ts": 1776649791341220, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791341206, "dur": 12, + "args": { + "External id": 7854, + "cbid": 211, "correlation": 62242 + } + }, + { + "ph": "s", "id": 62242, "pid": 435, "tid": 435, "ts": 1776649791341206, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649791341336, "dur": 511, + "args": { + "External id": 7862, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62256, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 62256, "pid": 0, "tid": 7, "ts": 1776649791341336, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791341283, "dur": 8, + "args": { + "External id": 7862, + "cbid": 211, "correlation": 62256 + } + }, + { + "ph": "s", "id": 62256, "pid": 435, "tid": 435, "ts": 1776649791341283, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649791573356, "dur": 7393, + "args": { + "External id": 7878, + "device": 0, "context": 1, + "stream": 7, "correlation": 62272, + "bytes": 67108864, "memory bandwidth (GB/s)": 9.076773818413571 + } + }, + { + "ph": "f", "id": 62272, "pid": 0, "tid": 7, "ts": 1776649791573356, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649791573220, "dur": 7506, + "args": { + "External id": 7878, + "cbid": 41, "correlation": 62272 + } + }, + { + "ph": "s", "id": 62272, "pid": 435, "tid": 435, "ts": 1776649791573220, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791580738, "dur": 23, + "args": { + "External id": 7878, + "cbid": 131, "correlation": 62273 + } + }, + { + "ph": "s", "id": 62273, "pid": 435, "tid": 435, "ts": 1776649791580738, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7, + "ts": 1776649791695781, "dur": 77, + "args": { + "External id": 7881, + "device": 0, "context": 1, + "stream": 7, "correlation": 62278, + "bytes": 67108864, "memory bandwidth (GB/s)": 870.5600684940392 + } + }, + { + "ph": "f", "id": 62278, "pid": 0, "tid": 7, "ts": 1776649791695781, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649791695705, "dur": 77, + "args": { + "External id": 7881, + "cbid": 41, "correlation": 62278 + } + }, + { + "ph": "s", "id": 62278, "pid": 435, "tid": 435, "ts": 1776649791695705, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649791696421, "dur": 4481, + "args": { + "External id": 7890, + "device": 0, "context": 1, + "stream": 7, "correlation": 62293, + "bytes": 16777216, "memory bandwidth (GB/s)": 3.743700925503609 + } + }, + { + "ph": "f", "id": 62293, "pid": 0, "tid": 7, "ts": 1776649791696421, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649791696074, "dur": 4785, + "args": { + "External id": 7890, + "cbid": 41, "correlation": 62293 + } + }, + { + "ph": "s", "id": 62293, "pid": 435, "tid": 435, "ts": 1776649791696074, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791700861, "dur": 51, + "args": { + "External id": 7890, + "cbid": 131, "correlation": 62294 + } + }, + { + "ph": "s", "id": 62294, "pid": 435, "tid": 435, "ts": 1776649791700861, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649791701193, "dur": 24, + "args": { + "External id": 7894, + "device": 0, "context": 1, + "stream": 7, "correlation": 62306, + "bytes": 524288, "memory bandwidth (GB/s)": 21.250324254215304 + } + }, + { + "ph": "f", "id": 62306, "pid": 0, "tid": 7, "ts": 1776649791701193, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649791701142, "dur": 48, + "args": { + "External id": 7894, + "cbid": 41, "correlation": 62306 + } + }, + { + "ph": "s", "id": 62306, "pid": 435, "tid": 435, "ts": 1776649791701142, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791701191, "dur": 33, + "args": { + "External id": 7894, + "cbid": 131, "correlation": 62307 + } + }, + { + "ph": "s", "id": 62307, "pid": 435, "tid": 435, "ts": 1776649791701191, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649791701259, "dur": 1, + "args": { + "External id": 7898, + "device": 0, "context": 1, + "stream": 7, "correlation": 62319, + "bytes": 4096, "memory bandwidth (GB/s)": 2.723404255319149 + } + }, + { + "ph": "f", "id": 62319, "pid": 0, "tid": 7, "ts": 1776649791701259, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649791701249, "dur": 6, + "args": { + "External id": 7898, + "cbid": 41, "correlation": 62319 + } + }, + { + "ph": "s", "id": 62319, "pid": 435, "tid": 435, "ts": 1776649791701249, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791701256, "dur": 9, + "args": { + "External id": 7898, + "cbid": 131, "correlation": 62320 + } + }, + { + "ph": "s", "id": 62320, "pid": 435, "tid": 435, "ts": 1776649791701256, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649791701446, "dur": 1036, + "args": { + "External id": 7913, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62332, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 62332, "pid": 0, "tid": 7, "ts": 1776649791701446, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791701425, "dur": 22, + "args": { + "External id": 7913, + "cbid": 211, "correlation": 62332 + } + }, + { + "ph": "s", "id": 62332, "pid": 435, "tid": 435, "ts": 1776649791701425, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649791702485, "dur": 7, + "args": { + "External id": 7919, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62346, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 4.740741, + "warps per SM": 18.962963, + "grid": [512, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 30 + } + }, + { + "ph": "f", "id": 62346, "pid": 0, "tid": 7, "ts": 1776649791702485, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791701486, "dur": 8, + "args": { + "External id": 7919, + "cbid": 211, "correlation": 62346 + } + }, + { + "ph": "s", "id": 62346, "pid": 435, "tid": 435, "ts": 1776649791701486, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649791701527, "dur": 2, + "args": { + "External id": 7921, + "cbid": 317, "correlation": 62356 + } + }, + { + "ph": "f", "id": 62356, "pid": 435, "tid": 435, "ts": 1776649791701527, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 435, "tid": 435, + "ts": 1776649791701530, "dur": 425, + "args": { + "External id": 7921, + "cbid": 20, "correlation": 62357 + } + }, + { + "ph": "f", "id": 62357, "pid": 435, "tid": 435, "ts": 1776649791701530, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649791702494, "dur": 247, + "args": { + "External id": 7921, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62361, + "registers per thread": 20, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 62361, "pid": 0, "tid": 7, "ts": 1776649791702494, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791701973, "dur": 13, + "args": { + "External id": 7921, + "cbid": 211, "correlation": 62361 + } + }, + { + "ph": "s", "id": 62361, "pid": 435, "tid": 435, "ts": 1776649791701973, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649791702012, "dur": 0, + "args": { + "External id": 7922, + "cbid": 317, "correlation": 62370 + } + }, + { + "ph": "f", "id": 62370, "pid": 435, "tid": 435, "ts": 1776649791702012, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 435, "tid": 435, + "ts": 1776649791702013, "dur": 272, + "args": { + "External id": 7922, + "cbid": 20, "correlation": 62371 + } + }, + { + "ph": "f", "id": 62371, "pid": 435, "tid": 435, "ts": 1776649791702013, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791702743, "dur": 76, + "args": { + "External id": 7922, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62375, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 62375, "pid": 0, "tid": 7, "ts": 1776649791702743, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791702299, "dur": 12, + "args": { + "External id": 7922, + "cbid": 211, "correlation": 62375 + } + }, + { + "ph": "s", "id": 62375, "pid": 435, "tid": 435, "ts": 1776649791702299, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649791702341, "dur": 1, + "args": { + "External id": 7925, + "cbid": 317, "correlation": 62384 + } + }, + { + "ph": "f", "id": 62384, "pid": 435, "tid": 435, "ts": 1776649791702341, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 435, "tid": 435, + "ts": 1776649791702342, "dur": 220, + "args": { + "External id": 7925, + "cbid": 20, "correlation": 62385 + } + }, + { + "ph": "f", "id": 62385, "pid": 435, "tid": 435, "ts": 1776649791702342, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649791702820, "dur": 193, + "args": { + "External id": 7926, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62392, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 62392, "pid": 0, "tid": 7, "ts": 1776649791702820, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791702581, "dur": 13, + "args": { + "External id": 7926, + "cbid": 211, "correlation": 62392 + } + }, + { + "ph": "s", "id": 62392, "pid": 435, "tid": 435, "ts": 1776649791702581, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, at::detail::Array >(int, at::native::CUDAFunctorOnSelf_add, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791703016, "dur": 150, + "args": { + "External id": 7927, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62399, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 62399, "pid": 0, "tid": 7, "ts": 1776649791703016, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791702636, "dur": 9, + "args": { + "External id": 7927, + "cbid": 211, "correlation": 62399 + } + }, + { + "ph": "s", "id": 62399, "pid": 435, "tid": 435, "ts": 1776649791702636, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649791702691, "dur": 1, + "args": { + "External id": 7928, + "cbid": 317, "correlation": 62408 + } + }, + { + "ph": "f", "id": 62408, "pid": 435, "tid": 435, "ts": 1776649791702691, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 435, "tid": 435, + "ts": 1776649791702692, "dur": 225, + "args": { + "External id": 7928, + "cbid": 20, "correlation": 62409 + } + }, + { + "ph": "f", "id": 62409, "pid": 435, "tid": 435, "ts": 1776649791702692, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1}, at::detail::Array >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791703168, "dur": 155, + "args": { + "External id": 7928, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62413, + "registers per thread": 22, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 62413, "pid": 0, "tid": 7, "ts": 1776649791703168, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791702933, "dur": 13, + "args": { + "External id": 7928, + "cbid": 211, "correlation": 62413 + } + }, + { + "ph": "s", "id": 62413, "pid": 435, "tid": 435, "ts": 1776649791702933, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649791703325, "dur": 198, + "args": { + "External id": 7938, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62426, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 1213.629639, + "warps per SM": 4854.518555, + "grid": [131072, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 62426, "pid": 0, "tid": 7, "ts": 1776649791703325, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791703012, "dur": 9, + "args": { + "External id": 7938, + "cbid": 211, "correlation": 62426 + } + }, + { + "ph": "s", "id": 62426, "pid": 435, "tid": 435, "ts": 1776649791703012, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649791703526, "dur": 1, + "args": { + "External id": 7948, + "device": 0, "context": 1, + "stream": 7, "correlation": 62441, + "bytes": 512, "memory bandwidth (GB/s)": 0.3902439024390244 + } + }, + { + "ph": "f", "id": 62441, "pid": 0, "tid": 7, "ts": 1776649791703526, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649791703110, "dur": 10, + "args": { + "External id": 7948, + "cbid": 41, "correlation": 62441 + } + }, + { + "ph": "s", "id": 62441, "pid": 435, "tid": 435, "ts": 1776649791703110, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791703120, "dur": 412, + "args": { + "External id": 7948, + "cbid": 131, "correlation": 62442 + } + }, + { + "ph": "s", "id": 62442, "pid": 435, "tid": 435, "ts": 1776649791703120, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649791703574, "dur": 1, + "args": { + "External id": 7950, + "cbid": 317, "correlation": 62450 + } + }, + { + "ph": "f", "id": 62450, "pid": 435, "tid": 435, "ts": 1776649791703574, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 435, "tid": 435, + "ts": 1776649791703575, "dur": 220, + "args": { + "External id": 7950, + "cbid": 20, "correlation": 62451 + } + }, + { + "ph": "f", "id": 62451, "pid": 435, "tid": 435, "ts": 1776649791703575, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649791703824, "dur": 9, + "args": { + "External id": 7950, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62455, + "registers per thread": 40, + "shared memory": 0, + "blocks per SM": 4.740741, + "warps per SM": 18.962963, + "grid": [512, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 30 + } + }, + { + "ph": "f", "id": 62455, "pid": 0, "tid": 7, "ts": 1776649791703824, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791703809, "dur": 14, + "args": { + "External id": 7950, + "cbid": 211, "correlation": 62455 + } + }, + { + "ph": "s", "id": 62455, "pid": 435, "tid": 435, "ts": 1776649791703809, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649791703906, "dur": 216, + "args": { + "External id": 7964, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62470, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 1213.629639, + "warps per SM": 4854.518555, + "grid": [131072, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 62470, "pid": 0, "tid": 7, "ts": 1776649791703906, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791703894, "dur": 9, + "args": { + "External id": 7964, + "cbid": 211, "correlation": 62470 + } + }, + { + "ph": "s", "id": 62470, "pid": 435, "tid": 435, "ts": 1776649791703894, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649791704125, "dur": 1, + "args": { + "External id": 7974, + "device": 0, "context": 1, + "stream": 7, "correlation": 62485, + "bytes": 8192, "memory bandwidth (GB/s)": 4.491228070175438 + } + }, + { + "ph": "f", "id": 62485, "pid": 0, "tid": 7, "ts": 1776649791704125, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649791703958, "dur": 10, + "args": { + "External id": 7974, + "cbid": 41, "correlation": 62485 + } + }, + { + "ph": "s", "id": 62485, "pid": 435, "tid": 435, "ts": 1776649791703958, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791703968, "dur": 163, + "args": { + "External id": 7974, + "cbid": 131, "correlation": 62486 + } + }, + { + "ph": "s", "id": 62486, "pid": 435, "tid": 435, "ts": 1776649791703968, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649791704164, "dur": 246, + "args": { + "External id": 7970, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62495, + "registers per thread": 40, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 75 + } + }, + { + "ph": "f", "id": 62495, "pid": 0, "tid": 7, "ts": 1776649791704164, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791704151, "dur": 10, + "args": { + "External id": 7970, + "cbid": 211, "correlation": 62495 + } + }, + { + "ph": "s", "id": 62495, "pid": 435, "tid": 435, "ts": 1776649791704151, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoH (Device -> Pageable)", "pid": 0, "tid": 7, + "ts": 1776649791704854, "dur": 91329, + "args": { + "External id": 7983, + "device": 0, "context": 1, + "stream": 7, "correlation": 62503, + "bytes": 134217728, "memory bandwidth (GB/s)": 1.4696005859756518 + } + }, + { + "ph": "f", "id": 62503, "pid": 0, "tid": 7, "ts": 1776649791704854, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649791704835, "dur": 92664, + "args": { + "External id": 7983, + "cbid": 41, "correlation": 62503 + } + }, + { + "ph": "s", "id": 62503, "pid": 435, "tid": 435, "ts": 1776649791704835, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791797506, "dur": 16, + "args": { + "External id": 7983, + "cbid": 131, "correlation": 62504 + } + }, + { + "ph": "s", "id": 62504, "pid": 435, "tid": 435, "ts": 1776649791797506, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649791929375, "dur": 1209, + "args": { + "External id": 7993, + "device": 0, "context": 1, + "stream": 7, "correlation": 62516, + "bytes": 16777216, "memory bandwidth (GB/s)": 13.869376411038152 + } + }, + { + "ph": "f", "id": 62516, "pid": 0, "tid": 7, "ts": 1776649791929375, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649791929211, "dur": 1323, + "args": { + "External id": 7993, + "cbid": 41, "correlation": 62516 + } + }, + { + "ph": "s", "id": 62516, "pid": 435, "tid": 435, "ts": 1776649791929211, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791930544, "dur": 51, + "args": { + "External id": 7993, + "cbid": 131, "correlation": 62517 + } + }, + { + "ph": "s", "id": 62517, "pid": 435, "tid": 435, "ts": 1776649791930544, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7, + "ts": 1776649791932555, "dur": 16, + "args": { + "External id": 7999, + "device": 0, "context": 1, + "stream": 7, "correlation": 62522, + "bytes": 16777216, "memory bandwidth (GB/s)": 1042.3220675944333 + } + }, + { + "ph": "f", "id": 62522, "pid": 0, "tid": 7, "ts": 1776649791932555, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649791932496, "dur": 57, + "args": { + "External id": 7999, + "cbid": 41, "correlation": 62522 + } + }, + { + "ph": "s", "id": 62522, "pid": 435, "tid": 435, "ts": 1776649791932496, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7, + "ts": 1776649791932623, "dur": 3, + "args": { + "External id": 8005, + "device": 0, "context": 1, + "stream": 7, "correlation": 62530, + "bytes": 524288, "memory bandwidth (GB/s)": 141.24137931034483 + } + }, + { + "ph": "f", "id": 62530, "pid": 0, "tid": 7, "ts": 1776649791932623, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649791932605, "dur": 15, + "args": { + "External id": 8005, + "cbid": 41, "correlation": 62530 + } + }, + { + "ph": "s", "id": 62530, "pid": 435, "tid": 435, "ts": 1776649791932605, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791934495, "dur": 1, + "args": { + "External id": 8010, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62553, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62553, "pid": 0, "tid": 7, "ts": 1776649791934495, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791934475, "dur": 25, + "args": { + "External id": 8010, + "cbid": 211, "correlation": 62553 + } + }, + { + "ph": "s", "id": 62553, "pid": 435, "tid": 435, "ts": 1776649791934475, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791934518, "dur": 1, + "args": { + "External id": 62559, + "cbid": 200, "correlation": 62559 + } + }, + { + "ph": "f", "id": 62559, "pid": 435, "tid": 435, "ts": 1776649791934518, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791934521, "dur": 3, + "args": { + "External id": 62560, + "cbid": 273, "correlation": 62560 + } + }, + { + "ph": "f", "id": 62560, "pid": 435, "tid": 435, "ts": 1776649791934521, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791934535, "dur": 19, + "args": { + "External id": 62561, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62561, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62561, "pid": 0, "tid": 7, "ts": 1776649791934535, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791934526, "dur": 7, + "args": { + "External id": 62561, + "cbid": 211, "correlation": 62561 + } + }, + { + "ph": "s", "id": 62561, "pid": 435, "tid": 435, "ts": 1776649791934526, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791934580, "dur": 1, + "args": { + "External id": 8015, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62579, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62579, "pid": 0, "tid": 7, "ts": 1776649791934580, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791934571, "dur": 7, + "args": { + "External id": 8015, + "cbid": 211, "correlation": 62579 + } + }, + { + "ph": "s", "id": 62579, "pid": 435, "tid": 435, "ts": 1776649791934571, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791934584, "dur": 0, + "args": { + "External id": 62585, + "cbid": 200, "correlation": 62585 + } + }, + { + "ph": "f", "id": 62585, "pid": 435, "tid": 435, "ts": 1776649791934584, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791934585, "dur": 0, + "args": { + "External id": 62586, + "cbid": 273, "correlation": 62586 + } + }, + { + "ph": "f", "id": 62586, "pid": 435, "tid": 435, "ts": 1776649791934585, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791934593, "dur": 19, + "args": { + "External id": 62587, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62587, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62587, "pid": 0, "tid": 7, "ts": 1776649791934593, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791934586, "dur": 4, + "args": { + "External id": 62587, + "cbid": 211, "correlation": 62587 + } + }, + { + "ph": "s", "id": 62587, "pid": 435, "tid": 435, "ts": 1776649791934586, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791934630, "dur": 1, + "args": { + "External id": 8020, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62605, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62605, "pid": 0, "tid": 7, "ts": 1776649791934630, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791934621, "dur": 6, + "args": { + "External id": 8020, + "cbid": 211, "correlation": 62605 + } + }, + { + "ph": "s", "id": 62605, "pid": 435, "tid": 435, "ts": 1776649791934621, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791934633, "dur": 0, + "args": { + "External id": 62611, + "cbid": 200, "correlation": 62611 + } + }, + { + "ph": "f", "id": 62611, "pid": 435, "tid": 435, "ts": 1776649791934633, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791934634, "dur": 0, + "args": { + "External id": 62612, + "cbid": 273, "correlation": 62612 + } + }, + { + "ph": "f", "id": 62612, "pid": 435, "tid": 435, "ts": 1776649791934634, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791934642, "dur": 19, + "args": { + "External id": 62613, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62613, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62613, "pid": 0, "tid": 7, "ts": 1776649791934642, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791934635, "dur": 4, + "args": { + "External id": 62613, + "cbid": 211, "correlation": 62613 + } + }, + { + "ph": "s", "id": 62613, "pid": 435, "tid": 435, "ts": 1776649791934635, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791934679, "dur": 1, + "args": { + "External id": 8025, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62631, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62631, "pid": 0, "tid": 7, "ts": 1776649791934679, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791934670, "dur": 7, + "args": { + "External id": 8025, + "cbid": 211, "correlation": 62631 + } + }, + { + "ph": "s", "id": 62631, "pid": 435, "tid": 435, "ts": 1776649791934670, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791934683, "dur": 0, + "args": { + "External id": 62637, + "cbid": 200, "correlation": 62637 + } + }, + { + "ph": "f", "id": 62637, "pid": 435, "tid": 435, "ts": 1776649791934683, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791934683, "dur": 0, + "args": { + "External id": 62638, + "cbid": 273, "correlation": 62638 + } + }, + { + "ph": "f", "id": 62638, "pid": 435, "tid": 435, "ts": 1776649791934683, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791934692, "dur": 19, + "args": { + "External id": 62639, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62639, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62639, "pid": 0, "tid": 7, "ts": 1776649791934692, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791934685, "dur": 4, + "args": { + "External id": 62639, + "cbid": 211, "correlation": 62639 + } + }, + { + "ph": "s", "id": 62639, "pid": 435, "tid": 435, "ts": 1776649791934685, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791934728, "dur": 1, + "args": { + "External id": 8030, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62657, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62657, "pid": 0, "tid": 7, "ts": 1776649791934728, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791934719, "dur": 6, + "args": { + "External id": 8030, + "cbid": 211, "correlation": 62657 + } + }, + { + "ph": "s", "id": 62657, "pid": 435, "tid": 435, "ts": 1776649791934719, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791934731, "dur": 0, + "args": { + "External id": 62663, + "cbid": 200, "correlation": 62663 + } + }, + { + "ph": "f", "id": 62663, "pid": 435, "tid": 435, "ts": 1776649791934731, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791934732, "dur": 0, + "args": { + "External id": 62664, + "cbid": 273, "correlation": 62664 + } + }, + { + "ph": "f", "id": 62664, "pid": 435, "tid": 435, "ts": 1776649791934732, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791934741, "dur": 19, + "args": { + "External id": 62665, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62665, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62665, "pid": 0, "tid": 7, "ts": 1776649791934741, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791934733, "dur": 4, + "args": { + "External id": 62665, + "cbid": 211, "correlation": 62665 + } + }, + { + "ph": "s", "id": 62665, "pid": 435, "tid": 435, "ts": 1776649791934733, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791934778, "dur": 1, + "args": { + "External id": 8035, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62683, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62683, "pid": 0, "tid": 7, "ts": 1776649791934778, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791934768, "dur": 6, + "args": { + "External id": 8035, + "cbid": 211, "correlation": 62683 + } + }, + { + "ph": "s", "id": 62683, "pid": 435, "tid": 435, "ts": 1776649791934768, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791934781, "dur": 0, + "args": { + "External id": 62689, + "cbid": 200, "correlation": 62689 + } + }, + { + "ph": "f", "id": 62689, "pid": 435, "tid": 435, "ts": 1776649791934781, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791934781, "dur": 0, + "args": { + "External id": 62690, + "cbid": 273, "correlation": 62690 + } + }, + { + "ph": "f", "id": 62690, "pid": 435, "tid": 435, "ts": 1776649791934781, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791934790, "dur": 19, + "args": { + "External id": 62691, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62691, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62691, "pid": 0, "tid": 7, "ts": 1776649791934790, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791934782, "dur": 4, + "args": { + "External id": 62691, + "cbid": 211, "correlation": 62691 + } + }, + { + "ph": "s", "id": 62691, "pid": 435, "tid": 435, "ts": 1776649791934782, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791934826, "dur": 1, + "args": { + "External id": 8040, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62709, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62709, "pid": 0, "tid": 7, "ts": 1776649791934826, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791934817, "dur": 6, + "args": { + "External id": 8040, + "cbid": 211, "correlation": 62709 + } + }, + { + "ph": "s", "id": 62709, "pid": 435, "tid": 435, "ts": 1776649791934817, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791934829, "dur": 0, + "args": { + "External id": 62715, + "cbid": 200, "correlation": 62715 + } + }, + { + "ph": "f", "id": 62715, "pid": 435, "tid": 435, "ts": 1776649791934829, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791934830, "dur": 0, + "args": { + "External id": 62716, + "cbid": 273, "correlation": 62716 + } + }, + { + "ph": "f", "id": 62716, "pid": 435, "tid": 435, "ts": 1776649791934830, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791934839, "dur": 19, + "args": { + "External id": 62717, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62717, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62717, "pid": 0, "tid": 7, "ts": 1776649791934839, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791934831, "dur": 5, + "args": { + "External id": 62717, + "cbid": 211, "correlation": 62717 + } + }, + { + "ph": "s", "id": 62717, "pid": 435, "tid": 435, "ts": 1776649791934831, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791934874, "dur": 1, + "args": { + "External id": 8045, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62735, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62735, "pid": 0, "tid": 7, "ts": 1776649791934874, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791934865, "dur": 6, + "args": { + "External id": 8045, + "cbid": 211, "correlation": 62735 + } + }, + { + "ph": "s", "id": 62735, "pid": 435, "tid": 435, "ts": 1776649791934865, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791934877, "dur": 0, + "args": { + "External id": 62741, + "cbid": 200, "correlation": 62741 + } + }, + { + "ph": "f", "id": 62741, "pid": 435, "tid": 435, "ts": 1776649791934877, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791934878, "dur": 0, + "args": { + "External id": 62742, + "cbid": 273, "correlation": 62742 + } + }, + { + "ph": "f", "id": 62742, "pid": 435, "tid": 435, "ts": 1776649791934878, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791934887, "dur": 19, + "args": { + "External id": 62743, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62743, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62743, "pid": 0, "tid": 7, "ts": 1776649791934887, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791934879, "dur": 4, + "args": { + "External id": 62743, + "cbid": 211, "correlation": 62743 + } + }, + { + "ph": "s", "id": 62743, "pid": 435, "tid": 435, "ts": 1776649791934879, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791934924, "dur": 1, + "args": { + "External id": 8050, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62761, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62761, "pid": 0, "tid": 7, "ts": 1776649791934924, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791934916, "dur": 6, + "args": { + "External id": 8050, + "cbid": 211, "correlation": 62761 + } + }, + { + "ph": "s", "id": 62761, "pid": 435, "tid": 435, "ts": 1776649791934916, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791934928, "dur": 0, + "args": { + "External id": 62767, + "cbid": 200, "correlation": 62767 + } + }, + { + "ph": "f", "id": 62767, "pid": 435, "tid": 435, "ts": 1776649791934928, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791934929, "dur": 0, + "args": { + "External id": 62768, + "cbid": 273, "correlation": 62768 + } + }, + { + "ph": "f", "id": 62768, "pid": 435, "tid": 435, "ts": 1776649791934929, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791934937, "dur": 19, + "args": { + "External id": 62769, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62769, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62769, "pid": 0, "tid": 7, "ts": 1776649791934937, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791934930, "dur": 4, + "args": { + "External id": 62769, + "cbid": 211, "correlation": 62769 + } + }, + { + "ph": "s", "id": 62769, "pid": 435, "tid": 435, "ts": 1776649791934930, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791934973, "dur": 1, + "args": { + "External id": 8055, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62787, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62787, "pid": 0, "tid": 7, "ts": 1776649791934973, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791934964, "dur": 6, + "args": { + "External id": 8055, + "cbid": 211, "correlation": 62787 + } + }, + { + "ph": "s", "id": 62787, "pid": 435, "tid": 435, "ts": 1776649791934964, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791934976, "dur": 0, + "args": { + "External id": 62793, + "cbid": 200, "correlation": 62793 + } + }, + { + "ph": "f", "id": 62793, "pid": 435, "tid": 435, "ts": 1776649791934976, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791934976, "dur": 0, + "args": { + "External id": 62794, + "cbid": 273, "correlation": 62794 + } + }, + { + "ph": "f", "id": 62794, "pid": 435, "tid": 435, "ts": 1776649791934976, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791934985, "dur": 19, + "args": { + "External id": 62795, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62795, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62795, "pid": 0, "tid": 7, "ts": 1776649791934985, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791934978, "dur": 4, + "args": { + "External id": 62795, + "cbid": 211, "correlation": 62795 + } + }, + { + "ph": "s", "id": 62795, "pid": 435, "tid": 435, "ts": 1776649791934978, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791935147, "dur": 14, + "args": { + "External id": 62798, + "cbid": 165, "correlation": 62798 + } + }, + { + "ph": "s", "id": 62798, "pid": 435, "tid": 435, "ts": 1776649791935147, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791935230, "dur": 8, + "args": { + "External id": 62807, + "cbid": 135, "correlation": 62807 + } + }, + { + "ph": "f", "id": 62807, "pid": 435, "tid": 435, "ts": 1776649791935230, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791935293, "dur": 1, + "args": { + "External id": 8060, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62827, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62827, "pid": 0, "tid": 7, "ts": 1776649791935293, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935281, "dur": 9, + "args": { + "External id": 8060, + "cbid": 211, "correlation": 62827 + } + }, + { + "ph": "s", "id": 62827, "pid": 435, "tid": 435, "ts": 1776649791935281, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935298, "dur": 0, + "args": { + "External id": 62833, + "cbid": 200, "correlation": 62833 + } + }, + { + "ph": "f", "id": 62833, "pid": 435, "tid": 435, "ts": 1776649791935298, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935299, "dur": 1, + "args": { + "External id": 62834, + "cbid": 273, "correlation": 62834 + } + }, + { + "ph": "f", "id": 62834, "pid": 435, "tid": 435, "ts": 1776649791935299, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791935308, "dur": 19, + "args": { + "External id": 62835, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62835, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62835, "pid": 0, "tid": 7, "ts": 1776649791935308, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935300, "dur": 5, + "args": { + "External id": 62835, + "cbid": 211, "correlation": 62835 + } + }, + { + "ph": "s", "id": 62835, "pid": 435, "tid": 435, "ts": 1776649791935300, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791935347, "dur": 1, + "args": { + "External id": 8065, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62853, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62853, "pid": 0, "tid": 7, "ts": 1776649791935347, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935338, "dur": 7, + "args": { + "External id": 8065, + "cbid": 211, "correlation": 62853 + } + }, + { + "ph": "s", "id": 62853, "pid": 435, "tid": 435, "ts": 1776649791935338, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935351, "dur": 0, + "args": { + "External id": 62859, + "cbid": 200, "correlation": 62859 + } + }, + { + "ph": "f", "id": 62859, "pid": 435, "tid": 435, "ts": 1776649791935351, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935352, "dur": 0, + "args": { + "External id": 62860, + "cbid": 273, "correlation": 62860 + } + }, + { + "ph": "f", "id": 62860, "pid": 435, "tid": 435, "ts": 1776649791935352, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791935361, "dur": 19, + "args": { + "External id": 62861, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62861, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62861, "pid": 0, "tid": 7, "ts": 1776649791935361, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935353, "dur": 4, + "args": { + "External id": 62861, + "cbid": 211, "correlation": 62861 + } + }, + { + "ph": "s", "id": 62861, "pid": 435, "tid": 435, "ts": 1776649791935353, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791935397, "dur": 1, + "args": { + "External id": 8070, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62879, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62879, "pid": 0, "tid": 7, "ts": 1776649791935397, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935387, "dur": 6, + "args": { + "External id": 8070, + "cbid": 211, "correlation": 62879 + } + }, + { + "ph": "s", "id": 62879, "pid": 435, "tid": 435, "ts": 1776649791935387, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935400, "dur": 0, + "args": { + "External id": 62885, + "cbid": 200, "correlation": 62885 + } + }, + { + "ph": "f", "id": 62885, "pid": 435, "tid": 435, "ts": 1776649791935400, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935401, "dur": 0, + "args": { + "External id": 62886, + "cbid": 273, "correlation": 62886 + } + }, + { + "ph": "f", "id": 62886, "pid": 435, "tid": 435, "ts": 1776649791935401, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791935409, "dur": 19, + "args": { + "External id": 62887, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62887, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62887, "pid": 0, "tid": 7, "ts": 1776649791935409, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935402, "dur": 5, + "args": { + "External id": 62887, + "cbid": 211, "correlation": 62887 + } + }, + { + "ph": "s", "id": 62887, "pid": 435, "tid": 435, "ts": 1776649791935402, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791935445, "dur": 1, + "args": { + "External id": 8075, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62905, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62905, "pid": 0, "tid": 7, "ts": 1776649791935445, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935436, "dur": 6, + "args": { + "External id": 8075, + "cbid": 211, "correlation": 62905 + } + }, + { + "ph": "s", "id": 62905, "pid": 435, "tid": 435, "ts": 1776649791935436, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935449, "dur": 0, + "args": { + "External id": 62911, + "cbid": 200, "correlation": 62911 + } + }, + { + "ph": "f", "id": 62911, "pid": 435, "tid": 435, "ts": 1776649791935449, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935449, "dur": 0, + "args": { + "External id": 62912, + "cbid": 273, "correlation": 62912 + } + }, + { + "ph": "f", "id": 62912, "pid": 435, "tid": 435, "ts": 1776649791935449, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791935458, "dur": 19, + "args": { + "External id": 62913, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62913, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62913, "pid": 0, "tid": 7, "ts": 1776649791935458, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935450, "dur": 4, + "args": { + "External id": 62913, + "cbid": 211, "correlation": 62913 + } + }, + { + "ph": "s", "id": 62913, "pid": 435, "tid": 435, "ts": 1776649791935450, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791935493, "dur": 1, + "args": { + "External id": 8080, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62931, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62931, "pid": 0, "tid": 7, "ts": 1776649791935493, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935484, "dur": 6, + "args": { + "External id": 8080, + "cbid": 211, "correlation": 62931 + } + }, + { + "ph": "s", "id": 62931, "pid": 435, "tid": 435, "ts": 1776649791935484, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935497, "dur": 0, + "args": { + "External id": 62937, + "cbid": 200, "correlation": 62937 + } + }, + { + "ph": "f", "id": 62937, "pid": 435, "tid": 435, "ts": 1776649791935497, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935497, "dur": 0, + "args": { + "External id": 62938, + "cbid": 273, "correlation": 62938 + } + }, + { + "ph": "f", "id": 62938, "pid": 435, "tid": 435, "ts": 1776649791935497, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791935506, "dur": 19, + "args": { + "External id": 62939, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62939, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62939, "pid": 0, "tid": 7, "ts": 1776649791935506, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935498, "dur": 4, + "args": { + "External id": 62939, + "cbid": 211, "correlation": 62939 + } + }, + { + "ph": "s", "id": 62939, "pid": 435, "tid": 435, "ts": 1776649791935498, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791935541, "dur": 1, + "args": { + "External id": 8085, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62957, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62957, "pid": 0, "tid": 7, "ts": 1776649791935541, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935532, "dur": 6, + "args": { + "External id": 8085, + "cbid": 211, "correlation": 62957 + } + }, + { + "ph": "s", "id": 62957, "pid": 435, "tid": 435, "ts": 1776649791935532, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935544, "dur": 0, + "args": { + "External id": 62963, + "cbid": 200, "correlation": 62963 + } + }, + { + "ph": "f", "id": 62963, "pid": 435, "tid": 435, "ts": 1776649791935544, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935545, "dur": 0, + "args": { + "External id": 62964, + "cbid": 273, "correlation": 62964 + } + }, + { + "ph": "f", "id": 62964, "pid": 435, "tid": 435, "ts": 1776649791935545, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791935553, "dur": 19, + "args": { + "External id": 62965, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62965, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62965, "pid": 0, "tid": 7, "ts": 1776649791935553, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935546, "dur": 4, + "args": { + "External id": 62965, + "cbid": 211, "correlation": 62965 + } + }, + { + "ph": "s", "id": 62965, "pid": 435, "tid": 435, "ts": 1776649791935546, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791935589, "dur": 1, + "args": { + "External id": 8090, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62983, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62983, "pid": 0, "tid": 7, "ts": 1776649791935589, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935580, "dur": 6, + "args": { + "External id": 8090, + "cbid": 211, "correlation": 62983 + } + }, + { + "ph": "s", "id": 62983, "pid": 435, "tid": 435, "ts": 1776649791935580, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935592, "dur": 0, + "args": { + "External id": 62989, + "cbid": 200, "correlation": 62989 + } + }, + { + "ph": "f", "id": 62989, "pid": 435, "tid": 435, "ts": 1776649791935592, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935592, "dur": 0, + "args": { + "External id": 62990, + "cbid": 273, "correlation": 62990 + } + }, + { + "ph": "f", "id": 62990, "pid": 435, "tid": 435, "ts": 1776649791935592, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791935601, "dur": 19, + "args": { + "External id": 62991, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62991, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 62991, "pid": 0, "tid": 7, "ts": 1776649791935601, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935593, "dur": 4, + "args": { + "External id": 62991, + "cbid": 211, "correlation": 62991 + } + }, + { + "ph": "s", "id": 62991, "pid": 435, "tid": 435, "ts": 1776649791935593, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791935636, "dur": 1, + "args": { + "External id": 8095, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63009, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63009, "pid": 0, "tid": 7, "ts": 1776649791935636, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935627, "dur": 6, + "args": { + "External id": 8095, + "cbid": 211, "correlation": 63009 + } + }, + { + "ph": "s", "id": 63009, "pid": 435, "tid": 435, "ts": 1776649791935627, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935640, "dur": 0, + "args": { + "External id": 63015, + "cbid": 200, "correlation": 63015 + } + }, + { + "ph": "f", "id": 63015, "pid": 435, "tid": 435, "ts": 1776649791935640, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935640, "dur": 0, + "args": { + "External id": 63016, + "cbid": 273, "correlation": 63016 + } + }, + { + "ph": "f", "id": 63016, "pid": 435, "tid": 435, "ts": 1776649791935640, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791935649, "dur": 19, + "args": { + "External id": 63017, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63017, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63017, "pid": 0, "tid": 7, "ts": 1776649791935649, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935641, "dur": 4, + "args": { + "External id": 63017, + "cbid": 211, "correlation": 63017 + } + }, + { + "ph": "s", "id": 63017, "pid": 435, "tid": 435, "ts": 1776649791935641, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791935686, "dur": 1, + "args": { + "External id": 8100, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63035, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63035, "pid": 0, "tid": 7, "ts": 1776649791935686, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935676, "dur": 6, + "args": { + "External id": 8100, + "cbid": 211, "correlation": 63035 + } + }, + { + "ph": "s", "id": 63035, "pid": 435, "tid": 435, "ts": 1776649791935676, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935689, "dur": 0, + "args": { + "External id": 63041, + "cbid": 200, "correlation": 63041 + } + }, + { + "ph": "f", "id": 63041, "pid": 435, "tid": 435, "ts": 1776649791935689, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935689, "dur": 0, + "args": { + "External id": 63042, + "cbid": 273, "correlation": 63042 + } + }, + { + "ph": "f", "id": 63042, "pid": 435, "tid": 435, "ts": 1776649791935689, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791935698, "dur": 19, + "args": { + "External id": 63043, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63043, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63043, "pid": 0, "tid": 7, "ts": 1776649791935698, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935691, "dur": 4, + "args": { + "External id": 63043, + "cbid": 211, "correlation": 63043 + } + }, + { + "ph": "s", "id": 63043, "pid": 435, "tid": 435, "ts": 1776649791935691, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791935733, "dur": 1, + "args": { + "External id": 8105, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63061, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63061, "pid": 0, "tid": 7, "ts": 1776649791935733, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935724, "dur": 6, + "args": { + "External id": 8105, + "cbid": 211, "correlation": 63061 + } + }, + { + "ph": "s", "id": 63061, "pid": 435, "tid": 435, "ts": 1776649791935724, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935737, "dur": 0, + "args": { + "External id": 63067, + "cbid": 200, "correlation": 63067 + } + }, + { + "ph": "f", "id": 63067, "pid": 435, "tid": 435, "ts": 1776649791935737, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935737, "dur": 0, + "args": { + "External id": 63068, + "cbid": 273, "correlation": 63068 + } + }, + { + "ph": "f", "id": 63068, "pid": 435, "tid": 435, "ts": 1776649791935737, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791935746, "dur": 19, + "args": { + "External id": 63069, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63069, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63069, "pid": 0, "tid": 7, "ts": 1776649791935746, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935738, "dur": 4, + "args": { + "External id": 63069, + "cbid": 211, "correlation": 63069 + } + }, + { + "ph": "s", "id": 63069, "pid": 435, "tid": 435, "ts": 1776649791935738, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791935781, "dur": 1, + "args": { + "External id": 8110, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63087, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63087, "pid": 0, "tid": 7, "ts": 1776649791935781, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935772, "dur": 6, + "args": { + "External id": 8110, + "cbid": 211, "correlation": 63087 + } + }, + { + "ph": "s", "id": 63087, "pid": 435, "tid": 435, "ts": 1776649791935772, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935784, "dur": 0, + "args": { + "External id": 63093, + "cbid": 200, "correlation": 63093 + } + }, + { + "ph": "f", "id": 63093, "pid": 435, "tid": 435, "ts": 1776649791935784, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935785, "dur": 0, + "args": { + "External id": 63094, + "cbid": 273, "correlation": 63094 + } + }, + { + "ph": "f", "id": 63094, "pid": 435, "tid": 435, "ts": 1776649791935785, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791935794, "dur": 19, + "args": { + "External id": 63095, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63095, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63095, "pid": 0, "tid": 7, "ts": 1776649791935794, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935786, "dur": 4, + "args": { + "External id": 63095, + "cbid": 211, "correlation": 63095 + } + }, + { + "ph": "s", "id": 63095, "pid": 435, "tid": 435, "ts": 1776649791935786, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791935830, "dur": 1, + "args": { + "External id": 8115, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63113, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63113, "pid": 0, "tid": 7, "ts": 1776649791935830, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935821, "dur": 6, + "args": { + "External id": 8115, + "cbid": 211, "correlation": 63113 + } + }, + { + "ph": "s", "id": 63113, "pid": 435, "tid": 435, "ts": 1776649791935821, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935833, "dur": 0, + "args": { + "External id": 63119, + "cbid": 200, "correlation": 63119 + } + }, + { + "ph": "f", "id": 63119, "pid": 435, "tid": 435, "ts": 1776649791935833, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935834, "dur": 0, + "args": { + "External id": 63120, + "cbid": 273, "correlation": 63120 + } + }, + { + "ph": "f", "id": 63120, "pid": 435, "tid": 435, "ts": 1776649791935834, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791935842, "dur": 19, + "args": { + "External id": 63121, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63121, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63121, "pid": 0, "tid": 7, "ts": 1776649791935842, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935835, "dur": 4, + "args": { + "External id": 63121, + "cbid": 211, "correlation": 63121 + } + }, + { + "ph": "s", "id": 63121, "pid": 435, "tid": 435, "ts": 1776649791935835, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791935878, "dur": 1, + "args": { + "External id": 8120, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63139, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63139, "pid": 0, "tid": 7, "ts": 1776649791935878, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935869, "dur": 6, + "args": { + "External id": 8120, + "cbid": 211, "correlation": 63139 + } + }, + { + "ph": "s", "id": 63139, "pid": 435, "tid": 435, "ts": 1776649791935869, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935881, "dur": 0, + "args": { + "External id": 63145, + "cbid": 200, "correlation": 63145 + } + }, + { + "ph": "f", "id": 63145, "pid": 435, "tid": 435, "ts": 1776649791935881, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935881, "dur": 0, + "args": { + "External id": 63146, + "cbid": 273, "correlation": 63146 + } + }, + { + "ph": "f", "id": 63146, "pid": 435, "tid": 435, "ts": 1776649791935881, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791935891, "dur": 19, + "args": { + "External id": 63147, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63147, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63147, "pid": 0, "tid": 7, "ts": 1776649791935891, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935883, "dur": 5, + "args": { + "External id": 63147, + "cbid": 211, "correlation": 63147 + } + }, + { + "ph": "s", "id": 63147, "pid": 435, "tid": 435, "ts": 1776649791935883, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791935926, "dur": 1, + "args": { + "External id": 8125, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63165, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63165, "pid": 0, "tid": 7, "ts": 1776649791935926, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935916, "dur": 6, + "args": { + "External id": 8125, + "cbid": 211, "correlation": 63165 + } + }, + { + "ph": "s", "id": 63165, "pid": 435, "tid": 435, "ts": 1776649791935916, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935929, "dur": 0, + "args": { + "External id": 63171, + "cbid": 200, "correlation": 63171 + } + }, + { + "ph": "f", "id": 63171, "pid": 435, "tid": 435, "ts": 1776649791935929, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935930, "dur": 0, + "args": { + "External id": 63172, + "cbid": 273, "correlation": 63172 + } + }, + { + "ph": "f", "id": 63172, "pid": 435, "tid": 435, "ts": 1776649791935930, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791935938, "dur": 19, + "args": { + "External id": 63173, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63173, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63173, "pid": 0, "tid": 7, "ts": 1776649791935938, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935931, "dur": 4, + "args": { + "External id": 63173, + "cbid": 211, "correlation": 63173 + } + }, + { + "ph": "s", "id": 63173, "pid": 435, "tid": 435, "ts": 1776649791935931, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791935974, "dur": 1, + "args": { + "External id": 8130, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63191, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63191, "pid": 0, "tid": 7, "ts": 1776649791935974, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935965, "dur": 6, + "args": { + "External id": 8130, + "cbid": 211, "correlation": 63191 + } + }, + { + "ph": "s", "id": 63191, "pid": 435, "tid": 435, "ts": 1776649791935965, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935977, "dur": 0, + "args": { + "External id": 63197, + "cbid": 200, "correlation": 63197 + } + }, + { + "ph": "f", "id": 63197, "pid": 435, "tid": 435, "ts": 1776649791935977, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791935977, "dur": 0, + "args": { + "External id": 63198, + "cbid": 273, "correlation": 63198 + } + }, + { + "ph": "f", "id": 63198, "pid": 435, "tid": 435, "ts": 1776649791935977, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791935986, "dur": 19, + "args": { + "External id": 63199, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63199, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63199, "pid": 0, "tid": 7, "ts": 1776649791935986, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791935978, "dur": 4, + "args": { + "External id": 63199, + "cbid": 211, "correlation": 63199 + } + }, + { + "ph": "s", "id": 63199, "pid": 435, "tid": 435, "ts": 1776649791935978, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791936021, "dur": 1, + "args": { + "External id": 8135, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63217, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63217, "pid": 0, "tid": 7, "ts": 1776649791936021, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936012, "dur": 6, + "args": { + "External id": 8135, + "cbid": 211, "correlation": 63217 + } + }, + { + "ph": "s", "id": 63217, "pid": 435, "tid": 435, "ts": 1776649791936012, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936024, "dur": 0, + "args": { + "External id": 63223, + "cbid": 200, "correlation": 63223 + } + }, + { + "ph": "f", "id": 63223, "pid": 435, "tid": 435, "ts": 1776649791936024, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936024, "dur": 0, + "args": { + "External id": 63224, + "cbid": 273, "correlation": 63224 + } + }, + { + "ph": "f", "id": 63224, "pid": 435, "tid": 435, "ts": 1776649791936024, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791936033, "dur": 19, + "args": { + "External id": 63225, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63225, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63225, "pid": 0, "tid": 7, "ts": 1776649791936033, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936026, "dur": 4, + "args": { + "External id": 63225, + "cbid": 211, "correlation": 63225 + } + }, + { + "ph": "s", "id": 63225, "pid": 435, "tid": 435, "ts": 1776649791936026, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791936068, "dur": 1, + "args": { + "External id": 8140, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63243, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63243, "pid": 0, "tid": 7, "ts": 1776649791936068, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936059, "dur": 6, + "args": { + "External id": 8140, + "cbid": 211, "correlation": 63243 + } + }, + { + "ph": "s", "id": 63243, "pid": 435, "tid": 435, "ts": 1776649791936059, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936071, "dur": 0, + "args": { + "External id": 63249, + "cbid": 200, "correlation": 63249 + } + }, + { + "ph": "f", "id": 63249, "pid": 435, "tid": 435, "ts": 1776649791936071, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936072, "dur": 0, + "args": { + "External id": 63250, + "cbid": 273, "correlation": 63250 + } + }, + { + "ph": "f", "id": 63250, "pid": 435, "tid": 435, "ts": 1776649791936072, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791936081, "dur": 19, + "args": { + "External id": 63251, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63251, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63251, "pid": 0, "tid": 7, "ts": 1776649791936081, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936073, "dur": 4, + "args": { + "External id": 63251, + "cbid": 211, "correlation": 63251 + } + }, + { + "ph": "s", "id": 63251, "pid": 435, "tid": 435, "ts": 1776649791936073, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791936116, "dur": 1, + "args": { + "External id": 8145, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63269, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63269, "pid": 0, "tid": 7, "ts": 1776649791936116, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936106, "dur": 6, + "args": { + "External id": 8145, + "cbid": 211, "correlation": 63269 + } + }, + { + "ph": "s", "id": 63269, "pid": 435, "tid": 435, "ts": 1776649791936106, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936119, "dur": 0, + "args": { + "External id": 63275, + "cbid": 200, "correlation": 63275 + } + }, + { + "ph": "f", "id": 63275, "pid": 435, "tid": 435, "ts": 1776649791936119, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936120, "dur": 0, + "args": { + "External id": 63276, + "cbid": 273, "correlation": 63276 + } + }, + { + "ph": "f", "id": 63276, "pid": 435, "tid": 435, "ts": 1776649791936120, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791936129, "dur": 19, + "args": { + "External id": 63277, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63277, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63277, "pid": 0, "tid": 7, "ts": 1776649791936129, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936121, "dur": 4, + "args": { + "External id": 63277, + "cbid": 211, "correlation": 63277 + } + }, + { + "ph": "s", "id": 63277, "pid": 435, "tid": 435, "ts": 1776649791936121, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791936164, "dur": 1, + "args": { + "External id": 8150, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63295, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63295, "pid": 0, "tid": 7, "ts": 1776649791936164, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936155, "dur": 6, + "args": { + "External id": 8150, + "cbid": 211, "correlation": 63295 + } + }, + { + "ph": "s", "id": 63295, "pid": 435, "tid": 435, "ts": 1776649791936155, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936167, "dur": 0, + "args": { + "External id": 63301, + "cbid": 200, "correlation": 63301 + } + }, + { + "ph": "f", "id": 63301, "pid": 435, "tid": 435, "ts": 1776649791936167, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936167, "dur": 0, + "args": { + "External id": 63302, + "cbid": 273, "correlation": 63302 + } + }, + { + "ph": "f", "id": 63302, "pid": 435, "tid": 435, "ts": 1776649791936167, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791936176, "dur": 19, + "args": { + "External id": 63303, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63303, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63303, "pid": 0, "tid": 7, "ts": 1776649791936176, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936168, "dur": 4, + "args": { + "External id": 63303, + "cbid": 211, "correlation": 63303 + } + }, + { + "ph": "s", "id": 63303, "pid": 435, "tid": 435, "ts": 1776649791936168, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791936212, "dur": 1, + "args": { + "External id": 8155, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63321, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63321, "pid": 0, "tid": 7, "ts": 1776649791936212, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936204, "dur": 6, + "args": { + "External id": 8155, + "cbid": 211, "correlation": 63321 + } + }, + { + "ph": "s", "id": 63321, "pid": 435, "tid": 435, "ts": 1776649791936204, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936237, "dur": 0, + "args": { + "External id": 63327, + "cbid": 200, "correlation": 63327 + } + }, + { + "ph": "f", "id": 63327, "pid": 435, "tid": 435, "ts": 1776649791936237, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936238, "dur": 0, + "args": { + "External id": 63328, + "cbid": 273, "correlation": 63328 + } + }, + { + "ph": "f", "id": 63328, "pid": 435, "tid": 435, "ts": 1776649791936238, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791936247, "dur": 19, + "args": { + "External id": 63329, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63329, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63329, "pid": 0, "tid": 7, "ts": 1776649791936247, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936239, "dur": 5, + "args": { + "External id": 63329, + "cbid": 211, "correlation": 63329 + } + }, + { + "ph": "s", "id": 63329, "pid": 435, "tid": 435, "ts": 1776649791936239, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791936287, "dur": 1, + "args": { + "External id": 8160, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63347, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63347, "pid": 0, "tid": 7, "ts": 1776649791936287, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936278, "dur": 6, + "args": { + "External id": 8160, + "cbid": 211, "correlation": 63347 + } + }, + { + "ph": "s", "id": 63347, "pid": 435, "tid": 435, "ts": 1776649791936278, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936291, "dur": 0, + "args": { + "External id": 63353, + "cbid": 200, "correlation": 63353 + } + }, + { + "ph": "f", "id": 63353, "pid": 435, "tid": 435, "ts": 1776649791936291, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936291, "dur": 0, + "args": { + "External id": 63354, + "cbid": 273, "correlation": 63354 + } + }, + { + "ph": "f", "id": 63354, "pid": 435, "tid": 435, "ts": 1776649791936291, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791936301, "dur": 19, + "args": { + "External id": 63355, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63355, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63355, "pid": 0, "tid": 7, "ts": 1776649791936301, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936292, "dur": 5, + "args": { + "External id": 63355, + "cbid": 211, "correlation": 63355 + } + }, + { + "ph": "s", "id": 63355, "pid": 435, "tid": 435, "ts": 1776649791936292, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791936336, "dur": 1, + "args": { + "External id": 8165, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63373, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63373, "pid": 0, "tid": 7, "ts": 1776649791936336, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936327, "dur": 6, + "args": { + "External id": 8165, + "cbid": 211, "correlation": 63373 + } + }, + { + "ph": "s", "id": 63373, "pid": 435, "tid": 435, "ts": 1776649791936327, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936339, "dur": 0, + "args": { + "External id": 63379, + "cbid": 200, "correlation": 63379 + } + }, + { + "ph": "f", "id": 63379, "pid": 435, "tid": 435, "ts": 1776649791936339, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936340, "dur": 0, + "args": { + "External id": 63380, + "cbid": 273, "correlation": 63380 + } + }, + { + "ph": "f", "id": 63380, "pid": 435, "tid": 435, "ts": 1776649791936340, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791936348, "dur": 19, + "args": { + "External id": 63381, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63381, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63381, "pid": 0, "tid": 7, "ts": 1776649791936348, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936341, "dur": 4, + "args": { + "External id": 63381, + "cbid": 211, "correlation": 63381 + } + }, + { + "ph": "s", "id": 63381, "pid": 435, "tid": 435, "ts": 1776649791936341, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791936384, "dur": 1, + "args": { + "External id": 8170, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63399, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63399, "pid": 0, "tid": 7, "ts": 1776649791936384, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936375, "dur": 6, + "args": { + "External id": 8170, + "cbid": 211, "correlation": 63399 + } + }, + { + "ph": "s", "id": 63399, "pid": 435, "tid": 435, "ts": 1776649791936375, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936388, "dur": 0, + "args": { + "External id": 63405, + "cbid": 200, "correlation": 63405 + } + }, + { + "ph": "f", "id": 63405, "pid": 435, "tid": 435, "ts": 1776649791936388, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936388, "dur": 0, + "args": { + "External id": 63406, + "cbid": 273, "correlation": 63406 + } + }, + { + "ph": "f", "id": 63406, "pid": 435, "tid": 435, "ts": 1776649791936388, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791936396, "dur": 19, + "args": { + "External id": 63407, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63407, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63407, "pid": 0, "tid": 7, "ts": 1776649791936396, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936389, "dur": 4, + "args": { + "External id": 63407, + "cbid": 211, "correlation": 63407 + } + }, + { + "ph": "s", "id": 63407, "pid": 435, "tid": 435, "ts": 1776649791936389, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791936432, "dur": 1, + "args": { + "External id": 8175, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63425, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63425, "pid": 0, "tid": 7, "ts": 1776649791936432, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936423, "dur": 6, + "args": { + "External id": 8175, + "cbid": 211, "correlation": 63425 + } + }, + { + "ph": "s", "id": 63425, "pid": 435, "tid": 435, "ts": 1776649791936423, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936435, "dur": 0, + "args": { + "External id": 63431, + "cbid": 200, "correlation": 63431 + } + }, + { + "ph": "f", "id": 63431, "pid": 435, "tid": 435, "ts": 1776649791936435, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936436, "dur": 0, + "args": { + "External id": 63432, + "cbid": 273, "correlation": 63432 + } + }, + { + "ph": "f", "id": 63432, "pid": 435, "tid": 435, "ts": 1776649791936436, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791936444, "dur": 19, + "args": { + "External id": 63433, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63433, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63433, "pid": 0, "tid": 7, "ts": 1776649791936444, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936437, "dur": 4, + "args": { + "External id": 63433, + "cbid": 211, "correlation": 63433 + } + }, + { + "ph": "s", "id": 63433, "pid": 435, "tid": 435, "ts": 1776649791936437, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791936480, "dur": 1, + "args": { + "External id": 8180, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63451, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63451, "pid": 0, "tid": 7, "ts": 1776649791936480, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936471, "dur": 6, + "args": { + "External id": 8180, + "cbid": 211, "correlation": 63451 + } + }, + { + "ph": "s", "id": 63451, "pid": 435, "tid": 435, "ts": 1776649791936471, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936483, "dur": 0, + "args": { + "External id": 63457, + "cbid": 200, "correlation": 63457 + } + }, + { + "ph": "f", "id": 63457, "pid": 435, "tid": 435, "ts": 1776649791936483, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936484, "dur": 0, + "args": { + "External id": 63458, + "cbid": 273, "correlation": 63458 + } + }, + { + "ph": "f", "id": 63458, "pid": 435, "tid": 435, "ts": 1776649791936484, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791936493, "dur": 19, + "args": { + "External id": 63459, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63459, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63459, "pid": 0, "tid": 7, "ts": 1776649791936493, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936485, "dur": 4, + "args": { + "External id": 63459, + "cbid": 211, "correlation": 63459 + } + }, + { + "ph": "s", "id": 63459, "pid": 435, "tid": 435, "ts": 1776649791936485, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791936528, "dur": 1, + "args": { + "External id": 8185, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63477, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63477, "pid": 0, "tid": 7, "ts": 1776649791936528, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936519, "dur": 7, + "args": { + "External id": 8185, + "cbid": 211, "correlation": 63477 + } + }, + { + "ph": "s", "id": 63477, "pid": 435, "tid": 435, "ts": 1776649791936519, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936531, "dur": 0, + "args": { + "External id": 63483, + "cbid": 200, "correlation": 63483 + } + }, + { + "ph": "f", "id": 63483, "pid": 435, "tid": 435, "ts": 1776649791936531, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936532, "dur": 0, + "args": { + "External id": 63484, + "cbid": 273, "correlation": 63484 + } + }, + { + "ph": "f", "id": 63484, "pid": 435, "tid": 435, "ts": 1776649791936532, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791936541, "dur": 19, + "args": { + "External id": 63485, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63485, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63485, "pid": 0, "tid": 7, "ts": 1776649791936541, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936533, "dur": 4, + "args": { + "External id": 63485, + "cbid": 211, "correlation": 63485 + } + }, + { + "ph": "s", "id": 63485, "pid": 435, "tid": 435, "ts": 1776649791936533, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791936575, "dur": 1, + "args": { + "External id": 8190, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63503, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63503, "pid": 0, "tid": 7, "ts": 1776649791936575, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936566, "dur": 6, + "args": { + "External id": 8190, + "cbid": 211, "correlation": 63503 + } + }, + { + "ph": "s", "id": 63503, "pid": 435, "tid": 435, "ts": 1776649791936566, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936579, "dur": 0, + "args": { + "External id": 63509, + "cbid": 200, "correlation": 63509 + } + }, + { + "ph": "f", "id": 63509, "pid": 435, "tid": 435, "ts": 1776649791936579, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936579, "dur": 0, + "args": { + "External id": 63510, + "cbid": 273, "correlation": 63510 + } + }, + { + "ph": "f", "id": 63510, "pid": 435, "tid": 435, "ts": 1776649791936579, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791936588, "dur": 19, + "args": { + "External id": 63511, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63511, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63511, "pid": 0, "tid": 7, "ts": 1776649791936588, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936580, "dur": 4, + "args": { + "External id": 63511, + "cbid": 211, "correlation": 63511 + } + }, + { + "ph": "s", "id": 63511, "pid": 435, "tid": 435, "ts": 1776649791936580, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791936633, "dur": 1, + "args": { + "External id": 8195, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63529, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63529, "pid": 0, "tid": 7, "ts": 1776649791936633, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936625, "dur": 6, + "args": { + "External id": 8195, + "cbid": 211, "correlation": 63529 + } + }, + { + "ph": "s", "id": 63529, "pid": 435, "tid": 435, "ts": 1776649791936625, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936637, "dur": 0, + "args": { + "External id": 63535, + "cbid": 200, "correlation": 63535 + } + }, + { + "ph": "f", "id": 63535, "pid": 435, "tid": 435, "ts": 1776649791936637, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936638, "dur": 0, + "args": { + "External id": 63536, + "cbid": 273, "correlation": 63536 + } + }, + { + "ph": "f", "id": 63536, "pid": 435, "tid": 435, "ts": 1776649791936638, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791936646, "dur": 19, + "args": { + "External id": 63537, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63537, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63537, "pid": 0, "tid": 7, "ts": 1776649791936646, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936639, "dur": 4, + "args": { + "External id": 63537, + "cbid": 211, "correlation": 63537 + } + }, + { + "ph": "s", "id": 63537, "pid": 435, "tid": 435, "ts": 1776649791936639, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791936681, "dur": 1, + "args": { + "External id": 8200, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63555, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63555, "pid": 0, "tid": 7, "ts": 1776649791936681, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936672, "dur": 6, + "args": { + "External id": 8200, + "cbid": 211, "correlation": 63555 + } + }, + { + "ph": "s", "id": 63555, "pid": 435, "tid": 435, "ts": 1776649791936672, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936684, "dur": 0, + "args": { + "External id": 63561, + "cbid": 200, "correlation": 63561 + } + }, + { + "ph": "f", "id": 63561, "pid": 435, "tid": 435, "ts": 1776649791936684, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936685, "dur": 0, + "args": { + "External id": 63562, + "cbid": 273, "correlation": 63562 + } + }, + { + "ph": "f", "id": 63562, "pid": 435, "tid": 435, "ts": 1776649791936685, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791936694, "dur": 19, + "args": { + "External id": 63563, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63563, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63563, "pid": 0, "tid": 7, "ts": 1776649791936694, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936686, "dur": 5, + "args": { + "External id": 63563, + "cbid": 211, "correlation": 63563 + } + }, + { + "ph": "s", "id": 63563, "pid": 435, "tid": 435, "ts": 1776649791936686, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791936729, "dur": 1, + "args": { + "External id": 8205, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63581, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63581, "pid": 0, "tid": 7, "ts": 1776649791936729, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936720, "dur": 6, + "args": { + "External id": 8205, + "cbid": 211, "correlation": 63581 + } + }, + { + "ph": "s", "id": 63581, "pid": 435, "tid": 435, "ts": 1776649791936720, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936732, "dur": 0, + "args": { + "External id": 63587, + "cbid": 200, "correlation": 63587 + } + }, + { + "ph": "f", "id": 63587, "pid": 435, "tid": 435, "ts": 1776649791936732, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936733, "dur": 0, + "args": { + "External id": 63588, + "cbid": 273, "correlation": 63588 + } + }, + { + "ph": "f", "id": 63588, "pid": 435, "tid": 435, "ts": 1776649791936733, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791936741, "dur": 19, + "args": { + "External id": 63589, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63589, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63589, "pid": 0, "tid": 7, "ts": 1776649791936741, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936734, "dur": 4, + "args": { + "External id": 63589, + "cbid": 211, "correlation": 63589 + } + }, + { + "ph": "s", "id": 63589, "pid": 435, "tid": 435, "ts": 1776649791936734, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791936776, "dur": 1, + "args": { + "External id": 8210, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63607, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63607, "pid": 0, "tid": 7, "ts": 1776649791936776, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936768, "dur": 6, + "args": { + "External id": 8210, + "cbid": 211, "correlation": 63607 + } + }, + { + "ph": "s", "id": 63607, "pid": 435, "tid": 435, "ts": 1776649791936768, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936780, "dur": 0, + "args": { + "External id": 63613, + "cbid": 200, "correlation": 63613 + } + }, + { + "ph": "f", "id": 63613, "pid": 435, "tid": 435, "ts": 1776649791936780, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936780, "dur": 0, + "args": { + "External id": 63614, + "cbid": 273, "correlation": 63614 + } + }, + { + "ph": "f", "id": 63614, "pid": 435, "tid": 435, "ts": 1776649791936780, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791936789, "dur": 19, + "args": { + "External id": 63615, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63615, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63615, "pid": 0, "tid": 7, "ts": 1776649791936789, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936781, "dur": 4, + "args": { + "External id": 63615, + "cbid": 211, "correlation": 63615 + } + }, + { + "ph": "s", "id": 63615, "pid": 435, "tid": 435, "ts": 1776649791936781, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791936823, "dur": 1, + "args": { + "External id": 8215, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63633, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63633, "pid": 0, "tid": 7, "ts": 1776649791936823, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936815, "dur": 6, + "args": { + "External id": 8215, + "cbid": 211, "correlation": 63633 + } + }, + { + "ph": "s", "id": 63633, "pid": 435, "tid": 435, "ts": 1776649791936815, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936827, "dur": 0, + "args": { + "External id": 63639, + "cbid": 200, "correlation": 63639 + } + }, + { + "ph": "f", "id": 63639, "pid": 435, "tid": 435, "ts": 1776649791936827, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936827, "dur": 0, + "args": { + "External id": 63640, + "cbid": 273, "correlation": 63640 + } + }, + { + "ph": "f", "id": 63640, "pid": 435, "tid": 435, "ts": 1776649791936827, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791936836, "dur": 19, + "args": { + "External id": 63641, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63641, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63641, "pid": 0, "tid": 7, "ts": 1776649791936836, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936828, "dur": 4, + "args": { + "External id": 63641, + "cbid": 211, "correlation": 63641 + } + }, + { + "ph": "s", "id": 63641, "pid": 435, "tid": 435, "ts": 1776649791936828, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791936871, "dur": 1, + "args": { + "External id": 8220, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63659, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63659, "pid": 0, "tid": 7, "ts": 1776649791936871, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936862, "dur": 6, + "args": { + "External id": 8220, + "cbid": 211, "correlation": 63659 + } + }, + { + "ph": "s", "id": 63659, "pid": 435, "tid": 435, "ts": 1776649791936862, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936875, "dur": 0, + "args": { + "External id": 63665, + "cbid": 200, "correlation": 63665 + } + }, + { + "ph": "f", "id": 63665, "pid": 435, "tid": 435, "ts": 1776649791936875, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936875, "dur": 0, + "args": { + "External id": 63666, + "cbid": 273, "correlation": 63666 + } + }, + { + "ph": "f", "id": 63666, "pid": 435, "tid": 435, "ts": 1776649791936875, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791936884, "dur": 19, + "args": { + "External id": 63667, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63667, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63667, "pid": 0, "tid": 7, "ts": 1776649791936884, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936876, "dur": 4, + "args": { + "External id": 63667, + "cbid": 211, "correlation": 63667 + } + }, + { + "ph": "s", "id": 63667, "pid": 435, "tid": 435, "ts": 1776649791936876, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791936920, "dur": 1, + "args": { + "External id": 8225, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63685, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63685, "pid": 0, "tid": 7, "ts": 1776649791936920, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936909, "dur": 7, + "args": { + "External id": 8225, + "cbid": 211, "correlation": 63685 + } + }, + { + "ph": "s", "id": 63685, "pid": 435, "tid": 435, "ts": 1776649791936909, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936923, "dur": 0, + "args": { + "External id": 63691, + "cbid": 200, "correlation": 63691 + } + }, + { + "ph": "f", "id": 63691, "pid": 435, "tid": 435, "ts": 1776649791936923, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936923, "dur": 0, + "args": { + "External id": 63692, + "cbid": 273, "correlation": 63692 + } + }, + { + "ph": "f", "id": 63692, "pid": 435, "tid": 435, "ts": 1776649791936923, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791936932, "dur": 19, + "args": { + "External id": 63693, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63693, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63693, "pid": 0, "tid": 7, "ts": 1776649791936932, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936925, "dur": 4, + "args": { + "External id": 63693, + "cbid": 211, "correlation": 63693 + } + }, + { + "ph": "s", "id": 63693, "pid": 435, "tid": 435, "ts": 1776649791936925, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791936968, "dur": 1, + "args": { + "External id": 8230, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63711, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63711, "pid": 0, "tid": 7, "ts": 1776649791936968, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936958, "dur": 6, + "args": { + "External id": 8230, + "cbid": 211, "correlation": 63711 + } + }, + { + "ph": "s", "id": 63711, "pid": 435, "tid": 435, "ts": 1776649791936958, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936971, "dur": 0, + "args": { + "External id": 63717, + "cbid": 200, "correlation": 63717 + } + }, + { + "ph": "f", "id": 63717, "pid": 435, "tid": 435, "ts": 1776649791936971, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791936972, "dur": 0, + "args": { + "External id": 63718, + "cbid": 273, "correlation": 63718 + } + }, + { + "ph": "f", "id": 63718, "pid": 435, "tid": 435, "ts": 1776649791936972, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791936981, "dur": 19, + "args": { + "External id": 63719, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63719, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63719, "pid": 0, "tid": 7, "ts": 1776649791936981, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791936973, "dur": 4, + "args": { + "External id": 63719, + "cbid": 211, "correlation": 63719 + } + }, + { + "ph": "s", "id": 63719, "pid": 435, "tid": 435, "ts": 1776649791936973, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791937016, "dur": 1, + "args": { + "External id": 8235, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63737, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63737, "pid": 0, "tid": 7, "ts": 1776649791937016, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937007, "dur": 6, + "args": { + "External id": 8235, + "cbid": 211, "correlation": 63737 + } + }, + { + "ph": "s", "id": 63737, "pid": 435, "tid": 435, "ts": 1776649791937007, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937020, "dur": 0, + "args": { + "External id": 63743, + "cbid": 200, "correlation": 63743 + } + }, + { + "ph": "f", "id": 63743, "pid": 435, "tid": 435, "ts": 1776649791937020, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937020, "dur": 0, + "args": { + "External id": 63744, + "cbid": 273, "correlation": 63744 + } + }, + { + "ph": "f", "id": 63744, "pid": 435, "tid": 435, "ts": 1776649791937020, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791937030, "dur": 19, + "args": { + "External id": 63745, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63745, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63745, "pid": 0, "tid": 7, "ts": 1776649791937030, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937021, "dur": 5, + "args": { + "External id": 63745, + "cbid": 211, "correlation": 63745 + } + }, + { + "ph": "s", "id": 63745, "pid": 435, "tid": 435, "ts": 1776649791937021, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791937065, "dur": 1, + "args": { + "External id": 8240, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63763, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63763, "pid": 0, "tid": 7, "ts": 1776649791937065, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937056, "dur": 7, + "args": { + "External id": 8240, + "cbid": 211, "correlation": 63763 + } + }, + { + "ph": "s", "id": 63763, "pid": 435, "tid": 435, "ts": 1776649791937056, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937069, "dur": 0, + "args": { + "External id": 63769, + "cbid": 200, "correlation": 63769 + } + }, + { + "ph": "f", "id": 63769, "pid": 435, "tid": 435, "ts": 1776649791937069, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937069, "dur": 0, + "args": { + "External id": 63770, + "cbid": 273, "correlation": 63770 + } + }, + { + "ph": "f", "id": 63770, "pid": 435, "tid": 435, "ts": 1776649791937069, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791937078, "dur": 19, + "args": { + "External id": 63771, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63771, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63771, "pid": 0, "tid": 7, "ts": 1776649791937078, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937070, "dur": 5, + "args": { + "External id": 63771, + "cbid": 211, "correlation": 63771 + } + }, + { + "ph": "s", "id": 63771, "pid": 435, "tid": 435, "ts": 1776649791937070, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791937120, "dur": 1, + "args": { + "External id": 8245, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63789, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63789, "pid": 0, "tid": 7, "ts": 1776649791937120, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937111, "dur": 7, + "args": { + "External id": 8245, + "cbid": 211, "correlation": 63789 + } + }, + { + "ph": "s", "id": 63789, "pid": 435, "tid": 435, "ts": 1776649791937111, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937124, "dur": 0, + "args": { + "External id": 63795, + "cbid": 200, "correlation": 63795 + } + }, + { + "ph": "f", "id": 63795, "pid": 435, "tid": 435, "ts": 1776649791937124, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937125, "dur": 0, + "args": { + "External id": 63796, + "cbid": 273, "correlation": 63796 + } + }, + { + "ph": "f", "id": 63796, "pid": 435, "tid": 435, "ts": 1776649791937125, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791937134, "dur": 19, + "args": { + "External id": 63797, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63797, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63797, "pid": 0, "tid": 7, "ts": 1776649791937134, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937126, "dur": 5, + "args": { + "External id": 63797, + "cbid": 211, "correlation": 63797 + } + }, + { + "ph": "s", "id": 63797, "pid": 435, "tid": 435, "ts": 1776649791937126, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791937170, "dur": 1, + "args": { + "External id": 8250, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63815, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63815, "pid": 0, "tid": 7, "ts": 1776649791937170, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937161, "dur": 6, + "args": { + "External id": 8250, + "cbid": 211, "correlation": 63815 + } + }, + { + "ph": "s", "id": 63815, "pid": 435, "tid": 435, "ts": 1776649791937161, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937173, "dur": 0, + "args": { + "External id": 63821, + "cbid": 200, "correlation": 63821 + } + }, + { + "ph": "f", "id": 63821, "pid": 435, "tid": 435, "ts": 1776649791937173, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937174, "dur": 0, + "args": { + "External id": 63822, + "cbid": 273, "correlation": 63822 + } + }, + { + "ph": "f", "id": 63822, "pid": 435, "tid": 435, "ts": 1776649791937174, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791937183, "dur": 19, + "args": { + "External id": 63823, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63823, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63823, "pid": 0, "tid": 7, "ts": 1776649791937183, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937175, "dur": 5, + "args": { + "External id": 63823, + "cbid": 211, "correlation": 63823 + } + }, + { + "ph": "s", "id": 63823, "pid": 435, "tid": 435, "ts": 1776649791937175, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791937218, "dur": 1, + "args": { + "External id": 8255, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63841, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63841, "pid": 0, "tid": 7, "ts": 1776649791937218, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937209, "dur": 6, + "args": { + "External id": 8255, + "cbid": 211, "correlation": 63841 + } + }, + { + "ph": "s", "id": 63841, "pid": 435, "tid": 435, "ts": 1776649791937209, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937221, "dur": 0, + "args": { + "External id": 63847, + "cbid": 200, "correlation": 63847 + } + }, + { + "ph": "f", "id": 63847, "pid": 435, "tid": 435, "ts": 1776649791937221, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937222, "dur": 0, + "args": { + "External id": 63848, + "cbid": 273, "correlation": 63848 + } + }, + { + "ph": "f", "id": 63848, "pid": 435, "tid": 435, "ts": 1776649791937222, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791937231, "dur": 19, + "args": { + "External id": 63849, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63849, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63849, "pid": 0, "tid": 7, "ts": 1776649791937231, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937223, "dur": 4, + "args": { + "External id": 63849, + "cbid": 211, "correlation": 63849 + } + }, + { + "ph": "s", "id": 63849, "pid": 435, "tid": 435, "ts": 1776649791937223, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791937266, "dur": 1, + "args": { + "External id": 8260, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63867, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63867, "pid": 0, "tid": 7, "ts": 1776649791937266, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937257, "dur": 6, + "args": { + "External id": 8260, + "cbid": 211, "correlation": 63867 + } + }, + { + "ph": "s", "id": 63867, "pid": 435, "tid": 435, "ts": 1776649791937257, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937269, "dur": 0, + "args": { + "External id": 63873, + "cbid": 200, "correlation": 63873 + } + }, + { + "ph": "f", "id": 63873, "pid": 435, "tid": 435, "ts": 1776649791937269, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937270, "dur": 0, + "args": { + "External id": 63874, + "cbid": 273, "correlation": 63874 + } + }, + { + "ph": "f", "id": 63874, "pid": 435, "tid": 435, "ts": 1776649791937270, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791937279, "dur": 19, + "args": { + "External id": 63875, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63875, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63875, "pid": 0, "tid": 7, "ts": 1776649791937279, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937271, "dur": 4, + "args": { + "External id": 63875, + "cbid": 211, "correlation": 63875 + } + }, + { + "ph": "s", "id": 63875, "pid": 435, "tid": 435, "ts": 1776649791937271, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791937315, "dur": 1, + "args": { + "External id": 8265, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63893, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63893, "pid": 0, "tid": 7, "ts": 1776649791937315, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937305, "dur": 7, + "args": { + "External id": 8265, + "cbid": 211, "correlation": 63893 + } + }, + { + "ph": "s", "id": 63893, "pid": 435, "tid": 435, "ts": 1776649791937305, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937318, "dur": 0, + "args": { + "External id": 63899, + "cbid": 200, "correlation": 63899 + } + }, + { + "ph": "f", "id": 63899, "pid": 435, "tid": 435, "ts": 1776649791937318, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937318, "dur": 0, + "args": { + "External id": 63900, + "cbid": 273, "correlation": 63900 + } + }, + { + "ph": "f", "id": 63900, "pid": 435, "tid": 435, "ts": 1776649791937318, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791937327, "dur": 19, + "args": { + "External id": 63901, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63901, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63901, "pid": 0, "tid": 7, "ts": 1776649791937327, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937319, "dur": 5, + "args": { + "External id": 63901, + "cbid": 211, "correlation": 63901 + } + }, + { + "ph": "s", "id": 63901, "pid": 435, "tid": 435, "ts": 1776649791937319, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791937363, "dur": 1, + "args": { + "External id": 8270, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63919, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63919, "pid": 0, "tid": 7, "ts": 1776649791937363, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937354, "dur": 6, + "args": { + "External id": 8270, + "cbid": 211, "correlation": 63919 + } + }, + { + "ph": "s", "id": 63919, "pid": 435, "tid": 435, "ts": 1776649791937354, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937366, "dur": 0, + "args": { + "External id": 63925, + "cbid": 200, "correlation": 63925 + } + }, + { + "ph": "f", "id": 63925, "pid": 435, "tid": 435, "ts": 1776649791937366, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937367, "dur": 0, + "args": { + "External id": 63926, + "cbid": 273, "correlation": 63926 + } + }, + { + "ph": "f", "id": 63926, "pid": 435, "tid": 435, "ts": 1776649791937367, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791937375, "dur": 19, + "args": { + "External id": 63927, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63927, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63927, "pid": 0, "tid": 7, "ts": 1776649791937375, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937368, "dur": 4, + "args": { + "External id": 63927, + "cbid": 211, "correlation": 63927 + } + }, + { + "ph": "s", "id": 63927, "pid": 435, "tid": 435, "ts": 1776649791937368, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791937411, "dur": 1, + "args": { + "External id": 8275, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63945, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63945, "pid": 0, "tid": 7, "ts": 1776649791937411, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937402, "dur": 6, + "args": { + "External id": 8275, + "cbid": 211, "correlation": 63945 + } + }, + { + "ph": "s", "id": 63945, "pid": 435, "tid": 435, "ts": 1776649791937402, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937414, "dur": 0, + "args": { + "External id": 63951, + "cbid": 200, "correlation": 63951 + } + }, + { + "ph": "f", "id": 63951, "pid": 435, "tid": 435, "ts": 1776649791937414, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937415, "dur": 0, + "args": { + "External id": 63952, + "cbid": 273, "correlation": 63952 + } + }, + { + "ph": "f", "id": 63952, "pid": 435, "tid": 435, "ts": 1776649791937415, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791937424, "dur": 19, + "args": { + "External id": 63953, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63953, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63953, "pid": 0, "tid": 7, "ts": 1776649791937424, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937416, "dur": 4, + "args": { + "External id": 63953, + "cbid": 211, "correlation": 63953 + } + }, + { + "ph": "s", "id": 63953, "pid": 435, "tid": 435, "ts": 1776649791937416, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791937459, "dur": 1, + "args": { + "External id": 8280, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63971, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63971, "pid": 0, "tid": 7, "ts": 1776649791937459, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937450, "dur": 6, + "args": { + "External id": 8280, + "cbid": 211, "correlation": 63971 + } + }, + { + "ph": "s", "id": 63971, "pid": 435, "tid": 435, "ts": 1776649791937450, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937463, "dur": 0, + "args": { + "External id": 63977, + "cbid": 200, "correlation": 63977 + } + }, + { + "ph": "f", "id": 63977, "pid": 435, "tid": 435, "ts": 1776649791937463, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937463, "dur": 0, + "args": { + "External id": 63978, + "cbid": 273, "correlation": 63978 + } + }, + { + "ph": "f", "id": 63978, "pid": 435, "tid": 435, "ts": 1776649791937463, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791937473, "dur": 19, + "args": { + "External id": 63979, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63979, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63979, "pid": 0, "tid": 7, "ts": 1776649791937473, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937464, "dur": 5, + "args": { + "External id": 63979, + "cbid": 211, "correlation": 63979 + } + }, + { + "ph": "s", "id": 63979, "pid": 435, "tid": 435, "ts": 1776649791937464, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791937508, "dur": 1, + "args": { + "External id": 8285, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 63997, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 63997, "pid": 0, "tid": 7, "ts": 1776649791937508, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937498, "dur": 6, + "args": { + "External id": 8285, + "cbid": 211, "correlation": 63997 + } + }, + { + "ph": "s", "id": 63997, "pid": 435, "tid": 435, "ts": 1776649791937498, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937511, "dur": 0, + "args": { + "External id": 64003, + "cbid": 200, "correlation": 64003 + } + }, + { + "ph": "f", "id": 64003, "pid": 435, "tid": 435, "ts": 1776649791937511, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937511, "dur": 0, + "args": { + "External id": 64004, + "cbid": 273, "correlation": 64004 + } + }, + { + "ph": "f", "id": 64004, "pid": 435, "tid": 435, "ts": 1776649791937511, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791937520, "dur": 19, + "args": { + "External id": 64005, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64005, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64005, "pid": 0, "tid": 7, "ts": 1776649791937520, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937512, "dur": 4, + "args": { + "External id": 64005, + "cbid": 211, "correlation": 64005 + } + }, + { + "ph": "s", "id": 64005, "pid": 435, "tid": 435, "ts": 1776649791937512, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791937557, "dur": 1, + "args": { + "External id": 8290, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64023, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64023, "pid": 0, "tid": 7, "ts": 1776649791937557, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937547, "dur": 6, + "args": { + "External id": 8290, + "cbid": 211, "correlation": 64023 + } + }, + { + "ph": "s", "id": 64023, "pid": 435, "tid": 435, "ts": 1776649791937547, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937560, "dur": 0, + "args": { + "External id": 64029, + "cbid": 200, "correlation": 64029 + } + }, + { + "ph": "f", "id": 64029, "pid": 435, "tid": 435, "ts": 1776649791937560, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937561, "dur": 0, + "args": { + "External id": 64030, + "cbid": 273, "correlation": 64030 + } + }, + { + "ph": "f", "id": 64030, "pid": 435, "tid": 435, "ts": 1776649791937561, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791937569, "dur": 19, + "args": { + "External id": 64031, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64031, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64031, "pid": 0, "tid": 7, "ts": 1776649791937569, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937562, "dur": 4, + "args": { + "External id": 64031, + "cbid": 211, "correlation": 64031 + } + }, + { + "ph": "s", "id": 64031, "pid": 435, "tid": 435, "ts": 1776649791937562, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791937605, "dur": 1, + "args": { + "External id": 8295, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64049, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64049, "pid": 0, "tid": 7, "ts": 1776649791937605, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937595, "dur": 6, + "args": { + "External id": 8295, + "cbid": 211, "correlation": 64049 + } + }, + { + "ph": "s", "id": 64049, "pid": 435, "tid": 435, "ts": 1776649791937595, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937608, "dur": 0, + "args": { + "External id": 64055, + "cbid": 200, "correlation": 64055 + } + }, + { + "ph": "f", "id": 64055, "pid": 435, "tid": 435, "ts": 1776649791937608, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937608, "dur": 0, + "args": { + "External id": 64056, + "cbid": 273, "correlation": 64056 + } + }, + { + "ph": "f", "id": 64056, "pid": 435, "tid": 435, "ts": 1776649791937608, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791937617, "dur": 19, + "args": { + "External id": 64057, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64057, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64057, "pid": 0, "tid": 7, "ts": 1776649791937617, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937609, "dur": 4, + "args": { + "External id": 64057, + "cbid": 211, "correlation": 64057 + } + }, + { + "ph": "s", "id": 64057, "pid": 435, "tid": 435, "ts": 1776649791937609, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791937653, "dur": 1, + "args": { + "External id": 8300, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64075, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64075, "pid": 0, "tid": 7, "ts": 1776649791937653, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937644, "dur": 6, + "args": { + "External id": 8300, + "cbid": 211, "correlation": 64075 + } + }, + { + "ph": "s", "id": 64075, "pid": 435, "tid": 435, "ts": 1776649791937644, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937656, "dur": 0, + "args": { + "External id": 64081, + "cbid": 200, "correlation": 64081 + } + }, + { + "ph": "f", "id": 64081, "pid": 435, "tid": 435, "ts": 1776649791937656, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937657, "dur": 0, + "args": { + "External id": 64082, + "cbid": 273, "correlation": 64082 + } + }, + { + "ph": "f", "id": 64082, "pid": 435, "tid": 435, "ts": 1776649791937657, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791937666, "dur": 19, + "args": { + "External id": 64083, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64083, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64083, "pid": 0, "tid": 7, "ts": 1776649791937666, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937658, "dur": 5, + "args": { + "External id": 64083, + "cbid": 211, "correlation": 64083 + } + }, + { + "ph": "s", "id": 64083, "pid": 435, "tid": 435, "ts": 1776649791937658, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791937702, "dur": 1, + "args": { + "External id": 8305, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64101, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64101, "pid": 0, "tid": 7, "ts": 1776649791937702, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937692, "dur": 7, + "args": { + "External id": 8305, + "cbid": 211, "correlation": 64101 + } + }, + { + "ph": "s", "id": 64101, "pid": 435, "tid": 435, "ts": 1776649791937692, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937705, "dur": 0, + "args": { + "External id": 64107, + "cbid": 200, "correlation": 64107 + } + }, + { + "ph": "f", "id": 64107, "pid": 435, "tid": 435, "ts": 1776649791937705, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937706, "dur": 0, + "args": { + "External id": 64108, + "cbid": 273, "correlation": 64108 + } + }, + { + "ph": "f", "id": 64108, "pid": 435, "tid": 435, "ts": 1776649791937706, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791937715, "dur": 19, + "args": { + "External id": 64109, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64109, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64109, "pid": 0, "tid": 7, "ts": 1776649791937715, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937707, "dur": 4, + "args": { + "External id": 64109, + "cbid": 211, "correlation": 64109 + } + }, + { + "ph": "s", "id": 64109, "pid": 435, "tid": 435, "ts": 1776649791937707, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791937738, "dur": 3, + "args": { + "External id": 64117, + "cbid": 135, "correlation": 64117 + } + }, + { + "ph": "f", "id": 64117, "pid": 435, "tid": 435, "ts": 1776649791937738, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791937754, "dur": 5, + "args": { + "External id": 64122, + "cbid": 165, "correlation": 64122 + } + }, + { + "ph": "s", "id": 64122, "pid": 435, "tid": 435, "ts": 1776649791937754, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649791937764, "dur": 7, + "args": { + "External id": 64125, + "cbid": 139, "correlation": 64125 + } + }, + { + "ph": "f", "id": 64125, "pid": 435, "tid": 435, "ts": 1776649791937764, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791937825, "dur": 1, + "args": { + "External id": 8310, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64153, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64153, "pid": 0, "tid": 7, "ts": 1776649791937825, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937814, "dur": 8, + "args": { + "External id": 8310, + "cbid": 211, "correlation": 64153 + } + }, + { + "ph": "s", "id": 64153, "pid": 435, "tid": 435, "ts": 1776649791937814, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937829, "dur": 0, + "args": { + "External id": 64159, + "cbid": 200, "correlation": 64159 + } + }, + { + "ph": "f", "id": 64159, "pid": 435, "tid": 435, "ts": 1776649791937829, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937830, "dur": 1, + "args": { + "External id": 64160, + "cbid": 273, "correlation": 64160 + } + }, + { + "ph": "f", "id": 64160, "pid": 435, "tid": 435, "ts": 1776649791937830, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791937839, "dur": 19, + "args": { + "External id": 64161, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64161, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64161, "pid": 0, "tid": 7, "ts": 1776649791937839, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937831, "dur": 5, + "args": { + "External id": 64161, + "cbid": 211, "correlation": 64161 + } + }, + { + "ph": "s", "id": 64161, "pid": 435, "tid": 435, "ts": 1776649791937831, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791937878, "dur": 1, + "args": { + "External id": 8315, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64179, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64179, "pid": 0, "tid": 7, "ts": 1776649791937878, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937869, "dur": 6, + "args": { + "External id": 8315, + "cbid": 211, "correlation": 64179 + } + }, + { + "ph": "s", "id": 64179, "pid": 435, "tid": 435, "ts": 1776649791937869, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937882, "dur": 0, + "args": { + "External id": 64185, + "cbid": 200, "correlation": 64185 + } + }, + { + "ph": "f", "id": 64185, "pid": 435, "tid": 435, "ts": 1776649791937882, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937882, "dur": 0, + "args": { + "External id": 64186, + "cbid": 273, "correlation": 64186 + } + }, + { + "ph": "f", "id": 64186, "pid": 435, "tid": 435, "ts": 1776649791937882, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791937891, "dur": 19, + "args": { + "External id": 64187, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64187, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64187, "pid": 0, "tid": 7, "ts": 1776649791937891, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937883, "dur": 4, + "args": { + "External id": 64187, + "cbid": 211, "correlation": 64187 + } + }, + { + "ph": "s", "id": 64187, "pid": 435, "tid": 435, "ts": 1776649791937883, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791937928, "dur": 1, + "args": { + "External id": 8320, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64205, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64205, "pid": 0, "tid": 7, "ts": 1776649791937928, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937919, "dur": 6, + "args": { + "External id": 8320, + "cbid": 211, "correlation": 64205 + } + }, + { + "ph": "s", "id": 64205, "pid": 435, "tid": 435, "ts": 1776649791937919, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937932, "dur": 0, + "args": { + "External id": 64211, + "cbid": 200, "correlation": 64211 + } + }, + { + "ph": "f", "id": 64211, "pid": 435, "tid": 435, "ts": 1776649791937932, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937933, "dur": 0, + "args": { + "External id": 64212, + "cbid": 273, "correlation": 64212 + } + }, + { + "ph": "f", "id": 64212, "pid": 435, "tid": 435, "ts": 1776649791937933, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791937942, "dur": 19, + "args": { + "External id": 64213, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64213, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64213, "pid": 0, "tid": 7, "ts": 1776649791937942, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937934, "dur": 5, + "args": { + "External id": 64213, + "cbid": 211, "correlation": 64213 + } + }, + { + "ph": "s", "id": 64213, "pid": 435, "tid": 435, "ts": 1776649791937934, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791937978, "dur": 1, + "args": { + "External id": 8325, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64231, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64231, "pid": 0, "tid": 7, "ts": 1776649791937978, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937969, "dur": 6, + "args": { + "External id": 8325, + "cbid": 211, "correlation": 64231 + } + }, + { + "ph": "s", "id": 64231, "pid": 435, "tid": 435, "ts": 1776649791937969, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937981, "dur": 0, + "args": { + "External id": 64237, + "cbid": 200, "correlation": 64237 + } + }, + { + "ph": "f", "id": 64237, "pid": 435, "tid": 435, "ts": 1776649791937981, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791937982, "dur": 0, + "args": { + "External id": 64238, + "cbid": 273, "correlation": 64238 + } + }, + { + "ph": "f", "id": 64238, "pid": 435, "tid": 435, "ts": 1776649791937982, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791937991, "dur": 19, + "args": { + "External id": 64239, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64239, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64239, "pid": 0, "tid": 7, "ts": 1776649791937991, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791937983, "dur": 4, + "args": { + "External id": 64239, + "cbid": 211, "correlation": 64239 + } + }, + { + "ph": "s", "id": 64239, "pid": 435, "tid": 435, "ts": 1776649791937983, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791938027, "dur": 1, + "args": { + "External id": 8330, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64257, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64257, "pid": 0, "tid": 7, "ts": 1776649791938027, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938018, "dur": 6, + "args": { + "External id": 8330, + "cbid": 211, "correlation": 64257 + } + }, + { + "ph": "s", "id": 64257, "pid": 435, "tid": 435, "ts": 1776649791938018, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938030, "dur": 0, + "args": { + "External id": 64263, + "cbid": 200, "correlation": 64263 + } + }, + { + "ph": "f", "id": 64263, "pid": 435, "tid": 435, "ts": 1776649791938030, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938031, "dur": 0, + "args": { + "External id": 64264, + "cbid": 273, "correlation": 64264 + } + }, + { + "ph": "f", "id": 64264, "pid": 435, "tid": 435, "ts": 1776649791938031, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791938040, "dur": 19, + "args": { + "External id": 64265, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64265, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64265, "pid": 0, "tid": 7, "ts": 1776649791938040, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938032, "dur": 5, + "args": { + "External id": 64265, + "cbid": 211, "correlation": 64265 + } + }, + { + "ph": "s", "id": 64265, "pid": 435, "tid": 435, "ts": 1776649791938032, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791938076, "dur": 1, + "args": { + "External id": 8335, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64283, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64283, "pid": 0, "tid": 7, "ts": 1776649791938076, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938067, "dur": 6, + "args": { + "External id": 8335, + "cbid": 211, "correlation": 64283 + } + }, + { + "ph": "s", "id": 64283, "pid": 435, "tid": 435, "ts": 1776649791938067, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938080, "dur": 0, + "args": { + "External id": 64289, + "cbid": 200, "correlation": 64289 + } + }, + { + "ph": "f", "id": 64289, "pid": 435, "tid": 435, "ts": 1776649791938080, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938080, "dur": 0, + "args": { + "External id": 64290, + "cbid": 273, "correlation": 64290 + } + }, + { + "ph": "f", "id": 64290, "pid": 435, "tid": 435, "ts": 1776649791938080, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791938089, "dur": 19, + "args": { + "External id": 64291, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64291, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64291, "pid": 0, "tid": 7, "ts": 1776649791938089, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938081, "dur": 4, + "args": { + "External id": 64291, + "cbid": 211, "correlation": 64291 + } + }, + { + "ph": "s", "id": 64291, "pid": 435, "tid": 435, "ts": 1776649791938081, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791938126, "dur": 1, + "args": { + "External id": 8340, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64309, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64309, "pid": 0, "tid": 7, "ts": 1776649791938126, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938116, "dur": 6, + "args": { + "External id": 8340, + "cbid": 211, "correlation": 64309 + } + }, + { + "ph": "s", "id": 64309, "pid": 435, "tid": 435, "ts": 1776649791938116, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938128, "dur": 0, + "args": { + "External id": 64315, + "cbid": 200, "correlation": 64315 + } + }, + { + "ph": "f", "id": 64315, "pid": 435, "tid": 435, "ts": 1776649791938128, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938129, "dur": 0, + "args": { + "External id": 64316, + "cbid": 273, "correlation": 64316 + } + }, + { + "ph": "f", "id": 64316, "pid": 435, "tid": 435, "ts": 1776649791938129, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791938138, "dur": 19, + "args": { + "External id": 64317, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64317, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64317, "pid": 0, "tid": 7, "ts": 1776649791938138, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938130, "dur": 4, + "args": { + "External id": 64317, + "cbid": 211, "correlation": 64317 + } + }, + { + "ph": "s", "id": 64317, "pid": 435, "tid": 435, "ts": 1776649791938130, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791938173, "dur": 1, + "args": { + "External id": 8345, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64335, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64335, "pid": 0, "tid": 7, "ts": 1776649791938173, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938164, "dur": 6, + "args": { + "External id": 8345, + "cbid": 211, "correlation": 64335 + } + }, + { + "ph": "s", "id": 64335, "pid": 435, "tid": 435, "ts": 1776649791938164, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938176, "dur": 0, + "args": { + "External id": 64341, + "cbid": 200, "correlation": 64341 + } + }, + { + "ph": "f", "id": 64341, "pid": 435, "tid": 435, "ts": 1776649791938176, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938177, "dur": 0, + "args": { + "External id": 64342, + "cbid": 273, "correlation": 64342 + } + }, + { + "ph": "f", "id": 64342, "pid": 435, "tid": 435, "ts": 1776649791938177, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791938186, "dur": 19, + "args": { + "External id": 64343, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64343, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64343, "pid": 0, "tid": 7, "ts": 1776649791938186, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938178, "dur": 4, + "args": { + "External id": 64343, + "cbid": 211, "correlation": 64343 + } + }, + { + "ph": "s", "id": 64343, "pid": 435, "tid": 435, "ts": 1776649791938178, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791938221, "dur": 1, + "args": { + "External id": 8350, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64361, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64361, "pid": 0, "tid": 7, "ts": 1776649791938221, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938212, "dur": 6, + "args": { + "External id": 8350, + "cbid": 211, "correlation": 64361 + } + }, + { + "ph": "s", "id": 64361, "pid": 435, "tid": 435, "ts": 1776649791938212, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938224, "dur": 0, + "args": { + "External id": 64367, + "cbid": 200, "correlation": 64367 + } + }, + { + "ph": "f", "id": 64367, "pid": 435, "tid": 435, "ts": 1776649791938224, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938225, "dur": 0, + "args": { + "External id": 64368, + "cbid": 273, "correlation": 64368 + } + }, + { + "ph": "f", "id": 64368, "pid": 435, "tid": 435, "ts": 1776649791938225, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791938234, "dur": 19, + "args": { + "External id": 64369, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64369, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64369, "pid": 0, "tid": 7, "ts": 1776649791938234, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938226, "dur": 4, + "args": { + "External id": 64369, + "cbid": 211, "correlation": 64369 + } + }, + { + "ph": "s", "id": 64369, "pid": 435, "tid": 435, "ts": 1776649791938226, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791938269, "dur": 1, + "args": { + "External id": 8355, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64387, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64387, "pid": 0, "tid": 7, "ts": 1776649791938269, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938260, "dur": 6, + "args": { + "External id": 8355, + "cbid": 211, "correlation": 64387 + } + }, + { + "ph": "s", "id": 64387, "pid": 435, "tid": 435, "ts": 1776649791938260, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938272, "dur": 0, + "args": { + "External id": 64393, + "cbid": 200, "correlation": 64393 + } + }, + { + "ph": "f", "id": 64393, "pid": 435, "tid": 435, "ts": 1776649791938272, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938273, "dur": 0, + "args": { + "External id": 64394, + "cbid": 273, "correlation": 64394 + } + }, + { + "ph": "f", "id": 64394, "pid": 435, "tid": 435, "ts": 1776649791938273, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791938282, "dur": 19, + "args": { + "External id": 64395, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64395, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64395, "pid": 0, "tid": 7, "ts": 1776649791938282, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938274, "dur": 4, + "args": { + "External id": 64395, + "cbid": 211, "correlation": 64395 + } + }, + { + "ph": "s", "id": 64395, "pid": 435, "tid": 435, "ts": 1776649791938274, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791938296, "dur": 10, + "args": { + "External id": 64398, + "cbid": 165, "correlation": 64398 + } + }, + { + "ph": "s", "id": 64398, "pid": 435, "tid": 435, "ts": 1776649791938296, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791938328, "dur": 3, + "args": { + "External id": 64407, + "cbid": 135, "correlation": 64407 + } + }, + { + "ph": "f", "id": 64407, "pid": 435, "tid": 435, "ts": 1776649791938328, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791938377, "dur": 1, + "args": { + "External id": 8360, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64427, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64427, "pid": 0, "tid": 7, "ts": 1776649791938377, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938366, "dur": 8, + "args": { + "External id": 8360, + "cbid": 211, "correlation": 64427 + } + }, + { + "ph": "s", "id": 64427, "pid": 435, "tid": 435, "ts": 1776649791938366, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938381, "dur": 0, + "args": { + "External id": 64433, + "cbid": 200, "correlation": 64433 + } + }, + { + "ph": "f", "id": 64433, "pid": 435, "tid": 435, "ts": 1776649791938381, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938382, "dur": 1, + "args": { + "External id": 64434, + "cbid": 273, "correlation": 64434 + } + }, + { + "ph": "f", "id": 64434, "pid": 435, "tid": 435, "ts": 1776649791938382, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791938392, "dur": 19, + "args": { + "External id": 64435, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64435, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64435, "pid": 0, "tid": 7, "ts": 1776649791938392, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938383, "dur": 5, + "args": { + "External id": 64435, + "cbid": 211, "correlation": 64435 + } + }, + { + "ph": "s", "id": 64435, "pid": 435, "tid": 435, "ts": 1776649791938383, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791938429, "dur": 1, + "args": { + "External id": 8365, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64453, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64453, "pid": 0, "tid": 7, "ts": 1776649791938429, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938420, "dur": 6, + "args": { + "External id": 8365, + "cbid": 211, "correlation": 64453 + } + }, + { + "ph": "s", "id": 64453, "pid": 435, "tid": 435, "ts": 1776649791938420, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938432, "dur": 0, + "args": { + "External id": 64459, + "cbid": 200, "correlation": 64459 + } + }, + { + "ph": "f", "id": 64459, "pid": 435, "tid": 435, "ts": 1776649791938432, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938433, "dur": 0, + "args": { + "External id": 64460, + "cbid": 273, "correlation": 64460 + } + }, + { + "ph": "f", "id": 64460, "pid": 435, "tid": 435, "ts": 1776649791938433, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791938442, "dur": 19, + "args": { + "External id": 64461, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64461, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64461, "pid": 0, "tid": 7, "ts": 1776649791938442, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938434, "dur": 4, + "args": { + "External id": 64461, + "cbid": 211, "correlation": 64461 + } + }, + { + "ph": "s", "id": 64461, "pid": 435, "tid": 435, "ts": 1776649791938434, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791938479, "dur": 1, + "args": { + "External id": 8370, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64479, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64479, "pid": 0, "tid": 7, "ts": 1776649791938479, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938470, "dur": 6, + "args": { + "External id": 8370, + "cbid": 211, "correlation": 64479 + } + }, + { + "ph": "s", "id": 64479, "pid": 435, "tid": 435, "ts": 1776649791938470, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938482, "dur": 0, + "args": { + "External id": 64485, + "cbid": 200, "correlation": 64485 + } + }, + { + "ph": "f", "id": 64485, "pid": 435, "tid": 435, "ts": 1776649791938482, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938483, "dur": 0, + "args": { + "External id": 64486, + "cbid": 273, "correlation": 64486 + } + }, + { + "ph": "f", "id": 64486, "pid": 435, "tid": 435, "ts": 1776649791938483, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791938492, "dur": 19, + "args": { + "External id": 64487, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64487, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64487, "pid": 0, "tid": 7, "ts": 1776649791938492, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938484, "dur": 4, + "args": { + "External id": 64487, + "cbid": 211, "correlation": 64487 + } + }, + { + "ph": "s", "id": 64487, "pid": 435, "tid": 435, "ts": 1776649791938484, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791938527, "dur": 1, + "args": { + "External id": 8375, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64505, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64505, "pid": 0, "tid": 7, "ts": 1776649791938527, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938518, "dur": 6, + "args": { + "External id": 8375, + "cbid": 211, "correlation": 64505 + } + }, + { + "ph": "s", "id": 64505, "pid": 435, "tid": 435, "ts": 1776649791938518, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938530, "dur": 0, + "args": { + "External id": 64511, + "cbid": 200, "correlation": 64511 + } + }, + { + "ph": "f", "id": 64511, "pid": 435, "tid": 435, "ts": 1776649791938530, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938531, "dur": 0, + "args": { + "External id": 64512, + "cbid": 273, "correlation": 64512 + } + }, + { + "ph": "f", "id": 64512, "pid": 435, "tid": 435, "ts": 1776649791938531, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791938540, "dur": 19, + "args": { + "External id": 64513, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64513, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64513, "pid": 0, "tid": 7, "ts": 1776649791938540, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938532, "dur": 4, + "args": { + "External id": 64513, + "cbid": 211, "correlation": 64513 + } + }, + { + "ph": "s", "id": 64513, "pid": 435, "tid": 435, "ts": 1776649791938532, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791938576, "dur": 1, + "args": { + "External id": 8380, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64531, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64531, "pid": 0, "tid": 7, "ts": 1776649791938576, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938567, "dur": 6, + "args": { + "External id": 8380, + "cbid": 211, "correlation": 64531 + } + }, + { + "ph": "s", "id": 64531, "pid": 435, "tid": 435, "ts": 1776649791938567, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938579, "dur": 0, + "args": { + "External id": 64537, + "cbid": 200, "correlation": 64537 + } + }, + { + "ph": "f", "id": 64537, "pid": 435, "tid": 435, "ts": 1776649791938579, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938580, "dur": 0, + "args": { + "External id": 64538, + "cbid": 273, "correlation": 64538 + } + }, + { + "ph": "f", "id": 64538, "pid": 435, "tid": 435, "ts": 1776649791938580, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791938589, "dur": 19, + "args": { + "External id": 64539, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64539, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64539, "pid": 0, "tid": 7, "ts": 1776649791938589, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938581, "dur": 4, + "args": { + "External id": 64539, + "cbid": 211, "correlation": 64539 + } + }, + { + "ph": "s", "id": 64539, "pid": 435, "tid": 435, "ts": 1776649791938581, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791938626, "dur": 1, + "args": { + "External id": 8385, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64557, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64557, "pid": 0, "tid": 7, "ts": 1776649791938626, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938616, "dur": 7, + "args": { + "External id": 8385, + "cbid": 211, "correlation": 64557 + } + }, + { + "ph": "s", "id": 64557, "pid": 435, "tid": 435, "ts": 1776649791938616, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938629, "dur": 0, + "args": { + "External id": 64563, + "cbid": 200, "correlation": 64563 + } + }, + { + "ph": "f", "id": 64563, "pid": 435, "tid": 435, "ts": 1776649791938629, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938630, "dur": 0, + "args": { + "External id": 64564, + "cbid": 273, "correlation": 64564 + } + }, + { + "ph": "f", "id": 64564, "pid": 435, "tid": 435, "ts": 1776649791938630, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791938638, "dur": 19, + "args": { + "External id": 64565, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64565, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64565, "pid": 0, "tid": 7, "ts": 1776649791938638, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938631, "dur": 4, + "args": { + "External id": 64565, + "cbid": 211, "correlation": 64565 + } + }, + { + "ph": "s", "id": 64565, "pid": 435, "tid": 435, "ts": 1776649791938631, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791938674, "dur": 1, + "args": { + "External id": 8390, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64583, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64583, "pid": 0, "tid": 7, "ts": 1776649791938674, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938665, "dur": 6, + "args": { + "External id": 8390, + "cbid": 211, "correlation": 64583 + } + }, + { + "ph": "s", "id": 64583, "pid": 435, "tid": 435, "ts": 1776649791938665, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938677, "dur": 0, + "args": { + "External id": 64589, + "cbid": 200, "correlation": 64589 + } + }, + { + "ph": "f", "id": 64589, "pid": 435, "tid": 435, "ts": 1776649791938677, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938678, "dur": 0, + "args": { + "External id": 64590, + "cbid": 273, "correlation": 64590 + } + }, + { + "ph": "f", "id": 64590, "pid": 435, "tid": 435, "ts": 1776649791938678, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791938687, "dur": 19, + "args": { + "External id": 64591, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64591, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64591, "pid": 0, "tid": 7, "ts": 1776649791938687, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938679, "dur": 4, + "args": { + "External id": 64591, + "cbid": 211, "correlation": 64591 + } + }, + { + "ph": "s", "id": 64591, "pid": 435, "tid": 435, "ts": 1776649791938679, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791938723, "dur": 1, + "args": { + "External id": 8395, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64609, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64609, "pid": 0, "tid": 7, "ts": 1776649791938723, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938714, "dur": 6, + "args": { + "External id": 8395, + "cbid": 211, "correlation": 64609 + } + }, + { + "ph": "s", "id": 64609, "pid": 435, "tid": 435, "ts": 1776649791938714, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938726, "dur": 0, + "args": { + "External id": 64615, + "cbid": 200, "correlation": 64615 + } + }, + { + "ph": "f", "id": 64615, "pid": 435, "tid": 435, "ts": 1776649791938726, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938727, "dur": 0, + "args": { + "External id": 64616, + "cbid": 273, "correlation": 64616 + } + }, + { + "ph": "f", "id": 64616, "pid": 435, "tid": 435, "ts": 1776649791938727, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791938736, "dur": 19, + "args": { + "External id": 64617, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64617, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64617, "pid": 0, "tid": 7, "ts": 1776649791938736, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938728, "dur": 4, + "args": { + "External id": 64617, + "cbid": 211, "correlation": 64617 + } + }, + { + "ph": "s", "id": 64617, "pid": 435, "tid": 435, "ts": 1776649791938728, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791938771, "dur": 1, + "args": { + "External id": 8400, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64635, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64635, "pid": 0, "tid": 7, "ts": 1776649791938771, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938762, "dur": 6, + "args": { + "External id": 8400, + "cbid": 211, "correlation": 64635 + } + }, + { + "ph": "s", "id": 64635, "pid": 435, "tid": 435, "ts": 1776649791938762, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938775, "dur": 0, + "args": { + "External id": 64641, + "cbid": 200, "correlation": 64641 + } + }, + { + "ph": "f", "id": 64641, "pid": 435, "tid": 435, "ts": 1776649791938775, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938775, "dur": 0, + "args": { + "External id": 64642, + "cbid": 273, "correlation": 64642 + } + }, + { + "ph": "f", "id": 64642, "pid": 435, "tid": 435, "ts": 1776649791938775, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791938785, "dur": 19, + "args": { + "External id": 64643, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64643, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64643, "pid": 0, "tid": 7, "ts": 1776649791938785, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938776, "dur": 5, + "args": { + "External id": 64643, + "cbid": 211, "correlation": 64643 + } + }, + { + "ph": "s", "id": 64643, "pid": 435, "tid": 435, "ts": 1776649791938776, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791938821, "dur": 1, + "args": { + "External id": 8405, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64661, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64661, "pid": 0, "tid": 7, "ts": 1776649791938821, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938811, "dur": 6, + "args": { + "External id": 8405, + "cbid": 211, "correlation": 64661 + } + }, + { + "ph": "s", "id": 64661, "pid": 435, "tid": 435, "ts": 1776649791938811, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938823, "dur": 0, + "args": { + "External id": 64667, + "cbid": 200, "correlation": 64667 + } + }, + { + "ph": "f", "id": 64667, "pid": 435, "tid": 435, "ts": 1776649791938823, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938824, "dur": 0, + "args": { + "External id": 64668, + "cbid": 273, "correlation": 64668 + } + }, + { + "ph": "f", "id": 64668, "pid": 435, "tid": 435, "ts": 1776649791938824, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791938833, "dur": 19, + "args": { + "External id": 64669, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64669, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64669, "pid": 0, "tid": 7, "ts": 1776649791938833, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938825, "dur": 4, + "args": { + "External id": 64669, + "cbid": 211, "correlation": 64669 + } + }, + { + "ph": "s", "id": 64669, "pid": 435, "tid": 435, "ts": 1776649791938825, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791938868, "dur": 1, + "args": { + "External id": 8410, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64687, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64687, "pid": 0, "tid": 7, "ts": 1776649791938868, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938859, "dur": 6, + "args": { + "External id": 8410, + "cbid": 211, "correlation": 64687 + } + }, + { + "ph": "s", "id": 64687, "pid": 435, "tid": 435, "ts": 1776649791938859, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938871, "dur": 0, + "args": { + "External id": 64693, + "cbid": 200, "correlation": 64693 + } + }, + { + "ph": "f", "id": 64693, "pid": 435, "tid": 435, "ts": 1776649791938871, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938872, "dur": 0, + "args": { + "External id": 64694, + "cbid": 273, "correlation": 64694 + } + }, + { + "ph": "f", "id": 64694, "pid": 435, "tid": 435, "ts": 1776649791938872, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791938881, "dur": 19, + "args": { + "External id": 64695, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64695, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64695, "pid": 0, "tid": 7, "ts": 1776649791938881, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938873, "dur": 4, + "args": { + "External id": 64695, + "cbid": 211, "correlation": 64695 + } + }, + { + "ph": "s", "id": 64695, "pid": 435, "tid": 435, "ts": 1776649791938873, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791938917, "dur": 1, + "args": { + "External id": 8415, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64713, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64713, "pid": 0, "tid": 7, "ts": 1776649791938917, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938907, "dur": 6, + "args": { + "External id": 8415, + "cbid": 211, "correlation": 64713 + } + }, + { + "ph": "s", "id": 64713, "pid": 435, "tid": 435, "ts": 1776649791938907, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938920, "dur": 0, + "args": { + "External id": 64719, + "cbid": 200, "correlation": 64719 + } + }, + { + "ph": "f", "id": 64719, "pid": 435, "tid": 435, "ts": 1776649791938920, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938920, "dur": 0, + "args": { + "External id": 64720, + "cbid": 273, "correlation": 64720 + } + }, + { + "ph": "f", "id": 64720, "pid": 435, "tid": 435, "ts": 1776649791938920, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791938929, "dur": 19, + "args": { + "External id": 64721, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64721, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64721, "pid": 0, "tid": 7, "ts": 1776649791938929, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938921, "dur": 4, + "args": { + "External id": 64721, + "cbid": 211, "correlation": 64721 + } + }, + { + "ph": "s", "id": 64721, "pid": 435, "tid": 435, "ts": 1776649791938921, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791938965, "dur": 1, + "args": { + "External id": 8420, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64739, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64739, "pid": 0, "tid": 7, "ts": 1776649791938965, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938956, "dur": 6, + "args": { + "External id": 8420, + "cbid": 211, "correlation": 64739 + } + }, + { + "ph": "s", "id": 64739, "pid": 435, "tid": 435, "ts": 1776649791938956, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938968, "dur": 0, + "args": { + "External id": 64745, + "cbid": 200, "correlation": 64745 + } + }, + { + "ph": "f", "id": 64745, "pid": 435, "tid": 435, "ts": 1776649791938968, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791938968, "dur": 0, + "args": { + "External id": 64746, + "cbid": 273, "correlation": 64746 + } + }, + { + "ph": "f", "id": 64746, "pid": 435, "tid": 435, "ts": 1776649791938968, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791938977, "dur": 19, + "args": { + "External id": 64747, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64747, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64747, "pid": 0, "tid": 7, "ts": 1776649791938977, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791938970, "dur": 4, + "args": { + "External id": 64747, + "cbid": 211, "correlation": 64747 + } + }, + { + "ph": "s", "id": 64747, "pid": 435, "tid": 435, "ts": 1776649791938970, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791939013, "dur": 1, + "args": { + "External id": 8425, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64765, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64765, "pid": 0, "tid": 7, "ts": 1776649791939013, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939003, "dur": 7, + "args": { + "External id": 8425, + "cbid": 211, "correlation": 64765 + } + }, + { + "ph": "s", "id": 64765, "pid": 435, "tid": 435, "ts": 1776649791939003, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939016, "dur": 0, + "args": { + "External id": 64771, + "cbid": 200, "correlation": 64771 + } + }, + { + "ph": "f", "id": 64771, "pid": 435, "tid": 435, "ts": 1776649791939016, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939017, "dur": 0, + "args": { + "External id": 64772, + "cbid": 273, "correlation": 64772 + } + }, + { + "ph": "f", "id": 64772, "pid": 435, "tid": 435, "ts": 1776649791939017, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791939026, "dur": 19, + "args": { + "External id": 64773, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64773, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64773, "pid": 0, "tid": 7, "ts": 1776649791939026, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939018, "dur": 4, + "args": { + "External id": 64773, + "cbid": 211, "correlation": 64773 + } + }, + { + "ph": "s", "id": 64773, "pid": 435, "tid": 435, "ts": 1776649791939018, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791939062, "dur": 1, + "args": { + "External id": 8430, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64791, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64791, "pid": 0, "tid": 7, "ts": 1776649791939062, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939053, "dur": 6, + "args": { + "External id": 8430, + "cbid": 211, "correlation": 64791 + } + }, + { + "ph": "s", "id": 64791, "pid": 435, "tid": 435, "ts": 1776649791939053, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939065, "dur": 0, + "args": { + "External id": 64797, + "cbid": 200, "correlation": 64797 + } + }, + { + "ph": "f", "id": 64797, "pid": 435, "tid": 435, "ts": 1776649791939065, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939066, "dur": 0, + "args": { + "External id": 64798, + "cbid": 273, "correlation": 64798 + } + }, + { + "ph": "f", "id": 64798, "pid": 435, "tid": 435, "ts": 1776649791939066, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791939075, "dur": 19, + "args": { + "External id": 64799, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64799, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64799, "pid": 0, "tid": 7, "ts": 1776649791939075, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939067, "dur": 4, + "args": { + "External id": 64799, + "cbid": 211, "correlation": 64799 + } + }, + { + "ph": "s", "id": 64799, "pid": 435, "tid": 435, "ts": 1776649791939067, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791939111, "dur": 1, + "args": { + "External id": 8435, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64817, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64817, "pid": 0, "tid": 7, "ts": 1776649791939111, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939101, "dur": 6, + "args": { + "External id": 8435, + "cbid": 211, "correlation": 64817 + } + }, + { + "ph": "s", "id": 64817, "pid": 435, "tid": 435, "ts": 1776649791939101, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939114, "dur": 0, + "args": { + "External id": 64823, + "cbid": 200, "correlation": 64823 + } + }, + { + "ph": "f", "id": 64823, "pid": 435, "tid": 435, "ts": 1776649791939114, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939115, "dur": 0, + "args": { + "External id": 64824, + "cbid": 273, "correlation": 64824 + } + }, + { + "ph": "f", "id": 64824, "pid": 435, "tid": 435, "ts": 1776649791939115, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791939123, "dur": 19, + "args": { + "External id": 64825, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64825, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64825, "pid": 0, "tid": 7, "ts": 1776649791939123, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939116, "dur": 4, + "args": { + "External id": 64825, + "cbid": 211, "correlation": 64825 + } + }, + { + "ph": "s", "id": 64825, "pid": 435, "tid": 435, "ts": 1776649791939116, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791939158, "dur": 1, + "args": { + "External id": 8440, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64843, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64843, "pid": 0, "tid": 7, "ts": 1776649791939158, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939149, "dur": 6, + "args": { + "External id": 8440, + "cbid": 211, "correlation": 64843 + } + }, + { + "ph": "s", "id": 64843, "pid": 435, "tid": 435, "ts": 1776649791939149, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939162, "dur": 0, + "args": { + "External id": 64849, + "cbid": 200, "correlation": 64849 + } + }, + { + "ph": "f", "id": 64849, "pid": 435, "tid": 435, "ts": 1776649791939162, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939162, "dur": 0, + "args": { + "External id": 64850, + "cbid": 273, "correlation": 64850 + } + }, + { + "ph": "f", "id": 64850, "pid": 435, "tid": 435, "ts": 1776649791939162, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791939172, "dur": 19, + "args": { + "External id": 64851, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64851, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64851, "pid": 0, "tid": 7, "ts": 1776649791939172, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939163, "dur": 5, + "args": { + "External id": 64851, + "cbid": 211, "correlation": 64851 + } + }, + { + "ph": "s", "id": 64851, "pid": 435, "tid": 435, "ts": 1776649791939163, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791939207, "dur": 1, + "args": { + "External id": 8445, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64869, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64869, "pid": 0, "tid": 7, "ts": 1776649791939207, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939198, "dur": 6, + "args": { + "External id": 8445, + "cbid": 211, "correlation": 64869 + } + }, + { + "ph": "s", "id": 64869, "pid": 435, "tid": 435, "ts": 1776649791939198, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939210, "dur": 0, + "args": { + "External id": 64875, + "cbid": 200, "correlation": 64875 + } + }, + { + "ph": "f", "id": 64875, "pid": 435, "tid": 435, "ts": 1776649791939210, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939211, "dur": 0, + "args": { + "External id": 64876, + "cbid": 273, "correlation": 64876 + } + }, + { + "ph": "f", "id": 64876, "pid": 435, "tid": 435, "ts": 1776649791939211, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791939220, "dur": 19, + "args": { + "External id": 64877, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64877, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64877, "pid": 0, "tid": 7, "ts": 1776649791939220, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939212, "dur": 4, + "args": { + "External id": 64877, + "cbid": 211, "correlation": 64877 + } + }, + { + "ph": "s", "id": 64877, "pid": 435, "tid": 435, "ts": 1776649791939212, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791939256, "dur": 1, + "args": { + "External id": 8450, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64895, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64895, "pid": 0, "tid": 7, "ts": 1776649791939256, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939247, "dur": 6, + "args": { + "External id": 8450, + "cbid": 211, "correlation": 64895 + } + }, + { + "ph": "s", "id": 64895, "pid": 435, "tid": 435, "ts": 1776649791939247, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939259, "dur": 0, + "args": { + "External id": 64901, + "cbid": 200, "correlation": 64901 + } + }, + { + "ph": "f", "id": 64901, "pid": 435, "tid": 435, "ts": 1776649791939259, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939259, "dur": 0, + "args": { + "External id": 64902, + "cbid": 273, "correlation": 64902 + } + }, + { + "ph": "f", "id": 64902, "pid": 435, "tid": 435, "ts": 1776649791939259, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791939268, "dur": 19, + "args": { + "External id": 64903, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64903, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64903, "pid": 0, "tid": 7, "ts": 1776649791939268, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939261, "dur": 4, + "args": { + "External id": 64903, + "cbid": 211, "correlation": 64903 + } + }, + { + "ph": "s", "id": 64903, "pid": 435, "tid": 435, "ts": 1776649791939261, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791939304, "dur": 1, + "args": { + "External id": 8455, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64921, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64921, "pid": 0, "tid": 7, "ts": 1776649791939304, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939295, "dur": 6, + "args": { + "External id": 8455, + "cbid": 211, "correlation": 64921 + } + }, + { + "ph": "s", "id": 64921, "pid": 435, "tid": 435, "ts": 1776649791939295, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939307, "dur": 0, + "args": { + "External id": 64927, + "cbid": 200, "correlation": 64927 + } + }, + { + "ph": "f", "id": 64927, "pid": 435, "tid": 435, "ts": 1776649791939307, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939308, "dur": 0, + "args": { + "External id": 64928, + "cbid": 273, "correlation": 64928 + } + }, + { + "ph": "f", "id": 64928, "pid": 435, "tid": 435, "ts": 1776649791939308, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791939317, "dur": 19, + "args": { + "External id": 64929, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64929, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64929, "pid": 0, "tid": 7, "ts": 1776649791939317, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939309, "dur": 5, + "args": { + "External id": 64929, + "cbid": 211, "correlation": 64929 + } + }, + { + "ph": "s", "id": 64929, "pid": 435, "tid": 435, "ts": 1776649791939309, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791939352, "dur": 1, + "args": { + "External id": 8460, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64947, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64947, "pid": 0, "tid": 7, "ts": 1776649791939352, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939343, "dur": 6, + "args": { + "External id": 8460, + "cbid": 211, "correlation": 64947 + } + }, + { + "ph": "s", "id": 64947, "pid": 435, "tid": 435, "ts": 1776649791939343, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939355, "dur": 0, + "args": { + "External id": 64953, + "cbid": 200, "correlation": 64953 + } + }, + { + "ph": "f", "id": 64953, "pid": 435, "tid": 435, "ts": 1776649791939355, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939356, "dur": 0, + "args": { + "External id": 64954, + "cbid": 273, "correlation": 64954 + } + }, + { + "ph": "f", "id": 64954, "pid": 435, "tid": 435, "ts": 1776649791939356, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791939365, "dur": 19, + "args": { + "External id": 64955, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64955, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64955, "pid": 0, "tid": 7, "ts": 1776649791939365, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939357, "dur": 4, + "args": { + "External id": 64955, + "cbid": 211, "correlation": 64955 + } + }, + { + "ph": "s", "id": 64955, "pid": 435, "tid": 435, "ts": 1776649791939357, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791939400, "dur": 1, + "args": { + "External id": 8465, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64973, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64973, "pid": 0, "tid": 7, "ts": 1776649791939400, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939390, "dur": 6, + "args": { + "External id": 8465, + "cbid": 211, "correlation": 64973 + } + }, + { + "ph": "s", "id": 64973, "pid": 435, "tid": 435, "ts": 1776649791939390, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939403, "dur": 0, + "args": { + "External id": 64979, + "cbid": 200, "correlation": 64979 + } + }, + { + "ph": "f", "id": 64979, "pid": 435, "tid": 435, "ts": 1776649791939403, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939404, "dur": 0, + "args": { + "External id": 64980, + "cbid": 273, "correlation": 64980 + } + }, + { + "ph": "f", "id": 64980, "pid": 435, "tid": 435, "ts": 1776649791939404, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791939412, "dur": 19, + "args": { + "External id": 64981, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64981, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64981, "pid": 0, "tid": 7, "ts": 1776649791939412, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939405, "dur": 4, + "args": { + "External id": 64981, + "cbid": 211, "correlation": 64981 + } + }, + { + "ph": "s", "id": 64981, "pid": 435, "tid": 435, "ts": 1776649791939405, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791939447, "dur": 1, + "args": { + "External id": 8470, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 64999, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 64999, "pid": 0, "tid": 7, "ts": 1776649791939447, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939438, "dur": 6, + "args": { + "External id": 8470, + "cbid": 211, "correlation": 64999 + } + }, + { + "ph": "s", "id": 64999, "pid": 435, "tid": 435, "ts": 1776649791939438, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939451, "dur": 0, + "args": { + "External id": 65005, + "cbid": 200, "correlation": 65005 + } + }, + { + "ph": "f", "id": 65005, "pid": 435, "tid": 435, "ts": 1776649791939451, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939451, "dur": 0, + "args": { + "External id": 65006, + "cbid": 273, "correlation": 65006 + } + }, + { + "ph": "f", "id": 65006, "pid": 435, "tid": 435, "ts": 1776649791939451, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791939460, "dur": 19, + "args": { + "External id": 65007, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65007, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65007, "pid": 0, "tid": 7, "ts": 1776649791939460, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939453, "dur": 5, + "args": { + "External id": 65007, + "cbid": 211, "correlation": 65007 + } + }, + { + "ph": "s", "id": 65007, "pid": 435, "tid": 435, "ts": 1776649791939453, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791939496, "dur": 1, + "args": { + "External id": 8475, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65025, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65025, "pid": 0, "tid": 7, "ts": 1776649791939496, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939487, "dur": 6, + "args": { + "External id": 8475, + "cbid": 211, "correlation": 65025 + } + }, + { + "ph": "s", "id": 65025, "pid": 435, "tid": 435, "ts": 1776649791939487, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939499, "dur": 0, + "args": { + "External id": 65031, + "cbid": 200, "correlation": 65031 + } + }, + { + "ph": "f", "id": 65031, "pid": 435, "tid": 435, "ts": 1776649791939499, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939500, "dur": 0, + "args": { + "External id": 65032, + "cbid": 273, "correlation": 65032 + } + }, + { + "ph": "f", "id": 65032, "pid": 435, "tid": 435, "ts": 1776649791939500, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791939508, "dur": 19, + "args": { + "External id": 65033, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65033, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65033, "pid": 0, "tid": 7, "ts": 1776649791939508, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939501, "dur": 4, + "args": { + "External id": 65033, + "cbid": 211, "correlation": 65033 + } + }, + { + "ph": "s", "id": 65033, "pid": 435, "tid": 435, "ts": 1776649791939501, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791939543, "dur": 1, + "args": { + "External id": 8480, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65051, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65051, "pid": 0, "tid": 7, "ts": 1776649791939543, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939534, "dur": 6, + "args": { + "External id": 8480, + "cbid": 211, "correlation": 65051 + } + }, + { + "ph": "s", "id": 65051, "pid": 435, "tid": 435, "ts": 1776649791939534, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939547, "dur": 0, + "args": { + "External id": 65057, + "cbid": 200, "correlation": 65057 + } + }, + { + "ph": "f", "id": 65057, "pid": 435, "tid": 435, "ts": 1776649791939547, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939547, "dur": 0, + "args": { + "External id": 65058, + "cbid": 273, "correlation": 65058 + } + }, + { + "ph": "f", "id": 65058, "pid": 435, "tid": 435, "ts": 1776649791939547, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791939557, "dur": 19, + "args": { + "External id": 65059, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65059, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65059, "pid": 0, "tid": 7, "ts": 1776649791939557, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939548, "dur": 5, + "args": { + "External id": 65059, + "cbid": 211, "correlation": 65059 + } + }, + { + "ph": "s", "id": 65059, "pid": 435, "tid": 435, "ts": 1776649791939548, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791939592, "dur": 1, + "args": { + "External id": 8485, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65077, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65077, "pid": 0, "tid": 7, "ts": 1776649791939592, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939583, "dur": 6, + "args": { + "External id": 8485, + "cbid": 211, "correlation": 65077 + } + }, + { + "ph": "s", "id": 65077, "pid": 435, "tid": 435, "ts": 1776649791939583, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939595, "dur": 0, + "args": { + "External id": 65083, + "cbid": 200, "correlation": 65083 + } + }, + { + "ph": "f", "id": 65083, "pid": 435, "tid": 435, "ts": 1776649791939595, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939596, "dur": 0, + "args": { + "External id": 65084, + "cbid": 273, "correlation": 65084 + } + }, + { + "ph": "f", "id": 65084, "pid": 435, "tid": 435, "ts": 1776649791939596, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791939605, "dur": 19, + "args": { + "External id": 65085, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65085, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65085, "pid": 0, "tid": 7, "ts": 1776649791939605, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939597, "dur": 4, + "args": { + "External id": 65085, + "cbid": 211, "correlation": 65085 + } + }, + { + "ph": "s", "id": 65085, "pid": 435, "tid": 435, "ts": 1776649791939597, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791939642, "dur": 1, + "args": { + "External id": 8490, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65103, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65103, "pid": 0, "tid": 7, "ts": 1776649791939642, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939632, "dur": 6, + "args": { + "External id": 8490, + "cbid": 211, "correlation": 65103 + } + }, + { + "ph": "s", "id": 65103, "pid": 435, "tid": 435, "ts": 1776649791939632, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939645, "dur": 0, + "args": { + "External id": 65109, + "cbid": 200, "correlation": 65109 + } + }, + { + "ph": "f", "id": 65109, "pid": 435, "tid": 435, "ts": 1776649791939645, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939646, "dur": 0, + "args": { + "External id": 65110, + "cbid": 273, "correlation": 65110 + } + }, + { + "ph": "f", "id": 65110, "pid": 435, "tid": 435, "ts": 1776649791939646, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791939654, "dur": 19, + "args": { + "External id": 65111, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65111, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65111, "pid": 0, "tid": 7, "ts": 1776649791939654, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939647, "dur": 4, + "args": { + "External id": 65111, + "cbid": 211, "correlation": 65111 + } + }, + { + "ph": "s", "id": 65111, "pid": 435, "tid": 435, "ts": 1776649791939647, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791939689, "dur": 1, + "args": { + "External id": 8495, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65129, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65129, "pid": 0, "tid": 7, "ts": 1776649791939689, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939680, "dur": 6, + "args": { + "External id": 8495, + "cbid": 211, "correlation": 65129 + } + }, + { + "ph": "s", "id": 65129, "pid": 435, "tid": 435, "ts": 1776649791939680, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939693, "dur": 0, + "args": { + "External id": 65135, + "cbid": 200, "correlation": 65135 + } + }, + { + "ph": "f", "id": 65135, "pid": 435, "tid": 435, "ts": 1776649791939693, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939693, "dur": 0, + "args": { + "External id": 65136, + "cbid": 273, "correlation": 65136 + } + }, + { + "ph": "f", "id": 65136, "pid": 435, "tid": 435, "ts": 1776649791939693, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791939702, "dur": 19, + "args": { + "External id": 65137, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65137, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65137, "pid": 0, "tid": 7, "ts": 1776649791939702, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939694, "dur": 4, + "args": { + "External id": 65137, + "cbid": 211, "correlation": 65137 + } + }, + { + "ph": "s", "id": 65137, "pid": 435, "tid": 435, "ts": 1776649791939694, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791939738, "dur": 1, + "args": { + "External id": 8500, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65155, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65155, "pid": 0, "tid": 7, "ts": 1776649791939738, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939729, "dur": 6, + "args": { + "External id": 8500, + "cbid": 211, "correlation": 65155 + } + }, + { + "ph": "s", "id": 65155, "pid": 435, "tid": 435, "ts": 1776649791939729, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939741, "dur": 0, + "args": { + "External id": 65161, + "cbid": 200, "correlation": 65161 + } + }, + { + "ph": "f", "id": 65161, "pid": 435, "tid": 435, "ts": 1776649791939741, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939742, "dur": 0, + "args": { + "External id": 65162, + "cbid": 273, "correlation": 65162 + } + }, + { + "ph": "f", "id": 65162, "pid": 435, "tid": 435, "ts": 1776649791939742, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791939751, "dur": 19, + "args": { + "External id": 65163, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65163, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65163, "pid": 0, "tid": 7, "ts": 1776649791939751, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939743, "dur": 4, + "args": { + "External id": 65163, + "cbid": 211, "correlation": 65163 + } + }, + { + "ph": "s", "id": 65163, "pid": 435, "tid": 435, "ts": 1776649791939743, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791939787, "dur": 1, + "args": { + "External id": 8505, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65181, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65181, "pid": 0, "tid": 7, "ts": 1776649791939787, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939777, "dur": 7, + "args": { + "External id": 8505, + "cbid": 211, "correlation": 65181 + } + }, + { + "ph": "s", "id": 65181, "pid": 435, "tid": 435, "ts": 1776649791939777, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939790, "dur": 0, + "args": { + "External id": 65187, + "cbid": 200, "correlation": 65187 + } + }, + { + "ph": "f", "id": 65187, "pid": 435, "tid": 435, "ts": 1776649791939790, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939791, "dur": 0, + "args": { + "External id": 65188, + "cbid": 273, "correlation": 65188 + } + }, + { + "ph": "f", "id": 65188, "pid": 435, "tid": 435, "ts": 1776649791939791, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791939799, "dur": 19, + "args": { + "External id": 65189, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65189, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65189, "pid": 0, "tid": 7, "ts": 1776649791939799, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939792, "dur": 4, + "args": { + "External id": 65189, + "cbid": 211, "correlation": 65189 + } + }, + { + "ph": "s", "id": 65189, "pid": 435, "tid": 435, "ts": 1776649791939792, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791939835, "dur": 1, + "args": { + "External id": 8510, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65207, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65207, "pid": 0, "tid": 7, "ts": 1776649791939835, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939826, "dur": 6, + "args": { + "External id": 8510, + "cbid": 211, "correlation": 65207 + } + }, + { + "ph": "s", "id": 65207, "pid": 435, "tid": 435, "ts": 1776649791939826, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939838, "dur": 0, + "args": { + "External id": 65213, + "cbid": 200, "correlation": 65213 + } + }, + { + "ph": "f", "id": 65213, "pid": 435, "tid": 435, "ts": 1776649791939838, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939839, "dur": 0, + "args": { + "External id": 65214, + "cbid": 273, "correlation": 65214 + } + }, + { + "ph": "f", "id": 65214, "pid": 435, "tid": 435, "ts": 1776649791939839, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791939847, "dur": 19, + "args": { + "External id": 65215, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65215, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65215, "pid": 0, "tid": 7, "ts": 1776649791939847, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939840, "dur": 4, + "args": { + "External id": 65215, + "cbid": 211, "correlation": 65215 + } + }, + { + "ph": "s", "id": 65215, "pid": 435, "tid": 435, "ts": 1776649791939840, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791939884, "dur": 1, + "args": { + "External id": 8515, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65233, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65233, "pid": 0, "tid": 7, "ts": 1776649791939884, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939875, "dur": 6, + "args": { + "External id": 8515, + "cbid": 211, "correlation": 65233 + } + }, + { + "ph": "s", "id": 65233, "pid": 435, "tid": 435, "ts": 1776649791939875, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939887, "dur": 0, + "args": { + "External id": 65239, + "cbid": 200, "correlation": 65239 + } + }, + { + "ph": "f", "id": 65239, "pid": 435, "tid": 435, "ts": 1776649791939887, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939887, "dur": 0, + "args": { + "External id": 65240, + "cbid": 273, "correlation": 65240 + } + }, + { + "ph": "f", "id": 65240, "pid": 435, "tid": 435, "ts": 1776649791939887, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791939896, "dur": 19, + "args": { + "External id": 65241, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65241, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65241, "pid": 0, "tid": 7, "ts": 1776649791939896, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939889, "dur": 4, + "args": { + "External id": 65241, + "cbid": 211, "correlation": 65241 + } + }, + { + "ph": "s", "id": 65241, "pid": 435, "tid": 435, "ts": 1776649791939889, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791939931, "dur": 1, + "args": { + "External id": 8520, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65259, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65259, "pid": 0, "tid": 7, "ts": 1776649791939931, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939922, "dur": 6, + "args": { + "External id": 8520, + "cbid": 211, "correlation": 65259 + } + }, + { + "ph": "s", "id": 65259, "pid": 435, "tid": 435, "ts": 1776649791939922, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939935, "dur": 0, + "args": { + "External id": 65265, + "cbid": 200, "correlation": 65265 + } + }, + { + "ph": "f", "id": 65265, "pid": 435, "tid": 435, "ts": 1776649791939935, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939935, "dur": 0, + "args": { + "External id": 65266, + "cbid": 273, "correlation": 65266 + } + }, + { + "ph": "f", "id": 65266, "pid": 435, "tid": 435, "ts": 1776649791939935, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791939945, "dur": 19, + "args": { + "External id": 65267, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65267, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65267, "pid": 0, "tid": 7, "ts": 1776649791939945, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939936, "dur": 5, + "args": { + "External id": 65267, + "cbid": 211, "correlation": 65267 + } + }, + { + "ph": "s", "id": 65267, "pid": 435, "tid": 435, "ts": 1776649791939936, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791939980, "dur": 1, + "args": { + "External id": 8525, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65285, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65285, "pid": 0, "tid": 7, "ts": 1776649791939980, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939971, "dur": 6, + "args": { + "External id": 8525, + "cbid": 211, "correlation": 65285 + } + }, + { + "ph": "s", "id": 65285, "pid": 435, "tid": 435, "ts": 1776649791939971, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939983, "dur": 0, + "args": { + "External id": 65291, + "cbid": 200, "correlation": 65291 + } + }, + { + "ph": "f", "id": 65291, "pid": 435, "tid": 435, "ts": 1776649791939983, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791939984, "dur": 0, + "args": { + "External id": 65292, + "cbid": 273, "correlation": 65292 + } + }, + { + "ph": "f", "id": 65292, "pid": 435, "tid": 435, "ts": 1776649791939984, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791939993, "dur": 19, + "args": { + "External id": 65293, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65293, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65293, "pid": 0, "tid": 7, "ts": 1776649791939993, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791939985, "dur": 4, + "args": { + "External id": 65293, + "cbid": 211, "correlation": 65293 + } + }, + { + "ph": "s", "id": 65293, "pid": 435, "tid": 435, "ts": 1776649791939985, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791940029, "dur": 1, + "args": { + "External id": 8530, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65311, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65311, "pid": 0, "tid": 7, "ts": 1776649791940029, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940020, "dur": 6, + "args": { + "External id": 8530, + "cbid": 211, "correlation": 65311 + } + }, + { + "ph": "s", "id": 65311, "pid": 435, "tid": 435, "ts": 1776649791940020, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940032, "dur": 0, + "args": { + "External id": 65317, + "cbid": 200, "correlation": 65317 + } + }, + { + "ph": "f", "id": 65317, "pid": 435, "tid": 435, "ts": 1776649791940032, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940033, "dur": 0, + "args": { + "External id": 65318, + "cbid": 273, "correlation": 65318 + } + }, + { + "ph": "f", "id": 65318, "pid": 435, "tid": 435, "ts": 1776649791940033, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791940042, "dur": 19, + "args": { + "External id": 65319, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65319, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65319, "pid": 0, "tid": 7, "ts": 1776649791940042, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940034, "dur": 4, + "args": { + "External id": 65319, + "cbid": 211, "correlation": 65319 + } + }, + { + "ph": "s", "id": 65319, "pid": 435, "tid": 435, "ts": 1776649791940034, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791940077, "dur": 1, + "args": { + "External id": 8535, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65337, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65337, "pid": 0, "tid": 7, "ts": 1776649791940077, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940068, "dur": 6, + "args": { + "External id": 8535, + "cbid": 211, "correlation": 65337 + } + }, + { + "ph": "s", "id": 65337, "pid": 435, "tid": 435, "ts": 1776649791940068, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940080, "dur": 0, + "args": { + "External id": 65343, + "cbid": 200, "correlation": 65343 + } + }, + { + "ph": "f", "id": 65343, "pid": 435, "tid": 435, "ts": 1776649791940080, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940081, "dur": 0, + "args": { + "External id": 65344, + "cbid": 273, "correlation": 65344 + } + }, + { + "ph": "f", "id": 65344, "pid": 435, "tid": 435, "ts": 1776649791940081, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791940089, "dur": 19, + "args": { + "External id": 65345, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65345, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65345, "pid": 0, "tid": 7, "ts": 1776649791940089, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940082, "dur": 5, + "args": { + "External id": 65345, + "cbid": 211, "correlation": 65345 + } + }, + { + "ph": "s", "id": 65345, "pid": 435, "tid": 435, "ts": 1776649791940082, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791940125, "dur": 1, + "args": { + "External id": 8540, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65363, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65363, "pid": 0, "tid": 7, "ts": 1776649791940125, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940116, "dur": 6, + "args": { + "External id": 8540, + "cbid": 211, "correlation": 65363 + } + }, + { + "ph": "s", "id": 65363, "pid": 435, "tid": 435, "ts": 1776649791940116, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940129, "dur": 0, + "args": { + "External id": 65369, + "cbid": 200, "correlation": 65369 + } + }, + { + "ph": "f", "id": 65369, "pid": 435, "tid": 435, "ts": 1776649791940129, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940129, "dur": 0, + "args": { + "External id": 65370, + "cbid": 273, "correlation": 65370 + } + }, + { + "ph": "f", "id": 65370, "pid": 435, "tid": 435, "ts": 1776649791940129, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791940138, "dur": 19, + "args": { + "External id": 65371, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65371, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65371, "pid": 0, "tid": 7, "ts": 1776649791940138, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940130, "dur": 4, + "args": { + "External id": 65371, + "cbid": 211, "correlation": 65371 + } + }, + { + "ph": "s", "id": 65371, "pid": 435, "tid": 435, "ts": 1776649791940130, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791940175, "dur": 1, + "args": { + "External id": 8545, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65389, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65389, "pid": 0, "tid": 7, "ts": 1776649791940175, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940164, "dur": 7, + "args": { + "External id": 8545, + "cbid": 211, "correlation": 65389 + } + }, + { + "ph": "s", "id": 65389, "pid": 435, "tid": 435, "ts": 1776649791940164, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940178, "dur": 0, + "args": { + "External id": 65395, + "cbid": 200, "correlation": 65395 + } + }, + { + "ph": "f", "id": 65395, "pid": 435, "tid": 435, "ts": 1776649791940178, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940178, "dur": 0, + "args": { + "External id": 65396, + "cbid": 273, "correlation": 65396 + } + }, + { + "ph": "f", "id": 65396, "pid": 435, "tid": 435, "ts": 1776649791940178, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791940187, "dur": 19, + "args": { + "External id": 65397, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65397, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65397, "pid": 0, "tid": 7, "ts": 1776649791940187, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940180, "dur": 4, + "args": { + "External id": 65397, + "cbid": 211, "correlation": 65397 + } + }, + { + "ph": "s", "id": 65397, "pid": 435, "tid": 435, "ts": 1776649791940180, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791940243, "dur": 1, + "args": { + "External id": 8550, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65415, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65415, "pid": 0, "tid": 7, "ts": 1776649791940243, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940213, "dur": 28, + "args": { + "External id": 8550, + "cbid": 211, "correlation": 65415 + } + }, + { + "ph": "s", "id": 65415, "pid": 435, "tid": 435, "ts": 1776649791940213, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940248, "dur": 0, + "args": { + "External id": 65421, + "cbid": 200, "correlation": 65421 + } + }, + { + "ph": "f", "id": 65421, "pid": 435, "tid": 435, "ts": 1776649791940248, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940249, "dur": 0, + "args": { + "External id": 65422, + "cbid": 273, "correlation": 65422 + } + }, + { + "ph": "f", "id": 65422, "pid": 435, "tid": 435, "ts": 1776649791940249, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791940258, "dur": 19, + "args": { + "External id": 65423, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65423, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65423, "pid": 0, "tid": 7, "ts": 1776649791940258, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940250, "dur": 5, + "args": { + "External id": 65423, + "cbid": 211, "correlation": 65423 + } + }, + { + "ph": "s", "id": 65423, "pid": 435, "tid": 435, "ts": 1776649791940250, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791940298, "dur": 1, + "args": { + "External id": 8555, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65441, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65441, "pid": 0, "tid": 7, "ts": 1776649791940298, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940288, "dur": 6, + "args": { + "External id": 8555, + "cbid": 211, "correlation": 65441 + } + }, + { + "ph": "s", "id": 65441, "pid": 435, "tid": 435, "ts": 1776649791940288, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940301, "dur": 0, + "args": { + "External id": 65447, + "cbid": 200, "correlation": 65447 + } + }, + { + "ph": "f", "id": 65447, "pid": 435, "tid": 435, "ts": 1776649791940301, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940302, "dur": 0, + "args": { + "External id": 65448, + "cbid": 273, "correlation": 65448 + } + }, + { + "ph": "f", "id": 65448, "pid": 435, "tid": 435, "ts": 1776649791940302, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791940311, "dur": 19, + "args": { + "External id": 65449, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65449, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65449, "pid": 0, "tid": 7, "ts": 1776649791940311, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940303, "dur": 4, + "args": { + "External id": 65449, + "cbid": 211, "correlation": 65449 + } + }, + { + "ph": "s", "id": 65449, "pid": 435, "tid": 435, "ts": 1776649791940303, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791940346, "dur": 1, + "args": { + "External id": 8560, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65467, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65467, "pid": 0, "tid": 7, "ts": 1776649791940346, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940337, "dur": 6, + "args": { + "External id": 8560, + "cbid": 211, "correlation": 65467 + } + }, + { + "ph": "s", "id": 65467, "pid": 435, "tid": 435, "ts": 1776649791940337, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940349, "dur": 0, + "args": { + "External id": 65473, + "cbid": 200, "correlation": 65473 + } + }, + { + "ph": "f", "id": 65473, "pid": 435, "tid": 435, "ts": 1776649791940349, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940350, "dur": 0, + "args": { + "External id": 65474, + "cbid": 273, "correlation": 65474 + } + }, + { + "ph": "f", "id": 65474, "pid": 435, "tid": 435, "ts": 1776649791940350, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791940359, "dur": 19, + "args": { + "External id": 65475, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65475, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65475, "pid": 0, "tid": 7, "ts": 1776649791940359, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940351, "dur": 5, + "args": { + "External id": 65475, + "cbid": 211, "correlation": 65475 + } + }, + { + "ph": "s", "id": 65475, "pid": 435, "tid": 435, "ts": 1776649791940351, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791940397, "dur": 1, + "args": { + "External id": 8565, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65493, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65493, "pid": 0, "tid": 7, "ts": 1776649791940397, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940388, "dur": 6, + "args": { + "External id": 8565, + "cbid": 211, "correlation": 65493 + } + }, + { + "ph": "s", "id": 65493, "pid": 435, "tid": 435, "ts": 1776649791940388, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940400, "dur": 0, + "args": { + "External id": 65499, + "cbid": 200, "correlation": 65499 + } + }, + { + "ph": "f", "id": 65499, "pid": 435, "tid": 435, "ts": 1776649791940400, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940400, "dur": 0, + "args": { + "External id": 65500, + "cbid": 273, "correlation": 65500 + } + }, + { + "ph": "f", "id": 65500, "pid": 435, "tid": 435, "ts": 1776649791940400, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791940409, "dur": 19, + "args": { + "External id": 65501, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65501, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65501, "pid": 0, "tid": 7, "ts": 1776649791940409, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940401, "dur": 4, + "args": { + "External id": 65501, + "cbid": 211, "correlation": 65501 + } + }, + { + "ph": "s", "id": 65501, "pid": 435, "tid": 435, "ts": 1776649791940401, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791940444, "dur": 1, + "args": { + "External id": 8570, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65519, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65519, "pid": 0, "tid": 7, "ts": 1776649791940444, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940435, "dur": 6, + "args": { + "External id": 8570, + "cbid": 211, "correlation": 65519 + } + }, + { + "ph": "s", "id": 65519, "pid": 435, "tid": 435, "ts": 1776649791940435, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940447, "dur": 0, + "args": { + "External id": 65525, + "cbid": 200, "correlation": 65525 + } + }, + { + "ph": "f", "id": 65525, "pid": 435, "tid": 435, "ts": 1776649791940447, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940448, "dur": 0, + "args": { + "External id": 65526, + "cbid": 273, "correlation": 65526 + } + }, + { + "ph": "f", "id": 65526, "pid": 435, "tid": 435, "ts": 1776649791940448, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791940457, "dur": 19, + "args": { + "External id": 65527, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65527, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65527, "pid": 0, "tid": 7, "ts": 1776649791940457, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940449, "dur": 5, + "args": { + "External id": 65527, + "cbid": 211, "correlation": 65527 + } + }, + { + "ph": "s", "id": 65527, "pid": 435, "tid": 435, "ts": 1776649791940449, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791940492, "dur": 1, + "args": { + "External id": 8575, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65545, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65545, "pid": 0, "tid": 7, "ts": 1776649791940492, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940483, "dur": 6, + "args": { + "External id": 8575, + "cbid": 211, "correlation": 65545 + } + }, + { + "ph": "s", "id": 65545, "pid": 435, "tid": 435, "ts": 1776649791940483, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940495, "dur": 0, + "args": { + "External id": 65551, + "cbid": 200, "correlation": 65551 + } + }, + { + "ph": "f", "id": 65551, "pid": 435, "tid": 435, "ts": 1776649791940495, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940495, "dur": 0, + "args": { + "External id": 65552, + "cbid": 273, "correlation": 65552 + } + }, + { + "ph": "f", "id": 65552, "pid": 435, "tid": 435, "ts": 1776649791940495, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791940504, "dur": 19, + "args": { + "External id": 65553, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65553, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65553, "pid": 0, "tid": 7, "ts": 1776649791940504, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940496, "dur": 5, + "args": { + "External id": 65553, + "cbid": 211, "correlation": 65553 + } + }, + { + "ph": "s", "id": 65553, "pid": 435, "tid": 435, "ts": 1776649791940496, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791940539, "dur": 1, + "args": { + "External id": 8580, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65571, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65571, "pid": 0, "tid": 7, "ts": 1776649791940539, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940530, "dur": 6, + "args": { + "External id": 8580, + "cbid": 211, "correlation": 65571 + } + }, + { + "ph": "s", "id": 65571, "pid": 435, "tid": 435, "ts": 1776649791940530, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940542, "dur": 0, + "args": { + "External id": 65577, + "cbid": 200, "correlation": 65577 + } + }, + { + "ph": "f", "id": 65577, "pid": 435, "tid": 435, "ts": 1776649791940542, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940543, "dur": 0, + "args": { + "External id": 65578, + "cbid": 273, "correlation": 65578 + } + }, + { + "ph": "f", "id": 65578, "pid": 435, "tid": 435, "ts": 1776649791940543, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791940552, "dur": 19, + "args": { + "External id": 65579, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65579, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65579, "pid": 0, "tid": 7, "ts": 1776649791940552, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940544, "dur": 4, + "args": { + "External id": 65579, + "cbid": 211, "correlation": 65579 + } + }, + { + "ph": "s", "id": 65579, "pid": 435, "tid": 435, "ts": 1776649791940544, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791940587, "dur": 1, + "args": { + "External id": 8585, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65597, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65597, "pid": 0, "tid": 7, "ts": 1776649791940587, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940578, "dur": 6, + "args": { + "External id": 8585, + "cbid": 211, "correlation": 65597 + } + }, + { + "ph": "s", "id": 65597, "pid": 435, "tid": 435, "ts": 1776649791940578, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940590, "dur": 0, + "args": { + "External id": 65603, + "cbid": 200, "correlation": 65603 + } + }, + { + "ph": "f", "id": 65603, "pid": 435, "tid": 435, "ts": 1776649791940590, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940591, "dur": 0, + "args": { + "External id": 65604, + "cbid": 273, "correlation": 65604 + } + }, + { + "ph": "f", "id": 65604, "pid": 435, "tid": 435, "ts": 1776649791940591, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791940599, "dur": 19, + "args": { + "External id": 65605, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65605, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65605, "pid": 0, "tid": 7, "ts": 1776649791940599, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940592, "dur": 5, + "args": { + "External id": 65605, + "cbid": 211, "correlation": 65605 + } + }, + { + "ph": "s", "id": 65605, "pid": 435, "tid": 435, "ts": 1776649791940592, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791940634, "dur": 1, + "args": { + "External id": 8590, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65623, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65623, "pid": 0, "tid": 7, "ts": 1776649791940634, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940625, "dur": 6, + "args": { + "External id": 8590, + "cbid": 211, "correlation": 65623 + } + }, + { + "ph": "s", "id": 65623, "pid": 435, "tid": 435, "ts": 1776649791940625, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940637, "dur": 0, + "args": { + "External id": 65629, + "cbid": 200, "correlation": 65629 + } + }, + { + "ph": "f", "id": 65629, "pid": 435, "tid": 435, "ts": 1776649791940637, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940638, "dur": 0, + "args": { + "External id": 65630, + "cbid": 273, "correlation": 65630 + } + }, + { + "ph": "f", "id": 65630, "pid": 435, "tid": 435, "ts": 1776649791940638, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791940647, "dur": 19, + "args": { + "External id": 65631, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65631, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65631, "pid": 0, "tid": 7, "ts": 1776649791940647, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940639, "dur": 5, + "args": { + "External id": 65631, + "cbid": 211, "correlation": 65631 + } + }, + { + "ph": "s", "id": 65631, "pid": 435, "tid": 435, "ts": 1776649791940639, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791940682, "dur": 1, + "args": { + "External id": 8595, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65649, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65649, "pid": 0, "tid": 7, "ts": 1776649791940682, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940672, "dur": 6, + "args": { + "External id": 8595, + "cbid": 211, "correlation": 65649 + } + }, + { + "ph": "s", "id": 65649, "pid": 435, "tid": 435, "ts": 1776649791940672, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940684, "dur": 0, + "args": { + "External id": 65655, + "cbid": 200, "correlation": 65655 + } + }, + { + "ph": "f", "id": 65655, "pid": 435, "tid": 435, "ts": 1776649791940684, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940685, "dur": 0, + "args": { + "External id": 65656, + "cbid": 273, "correlation": 65656 + } + }, + { + "ph": "f", "id": 65656, "pid": 435, "tid": 435, "ts": 1776649791940685, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791940694, "dur": 19, + "args": { + "External id": 65657, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65657, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65657, "pid": 0, "tid": 7, "ts": 1776649791940694, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940686, "dur": 5, + "args": { + "External id": 65657, + "cbid": 211, "correlation": 65657 + } + }, + { + "ph": "s", "id": 65657, "pid": 435, "tid": 435, "ts": 1776649791940686, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791940729, "dur": 1, + "args": { + "External id": 8600, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65675, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65675, "pid": 0, "tid": 7, "ts": 1776649791940729, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940720, "dur": 6, + "args": { + "External id": 8600, + "cbid": 211, "correlation": 65675 + } + }, + { + "ph": "s", "id": 65675, "pid": 435, "tid": 435, "ts": 1776649791940720, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940733, "dur": 0, + "args": { + "External id": 65681, + "cbid": 200, "correlation": 65681 + } + }, + { + "ph": "f", "id": 65681, "pid": 435, "tid": 435, "ts": 1776649791940733, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940733, "dur": 0, + "args": { + "External id": 65682, + "cbid": 273, "correlation": 65682 + } + }, + { + "ph": "f", "id": 65682, "pid": 435, "tid": 435, "ts": 1776649791940733, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791940742, "dur": 19, + "args": { + "External id": 65683, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65683, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65683, "pid": 0, "tid": 7, "ts": 1776649791940742, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940734, "dur": 4, + "args": { + "External id": 65683, + "cbid": 211, "correlation": 65683 + } + }, + { + "ph": "s", "id": 65683, "pid": 435, "tid": 435, "ts": 1776649791940734, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791940778, "dur": 1, + "args": { + "External id": 8605, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65701, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65701, "pid": 0, "tid": 7, "ts": 1776649791940778, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940768, "dur": 6, + "args": { + "External id": 8605, + "cbid": 211, "correlation": 65701 + } + }, + { + "ph": "s", "id": 65701, "pid": 435, "tid": 435, "ts": 1776649791940768, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940780, "dur": 0, + "args": { + "External id": 65707, + "cbid": 200, "correlation": 65707 + } + }, + { + "ph": "f", "id": 65707, "pid": 435, "tid": 435, "ts": 1776649791940780, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940781, "dur": 0, + "args": { + "External id": 65708, + "cbid": 273, "correlation": 65708 + } + }, + { + "ph": "f", "id": 65708, "pid": 435, "tid": 435, "ts": 1776649791940781, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791940790, "dur": 19, + "args": { + "External id": 65709, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65709, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65709, "pid": 0, "tid": 7, "ts": 1776649791940790, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940782, "dur": 4, + "args": { + "External id": 65709, + "cbid": 211, "correlation": 65709 + } + }, + { + "ph": "s", "id": 65709, "pid": 435, "tid": 435, "ts": 1776649791940782, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791940812, "dur": 3, + "args": { + "External id": 65717, + "cbid": 135, "correlation": 65717 + } + }, + { + "ph": "f", "id": 65717, "pid": 435, "tid": 435, "ts": 1776649791940812, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791940827, "dur": 5, + "args": { + "External id": 65722, + "cbid": 165, "correlation": 65722 + } + }, + { + "ph": "s", "id": 65722, "pid": 435, "tid": 435, "ts": 1776649791940827, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649791940836, "dur": 4, + "args": { + "External id": 65725, + "cbid": 139, "correlation": 65725 + } + }, + { + "ph": "f", "id": 65725, "pid": 435, "tid": 435, "ts": 1776649791940836, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791940892, "dur": 1, + "args": { + "External id": 8610, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65753, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65753, "pid": 0, "tid": 7, "ts": 1776649791940892, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940881, "dur": 8, + "args": { + "External id": 8610, + "cbid": 211, "correlation": 65753 + } + }, + { + "ph": "s", "id": 65753, "pid": 435, "tid": 435, "ts": 1776649791940881, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940896, "dur": 0, + "args": { + "External id": 65759, + "cbid": 200, "correlation": 65759 + } + }, + { + "ph": "f", "id": 65759, "pid": 435, "tid": 435, "ts": 1776649791940896, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940897, "dur": 0, + "args": { + "External id": 65760, + "cbid": 273, "correlation": 65760 + } + }, + { + "ph": "f", "id": 65760, "pid": 435, "tid": 435, "ts": 1776649791940897, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791940907, "dur": 19, + "args": { + "External id": 65761, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65761, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65761, "pid": 0, "tid": 7, "ts": 1776649791940907, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940898, "dur": 5, + "args": { + "External id": 65761, + "cbid": 211, "correlation": 65761 + } + }, + { + "ph": "s", "id": 65761, "pid": 435, "tid": 435, "ts": 1776649791940898, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791940944, "dur": 1, + "args": { + "External id": 8615, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65779, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65779, "pid": 0, "tid": 7, "ts": 1776649791940944, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940935, "dur": 6, + "args": { + "External id": 8615, + "cbid": 211, "correlation": 65779 + } + }, + { + "ph": "s", "id": 65779, "pid": 435, "tid": 435, "ts": 1776649791940935, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940947, "dur": 0, + "args": { + "External id": 65785, + "cbid": 200, "correlation": 65785 + } + }, + { + "ph": "f", "id": 65785, "pid": 435, "tid": 435, "ts": 1776649791940947, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940948, "dur": 0, + "args": { + "External id": 65786, + "cbid": 273, "correlation": 65786 + } + }, + { + "ph": "f", "id": 65786, "pid": 435, "tid": 435, "ts": 1776649791940948, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791940957, "dur": 19, + "args": { + "External id": 65787, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65787, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65787, "pid": 0, "tid": 7, "ts": 1776649791940957, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940949, "dur": 5, + "args": { + "External id": 65787, + "cbid": 211, "correlation": 65787 + } + }, + { + "ph": "s", "id": 65787, "pid": 435, "tid": 435, "ts": 1776649791940949, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791940992, "dur": 1, + "args": { + "External id": 8620, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65805, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65805, "pid": 0, "tid": 7, "ts": 1776649791940992, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940983, "dur": 6, + "args": { + "External id": 8620, + "cbid": 211, "correlation": 65805 + } + }, + { + "ph": "s", "id": 65805, "pid": 435, "tid": 435, "ts": 1776649791940983, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940995, "dur": 0, + "args": { + "External id": 65811, + "cbid": 200, "correlation": 65811 + } + }, + { + "ph": "f", "id": 65811, "pid": 435, "tid": 435, "ts": 1776649791940995, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791940996, "dur": 0, + "args": { + "External id": 65812, + "cbid": 273, "correlation": 65812 + } + }, + { + "ph": "f", "id": 65812, "pid": 435, "tid": 435, "ts": 1776649791940996, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791941005, "dur": 19, + "args": { + "External id": 65813, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65813, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65813, "pid": 0, "tid": 7, "ts": 1776649791941005, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791940997, "dur": 4, + "args": { + "External id": 65813, + "cbid": 211, "correlation": 65813 + } + }, + { + "ph": "s", "id": 65813, "pid": 435, "tid": 435, "ts": 1776649791940997, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791941041, "dur": 1, + "args": { + "External id": 8625, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65831, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65831, "pid": 0, "tid": 7, "ts": 1776649791941041, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941033, "dur": 6, + "args": { + "External id": 8625, + "cbid": 211, "correlation": 65831 + } + }, + { + "ph": "s", "id": 65831, "pid": 435, "tid": 435, "ts": 1776649791941033, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941045, "dur": 0, + "args": { + "External id": 65837, + "cbid": 200, "correlation": 65837 + } + }, + { + "ph": "f", "id": 65837, "pid": 435, "tid": 435, "ts": 1776649791941045, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941045, "dur": 0, + "args": { + "External id": 65838, + "cbid": 273, "correlation": 65838 + } + }, + { + "ph": "f", "id": 65838, "pid": 435, "tid": 435, "ts": 1776649791941045, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791941054, "dur": 19, + "args": { + "External id": 65839, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65839, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65839, "pid": 0, "tid": 7, "ts": 1776649791941054, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941047, "dur": 5, + "args": { + "External id": 65839, + "cbid": 211, "correlation": 65839 + } + }, + { + "ph": "s", "id": 65839, "pid": 435, "tid": 435, "ts": 1776649791941047, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791941090, "dur": 1, + "args": { + "External id": 8630, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65857, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65857, "pid": 0, "tid": 7, "ts": 1776649791941090, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941081, "dur": 6, + "args": { + "External id": 8630, + "cbid": 211, "correlation": 65857 + } + }, + { + "ph": "s", "id": 65857, "pid": 435, "tid": 435, "ts": 1776649791941081, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941093, "dur": 0, + "args": { + "External id": 65863, + "cbid": 200, "correlation": 65863 + } + }, + { + "ph": "f", "id": 65863, "pid": 435, "tid": 435, "ts": 1776649791941093, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941094, "dur": 0, + "args": { + "External id": 65864, + "cbid": 273, "correlation": 65864 + } + }, + { + "ph": "f", "id": 65864, "pid": 435, "tid": 435, "ts": 1776649791941094, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791941103, "dur": 19, + "args": { + "External id": 65865, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65865, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65865, "pid": 0, "tid": 7, "ts": 1776649791941103, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941095, "dur": 4, + "args": { + "External id": 65865, + "cbid": 211, "correlation": 65865 + } + }, + { + "ph": "s", "id": 65865, "pid": 435, "tid": 435, "ts": 1776649791941095, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791941138, "dur": 1, + "args": { + "External id": 8635, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65883, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65883, "pid": 0, "tid": 7, "ts": 1776649791941138, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941129, "dur": 6, + "args": { + "External id": 8635, + "cbid": 211, "correlation": 65883 + } + }, + { + "ph": "s", "id": 65883, "pid": 435, "tid": 435, "ts": 1776649791941129, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941141, "dur": 0, + "args": { + "External id": 65889, + "cbid": 200, "correlation": 65889 + } + }, + { + "ph": "f", "id": 65889, "pid": 435, "tid": 435, "ts": 1776649791941141, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941141, "dur": 0, + "args": { + "External id": 65890, + "cbid": 273, "correlation": 65890 + } + }, + { + "ph": "f", "id": 65890, "pid": 435, "tid": 435, "ts": 1776649791941141, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791941150, "dur": 19, + "args": { + "External id": 65891, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65891, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65891, "pid": 0, "tid": 7, "ts": 1776649791941150, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941142, "dur": 4, + "args": { + "External id": 65891, + "cbid": 211, "correlation": 65891 + } + }, + { + "ph": "s", "id": 65891, "pid": 435, "tid": 435, "ts": 1776649791941142, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791941184, "dur": 1, + "args": { + "External id": 8640, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65909, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65909, "pid": 0, "tid": 7, "ts": 1776649791941184, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941175, "dur": 6, + "args": { + "External id": 8640, + "cbid": 211, "correlation": 65909 + } + }, + { + "ph": "s", "id": 65909, "pid": 435, "tid": 435, "ts": 1776649791941175, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941188, "dur": 0, + "args": { + "External id": 65915, + "cbid": 200, "correlation": 65915 + } + }, + { + "ph": "f", "id": 65915, "pid": 435, "tid": 435, "ts": 1776649791941188, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941189, "dur": 0, + "args": { + "External id": 65916, + "cbid": 273, "correlation": 65916 + } + }, + { + "ph": "f", "id": 65916, "pid": 435, "tid": 435, "ts": 1776649791941189, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791941199, "dur": 19, + "args": { + "External id": 65917, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65917, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65917, "pid": 0, "tid": 7, "ts": 1776649791941199, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941190, "dur": 5, + "args": { + "External id": 65917, + "cbid": 211, "correlation": 65917 + } + }, + { + "ph": "s", "id": 65917, "pid": 435, "tid": 435, "ts": 1776649791941190, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791941236, "dur": 1, + "args": { + "External id": 8645, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65935, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65935, "pid": 0, "tid": 7, "ts": 1776649791941236, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941227, "dur": 6, + "args": { + "External id": 8645, + "cbid": 211, "correlation": 65935 + } + }, + { + "ph": "s", "id": 65935, "pid": 435, "tid": 435, "ts": 1776649791941227, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941239, "dur": 0, + "args": { + "External id": 65941, + "cbid": 200, "correlation": 65941 + } + }, + { + "ph": "f", "id": 65941, "pid": 435, "tid": 435, "ts": 1776649791941239, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941240, "dur": 0, + "args": { + "External id": 65942, + "cbid": 273, "correlation": 65942 + } + }, + { + "ph": "f", "id": 65942, "pid": 435, "tid": 435, "ts": 1776649791941240, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791941249, "dur": 19, + "args": { + "External id": 65943, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65943, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65943, "pid": 0, "tid": 7, "ts": 1776649791941249, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941241, "dur": 4, + "args": { + "External id": 65943, + "cbid": 211, "correlation": 65943 + } + }, + { + "ph": "s", "id": 65943, "pid": 435, "tid": 435, "ts": 1776649791941241, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791941283, "dur": 1, + "args": { + "External id": 8650, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65961, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65961, "pid": 0, "tid": 7, "ts": 1776649791941283, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941274, "dur": 6, + "args": { + "External id": 8650, + "cbid": 211, "correlation": 65961 + } + }, + { + "ph": "s", "id": 65961, "pid": 435, "tid": 435, "ts": 1776649791941274, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941286, "dur": 0, + "args": { + "External id": 65967, + "cbid": 200, "correlation": 65967 + } + }, + { + "ph": "f", "id": 65967, "pid": 435, "tid": 435, "ts": 1776649791941286, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941287, "dur": 0, + "args": { + "External id": 65968, + "cbid": 273, "correlation": 65968 + } + }, + { + "ph": "f", "id": 65968, "pid": 435, "tid": 435, "ts": 1776649791941287, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791941296, "dur": 19, + "args": { + "External id": 65969, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65969, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65969, "pid": 0, "tid": 7, "ts": 1776649791941296, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941288, "dur": 4, + "args": { + "External id": 65969, + "cbid": 211, "correlation": 65969 + } + }, + { + "ph": "s", "id": 65969, "pid": 435, "tid": 435, "ts": 1776649791941288, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791941330, "dur": 1, + "args": { + "External id": 8655, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65987, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65987, "pid": 0, "tid": 7, "ts": 1776649791941330, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941321, "dur": 6, + "args": { + "External id": 8655, + "cbid": 211, "correlation": 65987 + } + }, + { + "ph": "s", "id": 65987, "pid": 435, "tid": 435, "ts": 1776649791941321, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941333, "dur": 0, + "args": { + "External id": 65993, + "cbid": 200, "correlation": 65993 + } + }, + { + "ph": "f", "id": 65993, "pid": 435, "tid": 435, "ts": 1776649791941333, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941334, "dur": 0, + "args": { + "External id": 65994, + "cbid": 273, "correlation": 65994 + } + }, + { + "ph": "f", "id": 65994, "pid": 435, "tid": 435, "ts": 1776649791941334, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791941343, "dur": 19, + "args": { + "External id": 65995, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 65995, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 65995, "pid": 0, "tid": 7, "ts": 1776649791941343, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941335, "dur": 4, + "args": { + "External id": 65995, + "cbid": 211, "correlation": 65995 + } + }, + { + "ph": "s", "id": 65995, "pid": 435, "tid": 435, "ts": 1776649791941335, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791941357, "dur": 9, + "args": { + "External id": 65998, + "cbid": 165, "correlation": 65998 + } + }, + { + "ph": "s", "id": 65998, "pid": 435, "tid": 435, "ts": 1776649791941357, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791941387, "dur": 3, + "args": { + "External id": 66007, + "cbid": 135, "correlation": 66007 + } + }, + { + "ph": "f", "id": 66007, "pid": 435, "tid": 435, "ts": 1776649791941387, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791941434, "dur": 1, + "args": { + "External id": 8660, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66027, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66027, "pid": 0, "tid": 7, "ts": 1776649791941434, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941424, "dur": 8, + "args": { + "External id": 8660, + "cbid": 211, "correlation": 66027 + } + }, + { + "ph": "s", "id": 66027, "pid": 435, "tid": 435, "ts": 1776649791941424, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941438, "dur": 0, + "args": { + "External id": 66033, + "cbid": 200, "correlation": 66033 + } + }, + { + "ph": "f", "id": 66033, "pid": 435, "tid": 435, "ts": 1776649791941438, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941439, "dur": 0, + "args": { + "External id": 66034, + "cbid": 273, "correlation": 66034 + } + }, + { + "ph": "f", "id": 66034, "pid": 435, "tid": 435, "ts": 1776649791941439, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791941448, "dur": 19, + "args": { + "External id": 66035, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66035, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66035, "pid": 0, "tid": 7, "ts": 1776649791941448, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941440, "dur": 5, + "args": { + "External id": 66035, + "cbid": 211, "correlation": 66035 + } + }, + { + "ph": "s", "id": 66035, "pid": 435, "tid": 435, "ts": 1776649791941440, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791941485, "dur": 1, + "args": { + "External id": 8665, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66053, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66053, "pid": 0, "tid": 7, "ts": 1776649791941485, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941475, "dur": 6, + "args": { + "External id": 8665, + "cbid": 211, "correlation": 66053 + } + }, + { + "ph": "s", "id": 66053, "pid": 435, "tid": 435, "ts": 1776649791941475, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941488, "dur": 0, + "args": { + "External id": 66059, + "cbid": 200, "correlation": 66059 + } + }, + { + "ph": "f", "id": 66059, "pid": 435, "tid": 435, "ts": 1776649791941488, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941489, "dur": 0, + "args": { + "External id": 66060, + "cbid": 273, "correlation": 66060 + } + }, + { + "ph": "f", "id": 66060, "pid": 435, "tid": 435, "ts": 1776649791941489, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791941498, "dur": 19, + "args": { + "External id": 66061, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66061, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66061, "pid": 0, "tid": 7, "ts": 1776649791941498, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941490, "dur": 4, + "args": { + "External id": 66061, + "cbid": 211, "correlation": 66061 + } + }, + { + "ph": "s", "id": 66061, "pid": 435, "tid": 435, "ts": 1776649791941490, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791941533, "dur": 1, + "args": { + "External id": 8670, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66079, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66079, "pid": 0, "tid": 7, "ts": 1776649791941533, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941523, "dur": 6, + "args": { + "External id": 8670, + "cbid": 211, "correlation": 66079 + } + }, + { + "ph": "s", "id": 66079, "pid": 435, "tid": 435, "ts": 1776649791941523, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941536, "dur": 0, + "args": { + "External id": 66085, + "cbid": 200, "correlation": 66085 + } + }, + { + "ph": "f", "id": 66085, "pid": 435, "tid": 435, "ts": 1776649791941536, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941536, "dur": 0, + "args": { + "External id": 66086, + "cbid": 273, "correlation": 66086 + } + }, + { + "ph": "f", "id": 66086, "pid": 435, "tid": 435, "ts": 1776649791941536, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791941545, "dur": 19, + "args": { + "External id": 66087, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66087, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66087, "pid": 0, "tid": 7, "ts": 1776649791941545, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941537, "dur": 5, + "args": { + "External id": 66087, + "cbid": 211, "correlation": 66087 + } + }, + { + "ph": "s", "id": 66087, "pid": 435, "tid": 435, "ts": 1776649791941537, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791941580, "dur": 1, + "args": { + "External id": 8675, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66105, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66105, "pid": 0, "tid": 7, "ts": 1776649791941580, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941571, "dur": 6, + "args": { + "External id": 8675, + "cbid": 211, "correlation": 66105 + } + }, + { + "ph": "s", "id": 66105, "pid": 435, "tid": 435, "ts": 1776649791941571, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941583, "dur": 0, + "args": { + "External id": 66111, + "cbid": 200, "correlation": 66111 + } + }, + { + "ph": "f", "id": 66111, "pid": 435, "tid": 435, "ts": 1776649791941583, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941584, "dur": 0, + "args": { + "External id": 66112, + "cbid": 273, "correlation": 66112 + } + }, + { + "ph": "f", "id": 66112, "pid": 435, "tid": 435, "ts": 1776649791941584, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791941593, "dur": 19, + "args": { + "External id": 66113, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66113, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66113, "pid": 0, "tid": 7, "ts": 1776649791941593, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941585, "dur": 4, + "args": { + "External id": 66113, + "cbid": 211, "correlation": 66113 + } + }, + { + "ph": "s", "id": 66113, "pid": 435, "tid": 435, "ts": 1776649791941585, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791941628, "dur": 1, + "args": { + "External id": 8680, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66131, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66131, "pid": 0, "tid": 7, "ts": 1776649791941628, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941619, "dur": 6, + "args": { + "External id": 8680, + "cbid": 211, "correlation": 66131 + } + }, + { + "ph": "s", "id": 66131, "pid": 435, "tid": 435, "ts": 1776649791941619, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941631, "dur": 0, + "args": { + "External id": 66137, + "cbid": 200, "correlation": 66137 + } + }, + { + "ph": "f", "id": 66137, "pid": 435, "tid": 435, "ts": 1776649791941631, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941632, "dur": 0, + "args": { + "External id": 66138, + "cbid": 273, "correlation": 66138 + } + }, + { + "ph": "f", "id": 66138, "pid": 435, "tid": 435, "ts": 1776649791941632, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791941641, "dur": 19, + "args": { + "External id": 66139, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66139, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66139, "pid": 0, "tid": 7, "ts": 1776649791941641, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941633, "dur": 5, + "args": { + "External id": 66139, + "cbid": 211, "correlation": 66139 + } + }, + { + "ph": "s", "id": 66139, "pid": 435, "tid": 435, "ts": 1776649791941633, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791941676, "dur": 1, + "args": { + "External id": 8685, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66157, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66157, "pid": 0, "tid": 7, "ts": 1776649791941676, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941667, "dur": 6, + "args": { + "External id": 8685, + "cbid": 211, "correlation": 66157 + } + }, + { + "ph": "s", "id": 66157, "pid": 435, "tid": 435, "ts": 1776649791941667, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941679, "dur": 0, + "args": { + "External id": 66163, + "cbid": 200, "correlation": 66163 + } + }, + { + "ph": "f", "id": 66163, "pid": 435, "tid": 435, "ts": 1776649791941679, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941680, "dur": 0, + "args": { + "External id": 66164, + "cbid": 273, "correlation": 66164 + } + }, + { + "ph": "f", "id": 66164, "pid": 435, "tid": 435, "ts": 1776649791941680, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791941689, "dur": 19, + "args": { + "External id": 66165, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66165, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66165, "pid": 0, "tid": 7, "ts": 1776649791941689, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941681, "dur": 4, + "args": { + "External id": 66165, + "cbid": 211, "correlation": 66165 + } + }, + { + "ph": "s", "id": 66165, "pid": 435, "tid": 435, "ts": 1776649791941681, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791941723, "dur": 1, + "args": { + "External id": 8690, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66183, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66183, "pid": 0, "tid": 7, "ts": 1776649791941723, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941714, "dur": 6, + "args": { + "External id": 8690, + "cbid": 211, "correlation": 66183 + } + }, + { + "ph": "s", "id": 66183, "pid": 435, "tid": 435, "ts": 1776649791941714, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941726, "dur": 0, + "args": { + "External id": 66189, + "cbid": 200, "correlation": 66189 + } + }, + { + "ph": "f", "id": 66189, "pid": 435, "tid": 435, "ts": 1776649791941726, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941727, "dur": 0, + "args": { + "External id": 66190, + "cbid": 273, "correlation": 66190 + } + }, + { + "ph": "f", "id": 66190, "pid": 435, "tid": 435, "ts": 1776649791941727, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791941736, "dur": 19, + "args": { + "External id": 66191, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66191, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66191, "pid": 0, "tid": 7, "ts": 1776649791941736, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941728, "dur": 5, + "args": { + "External id": 66191, + "cbid": 211, "correlation": 66191 + } + }, + { + "ph": "s", "id": 66191, "pid": 435, "tid": 435, "ts": 1776649791941728, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791941771, "dur": 1, + "args": { + "External id": 8695, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66209, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66209, "pid": 0, "tid": 7, "ts": 1776649791941771, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941762, "dur": 6, + "args": { + "External id": 8695, + "cbid": 211, "correlation": 66209 + } + }, + { + "ph": "s", "id": 66209, "pid": 435, "tid": 435, "ts": 1776649791941762, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941774, "dur": 0, + "args": { + "External id": 66215, + "cbid": 200, "correlation": 66215 + } + }, + { + "ph": "f", "id": 66215, "pid": 435, "tid": 435, "ts": 1776649791941774, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941775, "dur": 0, + "args": { + "External id": 66216, + "cbid": 273, "correlation": 66216 + } + }, + { + "ph": "f", "id": 66216, "pid": 435, "tid": 435, "ts": 1776649791941775, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791941783, "dur": 19, + "args": { + "External id": 66217, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66217, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66217, "pid": 0, "tid": 7, "ts": 1776649791941783, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941776, "dur": 4, + "args": { + "External id": 66217, + "cbid": 211, "correlation": 66217 + } + }, + { + "ph": "s", "id": 66217, "pid": 435, "tid": 435, "ts": 1776649791941776, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791941819, "dur": 1, + "args": { + "External id": 8700, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66235, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66235, "pid": 0, "tid": 7, "ts": 1776649791941819, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941809, "dur": 6, + "args": { + "External id": 8700, + "cbid": 211, "correlation": 66235 + } + }, + { + "ph": "s", "id": 66235, "pid": 435, "tid": 435, "ts": 1776649791941809, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941822, "dur": 0, + "args": { + "External id": 66241, + "cbid": 200, "correlation": 66241 + } + }, + { + "ph": "f", "id": 66241, "pid": 435, "tid": 435, "ts": 1776649791941822, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941822, "dur": 0, + "args": { + "External id": 66242, + "cbid": 273, "correlation": 66242 + } + }, + { + "ph": "f", "id": 66242, "pid": 435, "tid": 435, "ts": 1776649791941822, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791941831, "dur": 19, + "args": { + "External id": 66243, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66243, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66243, "pid": 0, "tid": 7, "ts": 1776649791941831, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941823, "dur": 4, + "args": { + "External id": 66243, + "cbid": 211, "correlation": 66243 + } + }, + { + "ph": "s", "id": 66243, "pid": 435, "tid": 435, "ts": 1776649791941823, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791941876, "dur": 1, + "args": { + "External id": 8705, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66261, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66261, "pid": 0, "tid": 7, "ts": 1776649791941876, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941867, "dur": 6, + "args": { + "External id": 8705, + "cbid": 211, "correlation": 66261 + } + }, + { + "ph": "s", "id": 66261, "pid": 435, "tid": 435, "ts": 1776649791941867, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941879, "dur": 0, + "args": { + "External id": 66267, + "cbid": 200, "correlation": 66267 + } + }, + { + "ph": "f", "id": 66267, "pid": 435, "tid": 435, "ts": 1776649791941879, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941880, "dur": 0, + "args": { + "External id": 66268, + "cbid": 273, "correlation": 66268 + } + }, + { + "ph": "f", "id": 66268, "pid": 435, "tid": 435, "ts": 1776649791941880, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791941889, "dur": 19, + "args": { + "External id": 66269, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66269, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66269, "pid": 0, "tid": 7, "ts": 1776649791941889, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941881, "dur": 5, + "args": { + "External id": 66269, + "cbid": 211, "correlation": 66269 + } + }, + { + "ph": "s", "id": 66269, "pid": 435, "tid": 435, "ts": 1776649791941881, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791941925, "dur": 1, + "args": { + "External id": 8710, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66287, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66287, "pid": 0, "tid": 7, "ts": 1776649791941925, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941916, "dur": 6, + "args": { + "External id": 8710, + "cbid": 211, "correlation": 66287 + } + }, + { + "ph": "s", "id": 66287, "pid": 435, "tid": 435, "ts": 1776649791941916, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941929, "dur": 0, + "args": { + "External id": 66293, + "cbid": 200, "correlation": 66293 + } + }, + { + "ph": "f", "id": 66293, "pid": 435, "tid": 435, "ts": 1776649791941929, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941929, "dur": 0, + "args": { + "External id": 66294, + "cbid": 273, "correlation": 66294 + } + }, + { + "ph": "f", "id": 66294, "pid": 435, "tid": 435, "ts": 1776649791941929, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791941938, "dur": 19, + "args": { + "External id": 66295, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66295, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66295, "pid": 0, "tid": 7, "ts": 1776649791941938, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941930, "dur": 5, + "args": { + "External id": 66295, + "cbid": 211, "correlation": 66295 + } + }, + { + "ph": "s", "id": 66295, "pid": 435, "tid": 435, "ts": 1776649791941930, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791941974, "dur": 1, + "args": { + "External id": 8715, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66313, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66313, "pid": 0, "tid": 7, "ts": 1776649791941974, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941964, "dur": 6, + "args": { + "External id": 8715, + "cbid": 211, "correlation": 66313 + } + }, + { + "ph": "s", "id": 66313, "pid": 435, "tid": 435, "ts": 1776649791941964, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941977, "dur": 0, + "args": { + "External id": 66319, + "cbid": 200, "correlation": 66319 + } + }, + { + "ph": "f", "id": 66319, "pid": 435, "tid": 435, "ts": 1776649791941977, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791941977, "dur": 0, + "args": { + "External id": 66320, + "cbid": 273, "correlation": 66320 + } + }, + { + "ph": "f", "id": 66320, "pid": 435, "tid": 435, "ts": 1776649791941977, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791941987, "dur": 19, + "args": { + "External id": 66321, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66321, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66321, "pid": 0, "tid": 7, "ts": 1776649791941987, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791941979, "dur": 4, + "args": { + "External id": 66321, + "cbid": 211, "correlation": 66321 + } + }, + { + "ph": "s", "id": 66321, "pid": 435, "tid": 435, "ts": 1776649791941979, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791942022, "dur": 1, + "args": { + "External id": 8720, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66339, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66339, "pid": 0, "tid": 7, "ts": 1776649791942022, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942013, "dur": 7, + "args": { + "External id": 8720, + "cbid": 211, "correlation": 66339 + } + }, + { + "ph": "s", "id": 66339, "pid": 435, "tid": 435, "ts": 1776649791942013, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942026, "dur": 0, + "args": { + "External id": 66345, + "cbid": 200, "correlation": 66345 + } + }, + { + "ph": "f", "id": 66345, "pid": 435, "tid": 435, "ts": 1776649791942026, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942026, "dur": 0, + "args": { + "External id": 66346, + "cbid": 273, "correlation": 66346 + } + }, + { + "ph": "f", "id": 66346, "pid": 435, "tid": 435, "ts": 1776649791942026, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791942035, "dur": 19, + "args": { + "External id": 66347, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66347, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66347, "pid": 0, "tid": 7, "ts": 1776649791942035, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942027, "dur": 5, + "args": { + "External id": 66347, + "cbid": 211, "correlation": 66347 + } + }, + { + "ph": "s", "id": 66347, "pid": 435, "tid": 435, "ts": 1776649791942027, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791942070, "dur": 1, + "args": { + "External id": 8725, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66365, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66365, "pid": 0, "tid": 7, "ts": 1776649791942070, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942061, "dur": 6, + "args": { + "External id": 8725, + "cbid": 211, "correlation": 66365 + } + }, + { + "ph": "s", "id": 66365, "pid": 435, "tid": 435, "ts": 1776649791942061, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942073, "dur": 0, + "args": { + "External id": 66371, + "cbid": 200, "correlation": 66371 + } + }, + { + "ph": "f", "id": 66371, "pid": 435, "tid": 435, "ts": 1776649791942073, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942074, "dur": 0, + "args": { + "External id": 66372, + "cbid": 273, "correlation": 66372 + } + }, + { + "ph": "f", "id": 66372, "pid": 435, "tid": 435, "ts": 1776649791942074, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791942083, "dur": 19, + "args": { + "External id": 66373, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66373, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66373, "pid": 0, "tid": 7, "ts": 1776649791942083, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942075, "dur": 4, + "args": { + "External id": 66373, + "cbid": 211, "correlation": 66373 + } + }, + { + "ph": "s", "id": 66373, "pid": 435, "tid": 435, "ts": 1776649791942075, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791942119, "dur": 1, + "args": { + "External id": 8730, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66391, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66391, "pid": 0, "tid": 7, "ts": 1776649791942119, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942109, "dur": 6, + "args": { + "External id": 8730, + "cbid": 211, "correlation": 66391 + } + }, + { + "ph": "s", "id": 66391, "pid": 435, "tid": 435, "ts": 1776649791942109, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942121, "dur": 0, + "args": { + "External id": 66397, + "cbid": 200, "correlation": 66397 + } + }, + { + "ph": "f", "id": 66397, "pid": 435, "tid": 435, "ts": 1776649791942121, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942122, "dur": 0, + "args": { + "External id": 66398, + "cbid": 273, "correlation": 66398 + } + }, + { + "ph": "f", "id": 66398, "pid": 435, "tid": 435, "ts": 1776649791942122, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791942131, "dur": 19, + "args": { + "External id": 66399, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66399, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66399, "pid": 0, "tid": 7, "ts": 1776649791942131, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942123, "dur": 4, + "args": { + "External id": 66399, + "cbid": 211, "correlation": 66399 + } + }, + { + "ph": "s", "id": 66399, "pid": 435, "tid": 435, "ts": 1776649791942123, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791942166, "dur": 1, + "args": { + "External id": 8735, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66417, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66417, "pid": 0, "tid": 7, "ts": 1776649791942166, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942157, "dur": 6, + "args": { + "External id": 8735, + "cbid": 211, "correlation": 66417 + } + }, + { + "ph": "s", "id": 66417, "pid": 435, "tid": 435, "ts": 1776649791942157, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942170, "dur": 0, + "args": { + "External id": 66423, + "cbid": 200, "correlation": 66423 + } + }, + { + "ph": "f", "id": 66423, "pid": 435, "tid": 435, "ts": 1776649791942170, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942170, "dur": 0, + "args": { + "External id": 66424, + "cbid": 273, "correlation": 66424 + } + }, + { + "ph": "f", "id": 66424, "pid": 435, "tid": 435, "ts": 1776649791942170, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791942179, "dur": 19, + "args": { + "External id": 66425, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66425, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66425, "pid": 0, "tid": 7, "ts": 1776649791942179, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942171, "dur": 4, + "args": { + "External id": 66425, + "cbid": 211, "correlation": 66425 + } + }, + { + "ph": "s", "id": 66425, "pid": 435, "tid": 435, "ts": 1776649791942171, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791942215, "dur": 1, + "args": { + "External id": 8740, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66443, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66443, "pid": 0, "tid": 7, "ts": 1776649791942215, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942205, "dur": 6, + "args": { + "External id": 8740, + "cbid": 211, "correlation": 66443 + } + }, + { + "ph": "s", "id": 66443, "pid": 435, "tid": 435, "ts": 1776649791942205, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942218, "dur": 0, + "args": { + "External id": 66449, + "cbid": 200, "correlation": 66449 + } + }, + { + "ph": "f", "id": 66449, "pid": 435, "tid": 435, "ts": 1776649791942218, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942218, "dur": 0, + "args": { + "External id": 66450, + "cbid": 273, "correlation": 66450 + } + }, + { + "ph": "f", "id": 66450, "pid": 435, "tid": 435, "ts": 1776649791942218, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791942227, "dur": 19, + "args": { + "External id": 66451, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66451, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66451, "pid": 0, "tid": 7, "ts": 1776649791942227, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942220, "dur": 4, + "args": { + "External id": 66451, + "cbid": 211, "correlation": 66451 + } + }, + { + "ph": "s", "id": 66451, "pid": 435, "tid": 435, "ts": 1776649791942220, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791942263, "dur": 1, + "args": { + "External id": 8745, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66469, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66469, "pid": 0, "tid": 7, "ts": 1776649791942263, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942253, "dur": 7, + "args": { + "External id": 8745, + "cbid": 211, "correlation": 66469 + } + }, + { + "ph": "s", "id": 66469, "pid": 435, "tid": 435, "ts": 1776649791942253, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942266, "dur": 0, + "args": { + "External id": 66475, + "cbid": 200, "correlation": 66475 + } + }, + { + "ph": "f", "id": 66475, "pid": 435, "tid": 435, "ts": 1776649791942266, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942266, "dur": 0, + "args": { + "External id": 66476, + "cbid": 273, "correlation": 66476 + } + }, + { + "ph": "f", "id": 66476, "pid": 435, "tid": 435, "ts": 1776649791942266, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791942275, "dur": 19, + "args": { + "External id": 66477, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66477, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66477, "pid": 0, "tid": 7, "ts": 1776649791942275, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942267, "dur": 5, + "args": { + "External id": 66477, + "cbid": 211, "correlation": 66477 + } + }, + { + "ph": "s", "id": 66477, "pid": 435, "tid": 435, "ts": 1776649791942267, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791942311, "dur": 1, + "args": { + "External id": 8750, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66495, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66495, "pid": 0, "tid": 7, "ts": 1776649791942311, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942301, "dur": 6, + "args": { + "External id": 8750, + "cbid": 211, "correlation": 66495 + } + }, + { + "ph": "s", "id": 66495, "pid": 435, "tid": 435, "ts": 1776649791942301, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942314, "dur": 0, + "args": { + "External id": 66501, + "cbid": 200, "correlation": 66501 + } + }, + { + "ph": "f", "id": 66501, "pid": 435, "tid": 435, "ts": 1776649791942314, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942314, "dur": 0, + "args": { + "External id": 66502, + "cbid": 273, "correlation": 66502 + } + }, + { + "ph": "f", "id": 66502, "pid": 435, "tid": 435, "ts": 1776649791942314, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791942324, "dur": 19, + "args": { + "External id": 66503, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66503, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66503, "pid": 0, "tid": 7, "ts": 1776649791942324, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942316, "dur": 5, + "args": { + "External id": 66503, + "cbid": 211, "correlation": 66503 + } + }, + { + "ph": "s", "id": 66503, "pid": 435, "tid": 435, "ts": 1776649791942316, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791942359, "dur": 1, + "args": { + "External id": 8755, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66521, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66521, "pid": 0, "tid": 7, "ts": 1776649791942359, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942350, "dur": 6, + "args": { + "External id": 8755, + "cbid": 211, "correlation": 66521 + } + }, + { + "ph": "s", "id": 66521, "pid": 435, "tid": 435, "ts": 1776649791942350, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942362, "dur": 0, + "args": { + "External id": 66527, + "cbid": 200, "correlation": 66527 + } + }, + { + "ph": "f", "id": 66527, "pid": 435, "tid": 435, "ts": 1776649791942362, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942362, "dur": 0, + "args": { + "External id": 66528, + "cbid": 273, "correlation": 66528 + } + }, + { + "ph": "f", "id": 66528, "pid": 435, "tid": 435, "ts": 1776649791942362, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791942371, "dur": 19, + "args": { + "External id": 66529, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66529, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66529, "pid": 0, "tid": 7, "ts": 1776649791942371, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942364, "dur": 4, + "args": { + "External id": 66529, + "cbid": 211, "correlation": 66529 + } + }, + { + "ph": "s", "id": 66529, "pid": 435, "tid": 435, "ts": 1776649791942364, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791942407, "dur": 1, + "args": { + "External id": 8760, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66547, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66547, "pid": 0, "tid": 7, "ts": 1776649791942407, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942397, "dur": 6, + "args": { + "External id": 8760, + "cbid": 211, "correlation": 66547 + } + }, + { + "ph": "s", "id": 66547, "pid": 435, "tid": 435, "ts": 1776649791942397, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942410, "dur": 0, + "args": { + "External id": 66553, + "cbid": 200, "correlation": 66553 + } + }, + { + "ph": "f", "id": 66553, "pid": 435, "tid": 435, "ts": 1776649791942410, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942410, "dur": 0, + "args": { + "External id": 66554, + "cbid": 273, "correlation": 66554 + } + }, + { + "ph": "f", "id": 66554, "pid": 435, "tid": 435, "ts": 1776649791942410, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791942419, "dur": 19, + "args": { + "External id": 66555, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66555, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66555, "pid": 0, "tid": 7, "ts": 1776649791942419, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942411, "dur": 4, + "args": { + "External id": 66555, + "cbid": 211, "correlation": 66555 + } + }, + { + "ph": "s", "id": 66555, "pid": 435, "tid": 435, "ts": 1776649791942411, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791942455, "dur": 1, + "args": { + "External id": 8765, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66573, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66573, "pid": 0, "tid": 7, "ts": 1776649791942455, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942445, "dur": 6, + "args": { + "External id": 8765, + "cbid": 211, "correlation": 66573 + } + }, + { + "ph": "s", "id": 66573, "pid": 435, "tid": 435, "ts": 1776649791942445, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942458, "dur": 0, + "args": { + "External id": 66579, + "cbid": 200, "correlation": 66579 + } + }, + { + "ph": "f", "id": 66579, "pid": 435, "tid": 435, "ts": 1776649791942458, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942458, "dur": 0, + "args": { + "External id": 66580, + "cbid": 273, "correlation": 66580 + } + }, + { + "ph": "f", "id": 66580, "pid": 435, "tid": 435, "ts": 1776649791942458, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791942467, "dur": 19, + "args": { + "External id": 66581, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66581, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66581, "pid": 0, "tid": 7, "ts": 1776649791942467, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942459, "dur": 4, + "args": { + "External id": 66581, + "cbid": 211, "correlation": 66581 + } + }, + { + "ph": "s", "id": 66581, "pid": 435, "tid": 435, "ts": 1776649791942459, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791942503, "dur": 1, + "args": { + "External id": 8770, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66599, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66599, "pid": 0, "tid": 7, "ts": 1776649791942503, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942494, "dur": 6, + "args": { + "External id": 8770, + "cbid": 211, "correlation": 66599 + } + }, + { + "ph": "s", "id": 66599, "pid": 435, "tid": 435, "ts": 1776649791942494, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942506, "dur": 0, + "args": { + "External id": 66605, + "cbid": 200, "correlation": 66605 + } + }, + { + "ph": "f", "id": 66605, "pid": 435, "tid": 435, "ts": 1776649791942506, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942507, "dur": 0, + "args": { + "External id": 66606, + "cbid": 273, "correlation": 66606 + } + }, + { + "ph": "f", "id": 66606, "pid": 435, "tid": 435, "ts": 1776649791942507, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791942516, "dur": 19, + "args": { + "External id": 66607, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66607, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66607, "pid": 0, "tid": 7, "ts": 1776649791942516, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942508, "dur": 5, + "args": { + "External id": 66607, + "cbid": 211, "correlation": 66607 + } + }, + { + "ph": "s", "id": 66607, "pid": 435, "tid": 435, "ts": 1776649791942508, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791942551, "dur": 1, + "args": { + "External id": 8775, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66625, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66625, "pid": 0, "tid": 7, "ts": 1776649791942551, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942542, "dur": 6, + "args": { + "External id": 8775, + "cbid": 211, "correlation": 66625 + } + }, + { + "ph": "s", "id": 66625, "pid": 435, "tid": 435, "ts": 1776649791942542, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942554, "dur": 0, + "args": { + "External id": 66631, + "cbid": 200, "correlation": 66631 + } + }, + { + "ph": "f", "id": 66631, "pid": 435, "tid": 435, "ts": 1776649791942554, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942555, "dur": 0, + "args": { + "External id": 66632, + "cbid": 273, "correlation": 66632 + } + }, + { + "ph": "f", "id": 66632, "pid": 435, "tid": 435, "ts": 1776649791942555, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791942564, "dur": 19, + "args": { + "External id": 66633, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66633, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66633, "pid": 0, "tid": 7, "ts": 1776649791942564, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942556, "dur": 4, + "args": { + "External id": 66633, + "cbid": 211, "correlation": 66633 + } + }, + { + "ph": "s", "id": 66633, "pid": 435, "tid": 435, "ts": 1776649791942556, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791942599, "dur": 1, + "args": { + "External id": 8780, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66651, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66651, "pid": 0, "tid": 7, "ts": 1776649791942599, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942590, "dur": 6, + "args": { + "External id": 8780, + "cbid": 211, "correlation": 66651 + } + }, + { + "ph": "s", "id": 66651, "pid": 435, "tid": 435, "ts": 1776649791942590, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942602, "dur": 0, + "args": { + "External id": 66657, + "cbid": 200, "correlation": 66657 + } + }, + { + "ph": "f", "id": 66657, "pid": 435, "tid": 435, "ts": 1776649791942602, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942602, "dur": 0, + "args": { + "External id": 66658, + "cbid": 273, "correlation": 66658 + } + }, + { + "ph": "f", "id": 66658, "pid": 435, "tid": 435, "ts": 1776649791942602, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791942611, "dur": 19, + "args": { + "External id": 66659, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66659, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66659, "pid": 0, "tid": 7, "ts": 1776649791942611, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942604, "dur": 4, + "args": { + "External id": 66659, + "cbid": 211, "correlation": 66659 + } + }, + { + "ph": "s", "id": 66659, "pid": 435, "tid": 435, "ts": 1776649791942604, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791942647, "dur": 1, + "args": { + "External id": 8785, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66677, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66677, "pid": 0, "tid": 7, "ts": 1776649791942647, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942638, "dur": 6, + "args": { + "External id": 8785, + "cbid": 211, "correlation": 66677 + } + }, + { + "ph": "s", "id": 66677, "pid": 435, "tid": 435, "ts": 1776649791942638, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942650, "dur": 0, + "args": { + "External id": 66683, + "cbid": 200, "correlation": 66683 + } + }, + { + "ph": "f", "id": 66683, "pid": 435, "tid": 435, "ts": 1776649791942650, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942650, "dur": 0, + "args": { + "External id": 66684, + "cbid": 273, "correlation": 66684 + } + }, + { + "ph": "f", "id": 66684, "pid": 435, "tid": 435, "ts": 1776649791942650, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791942660, "dur": 19, + "args": { + "External id": 66685, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66685, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66685, "pid": 0, "tid": 7, "ts": 1776649791942660, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942652, "dur": 5, + "args": { + "External id": 66685, + "cbid": 211, "correlation": 66685 + } + }, + { + "ph": "s", "id": 66685, "pid": 435, "tid": 435, "ts": 1776649791942652, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791942695, "dur": 1, + "args": { + "External id": 8790, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66703, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66703, "pid": 0, "tid": 7, "ts": 1776649791942695, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942686, "dur": 6, + "args": { + "External id": 8790, + "cbid": 211, "correlation": 66703 + } + }, + { + "ph": "s", "id": 66703, "pid": 435, "tid": 435, "ts": 1776649791942686, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942698, "dur": 0, + "args": { + "External id": 66709, + "cbid": 200, "correlation": 66709 + } + }, + { + "ph": "f", "id": 66709, "pid": 435, "tid": 435, "ts": 1776649791942698, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942699, "dur": 0, + "args": { + "External id": 66710, + "cbid": 273, "correlation": 66710 + } + }, + { + "ph": "f", "id": 66710, "pid": 435, "tid": 435, "ts": 1776649791942699, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791942708, "dur": 19, + "args": { + "External id": 66711, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66711, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66711, "pid": 0, "tid": 7, "ts": 1776649791942708, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942700, "dur": 4, + "args": { + "External id": 66711, + "cbid": 211, "correlation": 66711 + } + }, + { + "ph": "s", "id": 66711, "pid": 435, "tid": 435, "ts": 1776649791942700, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791942743, "dur": 1, + "args": { + "External id": 8795, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66729, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66729, "pid": 0, "tid": 7, "ts": 1776649791942743, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942734, "dur": 6, + "args": { + "External id": 8795, + "cbid": 211, "correlation": 66729 + } + }, + { + "ph": "s", "id": 66729, "pid": 435, "tid": 435, "ts": 1776649791942734, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942746, "dur": 0, + "args": { + "External id": 66735, + "cbid": 200, "correlation": 66735 + } + }, + { + "ph": "f", "id": 66735, "pid": 435, "tid": 435, "ts": 1776649791942746, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942747, "dur": 0, + "args": { + "External id": 66736, + "cbid": 273, "correlation": 66736 + } + }, + { + "ph": "f", "id": 66736, "pid": 435, "tid": 435, "ts": 1776649791942747, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791942755, "dur": 19, + "args": { + "External id": 66737, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66737, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66737, "pid": 0, "tid": 7, "ts": 1776649791942755, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942748, "dur": 4, + "args": { + "External id": 66737, + "cbid": 211, "correlation": 66737 + } + }, + { + "ph": "s", "id": 66737, "pid": 435, "tid": 435, "ts": 1776649791942748, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791942790, "dur": 1, + "args": { + "External id": 8800, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66755, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66755, "pid": 0, "tid": 7, "ts": 1776649791942790, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942781, "dur": 6, + "args": { + "External id": 8800, + "cbid": 211, "correlation": 66755 + } + }, + { + "ph": "s", "id": 66755, "pid": 435, "tid": 435, "ts": 1776649791942781, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942793, "dur": 0, + "args": { + "External id": 66761, + "cbid": 200, "correlation": 66761 + } + }, + { + "ph": "f", "id": 66761, "pid": 435, "tid": 435, "ts": 1776649791942793, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942794, "dur": 0, + "args": { + "External id": 66762, + "cbid": 273, "correlation": 66762 + } + }, + { + "ph": "f", "id": 66762, "pid": 435, "tid": 435, "ts": 1776649791942794, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791942803, "dur": 19, + "args": { + "External id": 66763, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66763, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66763, "pid": 0, "tid": 7, "ts": 1776649791942803, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942795, "dur": 4, + "args": { + "External id": 66763, + "cbid": 211, "correlation": 66763 + } + }, + { + "ph": "s", "id": 66763, "pid": 435, "tid": 435, "ts": 1776649791942795, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791942838, "dur": 1, + "args": { + "External id": 8805, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66781, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66781, "pid": 0, "tid": 7, "ts": 1776649791942838, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942829, "dur": 7, + "args": { + "External id": 8805, + "cbid": 211, "correlation": 66781 + } + }, + { + "ph": "s", "id": 66781, "pid": 435, "tid": 435, "ts": 1776649791942829, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942842, "dur": 0, + "args": { + "External id": 66787, + "cbid": 200, "correlation": 66787 + } + }, + { + "ph": "f", "id": 66787, "pid": 435, "tid": 435, "ts": 1776649791942842, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942842, "dur": 0, + "args": { + "External id": 66788, + "cbid": 273, "correlation": 66788 + } + }, + { + "ph": "f", "id": 66788, "pid": 435, "tid": 435, "ts": 1776649791942842, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791942851, "dur": 19, + "args": { + "External id": 66789, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66789, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66789, "pid": 0, "tid": 7, "ts": 1776649791942851, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942843, "dur": 5, + "args": { + "External id": 66789, + "cbid": 211, "correlation": 66789 + } + }, + { + "ph": "s", "id": 66789, "pid": 435, "tid": 435, "ts": 1776649791942843, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791942887, "dur": 1, + "args": { + "External id": 8810, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66807, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66807, "pid": 0, "tid": 7, "ts": 1776649791942887, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942877, "dur": 6, + "args": { + "External id": 8810, + "cbid": 211, "correlation": 66807 + } + }, + { + "ph": "s", "id": 66807, "pid": 435, "tid": 435, "ts": 1776649791942877, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942890, "dur": 0, + "args": { + "External id": 66813, + "cbid": 200, "correlation": 66813 + } + }, + { + "ph": "f", "id": 66813, "pid": 435, "tid": 435, "ts": 1776649791942890, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942890, "dur": 0, + "args": { + "External id": 66814, + "cbid": 273, "correlation": 66814 + } + }, + { + "ph": "f", "id": 66814, "pid": 435, "tid": 435, "ts": 1776649791942890, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791942899, "dur": 19, + "args": { + "External id": 66815, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66815, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66815, "pid": 0, "tid": 7, "ts": 1776649791942899, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942891, "dur": 5, + "args": { + "External id": 66815, + "cbid": 211, "correlation": 66815 + } + }, + { + "ph": "s", "id": 66815, "pid": 435, "tid": 435, "ts": 1776649791942891, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791942935, "dur": 1, + "args": { + "External id": 8815, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66833, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66833, "pid": 0, "tid": 7, "ts": 1776649791942935, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942925, "dur": 6, + "args": { + "External id": 8815, + "cbid": 211, "correlation": 66833 + } + }, + { + "ph": "s", "id": 66833, "pid": 435, "tid": 435, "ts": 1776649791942925, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942938, "dur": 0, + "args": { + "External id": 66839, + "cbid": 200, "correlation": 66839 + } + }, + { + "ph": "f", "id": 66839, "pid": 435, "tid": 435, "ts": 1776649791942938, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942939, "dur": 0, + "args": { + "External id": 66840, + "cbid": 273, "correlation": 66840 + } + }, + { + "ph": "f", "id": 66840, "pid": 435, "tid": 435, "ts": 1776649791942939, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791942948, "dur": 19, + "args": { + "External id": 66841, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66841, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66841, "pid": 0, "tid": 7, "ts": 1776649791942948, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942940, "dur": 5, + "args": { + "External id": 66841, + "cbid": 211, "correlation": 66841 + } + }, + { + "ph": "s", "id": 66841, "pid": 435, "tid": 435, "ts": 1776649791942940, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791942983, "dur": 1, + "args": { + "External id": 8820, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66859, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66859, "pid": 0, "tid": 7, "ts": 1776649791942983, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942974, "dur": 6, + "args": { + "External id": 8820, + "cbid": 211, "correlation": 66859 + } + }, + { + "ph": "s", "id": 66859, "pid": 435, "tid": 435, "ts": 1776649791942974, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942986, "dur": 0, + "args": { + "External id": 66865, + "cbid": 200, "correlation": 66865 + } + }, + { + "ph": "f", "id": 66865, "pid": 435, "tid": 435, "ts": 1776649791942986, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791942987, "dur": 0, + "args": { + "External id": 66866, + "cbid": 273, "correlation": 66866 + } + }, + { + "ph": "f", "id": 66866, "pid": 435, "tid": 435, "ts": 1776649791942987, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791942996, "dur": 19, + "args": { + "External id": 66867, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66867, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66867, "pid": 0, "tid": 7, "ts": 1776649791942996, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791942988, "dur": 4, + "args": { + "External id": 66867, + "cbid": 211, "correlation": 66867 + } + }, + { + "ph": "s", "id": 66867, "pid": 435, "tid": 435, "ts": 1776649791942988, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791943031, "dur": 1, + "args": { + "External id": 8825, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66885, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66885, "pid": 0, "tid": 7, "ts": 1776649791943031, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943022, "dur": 6, + "args": { + "External id": 8825, + "cbid": 211, "correlation": 66885 + } + }, + { + "ph": "s", "id": 66885, "pid": 435, "tid": 435, "ts": 1776649791943022, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943034, "dur": 0, + "args": { + "External id": 66891, + "cbid": 200, "correlation": 66891 + } + }, + { + "ph": "f", "id": 66891, "pid": 435, "tid": 435, "ts": 1776649791943034, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943035, "dur": 0, + "args": { + "External id": 66892, + "cbid": 273, "correlation": 66892 + } + }, + { + "ph": "f", "id": 66892, "pid": 435, "tid": 435, "ts": 1776649791943035, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791943044, "dur": 19, + "args": { + "External id": 66893, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66893, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66893, "pid": 0, "tid": 7, "ts": 1776649791943044, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943036, "dur": 4, + "args": { + "External id": 66893, + "cbid": 211, "correlation": 66893 + } + }, + { + "ph": "s", "id": 66893, "pid": 435, "tid": 435, "ts": 1776649791943036, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791943079, "dur": 1, + "args": { + "External id": 8830, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66911, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66911, "pid": 0, "tid": 7, "ts": 1776649791943079, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943070, "dur": 6, + "args": { + "External id": 8830, + "cbid": 211, "correlation": 66911 + } + }, + { + "ph": "s", "id": 66911, "pid": 435, "tid": 435, "ts": 1776649791943070, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943082, "dur": 0, + "args": { + "External id": 66917, + "cbid": 200, "correlation": 66917 + } + }, + { + "ph": "f", "id": 66917, "pid": 435, "tid": 435, "ts": 1776649791943082, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943083, "dur": 0, + "args": { + "External id": 66918, + "cbid": 273, "correlation": 66918 + } + }, + { + "ph": "f", "id": 66918, "pid": 435, "tid": 435, "ts": 1776649791943083, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791943092, "dur": 19, + "args": { + "External id": 66919, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66919, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66919, "pid": 0, "tid": 7, "ts": 1776649791943092, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943084, "dur": 4, + "args": { + "External id": 66919, + "cbid": 211, "correlation": 66919 + } + }, + { + "ph": "s", "id": 66919, "pid": 435, "tid": 435, "ts": 1776649791943084, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791943126, "dur": 1, + "args": { + "External id": 8835, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66937, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66937, "pid": 0, "tid": 7, "ts": 1776649791943126, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943117, "dur": 6, + "args": { + "External id": 8835, + "cbid": 211, "correlation": 66937 + } + }, + { + "ph": "s", "id": 66937, "pid": 435, "tid": 435, "ts": 1776649791943117, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943130, "dur": 0, + "args": { + "External id": 66943, + "cbid": 200, "correlation": 66943 + } + }, + { + "ph": "f", "id": 66943, "pid": 435, "tid": 435, "ts": 1776649791943130, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943130, "dur": 0, + "args": { + "External id": 66944, + "cbid": 273, "correlation": 66944 + } + }, + { + "ph": "f", "id": 66944, "pid": 435, "tid": 435, "ts": 1776649791943130, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791943139, "dur": 19, + "args": { + "External id": 66945, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66945, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66945, "pid": 0, "tid": 7, "ts": 1776649791943139, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943131, "dur": 4, + "args": { + "External id": 66945, + "cbid": 211, "correlation": 66945 + } + }, + { + "ph": "s", "id": 66945, "pid": 435, "tid": 435, "ts": 1776649791943131, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791943174, "dur": 1, + "args": { + "External id": 8840, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66963, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66963, "pid": 0, "tid": 7, "ts": 1776649791943174, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943165, "dur": 6, + "args": { + "External id": 8840, + "cbid": 211, "correlation": 66963 + } + }, + { + "ph": "s", "id": 66963, "pid": 435, "tid": 435, "ts": 1776649791943165, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943177, "dur": 0, + "args": { + "External id": 66969, + "cbid": 200, "correlation": 66969 + } + }, + { + "ph": "f", "id": 66969, "pid": 435, "tid": 435, "ts": 1776649791943177, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943178, "dur": 0, + "args": { + "External id": 66970, + "cbid": 273, "correlation": 66970 + } + }, + { + "ph": "f", "id": 66970, "pid": 435, "tid": 435, "ts": 1776649791943178, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791943187, "dur": 19, + "args": { + "External id": 66971, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66971, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66971, "pid": 0, "tid": 7, "ts": 1776649791943187, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943179, "dur": 4, + "args": { + "External id": 66971, + "cbid": 211, "correlation": 66971 + } + }, + { + "ph": "s", "id": 66971, "pid": 435, "tid": 435, "ts": 1776649791943179, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791943222, "dur": 1, + "args": { + "External id": 8845, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66989, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66989, "pid": 0, "tid": 7, "ts": 1776649791943222, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943213, "dur": 6, + "args": { + "External id": 8845, + "cbid": 211, "correlation": 66989 + } + }, + { + "ph": "s", "id": 66989, "pid": 435, "tid": 435, "ts": 1776649791943213, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943226, "dur": 0, + "args": { + "External id": 66995, + "cbid": 200, "correlation": 66995 + } + }, + { + "ph": "f", "id": 66995, "pid": 435, "tid": 435, "ts": 1776649791943226, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943227, "dur": 0, + "args": { + "External id": 66996, + "cbid": 273, "correlation": 66996 + } + }, + { + "ph": "f", "id": 66996, "pid": 435, "tid": 435, "ts": 1776649791943227, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791943236, "dur": 19, + "args": { + "External id": 66997, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 66997, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 66997, "pid": 0, "tid": 7, "ts": 1776649791943236, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943228, "dur": 4, + "args": { + "External id": 66997, + "cbid": 211, "correlation": 66997 + } + }, + { + "ph": "s", "id": 66997, "pid": 435, "tid": 435, "ts": 1776649791943228, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791943272, "dur": 1, + "args": { + "External id": 8850, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67015, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67015, "pid": 0, "tid": 7, "ts": 1776649791943272, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943262, "dur": 6, + "args": { + "External id": 8850, + "cbid": 211, "correlation": 67015 + } + }, + { + "ph": "s", "id": 67015, "pid": 435, "tid": 435, "ts": 1776649791943262, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943274, "dur": 0, + "args": { + "External id": 67021, + "cbid": 200, "correlation": 67021 + } + }, + { + "ph": "f", "id": 67021, "pid": 435, "tid": 435, "ts": 1776649791943274, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943275, "dur": 0, + "args": { + "External id": 67022, + "cbid": 273, "correlation": 67022 + } + }, + { + "ph": "f", "id": 67022, "pid": 435, "tid": 435, "ts": 1776649791943275, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791943284, "dur": 19, + "args": { + "External id": 67023, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67023, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67023, "pid": 0, "tid": 7, "ts": 1776649791943284, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943276, "dur": 4, + "args": { + "External id": 67023, + "cbid": 211, "correlation": 67023 + } + }, + { + "ph": "s", "id": 67023, "pid": 435, "tid": 435, "ts": 1776649791943276, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791943319, "dur": 1, + "args": { + "External id": 8855, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67041, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67041, "pid": 0, "tid": 7, "ts": 1776649791943319, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943310, "dur": 6, + "args": { + "External id": 8855, + "cbid": 211, "correlation": 67041 + } + }, + { + "ph": "s", "id": 67041, "pid": 435, "tid": 435, "ts": 1776649791943310, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943322, "dur": 0, + "args": { + "External id": 67047, + "cbid": 200, "correlation": 67047 + } + }, + { + "ph": "f", "id": 67047, "pid": 435, "tid": 435, "ts": 1776649791943322, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943323, "dur": 0, + "args": { + "External id": 67048, + "cbid": 273, "correlation": 67048 + } + }, + { + "ph": "f", "id": 67048, "pid": 435, "tid": 435, "ts": 1776649791943323, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791943332, "dur": 19, + "args": { + "External id": 67049, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67049, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67049, "pid": 0, "tid": 7, "ts": 1776649791943332, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943324, "dur": 4, + "args": { + "External id": 67049, + "cbid": 211, "correlation": 67049 + } + }, + { + "ph": "s", "id": 67049, "pid": 435, "tid": 435, "ts": 1776649791943324, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791943368, "dur": 1, + "args": { + "External id": 8860, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67067, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67067, "pid": 0, "tid": 7, "ts": 1776649791943368, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943358, "dur": 6, + "args": { + "External id": 8860, + "cbid": 211, "correlation": 67067 + } + }, + { + "ph": "s", "id": 67067, "pid": 435, "tid": 435, "ts": 1776649791943358, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943370, "dur": 0, + "args": { + "External id": 67073, + "cbid": 200, "correlation": 67073 + } + }, + { + "ph": "f", "id": 67073, "pid": 435, "tid": 435, "ts": 1776649791943370, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943371, "dur": 0, + "args": { + "External id": 67074, + "cbid": 273, "correlation": 67074 + } + }, + { + "ph": "f", "id": 67074, "pid": 435, "tid": 435, "ts": 1776649791943371, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791943380, "dur": 19, + "args": { + "External id": 67075, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67075, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67075, "pid": 0, "tid": 7, "ts": 1776649791943380, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943372, "dur": 4, + "args": { + "External id": 67075, + "cbid": 211, "correlation": 67075 + } + }, + { + "ph": "s", "id": 67075, "pid": 435, "tid": 435, "ts": 1776649791943372, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791943415, "dur": 1, + "args": { + "External id": 8865, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67093, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67093, "pid": 0, "tid": 7, "ts": 1776649791943415, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943406, "dur": 6, + "args": { + "External id": 8865, + "cbid": 211, "correlation": 67093 + } + }, + { + "ph": "s", "id": 67093, "pid": 435, "tid": 435, "ts": 1776649791943406, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943418, "dur": 0, + "args": { + "External id": 67099, + "cbid": 200, "correlation": 67099 + } + }, + { + "ph": "f", "id": 67099, "pid": 435, "tid": 435, "ts": 1776649791943418, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943419, "dur": 0, + "args": { + "External id": 67100, + "cbid": 273, "correlation": 67100 + } + }, + { + "ph": "f", "id": 67100, "pid": 435, "tid": 435, "ts": 1776649791943419, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791943428, "dur": 19, + "args": { + "External id": 67101, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67101, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67101, "pid": 0, "tid": 7, "ts": 1776649791943428, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943420, "dur": 5, + "args": { + "External id": 67101, + "cbid": 211, "correlation": 67101 + } + }, + { + "ph": "s", "id": 67101, "pid": 435, "tid": 435, "ts": 1776649791943420, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791943463, "dur": 1, + "args": { + "External id": 8870, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67119, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67119, "pid": 0, "tid": 7, "ts": 1776649791943463, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943454, "dur": 6, + "args": { + "External id": 8870, + "cbid": 211, "correlation": 67119 + } + }, + { + "ph": "s", "id": 67119, "pid": 435, "tid": 435, "ts": 1776649791943454, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943467, "dur": 0, + "args": { + "External id": 67125, + "cbid": 200, "correlation": 67125 + } + }, + { + "ph": "f", "id": 67125, "pid": 435, "tid": 435, "ts": 1776649791943467, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943467, "dur": 0, + "args": { + "External id": 67126, + "cbid": 273, "correlation": 67126 + } + }, + { + "ph": "f", "id": 67126, "pid": 435, "tid": 435, "ts": 1776649791943467, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791943476, "dur": 19, + "args": { + "External id": 67127, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67127, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67127, "pid": 0, "tid": 7, "ts": 1776649791943476, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943469, "dur": 5, + "args": { + "External id": 67127, + "cbid": 211, "correlation": 67127 + } + }, + { + "ph": "s", "id": 67127, "pid": 435, "tid": 435, "ts": 1776649791943469, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791943512, "dur": 1, + "args": { + "External id": 8875, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67145, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67145, "pid": 0, "tid": 7, "ts": 1776649791943512, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943503, "dur": 6, + "args": { + "External id": 8875, + "cbid": 211, "correlation": 67145 + } + }, + { + "ph": "s", "id": 67145, "pid": 435, "tid": 435, "ts": 1776649791943503, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943515, "dur": 0, + "args": { + "External id": 67151, + "cbid": 200, "correlation": 67151 + } + }, + { + "ph": "f", "id": 67151, "pid": 435, "tid": 435, "ts": 1776649791943515, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943515, "dur": 0, + "args": { + "External id": 67152, + "cbid": 273, "correlation": 67152 + } + }, + { + "ph": "f", "id": 67152, "pid": 435, "tid": 435, "ts": 1776649791943515, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791943524, "dur": 19, + "args": { + "External id": 67153, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67153, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67153, "pid": 0, "tid": 7, "ts": 1776649791943524, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943517, "dur": 4, + "args": { + "External id": 67153, + "cbid": 211, "correlation": 67153 + } + }, + { + "ph": "s", "id": 67153, "pid": 435, "tid": 435, "ts": 1776649791943517, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791943559, "dur": 1, + "args": { + "External id": 8880, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67171, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67171, "pid": 0, "tid": 7, "ts": 1776649791943559, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943550, "dur": 6, + "args": { + "External id": 8880, + "cbid": 211, "correlation": 67171 + } + }, + { + "ph": "s", "id": 67171, "pid": 435, "tid": 435, "ts": 1776649791943550, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943563, "dur": 0, + "args": { + "External id": 67177, + "cbid": 200, "correlation": 67177 + } + }, + { + "ph": "f", "id": 67177, "pid": 435, "tid": 435, "ts": 1776649791943563, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943563, "dur": 0, + "args": { + "External id": 67178, + "cbid": 273, "correlation": 67178 + } + }, + { + "ph": "f", "id": 67178, "pid": 435, "tid": 435, "ts": 1776649791943563, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791943572, "dur": 19, + "args": { + "External id": 67179, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67179, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67179, "pid": 0, "tid": 7, "ts": 1776649791943572, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943564, "dur": 5, + "args": { + "External id": 67179, + "cbid": 211, "correlation": 67179 + } + }, + { + "ph": "s", "id": 67179, "pid": 435, "tid": 435, "ts": 1776649791943564, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791943607, "dur": 1, + "args": { + "External id": 8885, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67197, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67197, "pid": 0, "tid": 7, "ts": 1776649791943607, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943598, "dur": 6, + "args": { + "External id": 8885, + "cbid": 211, "correlation": 67197 + } + }, + { + "ph": "s", "id": 67197, "pid": 435, "tid": 435, "ts": 1776649791943598, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943610, "dur": 0, + "args": { + "External id": 67203, + "cbid": 200, "correlation": 67203 + } + }, + { + "ph": "f", "id": 67203, "pid": 435, "tid": 435, "ts": 1776649791943610, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943611, "dur": 0, + "args": { + "External id": 67204, + "cbid": 273, "correlation": 67204 + } + }, + { + "ph": "f", "id": 67204, "pid": 435, "tid": 435, "ts": 1776649791943611, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791943620, "dur": 19, + "args": { + "External id": 67205, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67205, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67205, "pid": 0, "tid": 7, "ts": 1776649791943620, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943612, "dur": 4, + "args": { + "External id": 67205, + "cbid": 211, "correlation": 67205 + } + }, + { + "ph": "s", "id": 67205, "pid": 435, "tid": 435, "ts": 1776649791943612, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791943655, "dur": 1, + "args": { + "External id": 8890, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67223, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67223, "pid": 0, "tid": 7, "ts": 1776649791943655, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943646, "dur": 6, + "args": { + "External id": 8890, + "cbid": 211, "correlation": 67223 + } + }, + { + "ph": "s", "id": 67223, "pid": 435, "tid": 435, "ts": 1776649791943646, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943658, "dur": 0, + "args": { + "External id": 67229, + "cbid": 200, "correlation": 67229 + } + }, + { + "ph": "f", "id": 67229, "pid": 435, "tid": 435, "ts": 1776649791943658, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943659, "dur": 0, + "args": { + "External id": 67230, + "cbid": 273, "correlation": 67230 + } + }, + { + "ph": "f", "id": 67230, "pid": 435, "tid": 435, "ts": 1776649791943659, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791943668, "dur": 19, + "args": { + "External id": 67231, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67231, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67231, "pid": 0, "tid": 7, "ts": 1776649791943668, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943660, "dur": 4, + "args": { + "External id": 67231, + "cbid": 211, "correlation": 67231 + } + }, + { + "ph": "s", "id": 67231, "pid": 435, "tid": 435, "ts": 1776649791943660, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791943703, "dur": 1, + "args": { + "External id": 8895, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67249, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67249, "pid": 0, "tid": 7, "ts": 1776649791943703, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943694, "dur": 6, + "args": { + "External id": 8895, + "cbid": 211, "correlation": 67249 + } + }, + { + "ph": "s", "id": 67249, "pid": 435, "tid": 435, "ts": 1776649791943694, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943706, "dur": 0, + "args": { + "External id": 67255, + "cbid": 200, "correlation": 67255 + } + }, + { + "ph": "f", "id": 67255, "pid": 435, "tid": 435, "ts": 1776649791943706, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943707, "dur": 0, + "args": { + "External id": 67256, + "cbid": 273, "correlation": 67256 + } + }, + { + "ph": "f", "id": 67256, "pid": 435, "tid": 435, "ts": 1776649791943707, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791943716, "dur": 19, + "args": { + "External id": 67257, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67257, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67257, "pid": 0, "tid": 7, "ts": 1776649791943716, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943708, "dur": 4, + "args": { + "External id": 67257, + "cbid": 211, "correlation": 67257 + } + }, + { + "ph": "s", "id": 67257, "pid": 435, "tid": 435, "ts": 1776649791943708, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791943751, "dur": 1, + "args": { + "External id": 8900, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67275, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67275, "pid": 0, "tid": 7, "ts": 1776649791943751, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943741, "dur": 6, + "args": { + "External id": 8900, + "cbid": 211, "correlation": 67275 + } + }, + { + "ph": "s", "id": 67275, "pid": 435, "tid": 435, "ts": 1776649791943741, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943754, "dur": 0, + "args": { + "External id": 67281, + "cbid": 200, "correlation": 67281 + } + }, + { + "ph": "f", "id": 67281, "pid": 435, "tid": 435, "ts": 1776649791943754, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943754, "dur": 0, + "args": { + "External id": 67282, + "cbid": 273, "correlation": 67282 + } + }, + { + "ph": "f", "id": 67282, "pid": 435, "tid": 435, "ts": 1776649791943754, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791943763, "dur": 19, + "args": { + "External id": 67283, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67283, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67283, "pid": 0, "tid": 7, "ts": 1776649791943763, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943755, "dur": 4, + "args": { + "External id": 67283, + "cbid": 211, "correlation": 67283 + } + }, + { + "ph": "s", "id": 67283, "pid": 435, "tid": 435, "ts": 1776649791943755, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791943798, "dur": 1, + "args": { + "External id": 8905, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67301, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67301, "pid": 0, "tid": 7, "ts": 1776649791943798, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943789, "dur": 6, + "args": { + "External id": 8905, + "cbid": 211, "correlation": 67301 + } + }, + { + "ph": "s", "id": 67301, "pid": 435, "tid": 435, "ts": 1776649791943789, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943801, "dur": 0, + "args": { + "External id": 67307, + "cbid": 200, "correlation": 67307 + } + }, + { + "ph": "f", "id": 67307, "pid": 435, "tid": 435, "ts": 1776649791943801, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943801, "dur": 0, + "args": { + "External id": 67308, + "cbid": 273, "correlation": 67308 + } + }, + { + "ph": "f", "id": 67308, "pid": 435, "tid": 435, "ts": 1776649791943801, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791943810, "dur": 19, + "args": { + "External id": 67309, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67309, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67309, "pid": 0, "tid": 7, "ts": 1776649791943810, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943803, "dur": 4, + "args": { + "External id": 67309, + "cbid": 211, "correlation": 67309 + } + }, + { + "ph": "s", "id": 67309, "pid": 435, "tid": 435, "ts": 1776649791943803, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791943832, "dur": 4, + "args": { + "External id": 67317, + "cbid": 135, "correlation": 67317 + } + }, + { + "ph": "f", "id": 67317, "pid": 435, "tid": 435, "ts": 1776649791943832, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791943846, "dur": 5, + "args": { + "External id": 67322, + "cbid": 165, "correlation": 67322 + } + }, + { + "ph": "s", "id": 67322, "pid": 435, "tid": 435, "ts": 1776649791943846, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649791943855, "dur": 4, + "args": { + "External id": 67325, + "cbid": 139, "correlation": 67325 + } + }, + { + "ph": "f", "id": 67325, "pid": 435, "tid": 435, "ts": 1776649791943855, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791943910, "dur": 1, + "args": { + "External id": 8910, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67353, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67353, "pid": 0, "tid": 7, "ts": 1776649791943910, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943899, "dur": 8, + "args": { + "External id": 8910, + "cbid": 211, "correlation": 67353 + } + }, + { + "ph": "s", "id": 67353, "pid": 435, "tid": 435, "ts": 1776649791943899, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943914, "dur": 0, + "args": { + "External id": 67359, + "cbid": 200, "correlation": 67359 + } + }, + { + "ph": "f", "id": 67359, "pid": 435, "tid": 435, "ts": 1776649791943914, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943915, "dur": 1, + "args": { + "External id": 67360, + "cbid": 273, "correlation": 67360 + } + }, + { + "ph": "f", "id": 67360, "pid": 435, "tid": 435, "ts": 1776649791943915, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791943924, "dur": 19, + "args": { + "External id": 67361, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67361, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67361, "pid": 0, "tid": 7, "ts": 1776649791943924, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943916, "dur": 5, + "args": { + "External id": 67361, + "cbid": 211, "correlation": 67361 + } + }, + { + "ph": "s", "id": 67361, "pid": 435, "tid": 435, "ts": 1776649791943916, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791943962, "dur": 1, + "args": { + "External id": 8915, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67379, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67379, "pid": 0, "tid": 7, "ts": 1776649791943962, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943952, "dur": 7, + "args": { + "External id": 8915, + "cbid": 211, "correlation": 67379 + } + }, + { + "ph": "s", "id": 67379, "pid": 435, "tid": 435, "ts": 1776649791943952, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943965, "dur": 0, + "args": { + "External id": 67385, + "cbid": 200, "correlation": 67385 + } + }, + { + "ph": "f", "id": 67385, "pid": 435, "tid": 435, "ts": 1776649791943965, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791943966, "dur": 0, + "args": { + "External id": 67386, + "cbid": 273, "correlation": 67386 + } + }, + { + "ph": "f", "id": 67386, "pid": 435, "tid": 435, "ts": 1776649791943966, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791943975, "dur": 19, + "args": { + "External id": 67387, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67387, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67387, "pid": 0, "tid": 7, "ts": 1776649791943975, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791943967, "dur": 5, + "args": { + "External id": 67387, + "cbid": 211, "correlation": 67387 + } + }, + { + "ph": "s", "id": 67387, "pid": 435, "tid": 435, "ts": 1776649791943967, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791944011, "dur": 1, + "args": { + "External id": 8920, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67405, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67405, "pid": 0, "tid": 7, "ts": 1776649791944011, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944001, "dur": 7, + "args": { + "External id": 8920, + "cbid": 211, "correlation": 67405 + } + }, + { + "ph": "s", "id": 67405, "pid": 435, "tid": 435, "ts": 1776649791944001, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944016, "dur": 0, + "args": { + "External id": 67411, + "cbid": 200, "correlation": 67411 + } + }, + { + "ph": "f", "id": 67411, "pid": 435, "tid": 435, "ts": 1776649791944016, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944016, "dur": 0, + "args": { + "External id": 67412, + "cbid": 273, "correlation": 67412 + } + }, + { + "ph": "f", "id": 67412, "pid": 435, "tid": 435, "ts": 1776649791944016, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791944025, "dur": 19, + "args": { + "External id": 67413, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67413, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67413, "pid": 0, "tid": 7, "ts": 1776649791944025, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944018, "dur": 4, + "args": { + "External id": 67413, + "cbid": 211, "correlation": 67413 + } + }, + { + "ph": "s", "id": 67413, "pid": 435, "tid": 435, "ts": 1776649791944018, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791944064, "dur": 1, + "args": { + "External id": 8925, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67431, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67431, "pid": 0, "tid": 7, "ts": 1776649791944064, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944054, "dur": 6, + "args": { + "External id": 8925, + "cbid": 211, "correlation": 67431 + } + }, + { + "ph": "s", "id": 67431, "pid": 435, "tid": 435, "ts": 1776649791944054, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944067, "dur": 0, + "args": { + "External id": 67437, + "cbid": 200, "correlation": 67437 + } + }, + { + "ph": "f", "id": 67437, "pid": 435, "tid": 435, "ts": 1776649791944067, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944068, "dur": 0, + "args": { + "External id": 67438, + "cbid": 273, "correlation": 67438 + } + }, + { + "ph": "f", "id": 67438, "pid": 435, "tid": 435, "ts": 1776649791944068, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791944077, "dur": 19, + "args": { + "External id": 67439, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67439, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67439, "pid": 0, "tid": 7, "ts": 1776649791944077, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944069, "dur": 4, + "args": { + "External id": 67439, + "cbid": 211, "correlation": 67439 + } + }, + { + "ph": "s", "id": 67439, "pid": 435, "tid": 435, "ts": 1776649791944069, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791944112, "dur": 1, + "args": { + "External id": 8930, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67457, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67457, "pid": 0, "tid": 7, "ts": 1776649791944112, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944103, "dur": 6, + "args": { + "External id": 8930, + "cbid": 211, "correlation": 67457 + } + }, + { + "ph": "s", "id": 67457, "pid": 435, "tid": 435, "ts": 1776649791944103, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944115, "dur": 0, + "args": { + "External id": 67463, + "cbid": 200, "correlation": 67463 + } + }, + { + "ph": "f", "id": 67463, "pid": 435, "tid": 435, "ts": 1776649791944115, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944115, "dur": 0, + "args": { + "External id": 67464, + "cbid": 273, "correlation": 67464 + } + }, + { + "ph": "f", "id": 67464, "pid": 435, "tid": 435, "ts": 1776649791944115, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791944124, "dur": 19, + "args": { + "External id": 67465, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67465, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67465, "pid": 0, "tid": 7, "ts": 1776649791944124, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944117, "dur": 4, + "args": { + "External id": 67465, + "cbid": 211, "correlation": 67465 + } + }, + { + "ph": "s", "id": 67465, "pid": 435, "tid": 435, "ts": 1776649791944117, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791944160, "dur": 1, + "args": { + "External id": 8935, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67483, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67483, "pid": 0, "tid": 7, "ts": 1776649791944160, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944151, "dur": 6, + "args": { + "External id": 8935, + "cbid": 211, "correlation": 67483 + } + }, + { + "ph": "s", "id": 67483, "pid": 435, "tid": 435, "ts": 1776649791944151, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944163, "dur": 0, + "args": { + "External id": 67489, + "cbid": 200, "correlation": 67489 + } + }, + { + "ph": "f", "id": 67489, "pid": 435, "tid": 435, "ts": 1776649791944163, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944163, "dur": 0, + "args": { + "External id": 67490, + "cbid": 273, "correlation": 67490 + } + }, + { + "ph": "f", "id": 67490, "pid": 435, "tid": 435, "ts": 1776649791944163, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791944172, "dur": 19, + "args": { + "External id": 67491, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67491, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67491, "pid": 0, "tid": 7, "ts": 1776649791944172, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944165, "dur": 4, + "args": { + "External id": 67491, + "cbid": 211, "correlation": 67491 + } + }, + { + "ph": "s", "id": 67491, "pid": 435, "tid": 435, "ts": 1776649791944165, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791944208, "dur": 1, + "args": { + "External id": 8940, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67509, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67509, "pid": 0, "tid": 7, "ts": 1776649791944208, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944199, "dur": 6, + "args": { + "External id": 8940, + "cbid": 211, "correlation": 67509 + } + }, + { + "ph": "s", "id": 67509, "pid": 435, "tid": 435, "ts": 1776649791944199, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944211, "dur": 0, + "args": { + "External id": 67515, + "cbid": 200, "correlation": 67515 + } + }, + { + "ph": "f", "id": 67515, "pid": 435, "tid": 435, "ts": 1776649791944211, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944212, "dur": 0, + "args": { + "External id": 67516, + "cbid": 273, "correlation": 67516 + } + }, + { + "ph": "f", "id": 67516, "pid": 435, "tid": 435, "ts": 1776649791944212, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791944238, "dur": 19, + "args": { + "External id": 67517, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67517, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67517, "pid": 0, "tid": 7, "ts": 1776649791944238, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944213, "dur": 22, + "args": { + "External id": 67517, + "cbid": 211, "correlation": 67517 + } + }, + { + "ph": "s", "id": 67517, "pid": 435, "tid": 435, "ts": 1776649791944213, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791944279, "dur": 1, + "args": { + "External id": 8945, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67535, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67535, "pid": 0, "tid": 7, "ts": 1776649791944279, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944269, "dur": 7, + "args": { + "External id": 8945, + "cbid": 211, "correlation": 67535 + } + }, + { + "ph": "s", "id": 67535, "pid": 435, "tid": 435, "ts": 1776649791944269, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944282, "dur": 0, + "args": { + "External id": 67541, + "cbid": 200, "correlation": 67541 + } + }, + { + "ph": "f", "id": 67541, "pid": 435, "tid": 435, "ts": 1776649791944282, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944283, "dur": 0, + "args": { + "External id": 67542, + "cbid": 273, "correlation": 67542 + } + }, + { + "ph": "f", "id": 67542, "pid": 435, "tid": 435, "ts": 1776649791944283, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791944292, "dur": 19, + "args": { + "External id": 67543, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67543, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67543, "pid": 0, "tid": 7, "ts": 1776649791944292, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944284, "dur": 4, + "args": { + "External id": 67543, + "cbid": 211, "correlation": 67543 + } + }, + { + "ph": "s", "id": 67543, "pid": 435, "tid": 435, "ts": 1776649791944284, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791944328, "dur": 1, + "args": { + "External id": 8950, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67561, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67561, "pid": 0, "tid": 7, "ts": 1776649791944328, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944318, "dur": 6, + "args": { + "External id": 8950, + "cbid": 211, "correlation": 67561 + } + }, + { + "ph": "s", "id": 67561, "pid": 435, "tid": 435, "ts": 1776649791944318, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944331, "dur": 0, + "args": { + "External id": 67567, + "cbid": 200, "correlation": 67567 + } + }, + { + "ph": "f", "id": 67567, "pid": 435, "tid": 435, "ts": 1776649791944331, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944332, "dur": 0, + "args": { + "External id": 67568, + "cbid": 273, "correlation": 67568 + } + }, + { + "ph": "f", "id": 67568, "pid": 435, "tid": 435, "ts": 1776649791944332, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791944341, "dur": 19, + "args": { + "External id": 67569, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67569, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67569, "pid": 0, "tid": 7, "ts": 1776649791944341, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944333, "dur": 4, + "args": { + "External id": 67569, + "cbid": 211, "correlation": 67569 + } + }, + { + "ph": "s", "id": 67569, "pid": 435, "tid": 435, "ts": 1776649791944333, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791944376, "dur": 1, + "args": { + "External id": 8955, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67587, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67587, "pid": 0, "tid": 7, "ts": 1776649791944376, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944367, "dur": 6, + "args": { + "External id": 8955, + "cbid": 211, "correlation": 67587 + } + }, + { + "ph": "s", "id": 67587, "pid": 435, "tid": 435, "ts": 1776649791944367, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944379, "dur": 0, + "args": { + "External id": 67593, + "cbid": 200, "correlation": 67593 + } + }, + { + "ph": "f", "id": 67593, "pid": 435, "tid": 435, "ts": 1776649791944379, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944380, "dur": 0, + "args": { + "External id": 67594, + "cbid": 273, "correlation": 67594 + } + }, + { + "ph": "f", "id": 67594, "pid": 435, "tid": 435, "ts": 1776649791944380, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791944389, "dur": 19, + "args": { + "External id": 67595, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67595, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67595, "pid": 0, "tid": 7, "ts": 1776649791944389, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944381, "dur": 4, + "args": { + "External id": 67595, + "cbid": 211, "correlation": 67595 + } + }, + { + "ph": "s", "id": 67595, "pid": 435, "tid": 435, "ts": 1776649791944381, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791944403, "dur": 9, + "args": { + "External id": 67598, + "cbid": 165, "correlation": 67598 + } + }, + { + "ph": "s", "id": 67598, "pid": 435, "tid": 435, "ts": 1776649791944403, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791944434, "dur": 3, + "args": { + "External id": 67607, + "cbid": 135, "correlation": 67607 + } + }, + { + "ph": "f", "id": 67607, "pid": 435, "tid": 435, "ts": 1776649791944434, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791944483, "dur": 1, + "args": { + "External id": 8960, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67627, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67627, "pid": 0, "tid": 7, "ts": 1776649791944483, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944472, "dur": 8, + "args": { + "External id": 8960, + "cbid": 211, "correlation": 67627 + } + }, + { + "ph": "s", "id": 67627, "pid": 435, "tid": 435, "ts": 1776649791944472, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944487, "dur": 0, + "args": { + "External id": 67633, + "cbid": 200, "correlation": 67633 + } + }, + { + "ph": "f", "id": 67633, "pid": 435, "tid": 435, "ts": 1776649791944487, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944488, "dur": 1, + "args": { + "External id": 67634, + "cbid": 273, "correlation": 67634 + } + }, + { + "ph": "f", "id": 67634, "pid": 435, "tid": 435, "ts": 1776649791944488, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791944497, "dur": 19, + "args": { + "External id": 67635, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67635, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67635, "pid": 0, "tid": 7, "ts": 1776649791944497, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944489, "dur": 5, + "args": { + "External id": 67635, + "cbid": 211, "correlation": 67635 + } + }, + { + "ph": "s", "id": 67635, "pid": 435, "tid": 435, "ts": 1776649791944489, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791944534, "dur": 1, + "args": { + "External id": 8965, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67653, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67653, "pid": 0, "tid": 7, "ts": 1776649791944534, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944525, "dur": 6, + "args": { + "External id": 8965, + "cbid": 211, "correlation": 67653 + } + }, + { + "ph": "s", "id": 67653, "pid": 435, "tid": 435, "ts": 1776649791944525, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944538, "dur": 0, + "args": { + "External id": 67659, + "cbid": 200, "correlation": 67659 + } + }, + { + "ph": "f", "id": 67659, "pid": 435, "tid": 435, "ts": 1776649791944538, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944538, "dur": 0, + "args": { + "External id": 67660, + "cbid": 273, "correlation": 67660 + } + }, + { + "ph": "f", "id": 67660, "pid": 435, "tid": 435, "ts": 1776649791944538, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791944547, "dur": 19, + "args": { + "External id": 67661, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67661, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67661, "pid": 0, "tid": 7, "ts": 1776649791944547, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944539, "dur": 5, + "args": { + "External id": 67661, + "cbid": 211, "correlation": 67661 + } + }, + { + "ph": "s", "id": 67661, "pid": 435, "tid": 435, "ts": 1776649791944539, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791944583, "dur": 1, + "args": { + "External id": 8970, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67679, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67679, "pid": 0, "tid": 7, "ts": 1776649791944583, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944574, "dur": 7, + "args": { + "External id": 8970, + "cbid": 211, "correlation": 67679 + } + }, + { + "ph": "s", "id": 67679, "pid": 435, "tid": 435, "ts": 1776649791944574, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944586, "dur": 0, + "args": { + "External id": 67685, + "cbid": 200, "correlation": 67685 + } + }, + { + "ph": "f", "id": 67685, "pid": 435, "tid": 435, "ts": 1776649791944586, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944587, "dur": 0, + "args": { + "External id": 67686, + "cbid": 273, "correlation": 67686 + } + }, + { + "ph": "f", "id": 67686, "pid": 435, "tid": 435, "ts": 1776649791944587, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791944596, "dur": 19, + "args": { + "External id": 67687, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67687, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67687, "pid": 0, "tid": 7, "ts": 1776649791944596, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944588, "dur": 5, + "args": { + "External id": 67687, + "cbid": 211, "correlation": 67687 + } + }, + { + "ph": "s", "id": 67687, "pid": 435, "tid": 435, "ts": 1776649791944588, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791944632, "dur": 1, + "args": { + "External id": 8975, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67705, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67705, "pid": 0, "tid": 7, "ts": 1776649791944632, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944623, "dur": 6, + "args": { + "External id": 8975, + "cbid": 211, "correlation": 67705 + } + }, + { + "ph": "s", "id": 67705, "pid": 435, "tid": 435, "ts": 1776649791944623, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944635, "dur": 0, + "args": { + "External id": 67711, + "cbid": 200, "correlation": 67711 + } + }, + { + "ph": "f", "id": 67711, "pid": 435, "tid": 435, "ts": 1776649791944635, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944636, "dur": 0, + "args": { + "External id": 67712, + "cbid": 273, "correlation": 67712 + } + }, + { + "ph": "f", "id": 67712, "pid": 435, "tid": 435, "ts": 1776649791944636, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791944645, "dur": 19, + "args": { + "External id": 67713, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67713, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67713, "pid": 0, "tid": 7, "ts": 1776649791944645, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944637, "dur": 4, + "args": { + "External id": 67713, + "cbid": 211, "correlation": 67713 + } + }, + { + "ph": "s", "id": 67713, "pid": 435, "tid": 435, "ts": 1776649791944637, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791944681, "dur": 1, + "args": { + "External id": 8980, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67731, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67731, "pid": 0, "tid": 7, "ts": 1776649791944681, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944671, "dur": 6, + "args": { + "External id": 8980, + "cbid": 211, "correlation": 67731 + } + }, + { + "ph": "s", "id": 67731, "pid": 435, "tid": 435, "ts": 1776649791944671, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944684, "dur": 0, + "args": { + "External id": 67737, + "cbid": 200, "correlation": 67737 + } + }, + { + "ph": "f", "id": 67737, "pid": 435, "tid": 435, "ts": 1776649791944684, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944684, "dur": 0, + "args": { + "External id": 67738, + "cbid": 273, "correlation": 67738 + } + }, + { + "ph": "f", "id": 67738, "pid": 435, "tid": 435, "ts": 1776649791944684, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791944693, "dur": 19, + "args": { + "External id": 67739, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67739, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67739, "pid": 0, "tid": 7, "ts": 1776649791944693, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944686, "dur": 5, + "args": { + "External id": 67739, + "cbid": 211, "correlation": 67739 + } + }, + { + "ph": "s", "id": 67739, "pid": 435, "tid": 435, "ts": 1776649791944686, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791944729, "dur": 1, + "args": { + "External id": 8985, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67757, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67757, "pid": 0, "tid": 7, "ts": 1776649791944729, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944720, "dur": 6, + "args": { + "External id": 8985, + "cbid": 211, "correlation": 67757 + } + }, + { + "ph": "s", "id": 67757, "pid": 435, "tid": 435, "ts": 1776649791944720, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944733, "dur": 0, + "args": { + "External id": 67763, + "cbid": 200, "correlation": 67763 + } + }, + { + "ph": "f", "id": 67763, "pid": 435, "tid": 435, "ts": 1776649791944733, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944733, "dur": 0, + "args": { + "External id": 67764, + "cbid": 273, "correlation": 67764 + } + }, + { + "ph": "f", "id": 67764, "pid": 435, "tid": 435, "ts": 1776649791944733, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791944742, "dur": 19, + "args": { + "External id": 67765, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67765, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67765, "pid": 0, "tid": 7, "ts": 1776649791944742, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944734, "dur": 5, + "args": { + "External id": 67765, + "cbid": 211, "correlation": 67765 + } + }, + { + "ph": "s", "id": 67765, "pid": 435, "tid": 435, "ts": 1776649791944734, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791944778, "dur": 1, + "args": { + "External id": 8990, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67783, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67783, "pid": 0, "tid": 7, "ts": 1776649791944778, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944769, "dur": 6, + "args": { + "External id": 8990, + "cbid": 211, "correlation": 67783 + } + }, + { + "ph": "s", "id": 67783, "pid": 435, "tid": 435, "ts": 1776649791944769, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944781, "dur": 0, + "args": { + "External id": 67789, + "cbid": 200, "correlation": 67789 + } + }, + { + "ph": "f", "id": 67789, "pid": 435, "tid": 435, "ts": 1776649791944781, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944782, "dur": 0, + "args": { + "External id": 67790, + "cbid": 273, "correlation": 67790 + } + }, + { + "ph": "f", "id": 67790, "pid": 435, "tid": 435, "ts": 1776649791944782, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791944790, "dur": 19, + "args": { + "External id": 67791, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67791, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67791, "pid": 0, "tid": 7, "ts": 1776649791944790, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944783, "dur": 4, + "args": { + "External id": 67791, + "cbid": 211, "correlation": 67791 + } + }, + { + "ph": "s", "id": 67791, "pid": 435, "tid": 435, "ts": 1776649791944783, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791944826, "dur": 1, + "args": { + "External id": 8995, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67809, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67809, "pid": 0, "tid": 7, "ts": 1776649791944826, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944817, "dur": 6, + "args": { + "External id": 8995, + "cbid": 211, "correlation": 67809 + } + }, + { + "ph": "s", "id": 67809, "pid": 435, "tid": 435, "ts": 1776649791944817, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944829, "dur": 0, + "args": { + "External id": 67815, + "cbid": 200, "correlation": 67815 + } + }, + { + "ph": "f", "id": 67815, "pid": 435, "tid": 435, "ts": 1776649791944829, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944830, "dur": 0, + "args": { + "External id": 67816, + "cbid": 273, "correlation": 67816 + } + }, + { + "ph": "f", "id": 67816, "pid": 435, "tid": 435, "ts": 1776649791944830, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791944839, "dur": 19, + "args": { + "External id": 67817, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67817, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67817, "pid": 0, "tid": 7, "ts": 1776649791944839, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944831, "dur": 5, + "args": { + "External id": 67817, + "cbid": 211, "correlation": 67817 + } + }, + { + "ph": "s", "id": 67817, "pid": 435, "tid": 435, "ts": 1776649791944831, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791944874, "dur": 1, + "args": { + "External id": 9000, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67835, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67835, "pid": 0, "tid": 7, "ts": 1776649791944874, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944865, "dur": 6, + "args": { + "External id": 9000, + "cbid": 211, "correlation": 67835 + } + }, + { + "ph": "s", "id": 67835, "pid": 435, "tid": 435, "ts": 1776649791944865, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944877, "dur": 0, + "args": { + "External id": 67841, + "cbid": 200, "correlation": 67841 + } + }, + { + "ph": "f", "id": 67841, "pid": 435, "tid": 435, "ts": 1776649791944877, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944878, "dur": 0, + "args": { + "External id": 67842, + "cbid": 273, "correlation": 67842 + } + }, + { + "ph": "f", "id": 67842, "pid": 435, "tid": 435, "ts": 1776649791944878, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791944887, "dur": 19, + "args": { + "External id": 67843, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67843, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67843, "pid": 0, "tid": 7, "ts": 1776649791944887, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944879, "dur": 5, + "args": { + "External id": 67843, + "cbid": 211, "correlation": 67843 + } + }, + { + "ph": "s", "id": 67843, "pid": 435, "tid": 435, "ts": 1776649791944879, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791944922, "dur": 1, + "args": { + "External id": 9005, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67861, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67861, "pid": 0, "tid": 7, "ts": 1776649791944922, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944913, "dur": 6, + "args": { + "External id": 9005, + "cbid": 211, "correlation": 67861 + } + }, + { + "ph": "s", "id": 67861, "pid": 435, "tid": 435, "ts": 1776649791944913, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944925, "dur": 0, + "args": { + "External id": 67867, + "cbid": 200, "correlation": 67867 + } + }, + { + "ph": "f", "id": 67867, "pid": 435, "tid": 435, "ts": 1776649791944925, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944926, "dur": 0, + "args": { + "External id": 67868, + "cbid": 273, "correlation": 67868 + } + }, + { + "ph": "f", "id": 67868, "pid": 435, "tid": 435, "ts": 1776649791944926, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791944935, "dur": 19, + "args": { + "External id": 67869, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67869, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67869, "pid": 0, "tid": 7, "ts": 1776649791944935, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944927, "dur": 4, + "args": { + "External id": 67869, + "cbid": 211, "correlation": 67869 + } + }, + { + "ph": "s", "id": 67869, "pid": 435, "tid": 435, "ts": 1776649791944927, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791944970, "dur": 1, + "args": { + "External id": 9010, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67887, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67887, "pid": 0, "tid": 7, "ts": 1776649791944970, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944961, "dur": 6, + "args": { + "External id": 9010, + "cbid": 211, "correlation": 67887 + } + }, + { + "ph": "s", "id": 67887, "pid": 435, "tid": 435, "ts": 1776649791944961, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944973, "dur": 0, + "args": { + "External id": 67893, + "cbid": 200, "correlation": 67893 + } + }, + { + "ph": "f", "id": 67893, "pid": 435, "tid": 435, "ts": 1776649791944973, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791944974, "dur": 0, + "args": { + "External id": 67894, + "cbid": 273, "correlation": 67894 + } + }, + { + "ph": "f", "id": 67894, "pid": 435, "tid": 435, "ts": 1776649791944974, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791944983, "dur": 19, + "args": { + "External id": 67895, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67895, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67895, "pid": 0, "tid": 7, "ts": 1776649791944983, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791944975, "dur": 4, + "args": { + "External id": 67895, + "cbid": 211, "correlation": 67895 + } + }, + { + "ph": "s", "id": 67895, "pid": 435, "tid": 435, "ts": 1776649791944975, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791945019, "dur": 1, + "args": { + "External id": 9015, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67913, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67913, "pid": 0, "tid": 7, "ts": 1776649791945019, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945009, "dur": 6, + "args": { + "External id": 9015, + "cbid": 211, "correlation": 67913 + } + }, + { + "ph": "s", "id": 67913, "pid": 435, "tid": 435, "ts": 1776649791945009, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945022, "dur": 0, + "args": { + "External id": 67919, + "cbid": 200, "correlation": 67919 + } + }, + { + "ph": "f", "id": 67919, "pid": 435, "tid": 435, "ts": 1776649791945022, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945022, "dur": 0, + "args": { + "External id": 67920, + "cbid": 273, "correlation": 67920 + } + }, + { + "ph": "f", "id": 67920, "pid": 435, "tid": 435, "ts": 1776649791945022, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791945031, "dur": 19, + "args": { + "External id": 67921, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67921, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67921, "pid": 0, "tid": 7, "ts": 1776649791945031, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945023, "dur": 5, + "args": { + "External id": 67921, + "cbid": 211, "correlation": 67921 + } + }, + { + "ph": "s", "id": 67921, "pid": 435, "tid": 435, "ts": 1776649791945023, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791945066, "dur": 1, + "args": { + "External id": 9020, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67939, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67939, "pid": 0, "tid": 7, "ts": 1776649791945066, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945057, "dur": 6, + "args": { + "External id": 9020, + "cbid": 211, "correlation": 67939 + } + }, + { + "ph": "s", "id": 67939, "pid": 435, "tid": 435, "ts": 1776649791945057, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945069, "dur": 0, + "args": { + "External id": 67945, + "cbid": 200, "correlation": 67945 + } + }, + { + "ph": "f", "id": 67945, "pid": 435, "tid": 435, "ts": 1776649791945069, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945070, "dur": 0, + "args": { + "External id": 67946, + "cbid": 273, "correlation": 67946 + } + }, + { + "ph": "f", "id": 67946, "pid": 435, "tid": 435, "ts": 1776649791945070, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791945079, "dur": 19, + "args": { + "External id": 67947, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67947, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67947, "pid": 0, "tid": 7, "ts": 1776649791945079, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945071, "dur": 4, + "args": { + "External id": 67947, + "cbid": 211, "correlation": 67947 + } + }, + { + "ph": "s", "id": 67947, "pid": 435, "tid": 435, "ts": 1776649791945071, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791945115, "dur": 1, + "args": { + "External id": 9025, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67965, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67965, "pid": 0, "tid": 7, "ts": 1776649791945115, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945105, "dur": 6, + "args": { + "External id": 9025, + "cbid": 211, "correlation": 67965 + } + }, + { + "ph": "s", "id": 67965, "pid": 435, "tid": 435, "ts": 1776649791945105, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945118, "dur": 0, + "args": { + "External id": 67971, + "cbid": 200, "correlation": 67971 + } + }, + { + "ph": "f", "id": 67971, "pid": 435, "tid": 435, "ts": 1776649791945118, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945118, "dur": 0, + "args": { + "External id": 67972, + "cbid": 273, "correlation": 67972 + } + }, + { + "ph": "f", "id": 67972, "pid": 435, "tid": 435, "ts": 1776649791945118, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791945127, "dur": 19, + "args": { + "External id": 67973, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67973, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67973, "pid": 0, "tid": 7, "ts": 1776649791945127, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945120, "dur": 4, + "args": { + "External id": 67973, + "cbid": 211, "correlation": 67973 + } + }, + { + "ph": "s", "id": 67973, "pid": 435, "tid": 435, "ts": 1776649791945120, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791945163, "dur": 1, + "args": { + "External id": 9030, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67991, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67991, "pid": 0, "tid": 7, "ts": 1776649791945163, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945153, "dur": 6, + "args": { + "External id": 9030, + "cbid": 211, "correlation": 67991 + } + }, + { + "ph": "s", "id": 67991, "pid": 435, "tid": 435, "ts": 1776649791945153, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945166, "dur": 0, + "args": { + "External id": 67997, + "cbid": 200, "correlation": 67997 + } + }, + { + "ph": "f", "id": 67997, "pid": 435, "tid": 435, "ts": 1776649791945166, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945167, "dur": 0, + "args": { + "External id": 67998, + "cbid": 273, "correlation": 67998 + } + }, + { + "ph": "f", "id": 67998, "pid": 435, "tid": 435, "ts": 1776649791945167, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791945176, "dur": 19, + "args": { + "External id": 67999, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 67999, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 67999, "pid": 0, "tid": 7, "ts": 1776649791945176, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945168, "dur": 4, + "args": { + "External id": 67999, + "cbid": 211, "correlation": 67999 + } + }, + { + "ph": "s", "id": 67999, "pid": 435, "tid": 435, "ts": 1776649791945168, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791945211, "dur": 1, + "args": { + "External id": 9035, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68017, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68017, "pid": 0, "tid": 7, "ts": 1776649791945211, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945202, "dur": 6, + "args": { + "External id": 9035, + "cbid": 211, "correlation": 68017 + } + }, + { + "ph": "s", "id": 68017, "pid": 435, "tid": 435, "ts": 1776649791945202, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945214, "dur": 0, + "args": { + "External id": 68023, + "cbid": 200, "correlation": 68023 + } + }, + { + "ph": "f", "id": 68023, "pid": 435, "tid": 435, "ts": 1776649791945214, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945215, "dur": 0, + "args": { + "External id": 68024, + "cbid": 273, "correlation": 68024 + } + }, + { + "ph": "f", "id": 68024, "pid": 435, "tid": 435, "ts": 1776649791945215, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791945224, "dur": 19, + "args": { + "External id": 68025, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68025, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68025, "pid": 0, "tid": 7, "ts": 1776649791945224, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945216, "dur": 5, + "args": { + "External id": 68025, + "cbid": 211, "correlation": 68025 + } + }, + { + "ph": "s", "id": 68025, "pid": 435, "tid": 435, "ts": 1776649791945216, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791945259, "dur": 1, + "args": { + "External id": 9040, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68043, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68043, "pid": 0, "tid": 7, "ts": 1776649791945259, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945250, "dur": 6, + "args": { + "External id": 9040, + "cbid": 211, "correlation": 68043 + } + }, + { + "ph": "s", "id": 68043, "pid": 435, "tid": 435, "ts": 1776649791945250, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945262, "dur": 0, + "args": { + "External id": 68049, + "cbid": 200, "correlation": 68049 + } + }, + { + "ph": "f", "id": 68049, "pid": 435, "tid": 435, "ts": 1776649791945262, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945263, "dur": 0, + "args": { + "External id": 68050, + "cbid": 273, "correlation": 68050 + } + }, + { + "ph": "f", "id": 68050, "pid": 435, "tid": 435, "ts": 1776649791945263, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791945272, "dur": 19, + "args": { + "External id": 68051, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68051, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68051, "pid": 0, "tid": 7, "ts": 1776649791945272, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945264, "dur": 5, + "args": { + "External id": 68051, + "cbid": 211, "correlation": 68051 + } + }, + { + "ph": "s", "id": 68051, "pid": 435, "tid": 435, "ts": 1776649791945264, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791945307, "dur": 1, + "args": { + "External id": 9045, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68069, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68069, "pid": 0, "tid": 7, "ts": 1776649791945307, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945298, "dur": 6, + "args": { + "External id": 9045, + "cbid": 211, "correlation": 68069 + } + }, + { + "ph": "s", "id": 68069, "pid": 435, "tid": 435, "ts": 1776649791945298, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945310, "dur": 0, + "args": { + "External id": 68075, + "cbid": 200, "correlation": 68075 + } + }, + { + "ph": "f", "id": 68075, "pid": 435, "tid": 435, "ts": 1776649791945310, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945311, "dur": 0, + "args": { + "External id": 68076, + "cbid": 273, "correlation": 68076 + } + }, + { + "ph": "f", "id": 68076, "pid": 435, "tid": 435, "ts": 1776649791945311, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791945320, "dur": 19, + "args": { + "External id": 68077, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68077, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68077, "pid": 0, "tid": 7, "ts": 1776649791945320, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945312, "dur": 4, + "args": { + "External id": 68077, + "cbid": 211, "correlation": 68077 + } + }, + { + "ph": "s", "id": 68077, "pid": 435, "tid": 435, "ts": 1776649791945312, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791945355, "dur": 1, + "args": { + "External id": 9050, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68095, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68095, "pid": 0, "tid": 7, "ts": 1776649791945355, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945346, "dur": 6, + "args": { + "External id": 9050, + "cbid": 211, "correlation": 68095 + } + }, + { + "ph": "s", "id": 68095, "pid": 435, "tid": 435, "ts": 1776649791945346, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945358, "dur": 0, + "args": { + "External id": 68101, + "cbid": 200, "correlation": 68101 + } + }, + { + "ph": "f", "id": 68101, "pid": 435, "tid": 435, "ts": 1776649791945358, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945359, "dur": 0, + "args": { + "External id": 68102, + "cbid": 273, "correlation": 68102 + } + }, + { + "ph": "f", "id": 68102, "pid": 435, "tid": 435, "ts": 1776649791945359, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791945368, "dur": 19, + "args": { + "External id": 68103, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68103, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68103, "pid": 0, "tid": 7, "ts": 1776649791945368, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945360, "dur": 4, + "args": { + "External id": 68103, + "cbid": 211, "correlation": 68103 + } + }, + { + "ph": "s", "id": 68103, "pid": 435, "tid": 435, "ts": 1776649791945360, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791945403, "dur": 1, + "args": { + "External id": 9055, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68121, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68121, "pid": 0, "tid": 7, "ts": 1776649791945403, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945394, "dur": 6, + "args": { + "External id": 9055, + "cbid": 211, "correlation": 68121 + } + }, + { + "ph": "s", "id": 68121, "pid": 435, "tid": 435, "ts": 1776649791945394, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945406, "dur": 0, + "args": { + "External id": 68127, + "cbid": 200, "correlation": 68127 + } + }, + { + "ph": "f", "id": 68127, "pid": 435, "tid": 435, "ts": 1776649791945406, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945407, "dur": 0, + "args": { + "External id": 68128, + "cbid": 273, "correlation": 68128 + } + }, + { + "ph": "f", "id": 68128, "pid": 435, "tid": 435, "ts": 1776649791945407, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791945416, "dur": 19, + "args": { + "External id": 68129, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68129, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68129, "pid": 0, "tid": 7, "ts": 1776649791945416, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945408, "dur": 4, + "args": { + "External id": 68129, + "cbid": 211, "correlation": 68129 + } + }, + { + "ph": "s", "id": 68129, "pid": 435, "tid": 435, "ts": 1776649791945408, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791945451, "dur": 1, + "args": { + "External id": 9060, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68147, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68147, "pid": 0, "tid": 7, "ts": 1776649791945451, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945441, "dur": 6, + "args": { + "External id": 9060, + "cbid": 211, "correlation": 68147 + } + }, + { + "ph": "s", "id": 68147, "pid": 435, "tid": 435, "ts": 1776649791945441, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945454, "dur": 0, + "args": { + "External id": 68153, + "cbid": 200, "correlation": 68153 + } + }, + { + "ph": "f", "id": 68153, "pid": 435, "tid": 435, "ts": 1776649791945454, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945454, "dur": 0, + "args": { + "External id": 68154, + "cbid": 273, "correlation": 68154 + } + }, + { + "ph": "f", "id": 68154, "pid": 435, "tid": 435, "ts": 1776649791945454, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791945463, "dur": 19, + "args": { + "External id": 68155, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68155, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68155, "pid": 0, "tid": 7, "ts": 1776649791945463, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945456, "dur": 5, + "args": { + "External id": 68155, + "cbid": 211, "correlation": 68155 + } + }, + { + "ph": "s", "id": 68155, "pid": 435, "tid": 435, "ts": 1776649791945456, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791945499, "dur": 1, + "args": { + "External id": 9065, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68173, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68173, "pid": 0, "tid": 7, "ts": 1776649791945499, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945490, "dur": 6, + "args": { + "External id": 9065, + "cbid": 211, "correlation": 68173 + } + }, + { + "ph": "s", "id": 68173, "pid": 435, "tid": 435, "ts": 1776649791945490, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945502, "dur": 0, + "args": { + "External id": 68179, + "cbid": 200, "correlation": 68179 + } + }, + { + "ph": "f", "id": 68179, "pid": 435, "tid": 435, "ts": 1776649791945502, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945502, "dur": 0, + "args": { + "External id": 68180, + "cbid": 273, "correlation": 68180 + } + }, + { + "ph": "f", "id": 68180, "pid": 435, "tid": 435, "ts": 1776649791945502, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791945511, "dur": 19, + "args": { + "External id": 68181, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68181, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68181, "pid": 0, "tid": 7, "ts": 1776649791945511, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945504, "dur": 4, + "args": { + "External id": 68181, + "cbid": 211, "correlation": 68181 + } + }, + { + "ph": "s", "id": 68181, "pid": 435, "tid": 435, "ts": 1776649791945504, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791945547, "dur": 1, + "args": { + "External id": 9070, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68199, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68199, "pid": 0, "tid": 7, "ts": 1776649791945547, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945538, "dur": 6, + "args": { + "External id": 9070, + "cbid": 211, "correlation": 68199 + } + }, + { + "ph": "s", "id": 68199, "pid": 435, "tid": 435, "ts": 1776649791945538, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945550, "dur": 0, + "args": { + "External id": 68205, + "cbid": 200, "correlation": 68205 + } + }, + { + "ph": "f", "id": 68205, "pid": 435, "tid": 435, "ts": 1776649791945550, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945551, "dur": 0, + "args": { + "External id": 68206, + "cbid": 273, "correlation": 68206 + } + }, + { + "ph": "f", "id": 68206, "pid": 435, "tid": 435, "ts": 1776649791945551, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791945560, "dur": 19, + "args": { + "External id": 68207, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68207, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68207, "pid": 0, "tid": 7, "ts": 1776649791945560, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945552, "dur": 4, + "args": { + "External id": 68207, + "cbid": 211, "correlation": 68207 + } + }, + { + "ph": "s", "id": 68207, "pid": 435, "tid": 435, "ts": 1776649791945552, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791945597, "dur": 1, + "args": { + "External id": 9075, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68225, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68225, "pid": 0, "tid": 7, "ts": 1776649791945597, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945588, "dur": 6, + "args": { + "External id": 9075, + "cbid": 211, "correlation": 68225 + } + }, + { + "ph": "s", "id": 68225, "pid": 435, "tid": 435, "ts": 1776649791945588, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945600, "dur": 0, + "args": { + "External id": 68231, + "cbid": 200, "correlation": 68231 + } + }, + { + "ph": "f", "id": 68231, "pid": 435, "tid": 435, "ts": 1776649791945600, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945601, "dur": 0, + "args": { + "External id": 68232, + "cbid": 273, "correlation": 68232 + } + }, + { + "ph": "f", "id": 68232, "pid": 435, "tid": 435, "ts": 1776649791945601, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791945610, "dur": 19, + "args": { + "External id": 68233, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68233, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68233, "pid": 0, "tid": 7, "ts": 1776649791945610, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945602, "dur": 4, + "args": { + "External id": 68233, + "cbid": 211, "correlation": 68233 + } + }, + { + "ph": "s", "id": 68233, "pid": 435, "tid": 435, "ts": 1776649791945602, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791945645, "dur": 1, + "args": { + "External id": 9080, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68251, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68251, "pid": 0, "tid": 7, "ts": 1776649791945645, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945636, "dur": 6, + "args": { + "External id": 9080, + "cbid": 211, "correlation": 68251 + } + }, + { + "ph": "s", "id": 68251, "pid": 435, "tid": 435, "ts": 1776649791945636, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945648, "dur": 0, + "args": { + "External id": 68257, + "cbid": 200, "correlation": 68257 + } + }, + { + "ph": "f", "id": 68257, "pid": 435, "tid": 435, "ts": 1776649791945648, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945649, "dur": 0, + "args": { + "External id": 68258, + "cbid": 273, "correlation": 68258 + } + }, + { + "ph": "f", "id": 68258, "pid": 435, "tid": 435, "ts": 1776649791945649, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791945658, "dur": 19, + "args": { + "External id": 68259, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68259, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68259, "pid": 0, "tid": 7, "ts": 1776649791945658, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945650, "dur": 5, + "args": { + "External id": 68259, + "cbid": 211, "correlation": 68259 + } + }, + { + "ph": "s", "id": 68259, "pid": 435, "tid": 435, "ts": 1776649791945650, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791945694, "dur": 1, + "args": { + "External id": 9085, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68277, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68277, "pid": 0, "tid": 7, "ts": 1776649791945694, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945684, "dur": 6, + "args": { + "External id": 9085, + "cbid": 211, "correlation": 68277 + } + }, + { + "ph": "s", "id": 68277, "pid": 435, "tid": 435, "ts": 1776649791945684, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945697, "dur": 0, + "args": { + "External id": 68283, + "cbid": 200, "correlation": 68283 + } + }, + { + "ph": "f", "id": 68283, "pid": 435, "tid": 435, "ts": 1776649791945697, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945697, "dur": 0, + "args": { + "External id": 68284, + "cbid": 273, "correlation": 68284 + } + }, + { + "ph": "f", "id": 68284, "pid": 435, "tid": 435, "ts": 1776649791945697, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791945706, "dur": 19, + "args": { + "External id": 68285, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68285, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68285, "pid": 0, "tid": 7, "ts": 1776649791945706, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945699, "dur": 4, + "args": { + "External id": 68285, + "cbid": 211, "correlation": 68285 + } + }, + { + "ph": "s", "id": 68285, "pid": 435, "tid": 435, "ts": 1776649791945699, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791945742, "dur": 1, + "args": { + "External id": 9090, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68303, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68303, "pid": 0, "tid": 7, "ts": 1776649791945742, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945732, "dur": 6, + "args": { + "External id": 9090, + "cbid": 211, "correlation": 68303 + } + }, + { + "ph": "s", "id": 68303, "pid": 435, "tid": 435, "ts": 1776649791945732, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945745, "dur": 0, + "args": { + "External id": 68309, + "cbid": 200, "correlation": 68309 + } + }, + { + "ph": "f", "id": 68309, "pid": 435, "tid": 435, "ts": 1776649791945745, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945745, "dur": 0, + "args": { + "External id": 68310, + "cbid": 273, "correlation": 68310 + } + }, + { + "ph": "f", "id": 68310, "pid": 435, "tid": 435, "ts": 1776649791945745, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791945754, "dur": 19, + "args": { + "External id": 68311, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68311, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68311, "pid": 0, "tid": 7, "ts": 1776649791945754, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945746, "dur": 5, + "args": { + "External id": 68311, + "cbid": 211, "correlation": 68311 + } + }, + { + "ph": "s", "id": 68311, "pid": 435, "tid": 435, "ts": 1776649791945746, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791945790, "dur": 1, + "args": { + "External id": 9095, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68329, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68329, "pid": 0, "tid": 7, "ts": 1776649791945790, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945780, "dur": 6, + "args": { + "External id": 9095, + "cbid": 211, "correlation": 68329 + } + }, + { + "ph": "s", "id": 68329, "pid": 435, "tid": 435, "ts": 1776649791945780, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945792, "dur": 0, + "args": { + "External id": 68335, + "cbid": 200, "correlation": 68335 + } + }, + { + "ph": "f", "id": 68335, "pid": 435, "tid": 435, "ts": 1776649791945792, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945793, "dur": 0, + "args": { + "External id": 68336, + "cbid": 273, "correlation": 68336 + } + }, + { + "ph": "f", "id": 68336, "pid": 435, "tid": 435, "ts": 1776649791945793, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791945802, "dur": 19, + "args": { + "External id": 68337, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68337, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68337, "pid": 0, "tid": 7, "ts": 1776649791945802, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945794, "dur": 4, + "args": { + "External id": 68337, + "cbid": 211, "correlation": 68337 + } + }, + { + "ph": "s", "id": 68337, "pid": 435, "tid": 435, "ts": 1776649791945794, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791945837, "dur": 1, + "args": { + "External id": 9100, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68355, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68355, "pid": 0, "tid": 7, "ts": 1776649791945837, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945828, "dur": 6, + "args": { + "External id": 9100, + "cbid": 211, "correlation": 68355 + } + }, + { + "ph": "s", "id": 68355, "pid": 435, "tid": 435, "ts": 1776649791945828, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945840, "dur": 0, + "args": { + "External id": 68361, + "cbid": 200, "correlation": 68361 + } + }, + { + "ph": "f", "id": 68361, "pid": 435, "tid": 435, "ts": 1776649791945840, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945840, "dur": 0, + "args": { + "External id": 68362, + "cbid": 273, "correlation": 68362 + } + }, + { + "ph": "f", "id": 68362, "pid": 435, "tid": 435, "ts": 1776649791945840, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791945849, "dur": 19, + "args": { + "External id": 68363, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68363, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68363, "pid": 0, "tid": 7, "ts": 1776649791945849, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945841, "dur": 5, + "args": { + "External id": 68363, + "cbid": 211, "correlation": 68363 + } + }, + { + "ph": "s", "id": 68363, "pid": 435, "tid": 435, "ts": 1776649791945841, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791945888, "dur": 1, + "args": { + "External id": 9105, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68381, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68381, "pid": 0, "tid": 7, "ts": 1776649791945888, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945879, "dur": 6, + "args": { + "External id": 9105, + "cbid": 211, "correlation": 68381 + } + }, + { + "ph": "s", "id": 68381, "pid": 435, "tid": 435, "ts": 1776649791945879, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945891, "dur": 0, + "args": { + "External id": 68387, + "cbid": 200, "correlation": 68387 + } + }, + { + "ph": "f", "id": 68387, "pid": 435, "tid": 435, "ts": 1776649791945891, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945892, "dur": 0, + "args": { + "External id": 68388, + "cbid": 273, "correlation": 68388 + } + }, + { + "ph": "f", "id": 68388, "pid": 435, "tid": 435, "ts": 1776649791945892, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791945900, "dur": 19, + "args": { + "External id": 68389, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68389, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68389, "pid": 0, "tid": 7, "ts": 1776649791945900, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945893, "dur": 4, + "args": { + "External id": 68389, + "cbid": 211, "correlation": 68389 + } + }, + { + "ph": "s", "id": 68389, "pid": 435, "tid": 435, "ts": 1776649791945893, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791945936, "dur": 1, + "args": { + "External id": 9110, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68407, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68407, "pid": 0, "tid": 7, "ts": 1776649791945936, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945927, "dur": 6, + "args": { + "External id": 9110, + "cbid": 211, "correlation": 68407 + } + }, + { + "ph": "s", "id": 68407, "pid": 435, "tid": 435, "ts": 1776649791945927, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945939, "dur": 0, + "args": { + "External id": 68413, + "cbid": 200, "correlation": 68413 + } + }, + { + "ph": "f", "id": 68413, "pid": 435, "tid": 435, "ts": 1776649791945939, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945940, "dur": 0, + "args": { + "External id": 68414, + "cbid": 273, "correlation": 68414 + } + }, + { + "ph": "f", "id": 68414, "pid": 435, "tid": 435, "ts": 1776649791945940, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791945949, "dur": 19, + "args": { + "External id": 68415, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68415, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68415, "pid": 0, "tid": 7, "ts": 1776649791945949, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945941, "dur": 4, + "args": { + "External id": 68415, + "cbid": 211, "correlation": 68415 + } + }, + { + "ph": "s", "id": 68415, "pid": 435, "tid": 435, "ts": 1776649791945941, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791945984, "dur": 1, + "args": { + "External id": 9115, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68433, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68433, "pid": 0, "tid": 7, "ts": 1776649791945984, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945975, "dur": 6, + "args": { + "External id": 9115, + "cbid": 211, "correlation": 68433 + } + }, + { + "ph": "s", "id": 68433, "pid": 435, "tid": 435, "ts": 1776649791945975, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945987, "dur": 0, + "args": { + "External id": 68439, + "cbid": 200, "correlation": 68439 + } + }, + { + "ph": "f", "id": 68439, "pid": 435, "tid": 435, "ts": 1776649791945987, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791945987, "dur": 0, + "args": { + "External id": 68440, + "cbid": 273, "correlation": 68440 + } + }, + { + "ph": "f", "id": 68440, "pid": 435, "tid": 435, "ts": 1776649791945987, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791945996, "dur": 19, + "args": { + "External id": 68441, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68441, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68441, "pid": 0, "tid": 7, "ts": 1776649791945996, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791945989, "dur": 4, + "args": { + "External id": 68441, + "cbid": 211, "correlation": 68441 + } + }, + { + "ph": "s", "id": 68441, "pid": 435, "tid": 435, "ts": 1776649791945989, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791946032, "dur": 1, + "args": { + "External id": 9120, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68459, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68459, "pid": 0, "tid": 7, "ts": 1776649791946032, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946022, "dur": 6, + "args": { + "External id": 9120, + "cbid": 211, "correlation": 68459 + } + }, + { + "ph": "s", "id": 68459, "pid": 435, "tid": 435, "ts": 1776649791946022, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946035, "dur": 0, + "args": { + "External id": 68465, + "cbid": 200, "correlation": 68465 + } + }, + { + "ph": "f", "id": 68465, "pid": 435, "tid": 435, "ts": 1776649791946035, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946035, "dur": 0, + "args": { + "External id": 68466, + "cbid": 273, "correlation": 68466 + } + }, + { + "ph": "f", "id": 68466, "pid": 435, "tid": 435, "ts": 1776649791946035, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791946044, "dur": 19, + "args": { + "External id": 68467, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68467, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68467, "pid": 0, "tid": 7, "ts": 1776649791946044, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946037, "dur": 5, + "args": { + "External id": 68467, + "cbid": 211, "correlation": 68467 + } + }, + { + "ph": "s", "id": 68467, "pid": 435, "tid": 435, "ts": 1776649791946037, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791946080, "dur": 1, + "args": { + "External id": 9125, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68485, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68485, "pid": 0, "tid": 7, "ts": 1776649791946080, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946070, "dur": 6, + "args": { + "External id": 9125, + "cbid": 211, "correlation": 68485 + } + }, + { + "ph": "s", "id": 68485, "pid": 435, "tid": 435, "ts": 1776649791946070, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946083, "dur": 0, + "args": { + "External id": 68491, + "cbid": 200, "correlation": 68491 + } + }, + { + "ph": "f", "id": 68491, "pid": 435, "tid": 435, "ts": 1776649791946083, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946083, "dur": 0, + "args": { + "External id": 68492, + "cbid": 273, "correlation": 68492 + } + }, + { + "ph": "f", "id": 68492, "pid": 435, "tid": 435, "ts": 1776649791946083, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791946093, "dur": 19, + "args": { + "External id": 68493, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68493, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68493, "pid": 0, "tid": 7, "ts": 1776649791946093, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946085, "dur": 5, + "args": { + "External id": 68493, + "cbid": 211, "correlation": 68493 + } + }, + { + "ph": "s", "id": 68493, "pid": 435, "tid": 435, "ts": 1776649791946085, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791946128, "dur": 1, + "args": { + "External id": 9130, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68511, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68511, "pid": 0, "tid": 7, "ts": 1776649791946128, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946119, "dur": 6, + "args": { + "External id": 9130, + "cbid": 211, "correlation": 68511 + } + }, + { + "ph": "s", "id": 68511, "pid": 435, "tid": 435, "ts": 1776649791946119, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946131, "dur": 0, + "args": { + "External id": 68517, + "cbid": 200, "correlation": 68517 + } + }, + { + "ph": "f", "id": 68517, "pid": 435, "tid": 435, "ts": 1776649791946131, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946132, "dur": 0, + "args": { + "External id": 68518, + "cbid": 273, "correlation": 68518 + } + }, + { + "ph": "f", "id": 68518, "pid": 435, "tid": 435, "ts": 1776649791946132, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791946140, "dur": 19, + "args": { + "External id": 68519, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68519, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68519, "pid": 0, "tid": 7, "ts": 1776649791946140, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946133, "dur": 4, + "args": { + "External id": 68519, + "cbid": 211, "correlation": 68519 + } + }, + { + "ph": "s", "id": 68519, "pid": 435, "tid": 435, "ts": 1776649791946133, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791946176, "dur": 1, + "args": { + "External id": 9135, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68537, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68537, "pid": 0, "tid": 7, "ts": 1776649791946176, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946166, "dur": 6, + "args": { + "External id": 9135, + "cbid": 211, "correlation": 68537 + } + }, + { + "ph": "s", "id": 68537, "pid": 435, "tid": 435, "ts": 1776649791946166, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946179, "dur": 0, + "args": { + "External id": 68543, + "cbid": 200, "correlation": 68543 + } + }, + { + "ph": "f", "id": 68543, "pid": 435, "tid": 435, "ts": 1776649791946179, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946179, "dur": 0, + "args": { + "External id": 68544, + "cbid": 273, "correlation": 68544 + } + }, + { + "ph": "f", "id": 68544, "pid": 435, "tid": 435, "ts": 1776649791946179, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791946188, "dur": 19, + "args": { + "External id": 68545, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68545, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68545, "pid": 0, "tid": 7, "ts": 1776649791946188, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946180, "dur": 4, + "args": { + "External id": 68545, + "cbid": 211, "correlation": 68545 + } + }, + { + "ph": "s", "id": 68545, "pid": 435, "tid": 435, "ts": 1776649791946180, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791946224, "dur": 1, + "args": { + "External id": 9140, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68563, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68563, "pid": 0, "tid": 7, "ts": 1776649791946224, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946214, "dur": 6, + "args": { + "External id": 9140, + "cbid": 211, "correlation": 68563 + } + }, + { + "ph": "s", "id": 68563, "pid": 435, "tid": 435, "ts": 1776649791946214, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946227, "dur": 0, + "args": { + "External id": 68569, + "cbid": 200, "correlation": 68569 + } + }, + { + "ph": "f", "id": 68569, "pid": 435, "tid": 435, "ts": 1776649791946227, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946228, "dur": 0, + "args": { + "External id": 68570, + "cbid": 273, "correlation": 68570 + } + }, + { + "ph": "f", "id": 68570, "pid": 435, "tid": 435, "ts": 1776649791946228, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791946237, "dur": 19, + "args": { + "External id": 68571, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68571, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68571, "pid": 0, "tid": 7, "ts": 1776649791946237, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946229, "dur": 5, + "args": { + "External id": 68571, + "cbid": 211, "correlation": 68571 + } + }, + { + "ph": "s", "id": 68571, "pid": 435, "tid": 435, "ts": 1776649791946229, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791946272, "dur": 1, + "args": { + "External id": 9145, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68589, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68589, "pid": 0, "tid": 7, "ts": 1776649791946272, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946262, "dur": 6, + "args": { + "External id": 9145, + "cbid": 211, "correlation": 68589 + } + }, + { + "ph": "s", "id": 68589, "pid": 435, "tid": 435, "ts": 1776649791946262, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946275, "dur": 0, + "args": { + "External id": 68595, + "cbid": 200, "correlation": 68595 + } + }, + { + "ph": "f", "id": 68595, "pid": 435, "tid": 435, "ts": 1776649791946275, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946276, "dur": 0, + "args": { + "External id": 68596, + "cbid": 273, "correlation": 68596 + } + }, + { + "ph": "f", "id": 68596, "pid": 435, "tid": 435, "ts": 1776649791946276, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791946284, "dur": 19, + "args": { + "External id": 68597, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68597, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68597, "pid": 0, "tid": 7, "ts": 1776649791946284, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946277, "dur": 5, + "args": { + "External id": 68597, + "cbid": 211, "correlation": 68597 + } + }, + { + "ph": "s", "id": 68597, "pid": 435, "tid": 435, "ts": 1776649791946277, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791946320, "dur": 1, + "args": { + "External id": 9150, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68615, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68615, "pid": 0, "tid": 7, "ts": 1776649791946320, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946311, "dur": 6, + "args": { + "External id": 9150, + "cbid": 211, "correlation": 68615 + } + }, + { + "ph": "s", "id": 68615, "pid": 435, "tid": 435, "ts": 1776649791946311, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946323, "dur": 0, + "args": { + "External id": 68621, + "cbid": 200, "correlation": 68621 + } + }, + { + "ph": "f", "id": 68621, "pid": 435, "tid": 435, "ts": 1776649791946323, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946324, "dur": 0, + "args": { + "External id": 68622, + "cbid": 273, "correlation": 68622 + } + }, + { + "ph": "f", "id": 68622, "pid": 435, "tid": 435, "ts": 1776649791946324, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791946333, "dur": 19, + "args": { + "External id": 68623, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68623, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68623, "pid": 0, "tid": 7, "ts": 1776649791946333, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946325, "dur": 5, + "args": { + "External id": 68623, + "cbid": 211, "correlation": 68623 + } + }, + { + "ph": "s", "id": 68623, "pid": 435, "tid": 435, "ts": 1776649791946325, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791946368, "dur": 1, + "args": { + "External id": 9155, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68641, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68641, "pid": 0, "tid": 7, "ts": 1776649791946368, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946359, "dur": 6, + "args": { + "External id": 9155, + "cbid": 211, "correlation": 68641 + } + }, + { + "ph": "s", "id": 68641, "pid": 435, "tid": 435, "ts": 1776649791946359, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946371, "dur": 0, + "args": { + "External id": 68647, + "cbid": 200, "correlation": 68647 + } + }, + { + "ph": "f", "id": 68647, "pid": 435, "tid": 435, "ts": 1776649791946371, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946372, "dur": 0, + "args": { + "External id": 68648, + "cbid": 273, "correlation": 68648 + } + }, + { + "ph": "f", "id": 68648, "pid": 435, "tid": 435, "ts": 1776649791946372, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791946381, "dur": 19, + "args": { + "External id": 68649, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68649, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68649, "pid": 0, "tid": 7, "ts": 1776649791946381, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946373, "dur": 5, + "args": { + "External id": 68649, + "cbid": 211, "correlation": 68649 + } + }, + { + "ph": "s", "id": 68649, "pid": 435, "tid": 435, "ts": 1776649791946373, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791946416, "dur": 1, + "args": { + "External id": 9160, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68667, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68667, "pid": 0, "tid": 7, "ts": 1776649791946416, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946407, "dur": 6, + "args": { + "External id": 9160, + "cbid": 211, "correlation": 68667 + } + }, + { + "ph": "s", "id": 68667, "pid": 435, "tid": 435, "ts": 1776649791946407, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946419, "dur": 0, + "args": { + "External id": 68673, + "cbid": 200, "correlation": 68673 + } + }, + { + "ph": "f", "id": 68673, "pid": 435, "tid": 435, "ts": 1776649791946419, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946420, "dur": 0, + "args": { + "External id": 68674, + "cbid": 273, "correlation": 68674 + } + }, + { + "ph": "f", "id": 68674, "pid": 435, "tid": 435, "ts": 1776649791946420, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791946429, "dur": 19, + "args": { + "External id": 68675, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68675, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68675, "pid": 0, "tid": 7, "ts": 1776649791946429, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946421, "dur": 4, + "args": { + "External id": 68675, + "cbid": 211, "correlation": 68675 + } + }, + { + "ph": "s", "id": 68675, "pid": 435, "tid": 435, "ts": 1776649791946421, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791946464, "dur": 1, + "args": { + "External id": 9165, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68693, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68693, "pid": 0, "tid": 7, "ts": 1776649791946464, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946454, "dur": 6, + "args": { + "External id": 9165, + "cbid": 211, "correlation": 68693 + } + }, + { + "ph": "s", "id": 68693, "pid": 435, "tid": 435, "ts": 1776649791946454, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946467, "dur": 0, + "args": { + "External id": 68699, + "cbid": 200, "correlation": 68699 + } + }, + { + "ph": "f", "id": 68699, "pid": 435, "tid": 435, "ts": 1776649791946467, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946468, "dur": 0, + "args": { + "External id": 68700, + "cbid": 273, "correlation": 68700 + } + }, + { + "ph": "f", "id": 68700, "pid": 435, "tid": 435, "ts": 1776649791946468, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791946477, "dur": 19, + "args": { + "External id": 68701, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68701, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68701, "pid": 0, "tid": 7, "ts": 1776649791946477, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946469, "dur": 5, + "args": { + "External id": 68701, + "cbid": 211, "correlation": 68701 + } + }, + { + "ph": "s", "id": 68701, "pid": 435, "tid": 435, "ts": 1776649791946469, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791946512, "dur": 1, + "args": { + "External id": 9170, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68719, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68719, "pid": 0, "tid": 7, "ts": 1776649791946512, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946503, "dur": 6, + "args": { + "External id": 9170, + "cbid": 211, "correlation": 68719 + } + }, + { + "ph": "s", "id": 68719, "pid": 435, "tid": 435, "ts": 1776649791946503, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946515, "dur": 0, + "args": { + "External id": 68725, + "cbid": 200, "correlation": 68725 + } + }, + { + "ph": "f", "id": 68725, "pid": 435, "tid": 435, "ts": 1776649791946515, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946516, "dur": 0, + "args": { + "External id": 68726, + "cbid": 273, "correlation": 68726 + } + }, + { + "ph": "f", "id": 68726, "pid": 435, "tid": 435, "ts": 1776649791946516, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791946525, "dur": 19, + "args": { + "External id": 68727, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68727, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68727, "pid": 0, "tid": 7, "ts": 1776649791946525, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946517, "dur": 5, + "args": { + "External id": 68727, + "cbid": 211, "correlation": 68727 + } + }, + { + "ph": "s", "id": 68727, "pid": 435, "tid": 435, "ts": 1776649791946517, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791946560, "dur": 1, + "args": { + "External id": 9175, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68745, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68745, "pid": 0, "tid": 7, "ts": 1776649791946560, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946550, "dur": 6, + "args": { + "External id": 9175, + "cbid": 211, "correlation": 68745 + } + }, + { + "ph": "s", "id": 68745, "pid": 435, "tid": 435, "ts": 1776649791946550, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946563, "dur": 0, + "args": { + "External id": 68751, + "cbid": 200, "correlation": 68751 + } + }, + { + "ph": "f", "id": 68751, "pid": 435, "tid": 435, "ts": 1776649791946563, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946563, "dur": 0, + "args": { + "External id": 68752, + "cbid": 273, "correlation": 68752 + } + }, + { + "ph": "f", "id": 68752, "pid": 435, "tid": 435, "ts": 1776649791946563, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791946572, "dur": 19, + "args": { + "External id": 68753, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68753, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68753, "pid": 0, "tid": 7, "ts": 1776649791946572, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946564, "dur": 4, + "args": { + "External id": 68753, + "cbid": 211, "correlation": 68753 + } + }, + { + "ph": "s", "id": 68753, "pid": 435, "tid": 435, "ts": 1776649791946564, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791946608, "dur": 1, + "args": { + "External id": 9180, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68771, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68771, "pid": 0, "tid": 7, "ts": 1776649791946608, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946599, "dur": 6, + "args": { + "External id": 9180, + "cbid": 211, "correlation": 68771 + } + }, + { + "ph": "s", "id": 68771, "pid": 435, "tid": 435, "ts": 1776649791946599, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946611, "dur": 0, + "args": { + "External id": 68777, + "cbid": 200, "correlation": 68777 + } + }, + { + "ph": "f", "id": 68777, "pid": 435, "tid": 435, "ts": 1776649791946611, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946612, "dur": 0, + "args": { + "External id": 68778, + "cbid": 273, "correlation": 68778 + } + }, + { + "ph": "f", "id": 68778, "pid": 435, "tid": 435, "ts": 1776649791946612, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791946621, "dur": 19, + "args": { + "External id": 68779, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68779, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68779, "pid": 0, "tid": 7, "ts": 1776649791946621, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946613, "dur": 4, + "args": { + "External id": 68779, + "cbid": 211, "correlation": 68779 + } + }, + { + "ph": "s", "id": 68779, "pid": 435, "tid": 435, "ts": 1776649791946613, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791946656, "dur": 1, + "args": { + "External id": 9185, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68797, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68797, "pid": 0, "tid": 7, "ts": 1776649791946656, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946647, "dur": 6, + "args": { + "External id": 9185, + "cbid": 211, "correlation": 68797 + } + }, + { + "ph": "s", "id": 68797, "pid": 435, "tid": 435, "ts": 1776649791946647, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946659, "dur": 0, + "args": { + "External id": 68803, + "cbid": 200, "correlation": 68803 + } + }, + { + "ph": "f", "id": 68803, "pid": 435, "tid": 435, "ts": 1776649791946659, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946660, "dur": 0, + "args": { + "External id": 68804, + "cbid": 273, "correlation": 68804 + } + }, + { + "ph": "f", "id": 68804, "pid": 435, "tid": 435, "ts": 1776649791946660, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791946669, "dur": 19, + "args": { + "External id": 68805, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68805, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68805, "pid": 0, "tid": 7, "ts": 1776649791946669, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946661, "dur": 4, + "args": { + "External id": 68805, + "cbid": 211, "correlation": 68805 + } + }, + { + "ph": "s", "id": 68805, "pid": 435, "tid": 435, "ts": 1776649791946661, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791946704, "dur": 1, + "args": { + "External id": 9190, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68823, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68823, "pid": 0, "tid": 7, "ts": 1776649791946704, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946695, "dur": 6, + "args": { + "External id": 9190, + "cbid": 211, "correlation": 68823 + } + }, + { + "ph": "s", "id": 68823, "pid": 435, "tid": 435, "ts": 1776649791946695, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946707, "dur": 0, + "args": { + "External id": 68829, + "cbid": 200, "correlation": 68829 + } + }, + { + "ph": "f", "id": 68829, "pid": 435, "tid": 435, "ts": 1776649791946707, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946708, "dur": 0, + "args": { + "External id": 68830, + "cbid": 273, "correlation": 68830 + } + }, + { + "ph": "f", "id": 68830, "pid": 435, "tid": 435, "ts": 1776649791946708, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791946717, "dur": 19, + "args": { + "External id": 68831, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68831, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68831, "pid": 0, "tid": 7, "ts": 1776649791946717, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946709, "dur": 4, + "args": { + "External id": 68831, + "cbid": 211, "correlation": 68831 + } + }, + { + "ph": "s", "id": 68831, "pid": 435, "tid": 435, "ts": 1776649791946709, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791946753, "dur": 1, + "args": { + "External id": 9195, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68849, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68849, "pid": 0, "tid": 7, "ts": 1776649791946753, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946744, "dur": 6, + "args": { + "External id": 9195, + "cbid": 211, "correlation": 68849 + } + }, + { + "ph": "s", "id": 68849, "pid": 435, "tid": 435, "ts": 1776649791946744, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946756, "dur": 0, + "args": { + "External id": 68855, + "cbid": 200, "correlation": 68855 + } + }, + { + "ph": "f", "id": 68855, "pid": 435, "tid": 435, "ts": 1776649791946756, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946757, "dur": 0, + "args": { + "External id": 68856, + "cbid": 273, "correlation": 68856 + } + }, + { + "ph": "f", "id": 68856, "pid": 435, "tid": 435, "ts": 1776649791946757, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791946766, "dur": 19, + "args": { + "External id": 68857, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68857, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68857, "pid": 0, "tid": 7, "ts": 1776649791946766, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946758, "dur": 5, + "args": { + "External id": 68857, + "cbid": 211, "correlation": 68857 + } + }, + { + "ph": "s", "id": 68857, "pid": 435, "tid": 435, "ts": 1776649791946758, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791946802, "dur": 1, + "args": { + "External id": 9200, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68875, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68875, "pid": 0, "tid": 7, "ts": 1776649791946802, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946793, "dur": 7, + "args": { + "External id": 9200, + "cbid": 211, "correlation": 68875 + } + }, + { + "ph": "s", "id": 68875, "pid": 435, "tid": 435, "ts": 1776649791946793, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946805, "dur": 0, + "args": { + "External id": 68881, + "cbid": 200, "correlation": 68881 + } + }, + { + "ph": "f", "id": 68881, "pid": 435, "tid": 435, "ts": 1776649791946805, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946806, "dur": 0, + "args": { + "External id": 68882, + "cbid": 273, "correlation": 68882 + } + }, + { + "ph": "f", "id": 68882, "pid": 435, "tid": 435, "ts": 1776649791946806, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791946815, "dur": 19, + "args": { + "External id": 68883, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68883, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68883, "pid": 0, "tid": 7, "ts": 1776649791946815, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946807, "dur": 5, + "args": { + "External id": 68883, + "cbid": 211, "correlation": 68883 + } + }, + { + "ph": "s", "id": 68883, "pid": 435, "tid": 435, "ts": 1776649791946807, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791946851, "dur": 1, + "args": { + "External id": 9205, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68901, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68901, "pid": 0, "tid": 7, "ts": 1776649791946851, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946841, "dur": 6, + "args": { + "External id": 9205, + "cbid": 211, "correlation": 68901 + } + }, + { + "ph": "s", "id": 68901, "pid": 435, "tid": 435, "ts": 1776649791946841, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946854, "dur": 0, + "args": { + "External id": 68907, + "cbid": 200, "correlation": 68907 + } + }, + { + "ph": "f", "id": 68907, "pid": 435, "tid": 435, "ts": 1776649791946854, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946854, "dur": 0, + "args": { + "External id": 68908, + "cbid": 273, "correlation": 68908 + } + }, + { + "ph": "f", "id": 68908, "pid": 435, "tid": 435, "ts": 1776649791946854, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791946864, "dur": 19, + "args": { + "External id": 68909, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68909, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68909, "pid": 0, "tid": 7, "ts": 1776649791946864, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946855, "dur": 5, + "args": { + "External id": 68909, + "cbid": 211, "correlation": 68909 + } + }, + { + "ph": "s", "id": 68909, "pid": 435, "tid": 435, "ts": 1776649791946855, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791946886, "dur": 3, + "args": { + "External id": 68917, + "cbid": 135, "correlation": 68917 + } + }, + { + "ph": "f", "id": 68917, "pid": 435, "tid": 435, "ts": 1776649791946886, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791946900, "dur": 5, + "args": { + "External id": 68922, + "cbid": 165, "correlation": 68922 + } + }, + { + "ph": "s", "id": 68922, "pid": 435, "tid": 435, "ts": 1776649791946900, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649791946909, "dur": 5, + "args": { + "External id": 68925, + "cbid": 139, "correlation": 68925 + } + }, + { + "ph": "f", "id": 68925, "pid": 435, "tid": 435, "ts": 1776649791946909, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791946964, "dur": 1, + "args": { + "External id": 9210, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68953, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68953, "pid": 0, "tid": 7, "ts": 1776649791946964, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946953, "dur": 8, + "args": { + "External id": 9210, + "cbid": 211, "correlation": 68953 + } + }, + { + "ph": "s", "id": 68953, "pid": 435, "tid": 435, "ts": 1776649791946953, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946969, "dur": 0, + "args": { + "External id": 68959, + "cbid": 200, "correlation": 68959 + } + }, + { + "ph": "f", "id": 68959, "pid": 435, "tid": 435, "ts": 1776649791946969, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791946970, "dur": 0, + "args": { + "External id": 68960, + "cbid": 273, "correlation": 68960 + } + }, + { + "ph": "f", "id": 68960, "pid": 435, "tid": 435, "ts": 1776649791946970, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791946979, "dur": 19, + "args": { + "External id": 68961, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68961, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68961, "pid": 0, "tid": 7, "ts": 1776649791946979, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791946971, "dur": 5, + "args": { + "External id": 68961, + "cbid": 211, "correlation": 68961 + } + }, + { + "ph": "s", "id": 68961, "pid": 435, "tid": 435, "ts": 1776649791946971, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791947017, "dur": 1, + "args": { + "External id": 9215, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68979, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68979, "pid": 0, "tid": 7, "ts": 1776649791947017, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947008, "dur": 6, + "args": { + "External id": 9215, + "cbid": 211, "correlation": 68979 + } + }, + { + "ph": "s", "id": 68979, "pid": 435, "tid": 435, "ts": 1776649791947008, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947020, "dur": 0, + "args": { + "External id": 68985, + "cbid": 200, "correlation": 68985 + } + }, + { + "ph": "f", "id": 68985, "pid": 435, "tid": 435, "ts": 1776649791947020, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947021, "dur": 0, + "args": { + "External id": 68986, + "cbid": 273, "correlation": 68986 + } + }, + { + "ph": "f", "id": 68986, "pid": 435, "tid": 435, "ts": 1776649791947021, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791947030, "dur": 19, + "args": { + "External id": 68987, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 68987, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 68987, "pid": 0, "tid": 7, "ts": 1776649791947030, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947022, "dur": 4, + "args": { + "External id": 68987, + "cbid": 211, "correlation": 68987 + } + }, + { + "ph": "s", "id": 68987, "pid": 435, "tid": 435, "ts": 1776649791947022, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791947077, "dur": 1, + "args": { + "External id": 9220, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69005, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69005, "pid": 0, "tid": 7, "ts": 1776649791947077, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947067, "dur": 7, + "args": { + "External id": 9220, + "cbid": 211, "correlation": 69005 + } + }, + { + "ph": "s", "id": 69005, "pid": 435, "tid": 435, "ts": 1776649791947067, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947080, "dur": 0, + "args": { + "External id": 69011, + "cbid": 200, "correlation": 69011 + } + }, + { + "ph": "f", "id": 69011, "pid": 435, "tid": 435, "ts": 1776649791947080, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947081, "dur": 0, + "args": { + "External id": 69012, + "cbid": 273, "correlation": 69012 + } + }, + { + "ph": "f", "id": 69012, "pid": 435, "tid": 435, "ts": 1776649791947081, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791947090, "dur": 19, + "args": { + "External id": 69013, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69013, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69013, "pid": 0, "tid": 7, "ts": 1776649791947090, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947082, "dur": 5, + "args": { + "External id": 69013, + "cbid": 211, "correlation": 69013 + } + }, + { + "ph": "s", "id": 69013, "pid": 435, "tid": 435, "ts": 1776649791947082, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791947134, "dur": 1, + "args": { + "External id": 9225, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69031, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69031, "pid": 0, "tid": 7, "ts": 1776649791947134, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947124, "dur": 7, + "args": { + "External id": 9225, + "cbid": 211, "correlation": 69031 + } + }, + { + "ph": "s", "id": 69031, "pid": 435, "tid": 435, "ts": 1776649791947124, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947137, "dur": 0, + "args": { + "External id": 69037, + "cbid": 200, "correlation": 69037 + } + }, + { + "ph": "f", "id": 69037, "pid": 435, "tid": 435, "ts": 1776649791947137, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947138, "dur": 0, + "args": { + "External id": 69038, + "cbid": 273, "correlation": 69038 + } + }, + { + "ph": "f", "id": 69038, "pid": 435, "tid": 435, "ts": 1776649791947138, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791947147, "dur": 19, + "args": { + "External id": 69039, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69039, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69039, "pid": 0, "tid": 7, "ts": 1776649791947147, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947139, "dur": 4, + "args": { + "External id": 69039, + "cbid": 211, "correlation": 69039 + } + }, + { + "ph": "s", "id": 69039, "pid": 435, "tid": 435, "ts": 1776649791947139, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791947183, "dur": 1, + "args": { + "External id": 9230, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69057, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69057, "pid": 0, "tid": 7, "ts": 1776649791947183, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947173, "dur": 6, + "args": { + "External id": 9230, + "cbid": 211, "correlation": 69057 + } + }, + { + "ph": "s", "id": 69057, "pid": 435, "tid": 435, "ts": 1776649791947173, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947186, "dur": 0, + "args": { + "External id": 69063, + "cbid": 200, "correlation": 69063 + } + }, + { + "ph": "f", "id": 69063, "pid": 435, "tid": 435, "ts": 1776649791947186, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947187, "dur": 0, + "args": { + "External id": 69064, + "cbid": 273, "correlation": 69064 + } + }, + { + "ph": "f", "id": 69064, "pid": 435, "tid": 435, "ts": 1776649791947187, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791947196, "dur": 19, + "args": { + "External id": 69065, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69065, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69065, "pid": 0, "tid": 7, "ts": 1776649791947196, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947188, "dur": 4, + "args": { + "External id": 69065, + "cbid": 211, "correlation": 69065 + } + }, + { + "ph": "s", "id": 69065, "pid": 435, "tid": 435, "ts": 1776649791947188, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791947231, "dur": 1, + "args": { + "External id": 9235, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69083, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69083, "pid": 0, "tid": 7, "ts": 1776649791947231, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947222, "dur": 6, + "args": { + "External id": 9235, + "cbid": 211, "correlation": 69083 + } + }, + { + "ph": "s", "id": 69083, "pid": 435, "tid": 435, "ts": 1776649791947222, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947235, "dur": 0, + "args": { + "External id": 69089, + "cbid": 200, "correlation": 69089 + } + }, + { + "ph": "f", "id": 69089, "pid": 435, "tid": 435, "ts": 1776649791947235, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947235, "dur": 0, + "args": { + "External id": 69090, + "cbid": 273, "correlation": 69090 + } + }, + { + "ph": "f", "id": 69090, "pid": 435, "tid": 435, "ts": 1776649791947235, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791947244, "dur": 19, + "args": { + "External id": 69091, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69091, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69091, "pid": 0, "tid": 7, "ts": 1776649791947244, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947236, "dur": 4, + "args": { + "External id": 69091, + "cbid": 211, "correlation": 69091 + } + }, + { + "ph": "s", "id": 69091, "pid": 435, "tid": 435, "ts": 1776649791947236, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791947279, "dur": 1, + "args": { + "External id": 9240, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69109, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69109, "pid": 0, "tid": 7, "ts": 1776649791947279, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947270, "dur": 6, + "args": { + "External id": 9240, + "cbid": 211, "correlation": 69109 + } + }, + { + "ph": "s", "id": 69109, "pid": 435, "tid": 435, "ts": 1776649791947270, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947282, "dur": 0, + "args": { + "External id": 69115, + "cbid": 200, "correlation": 69115 + } + }, + { + "ph": "f", "id": 69115, "pid": 435, "tid": 435, "ts": 1776649791947282, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947283, "dur": 0, + "args": { + "External id": 69116, + "cbid": 273, "correlation": 69116 + } + }, + { + "ph": "f", "id": 69116, "pid": 435, "tid": 435, "ts": 1776649791947283, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791947292, "dur": 19, + "args": { + "External id": 69117, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69117, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69117, "pid": 0, "tid": 7, "ts": 1776649791947292, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947284, "dur": 4, + "args": { + "External id": 69117, + "cbid": 211, "correlation": 69117 + } + }, + { + "ph": "s", "id": 69117, "pid": 435, "tid": 435, "ts": 1776649791947284, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791947327, "dur": 1, + "args": { + "External id": 9245, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69135, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69135, "pid": 0, "tid": 7, "ts": 1776649791947327, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947318, "dur": 6, + "args": { + "External id": 9245, + "cbid": 211, "correlation": 69135 + } + }, + { + "ph": "s", "id": 69135, "pid": 435, "tid": 435, "ts": 1776649791947318, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947330, "dur": 0, + "args": { + "External id": 69141, + "cbid": 200, "correlation": 69141 + } + }, + { + "ph": "f", "id": 69141, "pid": 435, "tid": 435, "ts": 1776649791947330, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947331, "dur": 0, + "args": { + "External id": 69142, + "cbid": 273, "correlation": 69142 + } + }, + { + "ph": "f", "id": 69142, "pid": 435, "tid": 435, "ts": 1776649791947331, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791947340, "dur": 19, + "args": { + "External id": 69143, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69143, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69143, "pid": 0, "tid": 7, "ts": 1776649791947340, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947332, "dur": 4, + "args": { + "External id": 69143, + "cbid": 211, "correlation": 69143 + } + }, + { + "ph": "s", "id": 69143, "pid": 435, "tid": 435, "ts": 1776649791947332, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791947375, "dur": 1, + "args": { + "External id": 9250, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69161, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69161, "pid": 0, "tid": 7, "ts": 1776649791947375, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947366, "dur": 6, + "args": { + "External id": 9250, + "cbid": 211, "correlation": 69161 + } + }, + { + "ph": "s", "id": 69161, "pid": 435, "tid": 435, "ts": 1776649791947366, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947378, "dur": 0, + "args": { + "External id": 69167, + "cbid": 200, "correlation": 69167 + } + }, + { + "ph": "f", "id": 69167, "pid": 435, "tid": 435, "ts": 1776649791947378, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947379, "dur": 0, + "args": { + "External id": 69168, + "cbid": 273, "correlation": 69168 + } + }, + { + "ph": "f", "id": 69168, "pid": 435, "tid": 435, "ts": 1776649791947379, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791947388, "dur": 19, + "args": { + "External id": 69169, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69169, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69169, "pid": 0, "tid": 7, "ts": 1776649791947388, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947380, "dur": 4, + "args": { + "External id": 69169, + "cbid": 211, "correlation": 69169 + } + }, + { + "ph": "s", "id": 69169, "pid": 435, "tid": 435, "ts": 1776649791947380, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791947423, "dur": 1, + "args": { + "External id": 9255, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69187, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69187, "pid": 0, "tid": 7, "ts": 1776649791947423, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947414, "dur": 6, + "args": { + "External id": 9255, + "cbid": 211, "correlation": 69187 + } + }, + { + "ph": "s", "id": 69187, "pid": 435, "tid": 435, "ts": 1776649791947414, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947426, "dur": 0, + "args": { + "External id": 69193, + "cbid": 200, "correlation": 69193 + } + }, + { + "ph": "f", "id": 69193, "pid": 435, "tid": 435, "ts": 1776649791947426, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947427, "dur": 0, + "args": { + "External id": 69194, + "cbid": 273, "correlation": 69194 + } + }, + { + "ph": "f", "id": 69194, "pid": 435, "tid": 435, "ts": 1776649791947427, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791947436, "dur": 19, + "args": { + "External id": 69195, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69195, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69195, "pid": 0, "tid": 7, "ts": 1776649791947436, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947428, "dur": 4, + "args": { + "External id": 69195, + "cbid": 211, "correlation": 69195 + } + }, + { + "ph": "s", "id": 69195, "pid": 435, "tid": 435, "ts": 1776649791947428, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791947449, "dur": 10, + "args": { + "External id": 69198, + "cbid": 165, "correlation": 69198 + } + }, + { + "ph": "s", "id": 69198, "pid": 435, "tid": 435, "ts": 1776649791947449, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791947480, "dur": 3, + "args": { + "External id": 69207, + "cbid": 135, "correlation": 69207 + } + }, + { + "ph": "f", "id": 69207, "pid": 435, "tid": 435, "ts": 1776649791947480, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791947529, "dur": 1, + "args": { + "External id": 9260, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69227, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69227, "pid": 0, "tid": 7, "ts": 1776649791947529, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947518, "dur": 8, + "args": { + "External id": 9260, + "cbid": 211, "correlation": 69227 + } + }, + { + "ph": "s", "id": 69227, "pid": 435, "tid": 435, "ts": 1776649791947518, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947533, "dur": 0, + "args": { + "External id": 69233, + "cbid": 200, "correlation": 69233 + } + }, + { + "ph": "f", "id": 69233, "pid": 435, "tid": 435, "ts": 1776649791947533, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947533, "dur": 0, + "args": { + "External id": 69234, + "cbid": 273, "correlation": 69234 + } + }, + { + "ph": "f", "id": 69234, "pid": 435, "tid": 435, "ts": 1776649791947533, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791947543, "dur": 19, + "args": { + "External id": 69235, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69235, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69235, "pid": 0, "tid": 7, "ts": 1776649791947543, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947535, "dur": 5, + "args": { + "External id": 69235, + "cbid": 211, "correlation": 69235 + } + }, + { + "ph": "s", "id": 69235, "pid": 435, "tid": 435, "ts": 1776649791947535, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791947580, "dur": 1, + "args": { + "External id": 9265, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69253, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69253, "pid": 0, "tid": 7, "ts": 1776649791947580, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947571, "dur": 7, + "args": { + "External id": 9265, + "cbid": 211, "correlation": 69253 + } + }, + { + "ph": "s", "id": 69253, "pid": 435, "tid": 435, "ts": 1776649791947571, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947584, "dur": 0, + "args": { + "External id": 69259, + "cbid": 200, "correlation": 69259 + } + }, + { + "ph": "f", "id": 69259, "pid": 435, "tid": 435, "ts": 1776649791947584, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947584, "dur": 0, + "args": { + "External id": 69260, + "cbid": 273, "correlation": 69260 + } + }, + { + "ph": "f", "id": 69260, "pid": 435, "tid": 435, "ts": 1776649791947584, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791947593, "dur": 19, + "args": { + "External id": 69261, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69261, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69261, "pid": 0, "tid": 7, "ts": 1776649791947593, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947585, "dur": 5, + "args": { + "External id": 69261, + "cbid": 211, "correlation": 69261 + } + }, + { + "ph": "s", "id": 69261, "pid": 435, "tid": 435, "ts": 1776649791947585, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791947629, "dur": 1, + "args": { + "External id": 9270, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69279, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69279, "pid": 0, "tid": 7, "ts": 1776649791947629, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947620, "dur": 6, + "args": { + "External id": 9270, + "cbid": 211, "correlation": 69279 + } + }, + { + "ph": "s", "id": 69279, "pid": 435, "tid": 435, "ts": 1776649791947620, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947632, "dur": 0, + "args": { + "External id": 69285, + "cbid": 200, "correlation": 69285 + } + }, + { + "ph": "f", "id": 69285, "pid": 435, "tid": 435, "ts": 1776649791947632, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947633, "dur": 0, + "args": { + "External id": 69286, + "cbid": 273, "correlation": 69286 + } + }, + { + "ph": "f", "id": 69286, "pid": 435, "tid": 435, "ts": 1776649791947633, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791947642, "dur": 19, + "args": { + "External id": 69287, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69287, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69287, "pid": 0, "tid": 7, "ts": 1776649791947642, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947634, "dur": 4, + "args": { + "External id": 69287, + "cbid": 211, "correlation": 69287 + } + }, + { + "ph": "s", "id": 69287, "pid": 435, "tid": 435, "ts": 1776649791947634, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791947678, "dur": 1, + "args": { + "External id": 9275, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69305, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69305, "pid": 0, "tid": 7, "ts": 1776649791947678, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947668, "dur": 6, + "args": { + "External id": 9275, + "cbid": 211, "correlation": 69305 + } + }, + { + "ph": "s", "id": 69305, "pid": 435, "tid": 435, "ts": 1776649791947668, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947681, "dur": 0, + "args": { + "External id": 69311, + "cbid": 200, "correlation": 69311 + } + }, + { + "ph": "f", "id": 69311, "pid": 435, "tid": 435, "ts": 1776649791947681, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947682, "dur": 0, + "args": { + "External id": 69312, + "cbid": 273, "correlation": 69312 + } + }, + { + "ph": "f", "id": 69312, "pid": 435, "tid": 435, "ts": 1776649791947682, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791947691, "dur": 19, + "args": { + "External id": 69313, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69313, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69313, "pid": 0, "tid": 7, "ts": 1776649791947691, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947683, "dur": 5, + "args": { + "External id": 69313, + "cbid": 211, "correlation": 69313 + } + }, + { + "ph": "s", "id": 69313, "pid": 435, "tid": 435, "ts": 1776649791947683, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791947726, "dur": 1, + "args": { + "External id": 9280, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69331, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69331, "pid": 0, "tid": 7, "ts": 1776649791947726, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947717, "dur": 6, + "args": { + "External id": 9280, + "cbid": 211, "correlation": 69331 + } + }, + { + "ph": "s", "id": 69331, "pid": 435, "tid": 435, "ts": 1776649791947717, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947729, "dur": 0, + "args": { + "External id": 69337, + "cbid": 200, "correlation": 69337 + } + }, + { + "ph": "f", "id": 69337, "pid": 435, "tid": 435, "ts": 1776649791947729, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947730, "dur": 0, + "args": { + "External id": 69338, + "cbid": 273, "correlation": 69338 + } + }, + { + "ph": "f", "id": 69338, "pid": 435, "tid": 435, "ts": 1776649791947730, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791947739, "dur": 19, + "args": { + "External id": 69339, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69339, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69339, "pid": 0, "tid": 7, "ts": 1776649791947739, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947731, "dur": 4, + "args": { + "External id": 69339, + "cbid": 211, "correlation": 69339 + } + }, + { + "ph": "s", "id": 69339, "pid": 435, "tid": 435, "ts": 1776649791947731, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791947775, "dur": 1, + "args": { + "External id": 9285, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69357, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69357, "pid": 0, "tid": 7, "ts": 1776649791947775, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947766, "dur": 6, + "args": { + "External id": 9285, + "cbid": 211, "correlation": 69357 + } + }, + { + "ph": "s", "id": 69357, "pid": 435, "tid": 435, "ts": 1776649791947766, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947778, "dur": 0, + "args": { + "External id": 69363, + "cbid": 200, "correlation": 69363 + } + }, + { + "ph": "f", "id": 69363, "pid": 435, "tid": 435, "ts": 1776649791947778, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947779, "dur": 0, + "args": { + "External id": 69364, + "cbid": 273, "correlation": 69364 + } + }, + { + "ph": "f", "id": 69364, "pid": 435, "tid": 435, "ts": 1776649791947779, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791947788, "dur": 19, + "args": { + "External id": 69365, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69365, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69365, "pid": 0, "tid": 7, "ts": 1776649791947788, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947780, "dur": 4, + "args": { + "External id": 69365, + "cbid": 211, "correlation": 69365 + } + }, + { + "ph": "s", "id": 69365, "pid": 435, "tid": 435, "ts": 1776649791947780, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791947824, "dur": 1, + "args": { + "External id": 9290, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69383, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69383, "pid": 0, "tid": 7, "ts": 1776649791947824, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947815, "dur": 6, + "args": { + "External id": 9290, + "cbid": 211, "correlation": 69383 + } + }, + { + "ph": "s", "id": 69383, "pid": 435, "tid": 435, "ts": 1776649791947815, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947827, "dur": 0, + "args": { + "External id": 69389, + "cbid": 200, "correlation": 69389 + } + }, + { + "ph": "f", "id": 69389, "pid": 435, "tid": 435, "ts": 1776649791947827, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947828, "dur": 0, + "args": { + "External id": 69390, + "cbid": 273, "correlation": 69390 + } + }, + { + "ph": "f", "id": 69390, "pid": 435, "tid": 435, "ts": 1776649791947828, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791947837, "dur": 19, + "args": { + "External id": 69391, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69391, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69391, "pid": 0, "tid": 7, "ts": 1776649791947837, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947829, "dur": 5, + "args": { + "External id": 69391, + "cbid": 211, "correlation": 69391 + } + }, + { + "ph": "s", "id": 69391, "pid": 435, "tid": 435, "ts": 1776649791947829, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791947872, "dur": 1, + "args": { + "External id": 9295, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69409, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69409, "pid": 0, "tid": 7, "ts": 1776649791947872, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947863, "dur": 6, + "args": { + "External id": 9295, + "cbid": 211, "correlation": 69409 + } + }, + { + "ph": "s", "id": 69409, "pid": 435, "tid": 435, "ts": 1776649791947863, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947875, "dur": 0, + "args": { + "External id": 69415, + "cbid": 200, "correlation": 69415 + } + }, + { + "ph": "f", "id": 69415, "pid": 435, "tid": 435, "ts": 1776649791947875, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947875, "dur": 0, + "args": { + "External id": 69416, + "cbid": 273, "correlation": 69416 + } + }, + { + "ph": "f", "id": 69416, "pid": 435, "tid": 435, "ts": 1776649791947875, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791947884, "dur": 19, + "args": { + "External id": 69417, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69417, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69417, "pid": 0, "tid": 7, "ts": 1776649791947884, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947877, "dur": 5, + "args": { + "External id": 69417, + "cbid": 211, "correlation": 69417 + } + }, + { + "ph": "s", "id": 69417, "pid": 435, "tid": 435, "ts": 1776649791947877, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791947920, "dur": 1, + "args": { + "External id": 9300, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69435, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69435, "pid": 0, "tid": 7, "ts": 1776649791947920, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947910, "dur": 6, + "args": { + "External id": 9300, + "cbid": 211, "correlation": 69435 + } + }, + { + "ph": "s", "id": 69435, "pid": 435, "tid": 435, "ts": 1776649791947910, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947923, "dur": 0, + "args": { + "External id": 69441, + "cbid": 200, "correlation": 69441 + } + }, + { + "ph": "f", "id": 69441, "pid": 435, "tid": 435, "ts": 1776649791947923, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947923, "dur": 0, + "args": { + "External id": 69442, + "cbid": 273, "correlation": 69442 + } + }, + { + "ph": "f", "id": 69442, "pid": 435, "tid": 435, "ts": 1776649791947923, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791947932, "dur": 19, + "args": { + "External id": 69443, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69443, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69443, "pid": 0, "tid": 7, "ts": 1776649791947932, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947924, "dur": 4, + "args": { + "External id": 69443, + "cbid": 211, "correlation": 69443 + } + }, + { + "ph": "s", "id": 69443, "pid": 435, "tid": 435, "ts": 1776649791947924, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791947968, "dur": 1, + "args": { + "External id": 9305, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69461, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69461, "pid": 0, "tid": 7, "ts": 1776649791947968, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947959, "dur": 6, + "args": { + "External id": 9305, + "cbid": 211, "correlation": 69461 + } + }, + { + "ph": "s", "id": 69461, "pid": 435, "tid": 435, "ts": 1776649791947959, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947971, "dur": 0, + "args": { + "External id": 69467, + "cbid": 200, "correlation": 69467 + } + }, + { + "ph": "f", "id": 69467, "pid": 435, "tid": 435, "ts": 1776649791947971, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791947972, "dur": 0, + "args": { + "External id": 69468, + "cbid": 273, "correlation": 69468 + } + }, + { + "ph": "f", "id": 69468, "pid": 435, "tid": 435, "ts": 1776649791947972, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791947981, "dur": 19, + "args": { + "External id": 69469, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69469, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69469, "pid": 0, "tid": 7, "ts": 1776649791947981, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791947973, "dur": 5, + "args": { + "External id": 69469, + "cbid": 211, "correlation": 69469 + } + }, + { + "ph": "s", "id": 69469, "pid": 435, "tid": 435, "ts": 1776649791947973, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791948017, "dur": 1, + "args": { + "External id": 9310, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69487, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69487, "pid": 0, "tid": 7, "ts": 1776649791948017, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948007, "dur": 6, + "args": { + "External id": 9310, + "cbid": 211, "correlation": 69487 + } + }, + { + "ph": "s", "id": 69487, "pid": 435, "tid": 435, "ts": 1776649791948007, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948020, "dur": 0, + "args": { + "External id": 69493, + "cbid": 200, "correlation": 69493 + } + }, + { + "ph": "f", "id": 69493, "pid": 435, "tid": 435, "ts": 1776649791948020, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948020, "dur": 0, + "args": { + "External id": 69494, + "cbid": 273, "correlation": 69494 + } + }, + { + "ph": "f", "id": 69494, "pid": 435, "tid": 435, "ts": 1776649791948020, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791948029, "dur": 19, + "args": { + "External id": 69495, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69495, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69495, "pid": 0, "tid": 7, "ts": 1776649791948029, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948022, "dur": 5, + "args": { + "External id": 69495, + "cbid": 211, "correlation": 69495 + } + }, + { + "ph": "s", "id": 69495, "pid": 435, "tid": 435, "ts": 1776649791948022, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791948067, "dur": 1, + "args": { + "External id": 9315, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69513, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69513, "pid": 0, "tid": 7, "ts": 1776649791948067, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948055, "dur": 8, + "args": { + "External id": 9315, + "cbid": 211, "correlation": 69513 + } + }, + { + "ph": "s", "id": 69513, "pid": 435, "tid": 435, "ts": 1776649791948055, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948070, "dur": 0, + "args": { + "External id": 69519, + "cbid": 200, "correlation": 69519 + } + }, + { + "ph": "f", "id": 69519, "pid": 435, "tid": 435, "ts": 1776649791948070, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948070, "dur": 0, + "args": { + "External id": 69520, + "cbid": 273, "correlation": 69520 + } + }, + { + "ph": "f", "id": 69520, "pid": 435, "tid": 435, "ts": 1776649791948070, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791948079, "dur": 19, + "args": { + "External id": 69521, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69521, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69521, "pid": 0, "tid": 7, "ts": 1776649791948079, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948071, "dur": 5, + "args": { + "External id": 69521, + "cbid": 211, "correlation": 69521 + } + }, + { + "ph": "s", "id": 69521, "pid": 435, "tid": 435, "ts": 1776649791948071, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791948115, "dur": 1, + "args": { + "External id": 9320, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69539, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69539, "pid": 0, "tid": 7, "ts": 1776649791948115, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948106, "dur": 6, + "args": { + "External id": 9320, + "cbid": 211, "correlation": 69539 + } + }, + { + "ph": "s", "id": 69539, "pid": 435, "tid": 435, "ts": 1776649791948106, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948118, "dur": 0, + "args": { + "External id": 69545, + "cbid": 200, "correlation": 69545 + } + }, + { + "ph": "f", "id": 69545, "pid": 435, "tid": 435, "ts": 1776649791948118, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948119, "dur": 0, + "args": { + "External id": 69546, + "cbid": 273, "correlation": 69546 + } + }, + { + "ph": "f", "id": 69546, "pid": 435, "tid": 435, "ts": 1776649791948119, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791948127, "dur": 19, + "args": { + "External id": 69547, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69547, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69547, "pid": 0, "tid": 7, "ts": 1776649791948127, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948120, "dur": 4, + "args": { + "External id": 69547, + "cbid": 211, "correlation": 69547 + } + }, + { + "ph": "s", "id": 69547, "pid": 435, "tid": 435, "ts": 1776649791948120, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791948163, "dur": 1, + "args": { + "External id": 9325, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69565, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69565, "pid": 0, "tid": 7, "ts": 1776649791948163, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948153, "dur": 6, + "args": { + "External id": 9325, + "cbid": 211, "correlation": 69565 + } + }, + { + "ph": "s", "id": 69565, "pid": 435, "tid": 435, "ts": 1776649791948153, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948165, "dur": 0, + "args": { + "External id": 69571, + "cbid": 200, "correlation": 69571 + } + }, + { + "ph": "f", "id": 69571, "pid": 435, "tid": 435, "ts": 1776649791948165, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948166, "dur": 0, + "args": { + "External id": 69572, + "cbid": 273, "correlation": 69572 + } + }, + { + "ph": "f", "id": 69572, "pid": 435, "tid": 435, "ts": 1776649791948166, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791948175, "dur": 19, + "args": { + "External id": 69573, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69573, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69573, "pid": 0, "tid": 7, "ts": 1776649791948175, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948167, "dur": 4, + "args": { + "External id": 69573, + "cbid": 211, "correlation": 69573 + } + }, + { + "ph": "s", "id": 69573, "pid": 435, "tid": 435, "ts": 1776649791948167, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791948210, "dur": 1, + "args": { + "External id": 9330, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69591, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69591, "pid": 0, "tid": 7, "ts": 1776649791948210, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948201, "dur": 6, + "args": { + "External id": 9330, + "cbid": 211, "correlation": 69591 + } + }, + { + "ph": "s", "id": 69591, "pid": 435, "tid": 435, "ts": 1776649791948201, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948213, "dur": 0, + "args": { + "External id": 69597, + "cbid": 200, "correlation": 69597 + } + }, + { + "ph": "f", "id": 69597, "pid": 435, "tid": 435, "ts": 1776649791948213, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948214, "dur": 0, + "args": { + "External id": 69598, + "cbid": 273, "correlation": 69598 + } + }, + { + "ph": "f", "id": 69598, "pid": 435, "tid": 435, "ts": 1776649791948214, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791948241, "dur": 19, + "args": { + "External id": 69599, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69599, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69599, "pid": 0, "tid": 7, "ts": 1776649791948241, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948215, "dur": 23, + "args": { + "External id": 69599, + "cbid": 211, "correlation": 69599 + } + }, + { + "ph": "s", "id": 69599, "pid": 435, "tid": 435, "ts": 1776649791948215, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791948282, "dur": 1, + "args": { + "External id": 9335, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69617, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69617, "pid": 0, "tid": 7, "ts": 1776649791948282, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948272, "dur": 7, + "args": { + "External id": 9335, + "cbid": 211, "correlation": 69617 + } + }, + { + "ph": "s", "id": 69617, "pid": 435, "tid": 435, "ts": 1776649791948272, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948285, "dur": 0, + "args": { + "External id": 69623, + "cbid": 200, "correlation": 69623 + } + }, + { + "ph": "f", "id": 69623, "pid": 435, "tid": 435, "ts": 1776649791948285, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948286, "dur": 0, + "args": { + "External id": 69624, + "cbid": 273, "correlation": 69624 + } + }, + { + "ph": "f", "id": 69624, "pid": 435, "tid": 435, "ts": 1776649791948286, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791948295, "dur": 19, + "args": { + "External id": 69625, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69625, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69625, "pid": 0, "tid": 7, "ts": 1776649791948295, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948287, "dur": 5, + "args": { + "External id": 69625, + "cbid": 211, "correlation": 69625 + } + }, + { + "ph": "s", "id": 69625, "pid": 435, "tid": 435, "ts": 1776649791948287, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791948330, "dur": 1, + "args": { + "External id": 9340, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69643, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69643, "pid": 0, "tid": 7, "ts": 1776649791948330, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948321, "dur": 6, + "args": { + "External id": 9340, + "cbid": 211, "correlation": 69643 + } + }, + { + "ph": "s", "id": 69643, "pid": 435, "tid": 435, "ts": 1776649791948321, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948333, "dur": 0, + "args": { + "External id": 69649, + "cbid": 200, "correlation": 69649 + } + }, + { + "ph": "f", "id": 69649, "pid": 435, "tid": 435, "ts": 1776649791948333, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948334, "dur": 0, + "args": { + "External id": 69650, + "cbid": 273, "correlation": 69650 + } + }, + { + "ph": "f", "id": 69650, "pid": 435, "tid": 435, "ts": 1776649791948334, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791948342, "dur": 19, + "args": { + "External id": 69651, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69651, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69651, "pid": 0, "tid": 7, "ts": 1776649791948342, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948335, "dur": 4, + "args": { + "External id": 69651, + "cbid": 211, "correlation": 69651 + } + }, + { + "ph": "s", "id": 69651, "pid": 435, "tid": 435, "ts": 1776649791948335, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791948377, "dur": 1, + "args": { + "External id": 9345, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69669, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69669, "pid": 0, "tid": 7, "ts": 1776649791948377, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948368, "dur": 6, + "args": { + "External id": 9345, + "cbid": 211, "correlation": 69669 + } + }, + { + "ph": "s", "id": 69669, "pid": 435, "tid": 435, "ts": 1776649791948368, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948381, "dur": 0, + "args": { + "External id": 69675, + "cbid": 200, "correlation": 69675 + } + }, + { + "ph": "f", "id": 69675, "pid": 435, "tid": 435, "ts": 1776649791948381, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948381, "dur": 0, + "args": { + "External id": 69676, + "cbid": 273, "correlation": 69676 + } + }, + { + "ph": "f", "id": 69676, "pid": 435, "tid": 435, "ts": 1776649791948381, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791948390, "dur": 19, + "args": { + "External id": 69677, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69677, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69677, "pid": 0, "tid": 7, "ts": 1776649791948390, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948382, "dur": 4, + "args": { + "External id": 69677, + "cbid": 211, "correlation": 69677 + } + }, + { + "ph": "s", "id": 69677, "pid": 435, "tid": 435, "ts": 1776649791948382, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791948425, "dur": 1, + "args": { + "External id": 9350, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69695, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69695, "pid": 0, "tid": 7, "ts": 1776649791948425, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948416, "dur": 6, + "args": { + "External id": 9350, + "cbid": 211, "correlation": 69695 + } + }, + { + "ph": "s", "id": 69695, "pid": 435, "tid": 435, "ts": 1776649791948416, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948428, "dur": 0, + "args": { + "External id": 69701, + "cbid": 200, "correlation": 69701 + } + }, + { + "ph": "f", "id": 69701, "pid": 435, "tid": 435, "ts": 1776649791948428, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948429, "dur": 0, + "args": { + "External id": 69702, + "cbid": 273, "correlation": 69702 + } + }, + { + "ph": "f", "id": 69702, "pid": 435, "tid": 435, "ts": 1776649791948429, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791948437, "dur": 19, + "args": { + "External id": 69703, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69703, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69703, "pid": 0, "tid": 7, "ts": 1776649791948437, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948430, "dur": 4, + "args": { + "External id": 69703, + "cbid": 211, "correlation": 69703 + } + }, + { + "ph": "s", "id": 69703, "pid": 435, "tid": 435, "ts": 1776649791948430, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791948472, "dur": 1, + "args": { + "External id": 9355, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69721, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69721, "pid": 0, "tid": 7, "ts": 1776649791948472, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948463, "dur": 6, + "args": { + "External id": 9355, + "cbid": 211, "correlation": 69721 + } + }, + { + "ph": "s", "id": 69721, "pid": 435, "tid": 435, "ts": 1776649791948463, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948475, "dur": 0, + "args": { + "External id": 69727, + "cbid": 200, "correlation": 69727 + } + }, + { + "ph": "f", "id": 69727, "pid": 435, "tid": 435, "ts": 1776649791948475, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948476, "dur": 0, + "args": { + "External id": 69728, + "cbid": 273, "correlation": 69728 + } + }, + { + "ph": "f", "id": 69728, "pid": 435, "tid": 435, "ts": 1776649791948476, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791948485, "dur": 19, + "args": { + "External id": 69729, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69729, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69729, "pid": 0, "tid": 7, "ts": 1776649791948485, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948477, "dur": 5, + "args": { + "External id": 69729, + "cbid": 211, "correlation": 69729 + } + }, + { + "ph": "s", "id": 69729, "pid": 435, "tid": 435, "ts": 1776649791948477, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791948520, "dur": 1, + "args": { + "External id": 9360, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69747, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69747, "pid": 0, "tid": 7, "ts": 1776649791948520, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948511, "dur": 6, + "args": { + "External id": 9360, + "cbid": 211, "correlation": 69747 + } + }, + { + "ph": "s", "id": 69747, "pid": 435, "tid": 435, "ts": 1776649791948511, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948523, "dur": 0, + "args": { + "External id": 69753, + "cbid": 200, "correlation": 69753 + } + }, + { + "ph": "f", "id": 69753, "pid": 435, "tid": 435, "ts": 1776649791948523, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948524, "dur": 0, + "args": { + "External id": 69754, + "cbid": 273, "correlation": 69754 + } + }, + { + "ph": "f", "id": 69754, "pid": 435, "tid": 435, "ts": 1776649791948524, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791948533, "dur": 19, + "args": { + "External id": 69755, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69755, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69755, "pid": 0, "tid": 7, "ts": 1776649791948533, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948525, "dur": 5, + "args": { + "External id": 69755, + "cbid": 211, "correlation": 69755 + } + }, + { + "ph": "s", "id": 69755, "pid": 435, "tid": 435, "ts": 1776649791948525, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791948567, "dur": 1, + "args": { + "External id": 9365, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69773, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69773, "pid": 0, "tid": 7, "ts": 1776649791948567, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948558, "dur": 6, + "args": { + "External id": 9365, + "cbid": 211, "correlation": 69773 + } + }, + { + "ph": "s", "id": 69773, "pid": 435, "tid": 435, "ts": 1776649791948558, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948571, "dur": 0, + "args": { + "External id": 69779, + "cbid": 200, "correlation": 69779 + } + }, + { + "ph": "f", "id": 69779, "pid": 435, "tid": 435, "ts": 1776649791948571, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948571, "dur": 0, + "args": { + "External id": 69780, + "cbid": 273, "correlation": 69780 + } + }, + { + "ph": "f", "id": 69780, "pid": 435, "tid": 435, "ts": 1776649791948571, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791948580, "dur": 19, + "args": { + "External id": 69781, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69781, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69781, "pid": 0, "tid": 7, "ts": 1776649791948580, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948572, "dur": 5, + "args": { + "External id": 69781, + "cbid": 211, "correlation": 69781 + } + }, + { + "ph": "s", "id": 69781, "pid": 435, "tid": 435, "ts": 1776649791948572, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791948615, "dur": 1, + "args": { + "External id": 9370, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69799, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69799, "pid": 0, "tid": 7, "ts": 1776649791948615, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948606, "dur": 6, + "args": { + "External id": 9370, + "cbid": 211, "correlation": 69799 + } + }, + { + "ph": "s", "id": 69799, "pid": 435, "tid": 435, "ts": 1776649791948606, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948618, "dur": 0, + "args": { + "External id": 69805, + "cbid": 200, "correlation": 69805 + } + }, + { + "ph": "f", "id": 69805, "pid": 435, "tid": 435, "ts": 1776649791948618, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948619, "dur": 0, + "args": { + "External id": 69806, + "cbid": 273, "correlation": 69806 + } + }, + { + "ph": "f", "id": 69806, "pid": 435, "tid": 435, "ts": 1776649791948619, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791948628, "dur": 19, + "args": { + "External id": 69807, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69807, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69807, "pid": 0, "tid": 7, "ts": 1776649791948628, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948620, "dur": 4, + "args": { + "External id": 69807, + "cbid": 211, "correlation": 69807 + } + }, + { + "ph": "s", "id": 69807, "pid": 435, "tid": 435, "ts": 1776649791948620, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791948662, "dur": 1, + "args": { + "External id": 9375, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69825, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69825, "pid": 0, "tid": 7, "ts": 1776649791948662, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948653, "dur": 6, + "args": { + "External id": 9375, + "cbid": 211, "correlation": 69825 + } + }, + { + "ph": "s", "id": 69825, "pid": 435, "tid": 435, "ts": 1776649791948653, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948666, "dur": 0, + "args": { + "External id": 69831, + "cbid": 200, "correlation": 69831 + } + }, + { + "ph": "f", "id": 69831, "pid": 435, "tid": 435, "ts": 1776649791948666, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948666, "dur": 0, + "args": { + "External id": 69832, + "cbid": 273, "correlation": 69832 + } + }, + { + "ph": "f", "id": 69832, "pid": 435, "tid": 435, "ts": 1776649791948666, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791948675, "dur": 19, + "args": { + "External id": 69833, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69833, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69833, "pid": 0, "tid": 7, "ts": 1776649791948675, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948667, "dur": 4, + "args": { + "External id": 69833, + "cbid": 211, "correlation": 69833 + } + }, + { + "ph": "s", "id": 69833, "pid": 435, "tid": 435, "ts": 1776649791948667, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791948710, "dur": 1, + "args": { + "External id": 9380, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69851, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69851, "pid": 0, "tid": 7, "ts": 1776649791948710, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948701, "dur": 6, + "args": { + "External id": 9380, + "cbid": 211, "correlation": 69851 + } + }, + { + "ph": "s", "id": 69851, "pid": 435, "tid": 435, "ts": 1776649791948701, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948713, "dur": 0, + "args": { + "External id": 69857, + "cbid": 200, "correlation": 69857 + } + }, + { + "ph": "f", "id": 69857, "pid": 435, "tid": 435, "ts": 1776649791948713, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948713, "dur": 0, + "args": { + "External id": 69858, + "cbid": 273, "correlation": 69858 + } + }, + { + "ph": "f", "id": 69858, "pid": 435, "tid": 435, "ts": 1776649791948713, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791948722, "dur": 19, + "args": { + "External id": 69859, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69859, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69859, "pid": 0, "tid": 7, "ts": 1776649791948722, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948714, "dur": 4, + "args": { + "External id": 69859, + "cbid": 211, "correlation": 69859 + } + }, + { + "ph": "s", "id": 69859, "pid": 435, "tid": 435, "ts": 1776649791948714, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791948757, "dur": 1, + "args": { + "External id": 9385, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69877, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69877, "pid": 0, "tid": 7, "ts": 1776649791948757, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948748, "dur": 6, + "args": { + "External id": 9385, + "cbid": 211, "correlation": 69877 + } + }, + { + "ph": "s", "id": 69877, "pid": 435, "tid": 435, "ts": 1776649791948748, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948761, "dur": 0, + "args": { + "External id": 69883, + "cbid": 200, "correlation": 69883 + } + }, + { + "ph": "f", "id": 69883, "pid": 435, "tid": 435, "ts": 1776649791948761, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948761, "dur": 0, + "args": { + "External id": 69884, + "cbid": 273, "correlation": 69884 + } + }, + { + "ph": "f", "id": 69884, "pid": 435, "tid": 435, "ts": 1776649791948761, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791948770, "dur": 19, + "args": { + "External id": 69885, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69885, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69885, "pid": 0, "tid": 7, "ts": 1776649791948770, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948762, "dur": 4, + "args": { + "External id": 69885, + "cbid": 211, "correlation": 69885 + } + }, + { + "ph": "s", "id": 69885, "pid": 435, "tid": 435, "ts": 1776649791948762, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791948805, "dur": 1, + "args": { + "External id": 9390, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69903, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69903, "pid": 0, "tid": 7, "ts": 1776649791948805, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948796, "dur": 6, + "args": { + "External id": 9390, + "cbid": 211, "correlation": 69903 + } + }, + { + "ph": "s", "id": 69903, "pid": 435, "tid": 435, "ts": 1776649791948796, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948808, "dur": 0, + "args": { + "External id": 69909, + "cbid": 200, "correlation": 69909 + } + }, + { + "ph": "f", "id": 69909, "pid": 435, "tid": 435, "ts": 1776649791948808, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948809, "dur": 0, + "args": { + "External id": 69910, + "cbid": 273, "correlation": 69910 + } + }, + { + "ph": "f", "id": 69910, "pid": 435, "tid": 435, "ts": 1776649791948809, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791948818, "dur": 19, + "args": { + "External id": 69911, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69911, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69911, "pid": 0, "tid": 7, "ts": 1776649791948818, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948810, "dur": 4, + "args": { + "External id": 69911, + "cbid": 211, "correlation": 69911 + } + }, + { + "ph": "s", "id": 69911, "pid": 435, "tid": 435, "ts": 1776649791948810, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791948852, "dur": 1, + "args": { + "External id": 9395, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69929, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69929, "pid": 0, "tid": 7, "ts": 1776649791948852, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948843, "dur": 6, + "args": { + "External id": 9395, + "cbid": 211, "correlation": 69929 + } + }, + { + "ph": "s", "id": 69929, "pid": 435, "tid": 435, "ts": 1776649791948843, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948855, "dur": 0, + "args": { + "External id": 69935, + "cbid": 200, "correlation": 69935 + } + }, + { + "ph": "f", "id": 69935, "pid": 435, "tid": 435, "ts": 1776649791948855, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948856, "dur": 0, + "args": { + "External id": 69936, + "cbid": 273, "correlation": 69936 + } + }, + { + "ph": "f", "id": 69936, "pid": 435, "tid": 435, "ts": 1776649791948856, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791948864, "dur": 19, + "args": { + "External id": 69937, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69937, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69937, "pid": 0, "tid": 7, "ts": 1776649791948864, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948857, "dur": 4, + "args": { + "External id": 69937, + "cbid": 211, "correlation": 69937 + } + }, + { + "ph": "s", "id": 69937, "pid": 435, "tid": 435, "ts": 1776649791948857, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791948899, "dur": 1, + "args": { + "External id": 9400, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69955, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69955, "pid": 0, "tid": 7, "ts": 1776649791948899, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948890, "dur": 6, + "args": { + "External id": 9400, + "cbid": 211, "correlation": 69955 + } + }, + { + "ph": "s", "id": 69955, "pid": 435, "tid": 435, "ts": 1776649791948890, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948902, "dur": 0, + "args": { + "External id": 69961, + "cbid": 200, "correlation": 69961 + } + }, + { + "ph": "f", "id": 69961, "pid": 435, "tid": 435, "ts": 1776649791948902, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948903, "dur": 0, + "args": { + "External id": 69962, + "cbid": 273, "correlation": 69962 + } + }, + { + "ph": "f", "id": 69962, "pid": 435, "tid": 435, "ts": 1776649791948903, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791948912, "dur": 19, + "args": { + "External id": 69963, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69963, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69963, "pid": 0, "tid": 7, "ts": 1776649791948912, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948904, "dur": 4, + "args": { + "External id": 69963, + "cbid": 211, "correlation": 69963 + } + }, + { + "ph": "s", "id": 69963, "pid": 435, "tid": 435, "ts": 1776649791948904, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791948946, "dur": 1, + "args": { + "External id": 9405, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69981, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69981, "pid": 0, "tid": 7, "ts": 1776649791948946, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948937, "dur": 6, + "args": { + "External id": 9405, + "cbid": 211, "correlation": 69981 + } + }, + { + "ph": "s", "id": 69981, "pid": 435, "tid": 435, "ts": 1776649791948937, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948949, "dur": 0, + "args": { + "External id": 69987, + "cbid": 200, "correlation": 69987 + } + }, + { + "ph": "f", "id": 69987, "pid": 435, "tid": 435, "ts": 1776649791948949, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948950, "dur": 0, + "args": { + "External id": 69988, + "cbid": 273, "correlation": 69988 + } + }, + { + "ph": "f", "id": 69988, "pid": 435, "tid": 435, "ts": 1776649791948950, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791948958, "dur": 19, + "args": { + "External id": 69989, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 69989, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 69989, "pid": 0, "tid": 7, "ts": 1776649791948958, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948951, "dur": 4, + "args": { + "External id": 69989, + "cbid": 211, "correlation": 69989 + } + }, + { + "ph": "s", "id": 69989, "pid": 435, "tid": 435, "ts": 1776649791948951, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791948993, "dur": 1, + "args": { + "External id": 9410, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70007, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70007, "pid": 0, "tid": 7, "ts": 1776649791948993, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948984, "dur": 6, + "args": { + "External id": 9410, + "cbid": 211, "correlation": 70007 + } + }, + { + "ph": "s", "id": 70007, "pid": 435, "tid": 435, "ts": 1776649791948984, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948996, "dur": 0, + "args": { + "External id": 70013, + "cbid": 200, "correlation": 70013 + } + }, + { + "ph": "f", "id": 70013, "pid": 435, "tid": 435, "ts": 1776649791948996, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791948997, "dur": 0, + "args": { + "External id": 70014, + "cbid": 273, "correlation": 70014 + } + }, + { + "ph": "f", "id": 70014, "pid": 435, "tid": 435, "ts": 1776649791948997, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791949006, "dur": 19, + "args": { + "External id": 70015, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70015, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70015, "pid": 0, "tid": 7, "ts": 1776649791949006, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791948998, "dur": 4, + "args": { + "External id": 70015, + "cbid": 211, "correlation": 70015 + } + }, + { + "ph": "s", "id": 70015, "pid": 435, "tid": 435, "ts": 1776649791948998, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791949040, "dur": 1, + "args": { + "External id": 9415, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70033, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70033, "pid": 0, "tid": 7, "ts": 1776649791949040, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949031, "dur": 6, + "args": { + "External id": 9415, + "cbid": 211, "correlation": 70033 + } + }, + { + "ph": "s", "id": 70033, "pid": 435, "tid": 435, "ts": 1776649791949031, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949043, "dur": 0, + "args": { + "External id": 70039, + "cbid": 200, "correlation": 70039 + } + }, + { + "ph": "f", "id": 70039, "pid": 435, "tid": 435, "ts": 1776649791949043, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949044, "dur": 0, + "args": { + "External id": 70040, + "cbid": 273, "correlation": 70040 + } + }, + { + "ph": "f", "id": 70040, "pid": 435, "tid": 435, "ts": 1776649791949044, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791949053, "dur": 19, + "args": { + "External id": 70041, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70041, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70041, "pid": 0, "tid": 7, "ts": 1776649791949053, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949045, "dur": 5, + "args": { + "External id": 70041, + "cbid": 211, "correlation": 70041 + } + }, + { + "ph": "s", "id": 70041, "pid": 435, "tid": 435, "ts": 1776649791949045, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791949088, "dur": 1, + "args": { + "External id": 9420, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70059, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70059, "pid": 0, "tid": 7, "ts": 1776649791949088, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949079, "dur": 6, + "args": { + "External id": 9420, + "cbid": 211, "correlation": 70059 + } + }, + { + "ph": "s", "id": 70059, "pid": 435, "tid": 435, "ts": 1776649791949079, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949091, "dur": 0, + "args": { + "External id": 70065, + "cbid": 200, "correlation": 70065 + } + }, + { + "ph": "f", "id": 70065, "pid": 435, "tid": 435, "ts": 1776649791949091, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949091, "dur": 0, + "args": { + "External id": 70066, + "cbid": 273, "correlation": 70066 + } + }, + { + "ph": "f", "id": 70066, "pid": 435, "tid": 435, "ts": 1776649791949091, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791949100, "dur": 19, + "args": { + "External id": 70067, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70067, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70067, "pid": 0, "tid": 7, "ts": 1776649791949100, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949092, "dur": 4, + "args": { + "External id": 70067, + "cbid": 211, "correlation": 70067 + } + }, + { + "ph": "s", "id": 70067, "pid": 435, "tid": 435, "ts": 1776649791949092, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791949135, "dur": 1, + "args": { + "External id": 9425, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70085, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70085, "pid": 0, "tid": 7, "ts": 1776649791949135, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949126, "dur": 6, + "args": { + "External id": 9425, + "cbid": 211, "correlation": 70085 + } + }, + { + "ph": "s", "id": 70085, "pid": 435, "tid": 435, "ts": 1776649791949126, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949138, "dur": 0, + "args": { + "External id": 70091, + "cbid": 200, "correlation": 70091 + } + }, + { + "ph": "f", "id": 70091, "pid": 435, "tid": 435, "ts": 1776649791949138, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949138, "dur": 0, + "args": { + "External id": 70092, + "cbid": 273, "correlation": 70092 + } + }, + { + "ph": "f", "id": 70092, "pid": 435, "tid": 435, "ts": 1776649791949138, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791949147, "dur": 19, + "args": { + "External id": 70093, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70093, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70093, "pid": 0, "tid": 7, "ts": 1776649791949147, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949139, "dur": 4, + "args": { + "External id": 70093, + "cbid": 211, "correlation": 70093 + } + }, + { + "ph": "s", "id": 70093, "pid": 435, "tid": 435, "ts": 1776649791949139, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791949182, "dur": 1, + "args": { + "External id": 9430, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70111, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70111, "pid": 0, "tid": 7, "ts": 1776649791949182, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949173, "dur": 6, + "args": { + "External id": 9430, + "cbid": 211, "correlation": 70111 + } + }, + { + "ph": "s", "id": 70111, "pid": 435, "tid": 435, "ts": 1776649791949173, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949185, "dur": 0, + "args": { + "External id": 70117, + "cbid": 200, "correlation": 70117 + } + }, + { + "ph": "f", "id": 70117, "pid": 435, "tid": 435, "ts": 1776649791949185, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949185, "dur": 0, + "args": { + "External id": 70118, + "cbid": 273, "correlation": 70118 + } + }, + { + "ph": "f", "id": 70118, "pid": 435, "tid": 435, "ts": 1776649791949185, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791949194, "dur": 19, + "args": { + "External id": 70119, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70119, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70119, "pid": 0, "tid": 7, "ts": 1776649791949194, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949186, "dur": 5, + "args": { + "External id": 70119, + "cbid": 211, "correlation": 70119 + } + }, + { + "ph": "s", "id": 70119, "pid": 435, "tid": 435, "ts": 1776649791949186, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791949229, "dur": 1, + "args": { + "External id": 9435, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70137, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70137, "pid": 0, "tid": 7, "ts": 1776649791949229, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949220, "dur": 6, + "args": { + "External id": 9435, + "cbid": 211, "correlation": 70137 + } + }, + { + "ph": "s", "id": 70137, "pid": 435, "tid": 435, "ts": 1776649791949220, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949232, "dur": 0, + "args": { + "External id": 70143, + "cbid": 200, "correlation": 70143 + } + }, + { + "ph": "f", "id": 70143, "pid": 435, "tid": 435, "ts": 1776649791949232, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949233, "dur": 0, + "args": { + "External id": 70144, + "cbid": 273, "correlation": 70144 + } + }, + { + "ph": "f", "id": 70144, "pid": 435, "tid": 435, "ts": 1776649791949233, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791949242, "dur": 19, + "args": { + "External id": 70145, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70145, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70145, "pid": 0, "tid": 7, "ts": 1776649791949242, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949234, "dur": 5, + "args": { + "External id": 70145, + "cbid": 211, "correlation": 70145 + } + }, + { + "ph": "s", "id": 70145, "pid": 435, "tid": 435, "ts": 1776649791949234, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791949277, "dur": 1, + "args": { + "External id": 9440, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70163, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70163, "pid": 0, "tid": 7, "ts": 1776649791949277, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949268, "dur": 6, + "args": { + "External id": 9440, + "cbid": 211, "correlation": 70163 + } + }, + { + "ph": "s", "id": 70163, "pid": 435, "tid": 435, "ts": 1776649791949268, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949280, "dur": 0, + "args": { + "External id": 70169, + "cbid": 200, "correlation": 70169 + } + }, + { + "ph": "f", "id": 70169, "pid": 435, "tid": 435, "ts": 1776649791949280, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949280, "dur": 0, + "args": { + "External id": 70170, + "cbid": 273, "correlation": 70170 + } + }, + { + "ph": "f", "id": 70170, "pid": 435, "tid": 435, "ts": 1776649791949280, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791949289, "dur": 19, + "args": { + "External id": 70171, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70171, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70171, "pid": 0, "tid": 7, "ts": 1776649791949289, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949281, "dur": 4, + "args": { + "External id": 70171, + "cbid": 211, "correlation": 70171 + } + }, + { + "ph": "s", "id": 70171, "pid": 435, "tid": 435, "ts": 1776649791949281, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791949323, "dur": 1, + "args": { + "External id": 9445, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70189, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70189, "pid": 0, "tid": 7, "ts": 1776649791949323, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949315, "dur": 6, + "args": { + "External id": 9445, + "cbid": 211, "correlation": 70189 + } + }, + { + "ph": "s", "id": 70189, "pid": 435, "tid": 435, "ts": 1776649791949315, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949327, "dur": 0, + "args": { + "External id": 70195, + "cbid": 200, "correlation": 70195 + } + }, + { + "ph": "f", "id": 70195, "pid": 435, "tid": 435, "ts": 1776649791949327, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949327, "dur": 0, + "args": { + "External id": 70196, + "cbid": 273, "correlation": 70196 + } + }, + { + "ph": "f", "id": 70196, "pid": 435, "tid": 435, "ts": 1776649791949327, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791949336, "dur": 19, + "args": { + "External id": 70197, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70197, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70197, "pid": 0, "tid": 7, "ts": 1776649791949336, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949328, "dur": 4, + "args": { + "External id": 70197, + "cbid": 211, "correlation": 70197 + } + }, + { + "ph": "s", "id": 70197, "pid": 435, "tid": 435, "ts": 1776649791949328, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791949372, "dur": 1, + "args": { + "External id": 9450, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70215, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70215, "pid": 0, "tid": 7, "ts": 1776649791949372, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949362, "dur": 6, + "args": { + "External id": 9450, + "cbid": 211, "correlation": 70215 + } + }, + { + "ph": "s", "id": 70215, "pid": 435, "tid": 435, "ts": 1776649791949362, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949375, "dur": 0, + "args": { + "External id": 70221, + "cbid": 200, "correlation": 70221 + } + }, + { + "ph": "f", "id": 70221, "pid": 435, "tid": 435, "ts": 1776649791949375, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949375, "dur": 0, + "args": { + "External id": 70222, + "cbid": 273, "correlation": 70222 + } + }, + { + "ph": "f", "id": 70222, "pid": 435, "tid": 435, "ts": 1776649791949375, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791949384, "dur": 19, + "args": { + "External id": 70223, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70223, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70223, "pid": 0, "tid": 7, "ts": 1776649791949384, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949376, "dur": 4, + "args": { + "External id": 70223, + "cbid": 211, "correlation": 70223 + } + }, + { + "ph": "s", "id": 70223, "pid": 435, "tid": 435, "ts": 1776649791949376, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791949419, "dur": 1, + "args": { + "External id": 9455, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70241, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70241, "pid": 0, "tid": 7, "ts": 1776649791949419, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949410, "dur": 6, + "args": { + "External id": 9455, + "cbid": 211, "correlation": 70241 + } + }, + { + "ph": "s", "id": 70241, "pid": 435, "tid": 435, "ts": 1776649791949410, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949422, "dur": 0, + "args": { + "External id": 70247, + "cbid": 200, "correlation": 70247 + } + }, + { + "ph": "f", "id": 70247, "pid": 435, "tid": 435, "ts": 1776649791949422, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949422, "dur": 0, + "args": { + "External id": 70248, + "cbid": 273, "correlation": 70248 + } + }, + { + "ph": "f", "id": 70248, "pid": 435, "tid": 435, "ts": 1776649791949422, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791949431, "dur": 19, + "args": { + "External id": 70249, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70249, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70249, "pid": 0, "tid": 7, "ts": 1776649791949431, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949423, "dur": 4, + "args": { + "External id": 70249, + "cbid": 211, "correlation": 70249 + } + }, + { + "ph": "s", "id": 70249, "pid": 435, "tid": 435, "ts": 1776649791949423, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791949466, "dur": 1, + "args": { + "External id": 9460, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70267, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70267, "pid": 0, "tid": 7, "ts": 1776649791949466, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949457, "dur": 6, + "args": { + "External id": 9460, + "cbid": 211, "correlation": 70267 + } + }, + { + "ph": "s", "id": 70267, "pid": 435, "tid": 435, "ts": 1776649791949457, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949469, "dur": 0, + "args": { + "External id": 70273, + "cbid": 200, "correlation": 70273 + } + }, + { + "ph": "f", "id": 70273, "pid": 435, "tid": 435, "ts": 1776649791949469, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949469, "dur": 0, + "args": { + "External id": 70274, + "cbid": 273, "correlation": 70274 + } + }, + { + "ph": "f", "id": 70274, "pid": 435, "tid": 435, "ts": 1776649791949469, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791949478, "dur": 19, + "args": { + "External id": 70275, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70275, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70275, "pid": 0, "tid": 7, "ts": 1776649791949478, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949471, "dur": 4, + "args": { + "External id": 70275, + "cbid": 211, "correlation": 70275 + } + }, + { + "ph": "s", "id": 70275, "pid": 435, "tid": 435, "ts": 1776649791949471, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791949513, "dur": 1, + "args": { + "External id": 9465, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70293, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70293, "pid": 0, "tid": 7, "ts": 1776649791949513, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949504, "dur": 6, + "args": { + "External id": 9465, + "cbid": 211, "correlation": 70293 + } + }, + { + "ph": "s", "id": 70293, "pid": 435, "tid": 435, "ts": 1776649791949504, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949516, "dur": 0, + "args": { + "External id": 70299, + "cbid": 200, "correlation": 70299 + } + }, + { + "ph": "f", "id": 70299, "pid": 435, "tid": 435, "ts": 1776649791949516, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949517, "dur": 0, + "args": { + "External id": 70300, + "cbid": 273, "correlation": 70300 + } + }, + { + "ph": "f", "id": 70300, "pid": 435, "tid": 435, "ts": 1776649791949517, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791949526, "dur": 19, + "args": { + "External id": 70301, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70301, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70301, "pid": 0, "tid": 7, "ts": 1776649791949526, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949518, "dur": 4, + "args": { + "External id": 70301, + "cbid": 211, "correlation": 70301 + } + }, + { + "ph": "s", "id": 70301, "pid": 435, "tid": 435, "ts": 1776649791949518, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791949561, "dur": 1, + "args": { + "External id": 9470, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70319, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70319, "pid": 0, "tid": 7, "ts": 1776649791949561, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949551, "dur": 6, + "args": { + "External id": 9470, + "cbid": 211, "correlation": 70319 + } + }, + { + "ph": "s", "id": 70319, "pid": 435, "tid": 435, "ts": 1776649791949551, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949563, "dur": 0, + "args": { + "External id": 70325, + "cbid": 200, "correlation": 70325 + } + }, + { + "ph": "f", "id": 70325, "pid": 435, "tid": 435, "ts": 1776649791949563, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949564, "dur": 0, + "args": { + "External id": 70326, + "cbid": 273, "correlation": 70326 + } + }, + { + "ph": "f", "id": 70326, "pid": 435, "tid": 435, "ts": 1776649791949564, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791949573, "dur": 19, + "args": { + "External id": 70327, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70327, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70327, "pid": 0, "tid": 7, "ts": 1776649791949573, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949565, "dur": 5, + "args": { + "External id": 70327, + "cbid": 211, "correlation": 70327 + } + }, + { + "ph": "s", "id": 70327, "pid": 435, "tid": 435, "ts": 1776649791949565, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791949608, "dur": 1, + "args": { + "External id": 9475, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70345, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70345, "pid": 0, "tid": 7, "ts": 1776649791949608, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949599, "dur": 6, + "args": { + "External id": 9475, + "cbid": 211, "correlation": 70345 + } + }, + { + "ph": "s", "id": 70345, "pid": 435, "tid": 435, "ts": 1776649791949599, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949610, "dur": 0, + "args": { + "External id": 70351, + "cbid": 200, "correlation": 70351 + } + }, + { + "ph": "f", "id": 70351, "pid": 435, "tid": 435, "ts": 1776649791949610, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949611, "dur": 0, + "args": { + "External id": 70352, + "cbid": 273, "correlation": 70352 + } + }, + { + "ph": "f", "id": 70352, "pid": 435, "tid": 435, "ts": 1776649791949611, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791949620, "dur": 19, + "args": { + "External id": 70353, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70353, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70353, "pid": 0, "tid": 7, "ts": 1776649791949620, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949612, "dur": 4, + "args": { + "External id": 70353, + "cbid": 211, "correlation": 70353 + } + }, + { + "ph": "s", "id": 70353, "pid": 435, "tid": 435, "ts": 1776649791949612, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791949654, "dur": 1, + "args": { + "External id": 9480, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70371, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70371, "pid": 0, "tid": 7, "ts": 1776649791949654, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949645, "dur": 6, + "args": { + "External id": 9480, + "cbid": 211, "correlation": 70371 + } + }, + { + "ph": "s", "id": 70371, "pid": 435, "tid": 435, "ts": 1776649791949645, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949657, "dur": 0, + "args": { + "External id": 70377, + "cbid": 200, "correlation": 70377 + } + }, + { + "ph": "f", "id": 70377, "pid": 435, "tid": 435, "ts": 1776649791949657, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949658, "dur": 0, + "args": { + "External id": 70378, + "cbid": 273, "correlation": 70378 + } + }, + { + "ph": "f", "id": 70378, "pid": 435, "tid": 435, "ts": 1776649791949658, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791949667, "dur": 19, + "args": { + "External id": 70379, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70379, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70379, "pid": 0, "tid": 7, "ts": 1776649791949667, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949659, "dur": 4, + "args": { + "External id": 70379, + "cbid": 211, "correlation": 70379 + } + }, + { + "ph": "s", "id": 70379, "pid": 435, "tid": 435, "ts": 1776649791949659, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791949702, "dur": 1, + "args": { + "External id": 9485, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70397, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70397, "pid": 0, "tid": 7, "ts": 1776649791949702, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949693, "dur": 6, + "args": { + "External id": 9485, + "cbid": 211, "correlation": 70397 + } + }, + { + "ph": "s", "id": 70397, "pid": 435, "tid": 435, "ts": 1776649791949693, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949705, "dur": 0, + "args": { + "External id": 70403, + "cbid": 200, "correlation": 70403 + } + }, + { + "ph": "f", "id": 70403, "pid": 435, "tid": 435, "ts": 1776649791949705, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949705, "dur": 0, + "args": { + "External id": 70404, + "cbid": 273, "correlation": 70404 + } + }, + { + "ph": "f", "id": 70404, "pid": 435, "tid": 435, "ts": 1776649791949705, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791949714, "dur": 19, + "args": { + "External id": 70405, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70405, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70405, "pid": 0, "tid": 7, "ts": 1776649791949714, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949706, "dur": 4, + "args": { + "External id": 70405, + "cbid": 211, "correlation": 70405 + } + }, + { + "ph": "s", "id": 70405, "pid": 435, "tid": 435, "ts": 1776649791949706, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791949749, "dur": 1, + "args": { + "External id": 9490, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70423, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70423, "pid": 0, "tid": 7, "ts": 1776649791949749, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949740, "dur": 6, + "args": { + "External id": 9490, + "cbid": 211, "correlation": 70423 + } + }, + { + "ph": "s", "id": 70423, "pid": 435, "tid": 435, "ts": 1776649791949740, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949752, "dur": 0, + "args": { + "External id": 70429, + "cbid": 200, "correlation": 70429 + } + }, + { + "ph": "f", "id": 70429, "pid": 435, "tid": 435, "ts": 1776649791949752, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949752, "dur": 0, + "args": { + "External id": 70430, + "cbid": 273, "correlation": 70430 + } + }, + { + "ph": "f", "id": 70430, "pid": 435, "tid": 435, "ts": 1776649791949752, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791949761, "dur": 19, + "args": { + "External id": 70431, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70431, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70431, "pid": 0, "tid": 7, "ts": 1776649791949761, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949753, "dur": 4, + "args": { + "External id": 70431, + "cbid": 211, "correlation": 70431 + } + }, + { + "ph": "s", "id": 70431, "pid": 435, "tid": 435, "ts": 1776649791949753, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791949796, "dur": 1, + "args": { + "External id": 9495, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70449, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70449, "pid": 0, "tid": 7, "ts": 1776649791949796, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949787, "dur": 6, + "args": { + "External id": 9495, + "cbid": 211, "correlation": 70449 + } + }, + { + "ph": "s", "id": 70449, "pid": 435, "tid": 435, "ts": 1776649791949787, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949799, "dur": 0, + "args": { + "External id": 70455, + "cbid": 200, "correlation": 70455 + } + }, + { + "ph": "f", "id": 70455, "pid": 435, "tid": 435, "ts": 1776649791949799, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949799, "dur": 0, + "args": { + "External id": 70456, + "cbid": 273, "correlation": 70456 + } + }, + { + "ph": "f", "id": 70456, "pid": 435, "tid": 435, "ts": 1776649791949799, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791949808, "dur": 19, + "args": { + "External id": 70457, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70457, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70457, "pid": 0, "tid": 7, "ts": 1776649791949808, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949800, "dur": 5, + "args": { + "External id": 70457, + "cbid": 211, "correlation": 70457 + } + }, + { + "ph": "s", "id": 70457, "pid": 435, "tid": 435, "ts": 1776649791949800, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791949843, "dur": 1, + "args": { + "External id": 9500, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70475, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70475, "pid": 0, "tid": 7, "ts": 1776649791949843, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949834, "dur": 6, + "args": { + "External id": 9500, + "cbid": 211, "correlation": 70475 + } + }, + { + "ph": "s", "id": 70475, "pid": 435, "tid": 435, "ts": 1776649791949834, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949846, "dur": 0, + "args": { + "External id": 70481, + "cbid": 200, "correlation": 70481 + } + }, + { + "ph": "f", "id": 70481, "pid": 435, "tid": 435, "ts": 1776649791949846, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949846, "dur": 0, + "args": { + "External id": 70482, + "cbid": 273, "correlation": 70482 + } + }, + { + "ph": "f", "id": 70482, "pid": 435, "tid": 435, "ts": 1776649791949846, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791949855, "dur": 19, + "args": { + "External id": 70483, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70483, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70483, "pid": 0, "tid": 7, "ts": 1776649791949855, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949847, "dur": 4, + "args": { + "External id": 70483, + "cbid": 211, "correlation": 70483 + } + }, + { + "ph": "s", "id": 70483, "pid": 435, "tid": 435, "ts": 1776649791949847, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791949889, "dur": 1, + "args": { + "External id": 9505, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70501, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70501, "pid": 0, "tid": 7, "ts": 1776649791949889, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949881, "dur": 6, + "args": { + "External id": 9505, + "cbid": 211, "correlation": 70501 + } + }, + { + "ph": "s", "id": 70501, "pid": 435, "tid": 435, "ts": 1776649791949881, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949892, "dur": 0, + "args": { + "External id": 70507, + "cbid": 200, "correlation": 70507 + } + }, + { + "ph": "f", "id": 70507, "pid": 435, "tid": 435, "ts": 1776649791949892, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791949893, "dur": 0, + "args": { + "External id": 70508, + "cbid": 273, "correlation": 70508 + } + }, + { + "ph": "f", "id": 70508, "pid": 435, "tid": 435, "ts": 1776649791949893, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791949902, "dur": 19, + "args": { + "External id": 70509, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70509, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70509, "pid": 0, "tid": 7, "ts": 1776649791949902, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791949894, "dur": 4, + "args": { + "External id": 70509, + "cbid": 211, "correlation": 70509 + } + }, + { + "ph": "s", "id": 70509, "pid": 435, "tid": 435, "ts": 1776649791949894, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791949924, "dur": 3, + "args": { + "External id": 70517, + "cbid": 135, "correlation": 70517 + } + }, + { + "ph": "f", "id": 70517, "pid": 435, "tid": 435, "ts": 1776649791949924, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791949939, "dur": 5, + "args": { + "External id": 70522, + "cbid": 165, "correlation": 70522 + } + }, + { + "ph": "s", "id": 70522, "pid": 435, "tid": 435, "ts": 1776649791949939, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649791949947, "dur": 5, + "args": { + "External id": 70525, + "cbid": 139, "correlation": 70525 + } + }, + { + "ph": "f", "id": 70525, "pid": 435, "tid": 435, "ts": 1776649791949947, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649791950254, "dur": 2, + "args": { + "External id": 9508, + "cbid": 317, "correlation": 70550 + } + }, + { + "ph": "f", "id": 70550, "pid": 435, "tid": 435, "ts": 1776649791950254, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649791950258, "dur": 0, + "args": { + "External id": 9508, + "cbid": 317, "correlation": 70554 + } + }, + { + "ph": "f", "id": 70554, "pid": 435, "tid": 435, "ts": 1776649791950258, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::distribution_elementwise_grid_stride_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::(anonymous namespace)::distribution_nullary_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2} const&, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(int, float)#1}>(int, at::PhiloxCudaState, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::(anonymous namespace)::distribution_nullary_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2} const&, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(int, float)#1})", "pid": 0, "tid": 7, + "ts": 1776649791950277, "dur": 3, + "args": { + "External id": 9508, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70556, + "registers per thread": 38, + "shared memory": 0, + "blocks per SM": 0.592593, + "warps per SM": 4.740741, + "grid": [64, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 7 + } + }, + { + "ph": "f", "id": 70556, "pid": 0, "tid": 7, "ts": 1776649791950277, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791950261, "dur": 14, + "args": { + "External id": 9508, + "cbid": 211, "correlation": 70556 + } + }, + { + "ph": "s", "id": 70556, "pid": 435, "tid": 435, "ts": 1776649791950261, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649791950313, "dur": 0, + "args": { + "External id": 70563, + "cbid": 317, "correlation": 70563 + } + }, + { + "ph": "f", "id": 70563, "pid": 435, "tid": 435, "ts": 1776649791950313, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649791950336, "dur": 0, + "args": { + "External id": 70567, + "cbid": 317, "correlation": 70567 + } + }, + { + "ph": "f", "id": 70567, "pid": 435, "tid": 435, "ts": 1776649791950336, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649791950362, "dur": 0, + "args": { + "External id": 9511, + "cbid": 317, "correlation": 70582 + } + }, + { + "ph": "f", "id": 70582, "pid": 435, "tid": 435, "ts": 1776649791950362, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649791950364, "dur": 0, + "args": { + "External id": 9511, + "cbid": 317, "correlation": 70586 + } + }, + { + "ph": "f", "id": 70586, "pid": 435, "tid": 435, "ts": 1776649791950364, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::distribution_elementwise_grid_stride_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::(anonymous namespace)::distribution_nullary_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2} const&, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(int, float)#1}>(int, at::PhiloxCudaState, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::(anonymous namespace)::distribution_nullary_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2} const&, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(int, float)#1})", "pid": 0, "tid": 7, + "ts": 1776649791950376, "dur": 165, + "args": { + "External id": 9511, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70588, + "registers per thread": 38, + "shared memory": 0, + "blocks per SM": 8.000000, + "warps per SM": 64.000000, + "grid": [864, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 75 + } + }, + { + "ph": "f", "id": 70588, "pid": 0, "tid": 7, "ts": 1776649791950376, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791950366, "dur": 8, + "args": { + "External id": 9511, + "cbid": 211, "correlation": 70588 + } + }, + { + "ph": "s", "id": 70588, "pid": 435, "tid": 435, "ts": 1776649791950366, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor >, at::detail::Array >(int, at::native::AUnaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791950543, "dur": 78, + "args": { + "External id": 9512, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70598, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 70598, "pid": 0, "tid": 7, "ts": 1776649791950543, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791950462, "dur": 11, + "args": { + "External id": 9512, + "cbid": 211, "correlation": 70598 + } + }, + { + "ph": "s", "id": 70598, "pid": 435, "tid": 435, "ts": 1776649791950462, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AbsFunctor, at::detail::Array >(int, at::native::AbsFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791950623, "dur": 77, + "args": { + "External id": 9517, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70618, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 70618, "pid": 0, "tid": 7, "ts": 1776649791950623, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791950573, "dur": 11, + "args": { + "External id": 9517, + "cbid": 211, "correlation": 70618 + } + }, + { + "ph": "s", "id": 70618, "pid": 435, "tid": 435, "ts": 1776649791950573, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<128, 4, at::native::ReduceOp >, unsigned int, c10::Half, 4> >(at::native::ReduceOp >, unsigned int, c10::Half, 4>)", "pid": 0, "tid": 7, + "ts": 1776649791950701, "dur": 72, + "args": { + "External id": 9519, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70631, + "registers per thread": 49, + "shared memory": 1040, + "blocks per SM": 18.962963, + "warps per SM": 75.851852, + "grid": [2048, 1, 1], + "block": [32, 4, 1], + "est. achieved occupancy %": 56 + } + }, + { + "ph": "f", "id": 70631, "pid": 0, "tid": 7, "ts": 1776649791950701, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791950653, "dur": 11, + "args": { + "External id": 9519, + "cbid": 211, "correlation": 70631 + } + }, + { + "ph": "s", "id": 70631, "pid": 435, "tid": 435, "ts": 1776649791950653, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BUnaryFunctor >, at::detail::Array >(int, at::native::BUnaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791950775, "dur": 3, + "args": { + "External id": 9521, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70641, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 4.740741, + "warps per SM": 18.962963, + "grid": [512, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 30 + } + }, + { + "ph": "f", "id": 70641, "pid": 0, "tid": 7, "ts": 1776649791950775, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791950697, "dur": 9, + "args": { + "External id": 9521, + "cbid": 211, "correlation": 70641 + } + }, + { + "ph": "s", "id": 70641, "pid": 435, "tid": 435, "ts": 1776649791950697, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#8}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#8}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791950780, "dur": 3, + "args": { + "External id": 9522, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70651, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 4.740741, + "warps per SM": 18.962963, + "grid": [512, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 30 + } + }, + { + "ph": "f", "id": 70651, "pid": 0, "tid": 7, "ts": 1776649791950780, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791950760, "dur": 9, + "args": { + "External id": 9522, + "cbid": 211, "correlation": 70651 + } + }, + { + "ph": "s", "id": 70651, "pid": 435, "tid": 435, "ts": 1776649791950760, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649791950836, "dur": 193, + "args": { + "External id": 9527, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70661, + "registers per thread": 20, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 70661, "pid": 0, "tid": 7, "ts": 1776649791950836, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791950824, "dur": 9, + "args": { + "External id": 9527, + "cbid": 211, "correlation": 70661 + } + }, + { + "ph": "s", "id": 70661, "pid": 435, "tid": 435, "ts": 1776649791950824, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791951030, "dur": 76, + "args": { + "External id": 9528, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70671, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 70671, "pid": 0, "tid": 7, "ts": 1776649791951030, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791950856, "dur": 7, + "args": { + "External id": 9528, + "cbid": 211, "correlation": 70671 + } + }, + { + "ph": "s", "id": 70671, "pid": 435, "tid": 435, "ts": 1776649791950856, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#8}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#8}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791951108, "dur": 79, + "args": { + "External id": 9529, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70681, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 70681, "pid": 0, "tid": 7, "ts": 1776649791951108, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791950881, "dur": 6, + "args": { + "External id": 9529, + "cbid": 211, "correlation": 70681 + } + }, + { + "ph": "s", "id": 70681, "pid": 435, "tid": 435, "ts": 1776649791950881, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649791951189, "dur": 183, + "args": { + "External id": 9534, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70694, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 70694, "pid": 0, "tid": 7, "ts": 1776649791951189, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791950934, "dur": 9, + "args": { + "External id": 9534, + "cbid": 211, "correlation": 70694 + } + }, + { + "ph": "s", "id": 70694, "pid": 435, "tid": 435, "ts": 1776649791950934, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, at::detail::Array >(int, at::native::CUDAFunctorOnSelf_add, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791951374, "dur": 63, + "args": { + "External id": 9535, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70705, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 70705, "pid": 0, "tid": 7, "ts": 1776649791951374, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791950979, "dur": 9, + "args": { + "External id": 9535, + "cbid": 211, "correlation": 70705 + } + }, + { + "ph": "s", "id": 70705, "pid": 435, "tid": 435, "ts": 1776649791950979, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#1}::operator()() const::{lambda(unsigned char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649791951438, "dur": 111, + "args": { + "External id": 9539, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70718, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 70718, "pid": 0, "tid": 7, "ts": 1776649791951438, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791951009, "dur": 7, + "args": { + "External id": 9539, + "cbid": 211, "correlation": 70718 + } + }, + { + "ph": "s", "id": 70718, "pid": 435, "tid": 435, "ts": 1776649791951009, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast >(at::TensorIteratorBase&, at::native::BUnaryFunctor const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast >(at::TensorIteratorBase&, at::native::BUnaryFunctor const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649791951551, "dur": 77, + "args": { + "External id": 9550, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70738, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 70738, "pid": 0, "tid": 7, "ts": 1776649791951551, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791951108, "dur": 11, + "args": { + "External id": 9550, + "cbid": 211, "correlation": 70738 + } + }, + { + "ph": "s", "id": 70738, "pid": 435, "tid": 435, "ts": 1776649791951108, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649791951630, "dur": 83, + "args": { + "External id": 9553, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70748, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 70748, "pid": 0, "tid": 7, "ts": 1776649791951630, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791951151, "dur": 8, + "args": { + "External id": 9553, + "cbid": 211, "correlation": 70748 + } + }, + { + "ph": "s", "id": 70748, "pid": 435, "tid": 435, "ts": 1776649791951151, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor >, at::detail::Array >(int, at::native::AUnaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791951715, "dur": 33, + "args": { + "External id": 9557, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70758, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 70758, "pid": 0, "tid": 7, "ts": 1776649791951715, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791951211, "dur": 8, + "args": { + "External id": 9557, + "cbid": 211, "correlation": 70758 + } + }, + { + "ph": "s", "id": 70758, "pid": 435, "tid": 435, "ts": 1776649791951211, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649791951750, "dur": 57, + "args": { + "External id": 9561, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70771, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 70771, "pid": 0, "tid": 7, "ts": 1776649791951750, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791951244, "dur": 7, + "args": { + "External id": 9561, + "cbid": 211, "correlation": 70771 + } + }, + { + "ph": "s", "id": 70771, "pid": 435, "tid": 435, "ts": 1776649791951244, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BUnaryFunctor, at::detail::Array >(int, at::native::BUnaryFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791951809, "dur": 33, + "args": { + "External id": 9562, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70785, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 70785, "pid": 0, "tid": 7, "ts": 1776649791951809, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791951280, "dur": 8, + "args": { + "External id": 9562, + "cbid": 211, "correlation": 70785 + } + }, + { + "ph": "s", "id": 70785, "pid": 435, "tid": 435, "ts": 1776649791951280, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649791951843, "dur": 57, + "args": { + "External id": 9567, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70798, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 70798, "pid": 0, "tid": 7, "ts": 1776649791951843, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791951307, "dur": 6, + "args": { + "External id": 9567, + "cbid": 211, "correlation": 70798 + } + }, + { + "ph": "s", "id": 70798, "pid": 435, "tid": 435, "ts": 1776649791951307, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy_contig, unsigned int, 3, 128, 1>(at::native::(anonymous namespace)::OpaqueType<1u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata, unsigned int, 128, 1>, at::native::(anonymous namespace)::TensorSizeStride, int, unsigned int)", "pid": 0, "tid": 7, + "ts": 1776649791951902, "dur": 127, + "args": { + "External id": 9573, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70810, + "registers per thread": 21, + "shared memory": 0, + "blocks per SM": 4.000000, + "warps per SM": 64.000000, + "grid": [216, 2, 1], + "block": [512, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 70810, "pid": 0, "tid": 7, "ts": 1776649791951902, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791951397, "dur": 17, + "args": { + "External id": 9573, + "cbid": 211, "correlation": 70810 + } + }, + { + "ph": "s", "id": 70810, "pid": 435, "tid": 435, "ts": 1776649791951397, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649791952031, "dur": 135, + "args": { + "External id": 9584, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70823, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 70823, "pid": 0, "tid": 7, "ts": 1776649791952031, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791951493, "dur": 10, + "args": { + "External id": 9584, + "cbid": 211, "correlation": 70823 + } + }, + { + "ph": "s", "id": 70823, "pid": 435, "tid": 435, "ts": 1776649791951493, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, at::detail::Array >(int, at::native::CUDAFunctorOnSelf_add, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791952168, "dur": 63, + "args": { + "External id": 9587, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70834, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 70834, "pid": 0, "tid": 7, "ts": 1776649791952168, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791951548, "dur": 8, + "args": { + "External id": 9587, + "cbid": 211, "correlation": 70834 + } + }, + { + "ph": "s", "id": 70834, "pid": 435, "tid": 435, "ts": 1776649791951548, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649791952233, "dur": 179, + "args": { + "External id": 9591, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70847, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 70847, "pid": 0, "tid": 7, "ts": 1776649791952233, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791951583, "dur": 8, + "args": { + "External id": 9591, + "cbid": 211, "correlation": 70847 + } + }, + { + "ph": "s", "id": 70847, "pid": 435, "tid": 435, "ts": 1776649791951583, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor >, at::detail::Array >(int, at::native::BinaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791952414, "dur": 114, + "args": { + "External id": 9592, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70858, + "registers per thread": 22, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 70858, "pid": 0, "tid": 7, "ts": 1776649791952414, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791951610, "dur": 7, + "args": { + "External id": 9592, + "cbid": 211, "correlation": 70858 + } + }, + { + "ph": "s", "id": 70858, "pid": 435, "tid": 435, "ts": 1776649791951610, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791951857, "dur": 7, + "args": { + "External id": 9594, + "cbid": 210, "correlation": 70885 + } + }, + { + "ph": "f", "id": 70885, "pid": 435, "tid": 435, "ts": 1776649791951857, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791951880, "dur": 0, + "args": { + "External id": 9594, + "cbid": 210, "correlation": 70886 + } + }, + { + "ph": "f", "id": 70886, "pid": 435, "tid": 435, "ts": 1776649791951880, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791952530, "dur": 49, + "args": { + "External id": 9594, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70887, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70887, "pid": 0, "tid": 7, "ts": 1776649791952530, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791951886, "dur": 14, + "args": { + "External id": 9594, + "cbid": 211, "correlation": 70887 + } + }, + { + "ph": "s", "id": 70887, "pid": 435, "tid": 435, "ts": 1776649791951886, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791951936, "dur": 0, + "args": { + "External id": 9596, + "cbid": 210, "correlation": 70914 + } + }, + { + "ph": "f", "id": 70914, "pid": 435, "tid": 435, "ts": 1776649791951936, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791952581, "dur": 44, + "args": { + "External id": 9596, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70915, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70915, "pid": 0, "tid": 7, "ts": 1776649791952581, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791951938, "dur": 7, + "args": { + "External id": 9596, + "cbid": 211, "correlation": 70915 + } + }, + { + "ph": "s", "id": 70915, "pid": 435, "tid": 435, "ts": 1776649791951938, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791951969, "dur": 0, + "args": { + "External id": 9598, + "cbid": 210, "correlation": 70942 + } + }, + { + "ph": "f", "id": 70942, "pid": 435, "tid": 435, "ts": 1776649791951969, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791952627, "dur": 51, + "args": { + "External id": 9598, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70943, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70943, "pid": 0, "tid": 7, "ts": 1776649791952627, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791951971, "dur": 6, + "args": { + "External id": 9598, + "cbid": 211, "correlation": 70943 + } + }, + { + "ph": "s", "id": 70943, "pid": 435, "tid": 435, "ts": 1776649791951971, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791952001, "dur": 0, + "args": { + "External id": 9600, + "cbid": 210, "correlation": 70970 + } + }, + { + "ph": "f", "id": 70970, "pid": 435, "tid": 435, "ts": 1776649791952001, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791952679, "dur": 54, + "args": { + "External id": 9600, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70971, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70971, "pid": 0, "tid": 7, "ts": 1776649791952679, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791952002, "dur": 6, + "args": { + "External id": 9600, + "cbid": 211, "correlation": 70971 + } + }, + { + "ph": "s", "id": 70971, "pid": 435, "tid": 435, "ts": 1776649791952002, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791952032, "dur": 0, + "args": { + "External id": 9602, + "cbid": 210, "correlation": 70998 + } + }, + { + "ph": "f", "id": 70998, "pid": 435, "tid": 435, "ts": 1776649791952032, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791952735, "dur": 49, + "args": { + "External id": 9602, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 70999, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 70999, "pid": 0, "tid": 7, "ts": 1776649791952735, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791952033, "dur": 6, + "args": { + "External id": 9602, + "cbid": 211, "correlation": 70999 + } + }, + { + "ph": "s", "id": 70999, "pid": 435, "tid": 435, "ts": 1776649791952033, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791952063, "dur": 0, + "args": { + "External id": 9604, + "cbid": 210, "correlation": 71026 + } + }, + { + "ph": "f", "id": 71026, "pid": 435, "tid": 435, "ts": 1776649791952063, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791952786, "dur": 51, + "args": { + "External id": 9604, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71027, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71027, "pid": 0, "tid": 7, "ts": 1776649791952786, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791952064, "dur": 6, + "args": { + "External id": 9604, + "cbid": 211, "correlation": 71027 + } + }, + { + "ph": "s", "id": 71027, "pid": 435, "tid": 435, "ts": 1776649791952064, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791952093, "dur": 0, + "args": { + "External id": 9606, + "cbid": 210, "correlation": 71054 + } + }, + { + "ph": "f", "id": 71054, "pid": 435, "tid": 435, "ts": 1776649791952093, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791952839, "dur": 52, + "args": { + "External id": 9606, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71055, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71055, "pid": 0, "tid": 7, "ts": 1776649791952839, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791952094, "dur": 6, + "args": { + "External id": 9606, + "cbid": 211, "correlation": 71055 + } + }, + { + "ph": "s", "id": 71055, "pid": 435, "tid": 435, "ts": 1776649791952094, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791952123, "dur": 0, + "args": { + "External id": 9608, + "cbid": 210, "correlation": 71082 + } + }, + { + "ph": "f", "id": 71082, "pid": 435, "tid": 435, "ts": 1776649791952123, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791952892, "dur": 49, + "args": { + "External id": 9608, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71083, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71083, "pid": 0, "tid": 7, "ts": 1776649791952892, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791952125, "dur": 5, + "args": { + "External id": 9608, + "cbid": 211, "correlation": 71083 + } + }, + { + "ph": "s", "id": 71083, "pid": 435, "tid": 435, "ts": 1776649791952125, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791952154, "dur": 0, + "args": { + "External id": 9610, + "cbid": 210, "correlation": 71110 + } + }, + { + "ph": "f", "id": 71110, "pid": 435, "tid": 435, "ts": 1776649791952154, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791952943, "dur": 44, + "args": { + "External id": 9610, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71111, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71111, "pid": 0, "tid": 7, "ts": 1776649791952943, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791952155, "dur": 6, + "args": { + "External id": 9610, + "cbid": 211, "correlation": 71111 + } + }, + { + "ph": "s", "id": 71111, "pid": 435, "tid": 435, "ts": 1776649791952155, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791952184, "dur": 0, + "args": { + "External id": 9612, + "cbid": 210, "correlation": 71138 + } + }, + { + "ph": "f", "id": 71138, "pid": 435, "tid": 435, "ts": 1776649791952184, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791952989, "dur": 50, + "args": { + "External id": 9612, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71139, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71139, "pid": 0, "tid": 7, "ts": 1776649791952989, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791952185, "dur": 6, + "args": { + "External id": 9612, + "cbid": 211, "correlation": 71139 + } + }, + { + "ph": "s", "id": 71139, "pid": 435, "tid": 435, "ts": 1776649791952185, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791952214, "dur": 830, + "args": { + "External id": 71145, + "cbid": 165, "correlation": 71145 + } + }, + { + "ph": "s", "id": 71145, "pid": 435, "tid": 435, "ts": 1776649791952214, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791953074, "dur": 4, + "args": { + "External id": 71154, + "cbid": 135, "correlation": 71154 + } + }, + { + "ph": "f", "id": 71154, "pid": 435, "tid": 435, "ts": 1776649791953074, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953108, "dur": 1, + "args": { + "External id": 9614, + "cbid": 210, "correlation": 71180 + } + }, + { + "ph": "f", "id": 71180, "pid": 435, "tid": 435, "ts": 1776649791953108, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791953121, "dur": 41, + "args": { + "External id": 9614, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71181, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71181, "pid": 0, "tid": 7, "ts": 1776649791953121, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953110, "dur": 8, + "args": { + "External id": 9614, + "cbid": 211, "correlation": 71181 + } + }, + { + "ph": "s", "id": 71181, "pid": 435, "tid": 435, "ts": 1776649791953110, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953145, "dur": 0, + "args": { + "External id": 9616, + "cbid": 210, "correlation": 71208 + } + }, + { + "ph": "f", "id": 71208, "pid": 435, "tid": 435, "ts": 1776649791953145, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791953164, "dur": 41, + "args": { + "External id": 9616, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71209, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71209, "pid": 0, "tid": 7, "ts": 1776649791953164, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953146, "dur": 6, + "args": { + "External id": 9616, + "cbid": 211, "correlation": 71209 + } + }, + { + "ph": "s", "id": 71209, "pid": 435, "tid": 435, "ts": 1776649791953146, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953176, "dur": 0, + "args": { + "External id": 9618, + "cbid": 210, "correlation": 71236 + } + }, + { + "ph": "f", "id": 71236, "pid": 435, "tid": 435, "ts": 1776649791953176, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791953206, "dur": 40, + "args": { + "External id": 9618, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71237, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71237, "pid": 0, "tid": 7, "ts": 1776649791953206, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953177, "dur": 6, + "args": { + "External id": 9618, + "cbid": 211, "correlation": 71237 + } + }, + { + "ph": "s", "id": 71237, "pid": 435, "tid": 435, "ts": 1776649791953177, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953206, "dur": 0, + "args": { + "External id": 9620, + "cbid": 210, "correlation": 71264 + } + }, + { + "ph": "f", "id": 71264, "pid": 435, "tid": 435, "ts": 1776649791953206, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791953248, "dur": 41, + "args": { + "External id": 9620, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71265, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71265, "pid": 0, "tid": 7, "ts": 1776649791953248, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953208, "dur": 5, + "args": { + "External id": 9620, + "cbid": 211, "correlation": 71265 + } + }, + { + "ph": "s", "id": 71265, "pid": 435, "tid": 435, "ts": 1776649791953208, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953237, "dur": 0, + "args": { + "External id": 9622, + "cbid": 210, "correlation": 71292 + } + }, + { + "ph": "f", "id": 71292, "pid": 435, "tid": 435, "ts": 1776649791953237, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791953291, "dur": 47, + "args": { + "External id": 9622, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71293, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71293, "pid": 0, "tid": 7, "ts": 1776649791953291, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953238, "dur": 5, + "args": { + "External id": 9622, + "cbid": 211, "correlation": 71293 + } + }, + { + "ph": "s", "id": 71293, "pid": 435, "tid": 435, "ts": 1776649791953238, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953267, "dur": 0, + "args": { + "External id": 9624, + "cbid": 210, "correlation": 71320 + } + }, + { + "ph": "f", "id": 71320, "pid": 435, "tid": 435, "ts": 1776649791953267, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791953339, "dur": 42, + "args": { + "External id": 9624, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71321, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71321, "pid": 0, "tid": 7, "ts": 1776649791953339, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953268, "dur": 5, + "args": { + "External id": 9624, + "cbid": 211, "correlation": 71321 + } + }, + { + "ph": "s", "id": 71321, "pid": 435, "tid": 435, "ts": 1776649791953268, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953297, "dur": 0, + "args": { + "External id": 9626, + "cbid": 210, "correlation": 71348 + } + }, + { + "ph": "f", "id": 71348, "pid": 435, "tid": 435, "ts": 1776649791953297, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791953383, "dur": 43, + "args": { + "External id": 9626, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71349, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71349, "pid": 0, "tid": 7, "ts": 1776649791953383, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953298, "dur": 6, + "args": { + "External id": 9626, + "cbid": 211, "correlation": 71349 + } + }, + { + "ph": "s", "id": 71349, "pid": 435, "tid": 435, "ts": 1776649791953298, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953327, "dur": 0, + "args": { + "External id": 9628, + "cbid": 210, "correlation": 71376 + } + }, + { + "ph": "f", "id": 71376, "pid": 435, "tid": 435, "ts": 1776649791953327, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791953427, "dur": 52, + "args": { + "External id": 9628, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71377, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71377, "pid": 0, "tid": 7, "ts": 1776649791953427, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953328, "dur": 5, + "args": { + "External id": 9628, + "cbid": 211, "correlation": 71377 + } + }, + { + "ph": "s", "id": 71377, "pid": 435, "tid": 435, "ts": 1776649791953328, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953357, "dur": 0, + "args": { + "External id": 9630, + "cbid": 210, "correlation": 71404 + } + }, + { + "ph": "f", "id": 71404, "pid": 435, "tid": 435, "ts": 1776649791953357, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791953481, "dur": 52, + "args": { + "External id": 9630, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71405, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71405, "pid": 0, "tid": 7, "ts": 1776649791953481, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953358, "dur": 6, + "args": { + "External id": 9630, + "cbid": 211, "correlation": 71405 + } + }, + { + "ph": "s", "id": 71405, "pid": 435, "tid": 435, "ts": 1776649791953358, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953387, "dur": 0, + "args": { + "External id": 9632, + "cbid": 210, "correlation": 71432 + } + }, + { + "ph": "f", "id": 71432, "pid": 435, "tid": 435, "ts": 1776649791953387, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791953535, "dur": 47, + "args": { + "External id": 9632, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71433, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71433, "pid": 0, "tid": 7, "ts": 1776649791953535, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953388, "dur": 6, + "args": { + "External id": 9632, + "cbid": 211, "correlation": 71433 + } + }, + { + "ph": "s", "id": 71433, "pid": 435, "tid": 435, "ts": 1776649791953388, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953417, "dur": 0, + "args": { + "External id": 9634, + "cbid": 210, "correlation": 71460 + } + }, + { + "ph": "f", "id": 71460, "pid": 435, "tid": 435, "ts": 1776649791953417, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791953583, "dur": 48, + "args": { + "External id": 9634, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71461, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71461, "pid": 0, "tid": 7, "ts": 1776649791953583, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953419, "dur": 5, + "args": { + "External id": 9634, + "cbid": 211, "correlation": 71461 + } + }, + { + "ph": "s", "id": 71461, "pid": 435, "tid": 435, "ts": 1776649791953419, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953447, "dur": 0, + "args": { + "External id": 9636, + "cbid": 210, "correlation": 71488 + } + }, + { + "ph": "f", "id": 71488, "pid": 435, "tid": 435, "ts": 1776649791953447, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791953633, "dur": 51, + "args": { + "External id": 9636, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71489, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71489, "pid": 0, "tid": 7, "ts": 1776649791953633, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953449, "dur": 6, + "args": { + "External id": 9636, + "cbid": 211, "correlation": 71489 + } + }, + { + "ph": "s", "id": 71489, "pid": 435, "tid": 435, "ts": 1776649791953449, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953477, "dur": 0, + "args": { + "External id": 9638, + "cbid": 210, "correlation": 71516 + } + }, + { + "ph": "f", "id": 71516, "pid": 435, "tid": 435, "ts": 1776649791953477, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791953687, "dur": 44, + "args": { + "External id": 9638, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71517, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71517, "pid": 0, "tid": 7, "ts": 1776649791953687, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953479, "dur": 6, + "args": { + "External id": 9638, + "cbid": 211, "correlation": 71517 + } + }, + { + "ph": "s", "id": 71517, "pid": 435, "tid": 435, "ts": 1776649791953479, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953510, "dur": 0, + "args": { + "External id": 9640, + "cbid": 210, "correlation": 71544 + } + }, + { + "ph": "f", "id": 71544, "pid": 435, "tid": 435, "ts": 1776649791953510, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791953733, "dur": 53, + "args": { + "External id": 9640, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71545, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71545, "pid": 0, "tid": 7, "ts": 1776649791953733, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953511, "dur": 6, + "args": { + "External id": 9640, + "cbid": 211, "correlation": 71545 + } + }, + { + "ph": "s", "id": 71545, "pid": 435, "tid": 435, "ts": 1776649791953511, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953540, "dur": 0, + "args": { + "External id": 9642, + "cbid": 210, "correlation": 71572 + } + }, + { + "ph": "f", "id": 71572, "pid": 435, "tid": 435, "ts": 1776649791953540, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791953787, "dur": 50, + "args": { + "External id": 9642, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71573, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71573, "pid": 0, "tid": 7, "ts": 1776649791953787, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953541, "dur": 6, + "args": { + "External id": 9642, + "cbid": 211, "correlation": 71573 + } + }, + { + "ph": "s", "id": 71573, "pid": 435, "tid": 435, "ts": 1776649791953541, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953571, "dur": 0, + "args": { + "External id": 9644, + "cbid": 210, "correlation": 71600 + } + }, + { + "ph": "f", "id": 71600, "pid": 435, "tid": 435, "ts": 1776649791953571, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791953839, "dur": 43, + "args": { + "External id": 9644, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71601, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71601, "pid": 0, "tid": 7, "ts": 1776649791953839, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953572, "dur": 5, + "args": { + "External id": 9644, + "cbid": 211, "correlation": 71601 + } + }, + { + "ph": "s", "id": 71601, "pid": 435, "tid": 435, "ts": 1776649791953572, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953601, "dur": 0, + "args": { + "External id": 9646, + "cbid": 210, "correlation": 71628 + } + }, + { + "ph": "f", "id": 71628, "pid": 435, "tid": 435, "ts": 1776649791953601, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791953884, "dur": 47, + "args": { + "External id": 9646, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71629, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71629, "pid": 0, "tid": 7, "ts": 1776649791953884, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953603, "dur": 5, + "args": { + "External id": 9646, + "cbid": 211, "correlation": 71629 + } + }, + { + "ph": "s", "id": 71629, "pid": 435, "tid": 435, "ts": 1776649791953603, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953632, "dur": 0, + "args": { + "External id": 9648, + "cbid": 210, "correlation": 71656 + } + }, + { + "ph": "f", "id": 71656, "pid": 435, "tid": 435, "ts": 1776649791953632, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791953933, "dur": 47, + "args": { + "External id": 9648, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71657, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71657, "pid": 0, "tid": 7, "ts": 1776649791953933, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953633, "dur": 6, + "args": { + "External id": 9648, + "cbid": 211, "correlation": 71657 + } + }, + { + "ph": "s", "id": 71657, "pid": 435, "tid": 435, "ts": 1776649791953633, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953663, "dur": 0, + "args": { + "External id": 9650, + "cbid": 210, "correlation": 71684 + } + }, + { + "ph": "f", "id": 71684, "pid": 435, "tid": 435, "ts": 1776649791953663, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791953981, "dur": 44, + "args": { + "External id": 9650, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71685, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71685, "pid": 0, "tid": 7, "ts": 1776649791953981, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953664, "dur": 5, + "args": { + "External id": 9650, + "cbid": 211, "correlation": 71685 + } + }, + { + "ph": "s", "id": 71685, "pid": 435, "tid": 435, "ts": 1776649791953664, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953693, "dur": 0, + "args": { + "External id": 9652, + "cbid": 210, "correlation": 71712 + } + }, + { + "ph": "f", "id": 71712, "pid": 435, "tid": 435, "ts": 1776649791953693, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791954027, "dur": 45, + "args": { + "External id": 9652, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71713, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71713, "pid": 0, "tid": 7, "ts": 1776649791954027, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953694, "dur": 6, + "args": { + "External id": 9652, + "cbid": 211, "correlation": 71713 + } + }, + { + "ph": "s", "id": 71713, "pid": 435, "tid": 435, "ts": 1776649791953694, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953724, "dur": 0, + "args": { + "External id": 9654, + "cbid": 210, "correlation": 71740 + } + }, + { + "ph": "f", "id": 71740, "pid": 435, "tid": 435, "ts": 1776649791953724, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791954074, "dur": 49, + "args": { + "External id": 9654, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71741, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71741, "pid": 0, "tid": 7, "ts": 1776649791954074, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953725, "dur": 6, + "args": { + "External id": 9654, + "cbid": 211, "correlation": 71741 + } + }, + { + "ph": "s", "id": 71741, "pid": 435, "tid": 435, "ts": 1776649791953725, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953754, "dur": 0, + "args": { + "External id": 9656, + "cbid": 210, "correlation": 71768 + } + }, + { + "ph": "f", "id": 71768, "pid": 435, "tid": 435, "ts": 1776649791953754, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791954126, "dur": 46, + "args": { + "External id": 9656, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71769, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71769, "pid": 0, "tid": 7, "ts": 1776649791954126, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953755, "dur": 6, + "args": { + "External id": 9656, + "cbid": 211, "correlation": 71769 + } + }, + { + "ph": "s", "id": 71769, "pid": 435, "tid": 435, "ts": 1776649791953755, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953785, "dur": 0, + "args": { + "External id": 9658, + "cbid": 210, "correlation": 71796 + } + }, + { + "ph": "f", "id": 71796, "pid": 435, "tid": 435, "ts": 1776649791953785, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791954173, "dur": 49, + "args": { + "External id": 9658, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71797, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71797, "pid": 0, "tid": 7, "ts": 1776649791954173, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953786, "dur": 6, + "args": { + "External id": 9658, + "cbid": 211, "correlation": 71797 + } + }, + { + "ph": "s", "id": 71797, "pid": 435, "tid": 435, "ts": 1776649791953786, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953815, "dur": 0, + "args": { + "External id": 9660, + "cbid": 210, "correlation": 71824 + } + }, + { + "ph": "f", "id": 71824, "pid": 435, "tid": 435, "ts": 1776649791953815, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791954224, "dur": 44, + "args": { + "External id": 9660, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71825, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71825, "pid": 0, "tid": 7, "ts": 1776649791954224, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953817, "dur": 6, + "args": { + "External id": 9660, + "cbid": 211, "correlation": 71825 + } + }, + { + "ph": "s", "id": 71825, "pid": 435, "tid": 435, "ts": 1776649791953817, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953846, "dur": 0, + "args": { + "External id": 9662, + "cbid": 210, "correlation": 71852 + } + }, + { + "ph": "f", "id": 71852, "pid": 435, "tid": 435, "ts": 1776649791953846, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791954269, "dur": 46, + "args": { + "External id": 9662, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71853, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71853, "pid": 0, "tid": 7, "ts": 1776649791954269, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953847, "dur": 5, + "args": { + "External id": 9662, + "cbid": 211, "correlation": 71853 + } + }, + { + "ph": "s", "id": 71853, "pid": 435, "tid": 435, "ts": 1776649791953847, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953876, "dur": 0, + "args": { + "External id": 9664, + "cbid": 210, "correlation": 71880 + } + }, + { + "ph": "f", "id": 71880, "pid": 435, "tid": 435, "ts": 1776649791953876, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791954317, "dur": 52, + "args": { + "External id": 9664, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71881, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71881, "pid": 0, "tid": 7, "ts": 1776649791954317, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953877, "dur": 6, + "args": { + "External id": 9664, + "cbid": 211, "correlation": 71881 + } + }, + { + "ph": "s", "id": 71881, "pid": 435, "tid": 435, "ts": 1776649791953877, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953907, "dur": 0, + "args": { + "External id": 9666, + "cbid": 210, "correlation": 71908 + } + }, + { + "ph": "f", "id": 71908, "pid": 435, "tid": 435, "ts": 1776649791953907, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791954371, "dur": 49, + "args": { + "External id": 9666, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71909, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71909, "pid": 0, "tid": 7, "ts": 1776649791954371, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953908, "dur": 6, + "args": { + "External id": 9666, + "cbid": 211, "correlation": 71909 + } + }, + { + "ph": "s", "id": 71909, "pid": 435, "tid": 435, "ts": 1776649791953908, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953937, "dur": 0, + "args": { + "External id": 9668, + "cbid": 210, "correlation": 71936 + } + }, + { + "ph": "f", "id": 71936, "pid": 435, "tid": 435, "ts": 1776649791953937, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791954422, "dur": 47, + "args": { + "External id": 9668, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71937, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71937, "pid": 0, "tid": 7, "ts": 1776649791954422, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953938, "dur": 6, + "args": { + "External id": 9668, + "cbid": 211, "correlation": 71937 + } + }, + { + "ph": "s", "id": 71937, "pid": 435, "tid": 435, "ts": 1776649791953938, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953968, "dur": 0, + "args": { + "External id": 9670, + "cbid": 210, "correlation": 71964 + } + }, + { + "ph": "f", "id": 71964, "pid": 435, "tid": 435, "ts": 1776649791953968, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791954470, "dur": 53, + "args": { + "External id": 9670, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71965, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71965, "pid": 0, "tid": 7, "ts": 1776649791954470, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953969, "dur": 5, + "args": { + "External id": 9670, + "cbid": 211, "correlation": 71965 + } + }, + { + "ph": "s", "id": 71965, "pid": 435, "tid": 435, "ts": 1776649791953969, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791953998, "dur": 0, + "args": { + "External id": 9672, + "cbid": 210, "correlation": 71992 + } + }, + { + "ph": "f", "id": 71992, "pid": 435, "tid": 435, "ts": 1776649791953998, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791954525, "dur": 50, + "args": { + "External id": 9672, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 71993, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 71993, "pid": 0, "tid": 7, "ts": 1776649791954525, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791953999, "dur": 6, + "args": { + "External id": 9672, + "cbid": 211, "correlation": 71993 + } + }, + { + "ph": "s", "id": 71993, "pid": 435, "tid": 435, "ts": 1776649791953999, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791954029, "dur": 0, + "args": { + "External id": 9674, + "cbid": 210, "correlation": 72020 + } + }, + { + "ph": "f", "id": 72020, "pid": 435, "tid": 435, "ts": 1776649791954029, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791954577, "dur": 47, + "args": { + "External id": 9674, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72021, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72021, "pid": 0, "tid": 7, "ts": 1776649791954577, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791954030, "dur": 6, + "args": { + "External id": 9674, + "cbid": 211, "correlation": 72021 + } + }, + { + "ph": "s", "id": 72021, "pid": 435, "tid": 435, "ts": 1776649791954030, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791954059, "dur": 0, + "args": { + "External id": 9676, + "cbid": 210, "correlation": 72048 + } + }, + { + "ph": "f", "id": 72048, "pid": 435, "tid": 435, "ts": 1776649791954059, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791954626, "dur": 47, + "args": { + "External id": 9676, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72049, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72049, "pid": 0, "tid": 7, "ts": 1776649791954626, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791954061, "dur": 5, + "args": { + "External id": 9676, + "cbid": 211, "correlation": 72049 + } + }, + { + "ph": "s", "id": 72049, "pid": 435, "tid": 435, "ts": 1776649791954061, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791954090, "dur": 0, + "args": { + "External id": 9678, + "cbid": 210, "correlation": 72076 + } + }, + { + "ph": "f", "id": 72076, "pid": 435, "tid": 435, "ts": 1776649791954090, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791954674, "dur": 47, + "args": { + "External id": 9678, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72077, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72077, "pid": 0, "tid": 7, "ts": 1776649791954674, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791954091, "dur": 6, + "args": { + "External id": 9678, + "cbid": 211, "correlation": 72077 + } + }, + { + "ph": "s", "id": 72077, "pid": 435, "tid": 435, "ts": 1776649791954091, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791954120, "dur": 0, + "args": { + "External id": 9680, + "cbid": 210, "correlation": 72104 + } + }, + { + "ph": "f", "id": 72104, "pid": 435, "tid": 435, "ts": 1776649791954120, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791954722, "dur": 47, + "args": { + "External id": 9680, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72105, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72105, "pid": 0, "tid": 7, "ts": 1776649791954722, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791954121, "dur": 6, + "args": { + "External id": 9680, + "cbid": 211, "correlation": 72105 + } + }, + { + "ph": "s", "id": 72105, "pid": 435, "tid": 435, "ts": 1776649791954121, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791954151, "dur": 0, + "args": { + "External id": 9682, + "cbid": 210, "correlation": 72132 + } + }, + { + "ph": "f", "id": 72132, "pid": 435, "tid": 435, "ts": 1776649791954151, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791954771, "dur": 47, + "args": { + "External id": 9682, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72133, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72133, "pid": 0, "tid": 7, "ts": 1776649791954771, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791954152, "dur": 5, + "args": { + "External id": 9682, + "cbid": 211, "correlation": 72133 + } + }, + { + "ph": "s", "id": 72133, "pid": 435, "tid": 435, "ts": 1776649791954152, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791954182, "dur": 0, + "args": { + "External id": 9684, + "cbid": 210, "correlation": 72160 + } + }, + { + "ph": "f", "id": 72160, "pid": 435, "tid": 435, "ts": 1776649791954182, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791954819, "dur": 49, + "args": { + "External id": 9684, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72161, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72161, "pid": 0, "tid": 7, "ts": 1776649791954819, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791954183, "dur": 6, + "args": { + "External id": 9684, + "cbid": 211, "correlation": 72161 + } + }, + { + "ph": "s", "id": 72161, "pid": 435, "tid": 435, "ts": 1776649791954183, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791954212, "dur": 0, + "args": { + "External id": 9686, + "cbid": 210, "correlation": 72188 + } + }, + { + "ph": "f", "id": 72188, "pid": 435, "tid": 435, "ts": 1776649791954212, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791954870, "dur": 44, + "args": { + "External id": 9686, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72189, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72189, "pid": 0, "tid": 7, "ts": 1776649791954870, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791954213, "dur": 6, + "args": { + "External id": 9686, + "cbid": 211, "correlation": 72189 + } + }, + { + "ph": "s", "id": 72189, "pid": 435, "tid": 435, "ts": 1776649791954213, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791954243, "dur": 0, + "args": { + "External id": 9688, + "cbid": 210, "correlation": 72216 + } + }, + { + "ph": "f", "id": 72216, "pid": 435, "tid": 435, "ts": 1776649791954243, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791954915, "dur": 44, + "args": { + "External id": 9688, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72217, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72217, "pid": 0, "tid": 7, "ts": 1776649791954915, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791954244, "dur": 6, + "args": { + "External id": 9688, + "cbid": 211, "correlation": 72217 + } + }, + { + "ph": "s", "id": 72217, "pid": 435, "tid": 435, "ts": 1776649791954244, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791954273, "dur": 0, + "args": { + "External id": 9690, + "cbid": 210, "correlation": 72244 + } + }, + { + "ph": "f", "id": 72244, "pid": 435, "tid": 435, "ts": 1776649791954273, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791954961, "dur": 47, + "args": { + "External id": 9690, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72245, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72245, "pid": 0, "tid": 7, "ts": 1776649791954961, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791954274, "dur": 6, + "args": { + "External id": 9690, + "cbid": 211, "correlation": 72245 + } + }, + { + "ph": "s", "id": 72245, "pid": 435, "tid": 435, "ts": 1776649791954274, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791954304, "dur": 0, + "args": { + "External id": 9692, + "cbid": 210, "correlation": 72272 + } + }, + { + "ph": "f", "id": 72272, "pid": 435, "tid": 435, "ts": 1776649791954304, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791955009, "dur": 49, + "args": { + "External id": 9692, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72273, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72273, "pid": 0, "tid": 7, "ts": 1776649791955009, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791954305, "dur": 6, + "args": { + "External id": 9692, + "cbid": 211, "correlation": 72273 + } + }, + { + "ph": "s", "id": 72273, "pid": 435, "tid": 435, "ts": 1776649791954305, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791954335, "dur": 0, + "args": { + "External id": 9694, + "cbid": 210, "correlation": 72300 + } + }, + { + "ph": "f", "id": 72300, "pid": 435, "tid": 435, "ts": 1776649791954335, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791955060, "dur": 49, + "args": { + "External id": 9694, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72301, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72301, "pid": 0, "tid": 7, "ts": 1776649791955060, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791954336, "dur": 6, + "args": { + "External id": 9694, + "cbid": 211, "correlation": 72301 + } + }, + { + "ph": "s", "id": 72301, "pid": 435, "tid": 435, "ts": 1776649791954336, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791954365, "dur": 0, + "args": { + "External id": 9696, + "cbid": 210, "correlation": 72328 + } + }, + { + "ph": "f", "id": 72328, "pid": 435, "tid": 435, "ts": 1776649791954365, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791955111, "dur": 52, + "args": { + "External id": 9696, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72329, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72329, "pid": 0, "tid": 7, "ts": 1776649791955111, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791954367, "dur": 6, + "args": { + "External id": 9696, + "cbid": 211, "correlation": 72329 + } + }, + { + "ph": "s", "id": 72329, "pid": 435, "tid": 435, "ts": 1776649791954367, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791954396, "dur": 0, + "args": { + "External id": 9698, + "cbid": 210, "correlation": 72356 + } + }, + { + "ph": "f", "id": 72356, "pid": 435, "tid": 435, "ts": 1776649791954396, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791955164, "dur": 46, + "args": { + "External id": 9698, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72357, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72357, "pid": 0, "tid": 7, "ts": 1776649791955164, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791954397, "dur": 5, + "args": { + "External id": 9698, + "cbid": 211, "correlation": 72357 + } + }, + { + "ph": "s", "id": 72357, "pid": 435, "tid": 435, "ts": 1776649791954397, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791954426, "dur": 0, + "args": { + "External id": 9700, + "cbid": 210, "correlation": 72384 + } + }, + { + "ph": "f", "id": 72384, "pid": 435, "tid": 435, "ts": 1776649791954426, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791955212, "dur": 43, + "args": { + "External id": 9700, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72385, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72385, "pid": 0, "tid": 7, "ts": 1776649791955212, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791954428, "dur": 6, + "args": { + "External id": 9700, + "cbid": 211, "correlation": 72385 + } + }, + { + "ph": "s", "id": 72385, "pid": 435, "tid": 435, "ts": 1776649791954428, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791954457, "dur": 0, + "args": { + "External id": 9702, + "cbid": 210, "correlation": 72412 + } + }, + { + "ph": "f", "id": 72412, "pid": 435, "tid": 435, "ts": 1776649791954457, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791955257, "dur": 46, + "args": { + "External id": 9702, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72413, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72413, "pid": 0, "tid": 7, "ts": 1776649791955257, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791954458, "dur": 5, + "args": { + "External id": 9702, + "cbid": 211, "correlation": 72413 + } + }, + { + "ph": "s", "id": 72413, "pid": 435, "tid": 435, "ts": 1776649791954458, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791954487, "dur": 0, + "args": { + "External id": 9704, + "cbid": 210, "correlation": 72440 + } + }, + { + "ph": "f", "id": 72440, "pid": 435, "tid": 435, "ts": 1776649791954487, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791955305, "dur": 50, + "args": { + "External id": 9704, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72441, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72441, "pid": 0, "tid": 7, "ts": 1776649791955305, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791954488, "dur": 6, + "args": { + "External id": 9704, + "cbid": 211, "correlation": 72441 + } + }, + { + "ph": "s", "id": 72441, "pid": 435, "tid": 435, "ts": 1776649791954488, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791954518, "dur": 0, + "args": { + "External id": 9706, + "cbid": 210, "correlation": 72468 + } + }, + { + "ph": "f", "id": 72468, "pid": 435, "tid": 435, "ts": 1776649791954518, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791955357, "dur": 47, + "args": { + "External id": 9706, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72469, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72469, "pid": 0, "tid": 7, "ts": 1776649791955357, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791954519, "dur": 5, + "args": { + "External id": 9706, + "cbid": 211, "correlation": 72469 + } + }, + { + "ph": "s", "id": 72469, "pid": 435, "tid": 435, "ts": 1776649791954519, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791954548, "dur": 0, + "args": { + "External id": 9708, + "cbid": 210, "correlation": 72496 + } + }, + { + "ph": "f", "id": 72496, "pid": 435, "tid": 435, "ts": 1776649791954548, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791955406, "dur": 44, + "args": { + "External id": 9708, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72497, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72497, "pid": 0, "tid": 7, "ts": 1776649791955406, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791954549, "dur": 6, + "args": { + "External id": 9708, + "cbid": 211, "correlation": 72497 + } + }, + { + "ph": "s", "id": 72497, "pid": 435, "tid": 435, "ts": 1776649791954549, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791954579, "dur": 0, + "args": { + "External id": 9710, + "cbid": 210, "correlation": 72524 + } + }, + { + "ph": "f", "id": 72524, "pid": 435, "tid": 435, "ts": 1776649791954579, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791955453, "dur": 49, + "args": { + "External id": 9710, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72525, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72525, "pid": 0, "tid": 7, "ts": 1776649791955453, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791954580, "dur": 5, + "args": { + "External id": 9710, + "cbid": 211, "correlation": 72525 + } + }, + { + "ph": "s", "id": 72525, "pid": 435, "tid": 435, "ts": 1776649791954580, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791954609, "dur": 0, + "args": { + "External id": 9712, + "cbid": 210, "correlation": 72552 + } + }, + { + "ph": "f", "id": 72552, "pid": 435, "tid": 435, "ts": 1776649791954609, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791955504, "dur": 44, + "args": { + "External id": 9712, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72553, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72553, "pid": 0, "tid": 7, "ts": 1776649791955504, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791954610, "dur": 6, + "args": { + "External id": 9712, + "cbid": 211, "correlation": 72553 + } + }, + { + "ph": "s", "id": 72553, "pid": 435, "tid": 435, "ts": 1776649791954610, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791954644, "dur": 4, + "args": { + "External id": 72564, + "cbid": 135, "correlation": 72564 + } + }, + { + "ph": "f", "id": 72564, "pid": 435, "tid": 435, "ts": 1776649791954644, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791954660, "dur": 896, + "args": { + "External id": 72569, + "cbid": 165, "correlation": 72569 + } + }, + { + "ph": "s", "id": 72569, "pid": 435, "tid": 435, "ts": 1776649791954660, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649791955560, "dur": 5, + "args": { + "External id": 72572, + "cbid": 139, "correlation": 72572 + } + }, + { + "ph": "f", "id": 72572, "pid": 435, "tid": 435, "ts": 1776649791955560, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791955598, "dur": 1, + "args": { + "External id": 9714, + "cbid": 210, "correlation": 72606 + } + }, + { + "ph": "f", "id": 72606, "pid": 435, "tid": 435, "ts": 1776649791955598, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791955611, "dur": 41, + "args": { + "External id": 9714, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72607, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72607, "pid": 0, "tid": 7, "ts": 1776649791955611, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791955600, "dur": 8, + "args": { + "External id": 9714, + "cbid": 211, "correlation": 72607 + } + }, + { + "ph": "s", "id": 72607, "pid": 435, "tid": 435, "ts": 1776649791955600, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791955634, "dur": 0, + "args": { + "External id": 9716, + "cbid": 210, "correlation": 72634 + } + }, + { + "ph": "f", "id": 72634, "pid": 435, "tid": 435, "ts": 1776649791955634, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791955654, "dur": 41, + "args": { + "External id": 9716, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72635, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72635, "pid": 0, "tid": 7, "ts": 1776649791955654, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791955636, "dur": 6, + "args": { + "External id": 9716, + "cbid": 211, "correlation": 72635 + } + }, + { + "ph": "s", "id": 72635, "pid": 435, "tid": 435, "ts": 1776649791955636, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791955666, "dur": 0, + "args": { + "External id": 9718, + "cbid": 210, "correlation": 72662 + } + }, + { + "ph": "f", "id": 72662, "pid": 435, "tid": 435, "ts": 1776649791955666, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791955697, "dur": 49, + "args": { + "External id": 9718, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72663, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72663, "pid": 0, "tid": 7, "ts": 1776649791955697, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791955668, "dur": 6, + "args": { + "External id": 9718, + "cbid": 211, "correlation": 72663 + } + }, + { + "ph": "s", "id": 72663, "pid": 435, "tid": 435, "ts": 1776649791955668, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791955697, "dur": 0, + "args": { + "External id": 9720, + "cbid": 210, "correlation": 72690 + } + }, + { + "ph": "f", "id": 72690, "pid": 435, "tid": 435, "ts": 1776649791955697, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791955747, "dur": 49, + "args": { + "External id": 9720, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72691, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72691, "pid": 0, "tid": 7, "ts": 1776649791955747, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791955699, "dur": 6, + "args": { + "External id": 9720, + "cbid": 211, "correlation": 72691 + } + }, + { + "ph": "s", "id": 72691, "pid": 435, "tid": 435, "ts": 1776649791955699, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791955728, "dur": 0, + "args": { + "External id": 9722, + "cbid": 210, "correlation": 72718 + } + }, + { + "ph": "f", "id": 72718, "pid": 435, "tid": 435, "ts": 1776649791955728, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791955798, "dur": 46, + "args": { + "External id": 9722, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72719, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72719, "pid": 0, "tid": 7, "ts": 1776649791955798, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791955730, "dur": 6, + "args": { + "External id": 9722, + "cbid": 211, "correlation": 72719 + } + }, + { + "ph": "s", "id": 72719, "pid": 435, "tid": 435, "ts": 1776649791955730, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791955759, "dur": 0, + "args": { + "External id": 9724, + "cbid": 210, "correlation": 72746 + } + }, + { + "ph": "f", "id": 72746, "pid": 435, "tid": 435, "ts": 1776649791955759, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791955846, "dur": 46, + "args": { + "External id": 9724, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72747, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72747, "pid": 0, "tid": 7, "ts": 1776649791955846, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791955760, "dur": 6, + "args": { + "External id": 9724, + "cbid": 211, "correlation": 72747 + } + }, + { + "ph": "s", "id": 72747, "pid": 435, "tid": 435, "ts": 1776649791955760, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791955790, "dur": 0, + "args": { + "External id": 9726, + "cbid": 210, "correlation": 72774 + } + }, + { + "ph": "f", "id": 72774, "pid": 435, "tid": 435, "ts": 1776649791955790, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791955894, "dur": 46, + "args": { + "External id": 9726, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72775, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72775, "pid": 0, "tid": 7, "ts": 1776649791955894, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791955791, "dur": 6, + "args": { + "External id": 9726, + "cbid": 211, "correlation": 72775 + } + }, + { + "ph": "s", "id": 72775, "pid": 435, "tid": 435, "ts": 1776649791955791, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791955821, "dur": 0, + "args": { + "External id": 9728, + "cbid": 210, "correlation": 72802 + } + }, + { + "ph": "f", "id": 72802, "pid": 435, "tid": 435, "ts": 1776649791955821, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791955941, "dur": 50, + "args": { + "External id": 9728, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72803, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72803, "pid": 0, "tid": 7, "ts": 1776649791955941, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791955822, "dur": 6, + "args": { + "External id": 9728, + "cbid": 211, "correlation": 72803 + } + }, + { + "ph": "s", "id": 72803, "pid": 435, "tid": 435, "ts": 1776649791955822, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791955863, "dur": 0, + "args": { + "External id": 9730, + "cbid": 210, "correlation": 72830 + } + }, + { + "ph": "f", "id": 72830, "pid": 435, "tid": 435, "ts": 1776649791955863, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791955993, "dur": 44, + "args": { + "External id": 9730, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72831, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72831, "pid": 0, "tid": 7, "ts": 1776649791955993, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791955864, "dur": 6, + "args": { + "External id": 9730, + "cbid": 211, "correlation": 72831 + } + }, + { + "ph": "s", "id": 72831, "pid": 435, "tid": 435, "ts": 1776649791955864, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791955894, "dur": 0, + "args": { + "External id": 9732, + "cbid": 210, "correlation": 72858 + } + }, + { + "ph": "f", "id": 72858, "pid": 435, "tid": 435, "ts": 1776649791955894, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791956039, "dur": 48, + "args": { + "External id": 9732, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72859, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72859, "pid": 0, "tid": 7, "ts": 1776649791956039, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791955895, "dur": 6, + "args": { + "External id": 9732, + "cbid": 211, "correlation": 72859 + } + }, + { + "ph": "s", "id": 72859, "pid": 435, "tid": 435, "ts": 1776649791955895, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791955921, "dur": 170, + "args": { + "External id": 72865, + "cbid": 165, "correlation": 72865 + } + }, + { + "ph": "s", "id": 72865, "pid": 435, "tid": 435, "ts": 1776649791955921, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791956113, "dur": 3, + "args": { + "External id": 72874, + "cbid": 135, "correlation": 72874 + } + }, + { + "ph": "f", "id": 72874, "pid": 435, "tid": 435, "ts": 1776649791956113, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956143, "dur": 1, + "args": { + "External id": 9734, + "cbid": 210, "correlation": 72900 + } + }, + { + "ph": "f", "id": 72900, "pid": 435, "tid": 435, "ts": 1776649791956143, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791956155, "dur": 40, + "args": { + "External id": 9734, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72901, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72901, "pid": 0, "tid": 7, "ts": 1776649791956155, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956145, "dur": 7, + "args": { + "External id": 9734, + "cbid": 211, "correlation": 72901 + } + }, + { + "ph": "s", "id": 72901, "pid": 435, "tid": 435, "ts": 1776649791956145, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956178, "dur": 0, + "args": { + "External id": 9736, + "cbid": 210, "correlation": 72928 + } + }, + { + "ph": "f", "id": 72928, "pid": 435, "tid": 435, "ts": 1776649791956178, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791956196, "dur": 43, + "args": { + "External id": 9736, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72929, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72929, "pid": 0, "tid": 7, "ts": 1776649791956196, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956179, "dur": 6, + "args": { + "External id": 9736, + "cbid": 211, "correlation": 72929 + } + }, + { + "ph": "s", "id": 72929, "pid": 435, "tid": 435, "ts": 1776649791956179, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956209, "dur": 0, + "args": { + "External id": 9738, + "cbid": 210, "correlation": 72956 + } + }, + { + "ph": "f", "id": 72956, "pid": 435, "tid": 435, "ts": 1776649791956209, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791956242, "dur": 46, + "args": { + "External id": 9738, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72957, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72957, "pid": 0, "tid": 7, "ts": 1776649791956242, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956210, "dur": 24, + "args": { + "External id": 9738, + "cbid": 211, "correlation": 72957 + } + }, + { + "ph": "s", "id": 72957, "pid": 435, "tid": 435, "ts": 1776649791956210, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956262, "dur": 0, + "args": { + "External id": 9740, + "cbid": 210, "correlation": 72984 + } + }, + { + "ph": "f", "id": 72984, "pid": 435, "tid": 435, "ts": 1776649791956262, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791956289, "dur": 46, + "args": { + "External id": 9740, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 72985, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 72985, "pid": 0, "tid": 7, "ts": 1776649791956289, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956264, "dur": 6, + "args": { + "External id": 9740, + "cbid": 211, "correlation": 72985 + } + }, + { + "ph": "s", "id": 72985, "pid": 435, "tid": 435, "ts": 1776649791956264, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956294, "dur": 0, + "args": { + "External id": 9742, + "cbid": 210, "correlation": 73012 + } + }, + { + "ph": "f", "id": 73012, "pid": 435, "tid": 435, "ts": 1776649791956294, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791956338, "dur": 41, + "args": { + "External id": 9742, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73013, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73013, "pid": 0, "tid": 7, "ts": 1776649791956338, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956296, "dur": 6, + "args": { + "External id": 9742, + "cbid": 211, "correlation": 73013 + } + }, + { + "ph": "s", "id": 73013, "pid": 435, "tid": 435, "ts": 1776649791956296, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956325, "dur": 0, + "args": { + "External id": 9744, + "cbid": 210, "correlation": 73040 + } + }, + { + "ph": "f", "id": 73040, "pid": 435, "tid": 435, "ts": 1776649791956325, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791956380, "dur": 47, + "args": { + "External id": 9744, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73041, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73041, "pid": 0, "tid": 7, "ts": 1776649791956380, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956327, "dur": 6, + "args": { + "External id": 9744, + "cbid": 211, "correlation": 73041 + } + }, + { + "ph": "s", "id": 73041, "pid": 435, "tid": 435, "ts": 1776649791956327, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956357, "dur": 0, + "args": { + "External id": 9746, + "cbid": 210, "correlation": 73068 + } + }, + { + "ph": "f", "id": 73068, "pid": 435, "tid": 435, "ts": 1776649791956357, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791956429, "dur": 52, + "args": { + "External id": 9746, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73069, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73069, "pid": 0, "tid": 7, "ts": 1776649791956429, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956358, "dur": 6, + "args": { + "External id": 9746, + "cbid": 211, "correlation": 73069 + } + }, + { + "ph": "s", "id": 73069, "pid": 435, "tid": 435, "ts": 1776649791956358, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956387, "dur": 0, + "args": { + "External id": 9748, + "cbid": 210, "correlation": 73096 + } + }, + { + "ph": "f", "id": 73096, "pid": 435, "tid": 435, "ts": 1776649791956387, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791956483, "dur": 49, + "args": { + "External id": 9748, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73097, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73097, "pid": 0, "tid": 7, "ts": 1776649791956483, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956389, "dur": 6, + "args": { + "External id": 9748, + "cbid": 211, "correlation": 73097 + } + }, + { + "ph": "s", "id": 73097, "pid": 435, "tid": 435, "ts": 1776649791956389, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956418, "dur": 0, + "args": { + "External id": 9750, + "cbid": 210, "correlation": 73124 + } + }, + { + "ph": "f", "id": 73124, "pid": 435, "tid": 435, "ts": 1776649791956418, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791956534, "dur": 46, + "args": { + "External id": 9750, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73125, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73125, "pid": 0, "tid": 7, "ts": 1776649791956534, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956420, "dur": 6, + "args": { + "External id": 9750, + "cbid": 211, "correlation": 73125 + } + }, + { + "ph": "s", "id": 73125, "pid": 435, "tid": 435, "ts": 1776649791956420, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956449, "dur": 0, + "args": { + "External id": 9752, + "cbid": 210, "correlation": 73152 + } + }, + { + "ph": "f", "id": 73152, "pid": 435, "tid": 435, "ts": 1776649791956449, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791956583, "dur": 44, + "args": { + "External id": 9752, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73153, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73153, "pid": 0, "tid": 7, "ts": 1776649791956583, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956450, "dur": 6, + "args": { + "External id": 9752, + "cbid": 211, "correlation": 73153 + } + }, + { + "ph": "s", "id": 73153, "pid": 435, "tid": 435, "ts": 1776649791956450, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956480, "dur": 0, + "args": { + "External id": 9754, + "cbid": 210, "correlation": 73180 + } + }, + { + "ph": "f", "id": 73180, "pid": 435, "tid": 435, "ts": 1776649791956480, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791956628, "dur": 46, + "args": { + "External id": 9754, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73181, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73181, "pid": 0, "tid": 7, "ts": 1776649791956628, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956481, "dur": 6, + "args": { + "External id": 9754, + "cbid": 211, "correlation": 73181 + } + }, + { + "ph": "s", "id": 73181, "pid": 435, "tid": 435, "ts": 1776649791956481, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956511, "dur": 0, + "args": { + "External id": 9756, + "cbid": 210, "correlation": 73208 + } + }, + { + "ph": "f", "id": 73208, "pid": 435, "tid": 435, "ts": 1776649791956511, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791956676, "dur": 49, + "args": { + "External id": 9756, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73209, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73209, "pid": 0, "tid": 7, "ts": 1776649791956676, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956512, "dur": 6, + "args": { + "External id": 9756, + "cbid": 211, "correlation": 73209 + } + }, + { + "ph": "s", "id": 73209, "pid": 435, "tid": 435, "ts": 1776649791956512, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956542, "dur": 0, + "args": { + "External id": 9758, + "cbid": 210, "correlation": 73236 + } + }, + { + "ph": "f", "id": 73236, "pid": 435, "tid": 435, "ts": 1776649791956542, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791956727, "dur": 49, + "args": { + "External id": 9758, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73237, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73237, "pid": 0, "tid": 7, "ts": 1776649791956727, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956543, "dur": 6, + "args": { + "External id": 9758, + "cbid": 211, "correlation": 73237 + } + }, + { + "ph": "s", "id": 73237, "pid": 435, "tid": 435, "ts": 1776649791956543, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956573, "dur": 0, + "args": { + "External id": 9760, + "cbid": 210, "correlation": 73264 + } + }, + { + "ph": "f", "id": 73264, "pid": 435, "tid": 435, "ts": 1776649791956573, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791956777, "dur": 49, + "args": { + "External id": 9760, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73265, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73265, "pid": 0, "tid": 7, "ts": 1776649791956777, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956574, "dur": 6, + "args": { + "External id": 9760, + "cbid": 211, "correlation": 73265 + } + }, + { + "ph": "s", "id": 73265, "pid": 435, "tid": 435, "ts": 1776649791956574, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956604, "dur": 0, + "args": { + "External id": 9762, + "cbid": 210, "correlation": 73292 + } + }, + { + "ph": "f", "id": 73292, "pid": 435, "tid": 435, "ts": 1776649791956604, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791956828, "dur": 46, + "args": { + "External id": 9762, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73293, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73293, "pid": 0, "tid": 7, "ts": 1776649791956828, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956605, "dur": 6, + "args": { + "External id": 9762, + "cbid": 211, "correlation": 73293 + } + }, + { + "ph": "s", "id": 73293, "pid": 435, "tid": 435, "ts": 1776649791956605, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956634, "dur": 0, + "args": { + "External id": 9764, + "cbid": 210, "correlation": 73320 + } + }, + { + "ph": "f", "id": 73320, "pid": 435, "tid": 435, "ts": 1776649791956634, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791956876, "dur": 47, + "args": { + "External id": 9764, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73321, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73321, "pid": 0, "tid": 7, "ts": 1776649791956876, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956635, "dur": 6, + "args": { + "External id": 9764, + "cbid": 211, "correlation": 73321 + } + }, + { + "ph": "s", "id": 73321, "pid": 435, "tid": 435, "ts": 1776649791956635, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956664, "dur": 0, + "args": { + "External id": 9766, + "cbid": 210, "correlation": 73348 + } + }, + { + "ph": "f", "id": 73348, "pid": 435, "tid": 435, "ts": 1776649791956664, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791956924, "dur": 47, + "args": { + "External id": 9766, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73349, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73349, "pid": 0, "tid": 7, "ts": 1776649791956924, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956665, "dur": 6, + "args": { + "External id": 9766, + "cbid": 211, "correlation": 73349 + } + }, + { + "ph": "s", "id": 73349, "pid": 435, "tid": 435, "ts": 1776649791956665, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956696, "dur": 0, + "args": { + "External id": 9768, + "cbid": 210, "correlation": 73376 + } + }, + { + "ph": "f", "id": 73376, "pid": 435, "tid": 435, "ts": 1776649791956696, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791956973, "dur": 47, + "args": { + "External id": 9768, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73377, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73377, "pid": 0, "tid": 7, "ts": 1776649791956973, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956698, "dur": 6, + "args": { + "External id": 9768, + "cbid": 211, "correlation": 73377 + } + }, + { + "ph": "s", "id": 73377, "pid": 435, "tid": 435, "ts": 1776649791956698, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956727, "dur": 0, + "args": { + "External id": 9770, + "cbid": 210, "correlation": 73404 + } + }, + { + "ph": "f", "id": 73404, "pid": 435, "tid": 435, "ts": 1776649791956727, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791957022, "dur": 47, + "args": { + "External id": 9770, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73405, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73405, "pid": 0, "tid": 7, "ts": 1776649791957022, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956728, "dur": 6, + "args": { + "External id": 9770, + "cbid": 211, "correlation": 73405 + } + }, + { + "ph": "s", "id": 73405, "pid": 435, "tid": 435, "ts": 1776649791956728, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956757, "dur": 0, + "args": { + "External id": 9772, + "cbid": 210, "correlation": 73432 + } + }, + { + "ph": "f", "id": 73432, "pid": 435, "tid": 435, "ts": 1776649791956757, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791957070, "dur": 44, + "args": { + "External id": 9772, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73433, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73433, "pid": 0, "tid": 7, "ts": 1776649791957070, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956759, "dur": 6, + "args": { + "External id": 9772, + "cbid": 211, "correlation": 73433 + } + }, + { + "ph": "s", "id": 73433, "pid": 435, "tid": 435, "ts": 1776649791956759, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956788, "dur": 0, + "args": { + "External id": 9774, + "cbid": 210, "correlation": 73460 + } + }, + { + "ph": "f", "id": 73460, "pid": 435, "tid": 435, "ts": 1776649791956788, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791957117, "dur": 47, + "args": { + "External id": 9774, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73461, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73461, "pid": 0, "tid": 7, "ts": 1776649791957117, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956789, "dur": 5, + "args": { + "External id": 9774, + "cbid": 211, "correlation": 73461 + } + }, + { + "ph": "s", "id": 73461, "pid": 435, "tid": 435, "ts": 1776649791956789, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956818, "dur": 0, + "args": { + "External id": 9776, + "cbid": 210, "correlation": 73488 + } + }, + { + "ph": "f", "id": 73488, "pid": 435, "tid": 435, "ts": 1776649791956818, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791957165, "dur": 46, + "args": { + "External id": 9776, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73489, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73489, "pid": 0, "tid": 7, "ts": 1776649791957165, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956819, "dur": 6, + "args": { + "External id": 9776, + "cbid": 211, "correlation": 73489 + } + }, + { + "ph": "s", "id": 73489, "pid": 435, "tid": 435, "ts": 1776649791956819, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956848, "dur": 0, + "args": { + "External id": 9778, + "cbid": 210, "correlation": 73516 + } + }, + { + "ph": "f", "id": 73516, "pid": 435, "tid": 435, "ts": 1776649791956848, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791957213, "dur": 52, + "args": { + "External id": 9778, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73517, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73517, "pid": 0, "tid": 7, "ts": 1776649791957213, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956850, "dur": 5, + "args": { + "External id": 9778, + "cbid": 211, "correlation": 73517 + } + }, + { + "ph": "s", "id": 73517, "pid": 435, "tid": 435, "ts": 1776649791956850, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956879, "dur": 0, + "args": { + "External id": 9780, + "cbid": 210, "correlation": 73544 + } + }, + { + "ph": "f", "id": 73544, "pid": 435, "tid": 435, "ts": 1776649791956879, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791957267, "dur": 46, + "args": { + "External id": 9780, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73545, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73545, "pid": 0, "tid": 7, "ts": 1776649791957267, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956880, "dur": 6, + "args": { + "External id": 9780, + "cbid": 211, "correlation": 73545 + } + }, + { + "ph": "s", "id": 73545, "pid": 435, "tid": 435, "ts": 1776649791956880, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956909, "dur": 0, + "args": { + "External id": 9782, + "cbid": 210, "correlation": 73572 + } + }, + { + "ph": "f", "id": 73572, "pid": 435, "tid": 435, "ts": 1776649791956909, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791957314, "dur": 47, + "args": { + "External id": 9782, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73573, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73573, "pid": 0, "tid": 7, "ts": 1776649791957314, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956910, "dur": 6, + "args": { + "External id": 9782, + "cbid": 211, "correlation": 73573 + } + }, + { + "ph": "s", "id": 73573, "pid": 435, "tid": 435, "ts": 1776649791956910, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956940, "dur": 0, + "args": { + "External id": 9784, + "cbid": 210, "correlation": 73600 + } + }, + { + "ph": "f", "id": 73600, "pid": 435, "tid": 435, "ts": 1776649791956940, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791957363, "dur": 51, + "args": { + "External id": 9784, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73601, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73601, "pid": 0, "tid": 7, "ts": 1776649791957363, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956941, "dur": 6, + "args": { + "External id": 9784, + "cbid": 211, "correlation": 73601 + } + }, + { + "ph": "s", "id": 73601, "pid": 435, "tid": 435, "ts": 1776649791956941, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791956971, "dur": 0, + "args": { + "External id": 9786, + "cbid": 210, "correlation": 73628 + } + }, + { + "ph": "f", "id": 73628, "pid": 435, "tid": 435, "ts": 1776649791956971, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791957416, "dur": 44, + "args": { + "External id": 9786, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73629, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73629, "pid": 0, "tid": 7, "ts": 1776649791957416, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791956972, "dur": 6, + "args": { + "External id": 9786, + "cbid": 211, "correlation": 73629 + } + }, + { + "ph": "s", "id": 73629, "pid": 435, "tid": 435, "ts": 1776649791956972, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791957001, "dur": 0, + "args": { + "External id": 9788, + "cbid": 210, "correlation": 73656 + } + }, + { + "ph": "f", "id": 73656, "pid": 435, "tid": 435, "ts": 1776649791957001, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791957461, "dur": 46, + "args": { + "External id": 9788, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73657, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73657, "pid": 0, "tid": 7, "ts": 1776649791957461, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791957002, "dur": 6, + "args": { + "External id": 9788, + "cbid": 211, "correlation": 73657 + } + }, + { + "ph": "s", "id": 73657, "pid": 435, "tid": 435, "ts": 1776649791957002, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791957032, "dur": 0, + "args": { + "External id": 9790, + "cbid": 210, "correlation": 73684 + } + }, + { + "ph": "f", "id": 73684, "pid": 435, "tid": 435, "ts": 1776649791957032, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791957509, "dur": 46, + "args": { + "External id": 9790, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73685, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73685, "pid": 0, "tid": 7, "ts": 1776649791957509, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791957033, "dur": 6, + "args": { + "External id": 9790, + "cbid": 211, "correlation": 73685 + } + }, + { + "ph": "s", "id": 73685, "pid": 435, "tid": 435, "ts": 1776649791957033, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791957062, "dur": 0, + "args": { + "External id": 9792, + "cbid": 210, "correlation": 73712 + } + }, + { + "ph": "f", "id": 73712, "pid": 435, "tid": 435, "ts": 1776649791957062, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791957557, "dur": 50, + "args": { + "External id": 9792, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73713, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73713, "pid": 0, "tid": 7, "ts": 1776649791957557, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791957064, "dur": 6, + "args": { + "External id": 9792, + "cbid": 211, "correlation": 73713 + } + }, + { + "ph": "s", "id": 73713, "pid": 435, "tid": 435, "ts": 1776649791957064, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791957093, "dur": 0, + "args": { + "External id": 9794, + "cbid": 210, "correlation": 73740 + } + }, + { + "ph": "f", "id": 73740, "pid": 435, "tid": 435, "ts": 1776649791957093, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791957609, "dur": 46, + "args": { + "External id": 9794, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73741, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73741, "pid": 0, "tid": 7, "ts": 1776649791957609, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791957094, "dur": 6, + "args": { + "External id": 9794, + "cbid": 211, "correlation": 73741 + } + }, + { + "ph": "s", "id": 73741, "pid": 435, "tid": 435, "ts": 1776649791957094, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791957130, "dur": 1, + "args": { + "External id": 9796, + "cbid": 210, "correlation": 73768 + } + }, + { + "ph": "f", "id": 73768, "pid": 435, "tid": 435, "ts": 1776649791957130, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791957656, "dur": 46, + "args": { + "External id": 9796, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73769, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73769, "pid": 0, "tid": 7, "ts": 1776649791957656, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791957132, "dur": 6, + "args": { + "External id": 9796, + "cbid": 211, "correlation": 73769 + } + }, + { + "ph": "s", "id": 73769, "pid": 435, "tid": 435, "ts": 1776649791957132, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791957162, "dur": 0, + "args": { + "External id": 9798, + "cbid": 210, "correlation": 73796 + } + }, + { + "ph": "f", "id": 73796, "pid": 435, "tid": 435, "ts": 1776649791957162, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791957704, "dur": 47, + "args": { + "External id": 9798, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73797, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73797, "pid": 0, "tid": 7, "ts": 1776649791957704, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791957163, "dur": 6, + "args": { + "External id": 9798, + "cbid": 211, "correlation": 73797 + } + }, + { + "ph": "s", "id": 73797, "pid": 435, "tid": 435, "ts": 1776649791957163, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791957192, "dur": 0, + "args": { + "External id": 9800, + "cbid": 210, "correlation": 73824 + } + }, + { + "ph": "f", "id": 73824, "pid": 435, "tid": 435, "ts": 1776649791957192, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791957753, "dur": 50, + "args": { + "External id": 9800, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73825, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73825, "pid": 0, "tid": 7, "ts": 1776649791957753, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791957194, "dur": 6, + "args": { + "External id": 9800, + "cbid": 211, "correlation": 73825 + } + }, + { + "ph": "s", "id": 73825, "pid": 435, "tid": 435, "ts": 1776649791957194, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791957223, "dur": 0, + "args": { + "External id": 9802, + "cbid": 210, "correlation": 73852 + } + }, + { + "ph": "f", "id": 73852, "pid": 435, "tid": 435, "ts": 1776649791957223, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791957804, "dur": 46, + "args": { + "External id": 9802, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73853, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73853, "pid": 0, "tid": 7, "ts": 1776649791957804, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791957224, "dur": 6, + "args": { + "External id": 9802, + "cbid": 211, "correlation": 73853 + } + }, + { + "ph": "s", "id": 73853, "pid": 435, "tid": 435, "ts": 1776649791957224, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791957254, "dur": 0, + "args": { + "External id": 9804, + "cbid": 210, "correlation": 73880 + } + }, + { + "ph": "f", "id": 73880, "pid": 435, "tid": 435, "ts": 1776649791957254, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791957852, "dur": 50, + "args": { + "External id": 9804, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73881, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73881, "pid": 0, "tid": 7, "ts": 1776649791957852, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791957255, "dur": 6, + "args": { + "External id": 9804, + "cbid": 211, "correlation": 73881 + } + }, + { + "ph": "s", "id": 73881, "pid": 435, "tid": 435, "ts": 1776649791957255, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791957285, "dur": 0, + "args": { + "External id": 9806, + "cbid": 210, "correlation": 73908 + } + }, + { + "ph": "f", "id": 73908, "pid": 435, "tid": 435, "ts": 1776649791957285, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791957903, "dur": 44, + "args": { + "External id": 9806, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73909, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73909, "pid": 0, "tid": 7, "ts": 1776649791957903, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791957286, "dur": 6, + "args": { + "External id": 9806, + "cbid": 211, "correlation": 73909 + } + }, + { + "ph": "s", "id": 73909, "pid": 435, "tid": 435, "ts": 1776649791957286, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791957315, "dur": 0, + "args": { + "External id": 9808, + "cbid": 210, "correlation": 73936 + } + }, + { + "ph": "f", "id": 73936, "pid": 435, "tid": 435, "ts": 1776649791957315, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791957949, "dur": 44, + "args": { + "External id": 9808, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73937, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73937, "pid": 0, "tid": 7, "ts": 1776649791957949, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791957316, "dur": 6, + "args": { + "External id": 9808, + "cbid": 211, "correlation": 73937 + } + }, + { + "ph": "s", "id": 73937, "pid": 435, "tid": 435, "ts": 1776649791957316, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791957346, "dur": 0, + "args": { + "External id": 9810, + "cbid": 210, "correlation": 73964 + } + }, + { + "ph": "f", "id": 73964, "pid": 435, "tid": 435, "ts": 1776649791957346, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791957995, "dur": 50, + "args": { + "External id": 9810, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73965, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73965, "pid": 0, "tid": 7, "ts": 1776649791957995, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791957347, "dur": 6, + "args": { + "External id": 9810, + "cbid": 211, "correlation": 73965 + } + }, + { + "ph": "s", "id": 73965, "pid": 435, "tid": 435, "ts": 1776649791957347, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791957376, "dur": 0, + "args": { + "External id": 9812, + "cbid": 210, "correlation": 73992 + } + }, + { + "ph": "f", "id": 73992, "pid": 435, "tid": 435, "ts": 1776649791957376, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791958047, "dur": 47, + "args": { + "External id": 9812, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 73993, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 73993, "pid": 0, "tid": 7, "ts": 1776649791958047, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791957377, "dur": 6, + "args": { + "External id": 9812, + "cbid": 211, "correlation": 73993 + } + }, + { + "ph": "s", "id": 73993, "pid": 435, "tid": 435, "ts": 1776649791957377, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791957406, "dur": 0, + "args": { + "External id": 9814, + "cbid": 210, "correlation": 74020 + } + }, + { + "ph": "f", "id": 74020, "pid": 435, "tid": 435, "ts": 1776649791957406, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791958095, "dur": 49, + "args": { + "External id": 9814, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74021, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74021, "pid": 0, "tid": 7, "ts": 1776649791958095, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791957408, "dur": 5, + "args": { + "External id": 9814, + "cbid": 211, "correlation": 74021 + } + }, + { + "ph": "s", "id": 74021, "pid": 435, "tid": 435, "ts": 1776649791957408, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791957436, "dur": 0, + "args": { + "External id": 9816, + "cbid": 210, "correlation": 74048 + } + }, + { + "ph": "f", "id": 74048, "pid": 435, "tid": 435, "ts": 1776649791957436, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791958146, "dur": 46, + "args": { + "External id": 9816, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74049, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74049, "pid": 0, "tid": 7, "ts": 1776649791958146, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791957437, "dur": 6, + "args": { + "External id": 9816, + "cbid": 211, "correlation": 74049 + } + }, + { + "ph": "s", "id": 74049, "pid": 435, "tid": 435, "ts": 1776649791957437, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791957467, "dur": 0, + "args": { + "External id": 9818, + "cbid": 210, "correlation": 74076 + } + }, + { + "ph": "f", "id": 74076, "pid": 435, "tid": 435, "ts": 1776649791957467, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791958194, "dur": 49, + "args": { + "External id": 9818, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74077, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74077, "pid": 0, "tid": 7, "ts": 1776649791958194, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791957468, "dur": 6, + "args": { + "External id": 9818, + "cbid": 211, "correlation": 74077 + } + }, + { + "ph": "s", "id": 74077, "pid": 435, "tid": 435, "ts": 1776649791957468, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791957497, "dur": 0, + "args": { + "External id": 9820, + "cbid": 210, "correlation": 74104 + } + }, + { + "ph": "f", "id": 74104, "pid": 435, "tid": 435, "ts": 1776649791957497, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791958244, "dur": 44, + "args": { + "External id": 9820, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74105, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74105, "pid": 0, "tid": 7, "ts": 1776649791958244, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791957498, "dur": 6, + "args": { + "External id": 9820, + "cbid": 211, "correlation": 74105 + } + }, + { + "ph": "s", "id": 74105, "pid": 435, "tid": 435, "ts": 1776649791957498, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791957527, "dur": 0, + "args": { + "External id": 9822, + "cbid": 210, "correlation": 74132 + } + }, + { + "ph": "f", "id": 74132, "pid": 435, "tid": 435, "ts": 1776649791957527, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791958290, "dur": 49, + "args": { + "External id": 9822, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74133, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74133, "pid": 0, "tid": 7, "ts": 1776649791958290, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791957528, "dur": 6, + "args": { + "External id": 9822, + "cbid": 211, "correlation": 74133 + } + }, + { + "ph": "s", "id": 74133, "pid": 435, "tid": 435, "ts": 1776649791957528, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791957557, "dur": 0, + "args": { + "External id": 9824, + "cbid": 210, "correlation": 74160 + } + }, + { + "ph": "f", "id": 74160, "pid": 435, "tid": 435, "ts": 1776649791957557, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791958341, "dur": 44, + "args": { + "External id": 9824, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74161, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74161, "pid": 0, "tid": 7, "ts": 1776649791958341, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791957558, "dur": 6, + "args": { + "External id": 9824, + "cbid": 211, "correlation": 74161 + } + }, + { + "ph": "s", "id": 74161, "pid": 435, "tid": 435, "ts": 1776649791957558, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791957587, "dur": 0, + "args": { + "External id": 9826, + "cbid": 210, "correlation": 74188 + } + }, + { + "ph": "f", "id": 74188, "pid": 435, "tid": 435, "ts": 1776649791957587, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791958386, "dur": 47, + "args": { + "External id": 9826, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74189, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74189, "pid": 0, "tid": 7, "ts": 1776649791958386, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791957589, "dur": 5, + "args": { + "External id": 9826, + "cbid": 211, "correlation": 74189 + } + }, + { + "ph": "s", "id": 74189, "pid": 435, "tid": 435, "ts": 1776649791957589, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791957617, "dur": 0, + "args": { + "External id": 9828, + "cbid": 210, "correlation": 74216 + } + }, + { + "ph": "f", "id": 74216, "pid": 435, "tid": 435, "ts": 1776649791957617, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791958435, "dur": 47, + "args": { + "External id": 9828, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74217, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74217, "pid": 0, "tid": 7, "ts": 1776649791958435, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791957619, "dur": 6, + "args": { + "External id": 9828, + "cbid": 211, "correlation": 74217 + } + }, + { + "ph": "s", "id": 74217, "pid": 435, "tid": 435, "ts": 1776649791957619, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791957648, "dur": 0, + "args": { + "External id": 9830, + "cbid": 210, "correlation": 74244 + } + }, + { + "ph": "f", "id": 74244, "pid": 435, "tid": 435, "ts": 1776649791957648, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791958484, "dur": 52, + "args": { + "External id": 9830, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74245, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74245, "pid": 0, "tid": 7, "ts": 1776649791958484, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791957649, "dur": 6, + "args": { + "External id": 9830, + "cbid": 211, "correlation": 74245 + } + }, + { + "ph": "s", "id": 74245, "pid": 435, "tid": 435, "ts": 1776649791957649, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791957678, "dur": 0, + "args": { + "External id": 9832, + "cbid": 210, "correlation": 74272 + } + }, + { + "ph": "f", "id": 74272, "pid": 435, "tid": 435, "ts": 1776649791957678, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791958538, "dur": 46, + "args": { + "External id": 9832, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74273, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74273, "pid": 0, "tid": 7, "ts": 1776649791958538, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791957679, "dur": 6, + "args": { + "External id": 9832, + "cbid": 211, "correlation": 74273 + } + }, + { + "ph": "s", "id": 74273, "pid": 435, "tid": 435, "ts": 1776649791957679, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791957713, "dur": 4, + "args": { + "External id": 74284, + "cbid": 135, "correlation": 74284 + } + }, + { + "ph": "f", "id": 74284, "pid": 435, "tid": 435, "ts": 1776649791957713, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791957729, "dur": 865, + "args": { + "External id": 74289, + "cbid": 165, "correlation": 74289 + } + }, + { + "ph": "s", "id": 74289, "pid": 435, "tid": 435, "ts": 1776649791957729, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649791958597, "dur": 4, + "args": { + "External id": 74292, + "cbid": 139, "correlation": 74292 + } + }, + { + "ph": "f", "id": 74292, "pid": 435, "tid": 435, "ts": 1776649791958597, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791958635, "dur": 1, + "args": { + "External id": 9834, + "cbid": 210, "correlation": 74326 + } + }, + { + "ph": "f", "id": 74326, "pid": 435, "tid": 435, "ts": 1776649791958635, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791958648, "dur": 42, + "args": { + "External id": 9834, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74327, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74327, "pid": 0, "tid": 7, "ts": 1776649791958648, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791958637, "dur": 8, + "args": { + "External id": 9834, + "cbid": 211, "correlation": 74327 + } + }, + { + "ph": "s", "id": 74327, "pid": 435, "tid": 435, "ts": 1776649791958637, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791958671, "dur": 0, + "args": { + "External id": 9836, + "cbid": 210, "correlation": 74354 + } + }, + { + "ph": "f", "id": 74354, "pid": 435, "tid": 435, "ts": 1776649791958671, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791958691, "dur": 44, + "args": { + "External id": 9836, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74355, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74355, "pid": 0, "tid": 7, "ts": 1776649791958691, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791958672, "dur": 6, + "args": { + "External id": 9836, + "cbid": 211, "correlation": 74355 + } + }, + { + "ph": "s", "id": 74355, "pid": 435, "tid": 435, "ts": 1776649791958672, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791958702, "dur": 0, + "args": { + "External id": 9838, + "cbid": 210, "correlation": 74382 + } + }, + { + "ph": "f", "id": 74382, "pid": 435, "tid": 435, "ts": 1776649791958702, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791958737, "dur": 46, + "args": { + "External id": 9838, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74383, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74383, "pid": 0, "tid": 7, "ts": 1776649791958737, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791958703, "dur": 6, + "args": { + "External id": 9838, + "cbid": 211, "correlation": 74383 + } + }, + { + "ph": "s", "id": 74383, "pid": 435, "tid": 435, "ts": 1776649791958703, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791958733, "dur": 0, + "args": { + "External id": 9840, + "cbid": 210, "correlation": 74410 + } + }, + { + "ph": "f", "id": 74410, "pid": 435, "tid": 435, "ts": 1776649791958733, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791958784, "dur": 50, + "args": { + "External id": 9840, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74411, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74411, "pid": 0, "tid": 7, "ts": 1776649791958784, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791958734, "dur": 6, + "args": { + "External id": 9840, + "cbid": 211, "correlation": 74411 + } + }, + { + "ph": "s", "id": 74411, "pid": 435, "tid": 435, "ts": 1776649791958734, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791958763, "dur": 0, + "args": { + "External id": 9842, + "cbid": 210, "correlation": 74438 + } + }, + { + "ph": "f", "id": 74438, "pid": 435, "tid": 435, "ts": 1776649791958763, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791958836, "dur": 47, + "args": { + "External id": 9842, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74439, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74439, "pid": 0, "tid": 7, "ts": 1776649791958836, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791958764, "dur": 6, + "args": { + "External id": 9842, + "cbid": 211, "correlation": 74439 + } + }, + { + "ph": "s", "id": 74439, "pid": 435, "tid": 435, "ts": 1776649791958764, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791958793, "dur": 0, + "args": { + "External id": 9844, + "cbid": 210, "correlation": 74466 + } + }, + { + "ph": "f", "id": 74466, "pid": 435, "tid": 435, "ts": 1776649791958793, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791958884, "dur": 47, + "args": { + "External id": 9844, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74467, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74467, "pid": 0, "tid": 7, "ts": 1776649791958884, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791958794, "dur": 6, + "args": { + "External id": 9844, + "cbid": 211, "correlation": 74467 + } + }, + { + "ph": "s", "id": 74467, "pid": 435, "tid": 435, "ts": 1776649791958794, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791958824, "dur": 0, + "args": { + "External id": 9846, + "cbid": 210, "correlation": 74494 + } + }, + { + "ph": "f", "id": 74494, "pid": 435, "tid": 435, "ts": 1776649791958824, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791958933, "dur": 48, + "args": { + "External id": 9846, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74495, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74495, "pid": 0, "tid": 7, "ts": 1776649791958933, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791958825, "dur": 6, + "args": { + "External id": 9846, + "cbid": 211, "correlation": 74495 + } + }, + { + "ph": "s", "id": 74495, "pid": 435, "tid": 435, "ts": 1776649791958825, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791958854, "dur": 0, + "args": { + "External id": 9848, + "cbid": 210, "correlation": 74522 + } + }, + { + "ph": "f", "id": 74522, "pid": 435, "tid": 435, "ts": 1776649791958854, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791958982, "dur": 48, + "args": { + "External id": 9848, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74523, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74523, "pid": 0, "tid": 7, "ts": 1776649791958982, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791958856, "dur": 6, + "args": { + "External id": 9848, + "cbid": 211, "correlation": 74523 + } + }, + { + "ph": "s", "id": 74523, "pid": 435, "tid": 435, "ts": 1776649791958856, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791958885, "dur": 0, + "args": { + "External id": 9850, + "cbid": 210, "correlation": 74550 + } + }, + { + "ph": "f", "id": 74550, "pid": 435, "tid": 435, "ts": 1776649791958885, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791959032, "dur": 47, + "args": { + "External id": 9850, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74551, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74551, "pid": 0, "tid": 7, "ts": 1776649791959032, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791958886, "dur": 6, + "args": { + "External id": 9850, + "cbid": 211, "correlation": 74551 + } + }, + { + "ph": "s", "id": 74551, "pid": 435, "tid": 435, "ts": 1776649791958886, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791958915, "dur": 0, + "args": { + "External id": 9852, + "cbid": 210, "correlation": 74578 + } + }, + { + "ph": "f", "id": 74578, "pid": 435, "tid": 435, "ts": 1776649791958915, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791959081, "dur": 47, + "args": { + "External id": 9852, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74579, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74579, "pid": 0, "tid": 7, "ts": 1776649791959081, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791958916, "dur": 6, + "args": { + "External id": 9852, + "cbid": 211, "correlation": 74579 + } + }, + { + "ph": "s", "id": 74579, "pid": 435, "tid": 435, "ts": 1776649791958916, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791958942, "dur": 190, + "args": { + "External id": 74585, + "cbid": 165, "correlation": 74585 + } + }, + { + "ph": "s", "id": 74585, "pid": 435, "tid": 435, "ts": 1776649791958942, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791959152, "dur": 3, + "args": { + "External id": 74594, + "cbid": 135, "correlation": 74594 + } + }, + { + "ph": "f", "id": 74594, "pid": 435, "tid": 435, "ts": 1776649791959152, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959182, "dur": 1, + "args": { + "External id": 9854, + "cbid": 210, "correlation": 74620 + } + }, + { + "ph": "f", "id": 74620, "pid": 435, "tid": 435, "ts": 1776649791959182, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791959194, "dur": 41, + "args": { + "External id": 9854, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74621, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74621, "pid": 0, "tid": 7, "ts": 1776649791959194, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959184, "dur": 7, + "args": { + "External id": 9854, + "cbid": 211, "correlation": 74621 + } + }, + { + "ph": "s", "id": 74621, "pid": 435, "tid": 435, "ts": 1776649791959184, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959217, "dur": 0, + "args": { + "External id": 9856, + "cbid": 210, "correlation": 74648 + } + }, + { + "ph": "f", "id": 74648, "pid": 435, "tid": 435, "ts": 1776649791959217, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791959237, "dur": 43, + "args": { + "External id": 9856, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74649, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74649, "pid": 0, "tid": 7, "ts": 1776649791959237, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959219, "dur": 6, + "args": { + "External id": 9856, + "cbid": 211, "correlation": 74649 + } + }, + { + "ph": "s", "id": 74649, "pid": 435, "tid": 435, "ts": 1776649791959219, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959249, "dur": 0, + "args": { + "External id": 9858, + "cbid": 210, "correlation": 74676 + } + }, + { + "ph": "f", "id": 74676, "pid": 435, "tid": 435, "ts": 1776649791959249, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791959283, "dur": 46, + "args": { + "External id": 9858, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74677, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74677, "pid": 0, "tid": 7, "ts": 1776649791959283, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959250, "dur": 6, + "args": { + "External id": 9858, + "cbid": 211, "correlation": 74677 + } + }, + { + "ph": "s", "id": 74677, "pid": 435, "tid": 435, "ts": 1776649791959250, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959280, "dur": 0, + "args": { + "External id": 9860, + "cbid": 210, "correlation": 74704 + } + }, + { + "ph": "f", "id": 74704, "pid": 435, "tid": 435, "ts": 1776649791959280, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791959331, "dur": 52, + "args": { + "External id": 9860, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74705, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74705, "pid": 0, "tid": 7, "ts": 1776649791959331, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959281, "dur": 6, + "args": { + "External id": 9860, + "cbid": 211, "correlation": 74705 + } + }, + { + "ph": "s", "id": 74705, "pid": 435, "tid": 435, "ts": 1776649791959281, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959310, "dur": 0, + "args": { + "External id": 9862, + "cbid": 210, "correlation": 74732 + } + }, + { + "ph": "f", "id": 74732, "pid": 435, "tid": 435, "ts": 1776649791959310, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791959385, "dur": 46, + "args": { + "External id": 9862, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74733, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74733, "pid": 0, "tid": 7, "ts": 1776649791959385, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959311, "dur": 6, + "args": { + "External id": 9862, + "cbid": 211, "correlation": 74733 + } + }, + { + "ph": "s", "id": 74733, "pid": 435, "tid": 435, "ts": 1776649791959311, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959341, "dur": 0, + "args": { + "External id": 9864, + "cbid": 210, "correlation": 74760 + } + }, + { + "ph": "f", "id": 74760, "pid": 435, "tid": 435, "ts": 1776649791959341, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791959433, "dur": 49, + "args": { + "External id": 9864, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74761, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74761, "pid": 0, "tid": 7, "ts": 1776649791959433, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959342, "dur": 6, + "args": { + "External id": 9864, + "cbid": 211, "correlation": 74761 + } + }, + { + "ph": "s", "id": 74761, "pid": 435, "tid": 435, "ts": 1776649791959342, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959371, "dur": 0, + "args": { + "External id": 9866, + "cbid": 210, "correlation": 74788 + } + }, + { + "ph": "f", "id": 74788, "pid": 435, "tid": 435, "ts": 1776649791959371, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791959484, "dur": 47, + "args": { + "External id": 9866, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74789, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74789, "pid": 0, "tid": 7, "ts": 1776649791959484, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959372, "dur": 6, + "args": { + "External id": 9866, + "cbid": 211, "correlation": 74789 + } + }, + { + "ph": "s", "id": 74789, "pid": 435, "tid": 435, "ts": 1776649791959372, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959402, "dur": 0, + "args": { + "External id": 9868, + "cbid": 210, "correlation": 74816 + } + }, + { + "ph": "f", "id": 74816, "pid": 435, "tid": 435, "ts": 1776649791959402, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791959533, "dur": 47, + "args": { + "External id": 9868, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74817, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74817, "pid": 0, "tid": 7, "ts": 1776649791959533, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959403, "dur": 6, + "args": { + "External id": 9868, + "cbid": 211, "correlation": 74817 + } + }, + { + "ph": "s", "id": 74817, "pid": 435, "tid": 435, "ts": 1776649791959403, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959432, "dur": 0, + "args": { + "External id": 9870, + "cbid": 210, "correlation": 74844 + } + }, + { + "ph": "f", "id": 74844, "pid": 435, "tid": 435, "ts": 1776649791959432, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791959582, "dur": 47, + "args": { + "External id": 9870, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74845, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74845, "pid": 0, "tid": 7, "ts": 1776649791959582, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959433, "dur": 6, + "args": { + "External id": 9870, + "cbid": 211, "correlation": 74845 + } + }, + { + "ph": "s", "id": 74845, "pid": 435, "tid": 435, "ts": 1776649791959433, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959462, "dur": 0, + "args": { + "External id": 9872, + "cbid": 210, "correlation": 74872 + } + }, + { + "ph": "f", "id": 74872, "pid": 435, "tid": 435, "ts": 1776649791959462, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791959631, "dur": 50, + "args": { + "External id": 9872, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74873, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74873, "pid": 0, "tid": 7, "ts": 1776649791959631, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959464, "dur": 6, + "args": { + "External id": 9872, + "cbid": 211, "correlation": 74873 + } + }, + { + "ph": "s", "id": 74873, "pid": 435, "tid": 435, "ts": 1776649791959464, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959493, "dur": 0, + "args": { + "External id": 9874, + "cbid": 210, "correlation": 74900 + } + }, + { + "ph": "f", "id": 74900, "pid": 435, "tid": 435, "ts": 1776649791959493, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791959682, "dur": 41, + "args": { + "External id": 9874, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74901, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74901, "pid": 0, "tid": 7, "ts": 1776649791959682, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959494, "dur": 6, + "args": { + "External id": 9874, + "cbid": 211, "correlation": 74901 + } + }, + { + "ph": "s", "id": 74901, "pid": 435, "tid": 435, "ts": 1776649791959494, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959523, "dur": 0, + "args": { + "External id": 9876, + "cbid": 210, "correlation": 74928 + } + }, + { + "ph": "f", "id": 74928, "pid": 435, "tid": 435, "ts": 1776649791959523, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791959725, "dur": 47, + "args": { + "External id": 9876, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74929, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74929, "pid": 0, "tid": 7, "ts": 1776649791959725, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959524, "dur": 6, + "args": { + "External id": 9876, + "cbid": 211, "correlation": 74929 + } + }, + { + "ph": "s", "id": 74929, "pid": 435, "tid": 435, "ts": 1776649791959524, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959553, "dur": 0, + "args": { + "External id": 9878, + "cbid": 210, "correlation": 74956 + } + }, + { + "ph": "f", "id": 74956, "pid": 435, "tid": 435, "ts": 1776649791959553, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791959774, "dur": 44, + "args": { + "External id": 9878, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74957, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74957, "pid": 0, "tid": 7, "ts": 1776649791959774, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959554, "dur": 6, + "args": { + "External id": 9878, + "cbid": 211, "correlation": 74957 + } + }, + { + "ph": "s", "id": 74957, "pid": 435, "tid": 435, "ts": 1776649791959554, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959583, "dur": 0, + "args": { + "External id": 9880, + "cbid": 210, "correlation": 74984 + } + }, + { + "ph": "f", "id": 74984, "pid": 435, "tid": 435, "ts": 1776649791959583, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791959819, "dur": 49, + "args": { + "External id": 9880, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 74985, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 74985, "pid": 0, "tid": 7, "ts": 1776649791959819, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959585, "dur": 6, + "args": { + "External id": 9880, + "cbid": 211, "correlation": 74985 + } + }, + { + "ph": "s", "id": 74985, "pid": 435, "tid": 435, "ts": 1776649791959585, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959614, "dur": 0, + "args": { + "External id": 9882, + "cbid": 210, "correlation": 75012 + } + }, + { + "ph": "f", "id": 75012, "pid": 435, "tid": 435, "ts": 1776649791959614, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791959870, "dur": 49, + "args": { + "External id": 9882, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75013, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75013, "pid": 0, "tid": 7, "ts": 1776649791959870, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959615, "dur": 5, + "args": { + "External id": 9882, + "cbid": 211, "correlation": 75013 + } + }, + { + "ph": "s", "id": 75013, "pid": 435, "tid": 435, "ts": 1776649791959615, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959644, "dur": 0, + "args": { + "External id": 9884, + "cbid": 210, "correlation": 75040 + } + }, + { + "ph": "f", "id": 75040, "pid": 435, "tid": 435, "ts": 1776649791959644, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791959921, "dur": 52, + "args": { + "External id": 9884, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75041, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75041, "pid": 0, "tid": 7, "ts": 1776649791959921, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959645, "dur": 6, + "args": { + "External id": 9884, + "cbid": 211, "correlation": 75041 + } + }, + { + "ph": "s", "id": 75041, "pid": 435, "tid": 435, "ts": 1776649791959645, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959674, "dur": 0, + "args": { + "External id": 9886, + "cbid": 210, "correlation": 75068 + } + }, + { + "ph": "f", "id": 75068, "pid": 435, "tid": 435, "ts": 1776649791959674, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791959974, "dur": 43, + "args": { + "External id": 9886, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75069, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75069, "pid": 0, "tid": 7, "ts": 1776649791959974, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959675, "dur": 6, + "args": { + "External id": 9886, + "cbid": 211, "correlation": 75069 + } + }, + { + "ph": "s", "id": 75069, "pid": 435, "tid": 435, "ts": 1776649791959675, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959704, "dur": 0, + "args": { + "External id": 9888, + "cbid": 210, "correlation": 75096 + } + }, + { + "ph": "f", "id": 75096, "pid": 435, "tid": 435, "ts": 1776649791959704, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791960019, "dur": 50, + "args": { + "External id": 9888, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75097, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75097, "pid": 0, "tid": 7, "ts": 1776649791960019, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959705, "dur": 6, + "args": { + "External id": 9888, + "cbid": 211, "correlation": 75097 + } + }, + { + "ph": "s", "id": 75097, "pid": 435, "tid": 435, "ts": 1776649791959705, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959734, "dur": 0, + "args": { + "External id": 9890, + "cbid": 210, "correlation": 75124 + } + }, + { + "ph": "f", "id": 75124, "pid": 435, "tid": 435, "ts": 1776649791959734, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791960070, "dur": 49, + "args": { + "External id": 9890, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75125, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75125, "pid": 0, "tid": 7, "ts": 1776649791960070, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959735, "dur": 6, + "args": { + "External id": 9890, + "cbid": 211, "correlation": 75125 + } + }, + { + "ph": "s", "id": 75125, "pid": 435, "tid": 435, "ts": 1776649791959735, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959764, "dur": 0, + "args": { + "External id": 9892, + "cbid": 210, "correlation": 75152 + } + }, + { + "ph": "f", "id": 75152, "pid": 435, "tid": 435, "ts": 1776649791959764, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791960121, "dur": 49, + "args": { + "External id": 9892, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75153, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75153, "pid": 0, "tid": 7, "ts": 1776649791960121, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959765, "dur": 6, + "args": { + "External id": 9892, + "cbid": 211, "correlation": 75153 + } + }, + { + "ph": "s", "id": 75153, "pid": 435, "tid": 435, "ts": 1776649791959765, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959794, "dur": 0, + "args": { + "External id": 9894, + "cbid": 210, "correlation": 75180 + } + }, + { + "ph": "f", "id": 75180, "pid": 435, "tid": 435, "ts": 1776649791959794, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791960172, "dur": 47, + "args": { + "External id": 9894, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75181, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75181, "pid": 0, "tid": 7, "ts": 1776649791960172, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959795, "dur": 5, + "args": { + "External id": 9894, + "cbid": 211, "correlation": 75181 + } + }, + { + "ph": "s", "id": 75181, "pid": 435, "tid": 435, "ts": 1776649791959795, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959826, "dur": 0, + "args": { + "External id": 9896, + "cbid": 210, "correlation": 75208 + } + }, + { + "ph": "f", "id": 75208, "pid": 435, "tid": 435, "ts": 1776649791959826, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791960220, "dur": 47, + "args": { + "External id": 9896, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75209, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75209, "pid": 0, "tid": 7, "ts": 1776649791960220, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959827, "dur": 6, + "args": { + "External id": 9896, + "cbid": 211, "correlation": 75209 + } + }, + { + "ph": "s", "id": 75209, "pid": 435, "tid": 435, "ts": 1776649791959827, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959856, "dur": 0, + "args": { + "External id": 9898, + "cbid": 210, "correlation": 75236 + } + }, + { + "ph": "f", "id": 75236, "pid": 435, "tid": 435, "ts": 1776649791959856, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791960268, "dur": 42, + "args": { + "External id": 9898, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75237, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75237, "pid": 0, "tid": 7, "ts": 1776649791960268, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959858, "dur": 6, + "args": { + "External id": 9898, + "cbid": 211, "correlation": 75237 + } + }, + { + "ph": "s", "id": 75237, "pid": 435, "tid": 435, "ts": 1776649791959858, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959887, "dur": 0, + "args": { + "External id": 9900, + "cbid": 210, "correlation": 75264 + } + }, + { + "ph": "f", "id": 75264, "pid": 435, "tid": 435, "ts": 1776649791959887, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791960312, "dur": 45, + "args": { + "External id": 9900, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75265, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75265, "pid": 0, "tid": 7, "ts": 1776649791960312, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959888, "dur": 6, + "args": { + "External id": 9900, + "cbid": 211, "correlation": 75265 + } + }, + { + "ph": "s", "id": 75265, "pid": 435, "tid": 435, "ts": 1776649791959888, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959917, "dur": 0, + "args": { + "External id": 9902, + "cbid": 210, "correlation": 75292 + } + }, + { + "ph": "f", "id": 75292, "pid": 435, "tid": 435, "ts": 1776649791959917, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791960358, "dur": 44, + "args": { + "External id": 9902, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75293, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75293, "pid": 0, "tid": 7, "ts": 1776649791960358, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959919, "dur": 6, + "args": { + "External id": 9902, + "cbid": 211, "correlation": 75293 + } + }, + { + "ph": "s", "id": 75293, "pid": 435, "tid": 435, "ts": 1776649791959919, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959948, "dur": 0, + "args": { + "External id": 9904, + "cbid": 210, "correlation": 75320 + } + }, + { + "ph": "f", "id": 75320, "pid": 435, "tid": 435, "ts": 1776649791959948, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791960404, "dur": 49, + "args": { + "External id": 9904, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75321, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75321, "pid": 0, "tid": 7, "ts": 1776649791960404, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959949, "dur": 6, + "args": { + "External id": 9904, + "cbid": 211, "correlation": 75321 + } + }, + { + "ph": "s", "id": 75321, "pid": 435, "tid": 435, "ts": 1776649791959949, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791959978, "dur": 0, + "args": { + "External id": 9906, + "cbid": 210, "correlation": 75348 + } + }, + { + "ph": "f", "id": 75348, "pid": 435, "tid": 435, "ts": 1776649791959978, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791960455, "dur": 47, + "args": { + "External id": 9906, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75349, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75349, "pid": 0, "tid": 7, "ts": 1776649791960455, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791959979, "dur": 6, + "args": { + "External id": 9906, + "cbid": 211, "correlation": 75349 + } + }, + { + "ph": "s", "id": 75349, "pid": 435, "tid": 435, "ts": 1776649791959979, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791960009, "dur": 0, + "args": { + "External id": 9908, + "cbid": 210, "correlation": 75376 + } + }, + { + "ph": "f", "id": 75376, "pid": 435, "tid": 435, "ts": 1776649791960009, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791960503, "dur": 49, + "args": { + "External id": 9908, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75377, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75377, "pid": 0, "tid": 7, "ts": 1776649791960503, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791960010, "dur": 5, + "args": { + "External id": 9908, + "cbid": 211, "correlation": 75377 + } + }, + { + "ph": "s", "id": 75377, "pid": 435, "tid": 435, "ts": 1776649791960010, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791960039, "dur": 0, + "args": { + "External id": 9910, + "cbid": 210, "correlation": 75404 + } + }, + { + "ph": "f", "id": 75404, "pid": 435, "tid": 435, "ts": 1776649791960039, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791960554, "dur": 45, + "args": { + "External id": 9910, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75405, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75405, "pid": 0, "tid": 7, "ts": 1776649791960554, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791960040, "dur": 5, + "args": { + "External id": 9910, + "cbid": 211, "correlation": 75405 + } + }, + { + "ph": "s", "id": 75405, "pid": 435, "tid": 435, "ts": 1776649791960040, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791960069, "dur": 0, + "args": { + "External id": 9912, + "cbid": 210, "correlation": 75432 + } + }, + { + "ph": "f", "id": 75432, "pid": 435, "tid": 435, "ts": 1776649791960069, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791960601, "dur": 48, + "args": { + "External id": 9912, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75433, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75433, "pid": 0, "tid": 7, "ts": 1776649791960601, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791960070, "dur": 6, + "args": { + "External id": 9912, + "cbid": 211, "correlation": 75433 + } + }, + { + "ph": "s", "id": 75433, "pid": 435, "tid": 435, "ts": 1776649791960070, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791960099, "dur": 0, + "args": { + "External id": 9914, + "cbid": 210, "correlation": 75460 + } + }, + { + "ph": "f", "id": 75460, "pid": 435, "tid": 435, "ts": 1776649791960099, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791960651, "dur": 49, + "args": { + "External id": 9914, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75461, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75461, "pid": 0, "tid": 7, "ts": 1776649791960651, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791960100, "dur": 5, + "args": { + "External id": 9914, + "cbid": 211, "correlation": 75461 + } + }, + { + "ph": "s", "id": 75461, "pid": 435, "tid": 435, "ts": 1776649791960100, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791960129, "dur": 0, + "args": { + "External id": 9916, + "cbid": 210, "correlation": 75488 + } + }, + { + "ph": "f", "id": 75488, "pid": 435, "tid": 435, "ts": 1776649791960129, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791960703, "dur": 49, + "args": { + "External id": 9916, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75489, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75489, "pid": 0, "tid": 7, "ts": 1776649791960703, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791960130, "dur": 6, + "args": { + "External id": 9916, + "cbid": 211, "correlation": 75489 + } + }, + { + "ph": "s", "id": 75489, "pid": 435, "tid": 435, "ts": 1776649791960130, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791960159, "dur": 0, + "args": { + "External id": 9918, + "cbid": 210, "correlation": 75516 + } + }, + { + "ph": "f", "id": 75516, "pid": 435, "tid": 435, "ts": 1776649791960159, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791960753, "dur": 49, + "args": { + "External id": 9918, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75517, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75517, "pid": 0, "tid": 7, "ts": 1776649791960753, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791960160, "dur": 5, + "args": { + "External id": 9918, + "cbid": 211, "correlation": 75517 + } + }, + { + "ph": "s", "id": 75517, "pid": 435, "tid": 435, "ts": 1776649791960160, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791960188, "dur": 0, + "args": { + "External id": 9920, + "cbid": 210, "correlation": 75544 + } + }, + { + "ph": "f", "id": 75544, "pid": 435, "tid": 435, "ts": 1776649791960188, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791960804, "dur": 46, + "args": { + "External id": 9920, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75545, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75545, "pid": 0, "tid": 7, "ts": 1776649791960804, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791960190, "dur": 6, + "args": { + "External id": 9920, + "cbid": 211, "correlation": 75545 + } + }, + { + "ph": "s", "id": 75545, "pid": 435, "tid": 435, "ts": 1776649791960190, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791960236, "dur": 0, + "args": { + "External id": 9922, + "cbid": 210, "correlation": 75572 + } + }, + { + "ph": "f", "id": 75572, "pid": 435, "tid": 435, "ts": 1776649791960236, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791960852, "dur": 48, + "args": { + "External id": 9922, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75573, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75573, "pid": 0, "tid": 7, "ts": 1776649791960852, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791960238, "dur": 7, + "args": { + "External id": 9922, + "cbid": 211, "correlation": 75573 + } + }, + { + "ph": "s", "id": 75573, "pid": 435, "tid": 435, "ts": 1776649791960238, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791960271, "dur": 0, + "args": { + "External id": 9924, + "cbid": 210, "correlation": 75600 + } + }, + { + "ph": "f", "id": 75600, "pid": 435, "tid": 435, "ts": 1776649791960271, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791960902, "dur": 46, + "args": { + "External id": 9924, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75601, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75601, "pid": 0, "tid": 7, "ts": 1776649791960902, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791960272, "dur": 6, + "args": { + "External id": 9924, + "cbid": 211, "correlation": 75601 + } + }, + { + "ph": "s", "id": 75601, "pid": 435, "tid": 435, "ts": 1776649791960272, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791960302, "dur": 0, + "args": { + "External id": 9926, + "cbid": 210, "correlation": 75628 + } + }, + { + "ph": "f", "id": 75628, "pid": 435, "tid": 435, "ts": 1776649791960302, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791960950, "dur": 44, + "args": { + "External id": 9926, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75629, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75629, "pid": 0, "tid": 7, "ts": 1776649791960950, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791960303, "dur": 6, + "args": { + "External id": 9926, + "cbid": 211, "correlation": 75629 + } + }, + { + "ph": "s", "id": 75629, "pid": 435, "tid": 435, "ts": 1776649791960303, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791960332, "dur": 0, + "args": { + "External id": 9928, + "cbid": 210, "correlation": 75656 + } + }, + { + "ph": "f", "id": 75656, "pid": 435, "tid": 435, "ts": 1776649791960332, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791960996, "dur": 48, + "args": { + "External id": 9928, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75657, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75657, "pid": 0, "tid": 7, "ts": 1776649791960996, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791960333, "dur": 6, + "args": { + "External id": 9928, + "cbid": 211, "correlation": 75657 + } + }, + { + "ph": "s", "id": 75657, "pid": 435, "tid": 435, "ts": 1776649791960333, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791960363, "dur": 0, + "args": { + "External id": 9930, + "cbid": 210, "correlation": 75684 + } + }, + { + "ph": "f", "id": 75684, "pid": 435, "tid": 435, "ts": 1776649791960363, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791961046, "dur": 44, + "args": { + "External id": 9930, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75685, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75685, "pid": 0, "tid": 7, "ts": 1776649791961046, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791960364, "dur": 6, + "args": { + "External id": 9930, + "cbid": 211, "correlation": 75685 + } + }, + { + "ph": "s", "id": 75685, "pid": 435, "tid": 435, "ts": 1776649791960364, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791960393, "dur": 0, + "args": { + "External id": 9932, + "cbid": 210, "correlation": 75712 + } + }, + { + "ph": "f", "id": 75712, "pid": 435, "tid": 435, "ts": 1776649791960393, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791961092, "dur": 50, + "args": { + "External id": 9932, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75713, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75713, "pid": 0, "tid": 7, "ts": 1776649791961092, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791960395, "dur": 6, + "args": { + "External id": 9932, + "cbid": 211, "correlation": 75713 + } + }, + { + "ph": "s", "id": 75713, "pid": 435, "tid": 435, "ts": 1776649791960395, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791960423, "dur": 0, + "args": { + "External id": 9934, + "cbid": 210, "correlation": 75740 + } + }, + { + "ph": "f", "id": 75740, "pid": 435, "tid": 435, "ts": 1776649791960423, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791961144, "dur": 47, + "args": { + "External id": 9934, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75741, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75741, "pid": 0, "tid": 7, "ts": 1776649791961144, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791960425, "dur": 6, + "args": { + "External id": 9934, + "cbid": 211, "correlation": 75741 + } + }, + { + "ph": "s", "id": 75741, "pid": 435, "tid": 435, "ts": 1776649791960425, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791960454, "dur": 0, + "args": { + "External id": 9936, + "cbid": 210, "correlation": 75768 + } + }, + { + "ph": "f", "id": 75768, "pid": 435, "tid": 435, "ts": 1776649791960454, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791961192, "dur": 44, + "args": { + "External id": 9936, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75769, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75769, "pid": 0, "tid": 7, "ts": 1776649791961192, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791960455, "dur": 5, + "args": { + "External id": 9936, + "cbid": 211, "correlation": 75769 + } + }, + { + "ph": "s", "id": 75769, "pid": 435, "tid": 435, "ts": 1776649791960455, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791960484, "dur": 0, + "args": { + "External id": 9938, + "cbid": 210, "correlation": 75796 + } + }, + { + "ph": "f", "id": 75796, "pid": 435, "tid": 435, "ts": 1776649791960484, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791961239, "dur": 47, + "args": { + "External id": 9938, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75797, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75797, "pid": 0, "tid": 7, "ts": 1776649791961239, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791960485, "dur": 5, + "args": { + "External id": 9938, + "cbid": 211, "correlation": 75797 + } + }, + { + "ph": "s", "id": 75797, "pid": 435, "tid": 435, "ts": 1776649791960485, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791960514, "dur": 0, + "args": { + "External id": 9940, + "cbid": 210, "correlation": 75824 + } + }, + { + "ph": "f", "id": 75824, "pid": 435, "tid": 435, "ts": 1776649791960514, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791961287, "dur": 46, + "args": { + "External id": 9940, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75825, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75825, "pid": 0, "tid": 7, "ts": 1776649791961287, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791960515, "dur": 5, + "args": { + "External id": 9940, + "cbid": 211, "correlation": 75825 + } + }, + { + "ph": "s", "id": 75825, "pid": 435, "tid": 435, "ts": 1776649791960515, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791960544, "dur": 0, + "args": { + "External id": 9942, + "cbid": 210, "correlation": 75852 + } + }, + { + "ph": "f", "id": 75852, "pid": 435, "tid": 435, "ts": 1776649791960544, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791961335, "dur": 47, + "args": { + "External id": 9942, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75853, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75853, "pid": 0, "tid": 7, "ts": 1776649791961335, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791960545, "dur": 5, + "args": { + "External id": 9942, + "cbid": 211, "correlation": 75853 + } + }, + { + "ph": "s", "id": 75853, "pid": 435, "tid": 435, "ts": 1776649791960545, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791960574, "dur": 0, + "args": { + "External id": 9944, + "cbid": 210, "correlation": 75880 + } + }, + { + "ph": "f", "id": 75880, "pid": 435, "tid": 435, "ts": 1776649791960574, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791961383, "dur": 50, + "args": { + "External id": 9944, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75881, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75881, "pid": 0, "tid": 7, "ts": 1776649791961383, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791960575, "dur": 6, + "args": { + "External id": 9944, + "cbid": 211, "correlation": 75881 + } + }, + { + "ph": "s", "id": 75881, "pid": 435, "tid": 435, "ts": 1776649791960575, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791960603, "dur": 0, + "args": { + "External id": 9946, + "cbid": 210, "correlation": 75908 + } + }, + { + "ph": "f", "id": 75908, "pid": 435, "tid": 435, "ts": 1776649791960603, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791961435, "dur": 48, + "args": { + "External id": 9946, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75909, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75909, "pid": 0, "tid": 7, "ts": 1776649791961435, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791960605, "dur": 6, + "args": { + "External id": 9946, + "cbid": 211, "correlation": 75909 + } + }, + { + "ph": "s", "id": 75909, "pid": 435, "tid": 435, "ts": 1776649791960605, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791960633, "dur": 0, + "args": { + "External id": 9948, + "cbid": 210, "correlation": 75936 + } + }, + { + "ph": "f", "id": 75936, "pid": 435, "tid": 435, "ts": 1776649791960633, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791961485, "dur": 50, + "args": { + "External id": 9948, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75937, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75937, "pid": 0, "tid": 7, "ts": 1776649791961485, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791960634, "dur": 6, + "args": { + "External id": 9948, + "cbid": 211, "correlation": 75937 + } + }, + { + "ph": "s", "id": 75937, "pid": 435, "tid": 435, "ts": 1776649791960634, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791960663, "dur": 0, + "args": { + "External id": 9950, + "cbid": 210, "correlation": 75964 + } + }, + { + "ph": "f", "id": 75964, "pid": 435, "tid": 435, "ts": 1776649791960663, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791961537, "dur": 48, + "args": { + "External id": 9950, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75965, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75965, "pid": 0, "tid": 7, "ts": 1776649791961537, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791960664, "dur": 6, + "args": { + "External id": 9950, + "cbid": 211, "correlation": 75965 + } + }, + { + "ph": "s", "id": 75965, "pid": 435, "tid": 435, "ts": 1776649791960664, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791960693, "dur": 0, + "args": { + "External id": 9952, + "cbid": 210, "correlation": 75992 + } + }, + { + "ph": "f", "id": 75992, "pid": 435, "tid": 435, "ts": 1776649791960693, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791961587, "dur": 46, + "args": { + "External id": 9952, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 75993, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 75993, "pid": 0, "tid": 7, "ts": 1776649791961587, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791960694, "dur": 5, + "args": { + "External id": 9952, + "cbid": 211, "correlation": 75993 + } + }, + { + "ph": "s", "id": 75993, "pid": 435, "tid": 435, "ts": 1776649791960694, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791960728, "dur": 4, + "args": { + "External id": 76004, + "cbid": 135, "correlation": 76004 + } + }, + { + "ph": "f", "id": 76004, "pid": 435, "tid": 435, "ts": 1776649791960728, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791960743, "dur": 899, + "args": { + "External id": 76009, + "cbid": 165, "correlation": 76009 + } + }, + { + "ph": "s", "id": 76009, "pid": 435, "tid": 435, "ts": 1776649791960743, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649791961646, "dur": 4, + "args": { + "External id": 76012, + "cbid": 139, "correlation": 76012 + } + }, + { + "ph": "f", "id": 76012, "pid": 435, "tid": 435, "ts": 1776649791961646, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791961683, "dur": 1, + "args": { + "External id": 9954, + "cbid": 210, "correlation": 76046 + } + }, + { + "ph": "f", "id": 76046, "pid": 435, "tid": 435, "ts": 1776649791961683, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791961696, "dur": 41, + "args": { + "External id": 9954, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76047, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76047, "pid": 0, "tid": 7, "ts": 1776649791961696, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791961685, "dur": 7, + "args": { + "External id": 9954, + "cbid": 211, "correlation": 76047 + } + }, + { + "ph": "s", "id": 76047, "pid": 435, "tid": 435, "ts": 1776649791961685, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791961718, "dur": 0, + "args": { + "External id": 9956, + "cbid": 210, "correlation": 76074 + } + }, + { + "ph": "f", "id": 76074, "pid": 435, "tid": 435, "ts": 1776649791961718, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791961738, "dur": 46, + "args": { + "External id": 9956, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76075, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76075, "pid": 0, "tid": 7, "ts": 1776649791961738, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791961720, "dur": 6, + "args": { + "External id": 9956, + "cbid": 211, "correlation": 76075 + } + }, + { + "ph": "s", "id": 76075, "pid": 435, "tid": 435, "ts": 1776649791961720, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791961749, "dur": 0, + "args": { + "External id": 9958, + "cbid": 210, "correlation": 76102 + } + }, + { + "ph": "f", "id": 76102, "pid": 435, "tid": 435, "ts": 1776649791961749, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791961785, "dur": 43, + "args": { + "External id": 9958, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76103, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76103, "pid": 0, "tid": 7, "ts": 1776649791961785, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791961750, "dur": 5, + "args": { + "External id": 9958, + "cbid": 211, "correlation": 76103 + } + }, + { + "ph": "s", "id": 76103, "pid": 435, "tid": 435, "ts": 1776649791961750, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791961779, "dur": 0, + "args": { + "External id": 9960, + "cbid": 210, "correlation": 76130 + } + }, + { + "ph": "f", "id": 76130, "pid": 435, "tid": 435, "ts": 1776649791961779, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791961830, "dur": 52, + "args": { + "External id": 9960, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76131, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76131, "pid": 0, "tid": 7, "ts": 1776649791961830, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791961780, "dur": 6, + "args": { + "External id": 9960, + "cbid": 211, "correlation": 76131 + } + }, + { + "ph": "s", "id": 76131, "pid": 435, "tid": 435, "ts": 1776649791961780, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791961809, "dur": 0, + "args": { + "External id": 9962, + "cbid": 210, "correlation": 76158 + } + }, + { + "ph": "f", "id": 76158, "pid": 435, "tid": 435, "ts": 1776649791961809, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791961884, "dur": 44, + "args": { + "External id": 9962, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76159, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76159, "pid": 0, "tid": 7, "ts": 1776649791961884, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791961810, "dur": 6, + "args": { + "External id": 9962, + "cbid": 211, "correlation": 76159 + } + }, + { + "ph": "s", "id": 76159, "pid": 435, "tid": 435, "ts": 1776649791961810, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791961839, "dur": 0, + "args": { + "External id": 9964, + "cbid": 210, "correlation": 76186 + } + }, + { + "ph": "f", "id": 76186, "pid": 435, "tid": 435, "ts": 1776649791961839, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791961929, "dur": 44, + "args": { + "External id": 9964, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76187, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76187, "pid": 0, "tid": 7, "ts": 1776649791961929, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791961840, "dur": 6, + "args": { + "External id": 9964, + "cbid": 211, "correlation": 76187 + } + }, + { + "ph": "s", "id": 76187, "pid": 435, "tid": 435, "ts": 1776649791961840, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791961869, "dur": 0, + "args": { + "External id": 9966, + "cbid": 210, "correlation": 76214 + } + }, + { + "ph": "f", "id": 76214, "pid": 435, "tid": 435, "ts": 1776649791961869, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791961975, "dur": 47, + "args": { + "External id": 9966, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76215, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76215, "pid": 0, "tid": 7, "ts": 1776649791961975, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791961870, "dur": 6, + "args": { + "External id": 9966, + "cbid": 211, "correlation": 76215 + } + }, + { + "ph": "s", "id": 76215, "pid": 435, "tid": 435, "ts": 1776649791961870, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791961899, "dur": 0, + "args": { + "External id": 9968, + "cbid": 210, "correlation": 76242 + } + }, + { + "ph": "f", "id": 76242, "pid": 435, "tid": 435, "ts": 1776649791961899, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791962024, "dur": 47, + "args": { + "External id": 9968, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76243, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76243, "pid": 0, "tid": 7, "ts": 1776649791962024, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791961900, "dur": 5, + "args": { + "External id": 9968, + "cbid": 211, "correlation": 76243 + } + }, + { + "ph": "s", "id": 76243, "pid": 435, "tid": 435, "ts": 1776649791961900, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791961928, "dur": 0, + "args": { + "External id": 9970, + "cbid": 210, "correlation": 76270 + } + }, + { + "ph": "f", "id": 76270, "pid": 435, "tid": 435, "ts": 1776649791961928, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791962073, "dur": 49, + "args": { + "External id": 9970, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76271, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76271, "pid": 0, "tid": 7, "ts": 1776649791962073, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791961930, "dur": 5, + "args": { + "External id": 9970, + "cbid": 211, "correlation": 76271 + } + }, + { + "ph": "s", "id": 76271, "pid": 435, "tid": 435, "ts": 1776649791961930, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791961958, "dur": 0, + "args": { + "External id": 9972, + "cbid": 210, "correlation": 76298 + } + }, + { + "ph": "f", "id": 76298, "pid": 435, "tid": 435, "ts": 1776649791961958, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791962124, "dur": 47, + "args": { + "External id": 9972, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76299, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76299, "pid": 0, "tid": 7, "ts": 1776649791962124, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791961959, "dur": 5, + "args": { + "External id": 9972, + "cbid": 211, "correlation": 76299 + } + }, + { + "ph": "s", "id": 76299, "pid": 435, "tid": 435, "ts": 1776649791961959, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791961984, "dur": 190, + "args": { + "External id": 76305, + "cbid": 165, "correlation": 76305 + } + }, + { + "ph": "s", "id": 76305, "pid": 435, "tid": 435, "ts": 1776649791961984, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791962195, "dur": 3, + "args": { + "External id": 76314, + "cbid": 135, "correlation": 76314 + } + }, + { + "ph": "f", "id": 76314, "pid": 435, "tid": 435, "ts": 1776649791962195, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791962225, "dur": 1, + "args": { + "External id": 9974, + "cbid": 210, "correlation": 76340 + } + }, + { + "ph": "f", "id": 76340, "pid": 435, "tid": 435, "ts": 1776649791962225, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791962236, "dur": 41, + "args": { + "External id": 9974, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76341, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76341, "pid": 0, "tid": 7, "ts": 1776649791962236, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791962227, "dur": 7, + "args": { + "External id": 9974, + "cbid": 211, "correlation": 76341 + } + }, + { + "ph": "s", "id": 76341, "pid": 435, "tid": 435, "ts": 1776649791962227, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791962259, "dur": 0, + "args": { + "External id": 9976, + "cbid": 210, "correlation": 76368 + } + }, + { + "ph": "f", "id": 76368, "pid": 435, "tid": 435, "ts": 1776649791962259, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791962280, "dur": 43, + "args": { + "External id": 9976, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76369, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76369, "pid": 0, "tid": 7, "ts": 1776649791962280, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791962260, "dur": 6, + "args": { + "External id": 9976, + "cbid": 211, "correlation": 76369 + } + }, + { + "ph": "s", "id": 76369, "pid": 435, "tid": 435, "ts": 1776649791962260, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791962294, "dur": 0, + "args": { + "External id": 9978, + "cbid": 210, "correlation": 76396 + } + }, + { + "ph": "f", "id": 76396, "pid": 435, "tid": 435, "ts": 1776649791962294, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791962325, "dur": 43, + "args": { + "External id": 9978, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76397, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76397, "pid": 0, "tid": 7, "ts": 1776649791962325, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791962295, "dur": 6, + "args": { + "External id": 9978, + "cbid": 211, "correlation": 76397 + } + }, + { + "ph": "s", "id": 76397, "pid": 435, "tid": 435, "ts": 1776649791962295, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791962324, "dur": 0, + "args": { + "External id": 9980, + "cbid": 210, "correlation": 76424 + } + }, + { + "ph": "f", "id": 76424, "pid": 435, "tid": 435, "ts": 1776649791962324, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791962370, "dur": 44, + "args": { + "External id": 9980, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76425, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76425, "pid": 0, "tid": 7, "ts": 1776649791962370, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791962326, "dur": 6, + "args": { + "External id": 9980, + "cbid": 211, "correlation": 76425 + } + }, + { + "ph": "s", "id": 76425, "pid": 435, "tid": 435, "ts": 1776649791962326, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791962355, "dur": 0, + "args": { + "External id": 9982, + "cbid": 210, "correlation": 76452 + } + }, + { + "ph": "f", "id": 76452, "pid": 435, "tid": 435, "ts": 1776649791962355, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791962416, "dur": 46, + "args": { + "External id": 9982, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76453, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76453, "pid": 0, "tid": 7, "ts": 1776649791962416, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791962356, "dur": 5, + "args": { + "External id": 9982, + "cbid": 211, "correlation": 76453 + } + }, + { + "ph": "s", "id": 76453, "pid": 435, "tid": 435, "ts": 1776649791962356, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791962384, "dur": 0, + "args": { + "External id": 9984, + "cbid": 210, "correlation": 76480 + } + }, + { + "ph": "f", "id": 76480, "pid": 435, "tid": 435, "ts": 1776649791962384, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791962463, "dur": 51, + "args": { + "External id": 9984, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76481, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76481, "pid": 0, "tid": 7, "ts": 1776649791962463, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791962386, "dur": 5, + "args": { + "External id": 9984, + "cbid": 211, "correlation": 76481 + } + }, + { + "ph": "s", "id": 76481, "pid": 435, "tid": 435, "ts": 1776649791962386, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791962414, "dur": 0, + "args": { + "External id": 9986, + "cbid": 210, "correlation": 76508 + } + }, + { + "ph": "f", "id": 76508, "pid": 435, "tid": 435, "ts": 1776649791962414, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791962516, "dur": 48, + "args": { + "External id": 9986, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76509, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76509, "pid": 0, "tid": 7, "ts": 1776649791962516, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791962416, "dur": 6, + "args": { + "External id": 9986, + "cbid": 211, "correlation": 76509 + } + }, + { + "ph": "s", "id": 76509, "pid": 435, "tid": 435, "ts": 1776649791962416, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791962444, "dur": 0, + "args": { + "External id": 9988, + "cbid": 210, "correlation": 76536 + } + }, + { + "ph": "f", "id": 76536, "pid": 435, "tid": 435, "ts": 1776649791962444, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791962566, "dur": 47, + "args": { + "External id": 9988, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76537, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76537, "pid": 0, "tid": 7, "ts": 1776649791962566, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791962446, "dur": 5, + "args": { + "External id": 9988, + "cbid": 211, "correlation": 76537 + } + }, + { + "ph": "s", "id": 76537, "pid": 435, "tid": 435, "ts": 1776649791962446, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791962474, "dur": 0, + "args": { + "External id": 9990, + "cbid": 210, "correlation": 76564 + } + }, + { + "ph": "f", "id": 76564, "pid": 435, "tid": 435, "ts": 1776649791962474, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791962615, "dur": 46, + "args": { + "External id": 9990, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76565, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76565, "pid": 0, "tid": 7, "ts": 1776649791962615, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791962475, "dur": 6, + "args": { + "External id": 9990, + "cbid": 211, "correlation": 76565 + } + }, + { + "ph": "s", "id": 76565, "pid": 435, "tid": 435, "ts": 1776649791962475, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791962504, "dur": 0, + "args": { + "External id": 9992, + "cbid": 210, "correlation": 76592 + } + }, + { + "ph": "f", "id": 76592, "pid": 435, "tid": 435, "ts": 1776649791962504, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791962663, "dur": 51, + "args": { + "External id": 9992, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76593, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76593, "pid": 0, "tid": 7, "ts": 1776649791962663, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791962506, "dur": 5, + "args": { + "External id": 9992, + "cbid": 211, "correlation": 76593 + } + }, + { + "ph": "s", "id": 76593, "pid": 435, "tid": 435, "ts": 1776649791962506, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791962534, "dur": 0, + "args": { + "External id": 9994, + "cbid": 210, "correlation": 76620 + } + }, + { + "ph": "f", "id": 76620, "pid": 435, "tid": 435, "ts": 1776649791962534, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791962716, "dur": 47, + "args": { + "External id": 9994, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76621, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76621, "pid": 0, "tid": 7, "ts": 1776649791962716, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791962535, "dur": 6, + "args": { + "External id": 9994, + "cbid": 211, "correlation": 76621 + } + }, + { + "ph": "s", "id": 76621, "pid": 435, "tid": 435, "ts": 1776649791962535, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791962564, "dur": 0, + "args": { + "External id": 9996, + "cbid": 210, "correlation": 76648 + } + }, + { + "ph": "f", "id": 76648, "pid": 435, "tid": 435, "ts": 1776649791962564, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791962765, "dur": 44, + "args": { + "External id": 9996, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76649, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76649, "pid": 0, "tid": 7, "ts": 1776649791962765, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791962565, "dur": 5, + "args": { + "External id": 9996, + "cbid": 211, "correlation": 76649 + } + }, + { + "ph": "s", "id": 76649, "pid": 435, "tid": 435, "ts": 1776649791962565, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791962594, "dur": 0, + "args": { + "External id": 9998, + "cbid": 210, "correlation": 76676 + } + }, + { + "ph": "f", "id": 76676, "pid": 435, "tid": 435, "ts": 1776649791962594, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791962810, "dur": 52, + "args": { + "External id": 9998, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76677, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76677, "pid": 0, "tid": 7, "ts": 1776649791962810, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791962595, "dur": 6, + "args": { + "External id": 9998, + "cbid": 211, "correlation": 76677 + } + }, + { + "ph": "s", "id": 76677, "pid": 435, "tid": 435, "ts": 1776649791962595, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791962624, "dur": 0, + "args": { + "External id": 10000, + "cbid": 210, "correlation": 76704 + } + }, + { + "ph": "f", "id": 76704, "pid": 435, "tid": 435, "ts": 1776649791962624, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791962863, "dur": 46, + "args": { + "External id": 10000, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76705, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76705, "pid": 0, "tid": 7, "ts": 1776649791962863, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791962625, "dur": 6, + "args": { + "External id": 10000, + "cbid": 211, "correlation": 76705 + } + }, + { + "ph": "s", "id": 76705, "pid": 435, "tid": 435, "ts": 1776649791962625, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791962653, "dur": 0, + "args": { + "External id": 10002, + "cbid": 210, "correlation": 76732 + } + }, + { + "ph": "f", "id": 76732, "pid": 435, "tid": 435, "ts": 1776649791962653, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791962912, "dur": 45, + "args": { + "External id": 10002, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76733, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76733, "pid": 0, "tid": 7, "ts": 1776649791962912, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791962655, "dur": 6, + "args": { + "External id": 10002, + "cbid": 211, "correlation": 76733 + } + }, + { + "ph": "s", "id": 76733, "pid": 435, "tid": 435, "ts": 1776649791962655, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791962683, "dur": 0, + "args": { + "External id": 10004, + "cbid": 210, "correlation": 76760 + } + }, + { + "ph": "f", "id": 76760, "pid": 435, "tid": 435, "ts": 1776649791962683, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791962958, "dur": 44, + "args": { + "External id": 10004, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76761, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76761, "pid": 0, "tid": 7, "ts": 1776649791962958, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791962684, "dur": 6, + "args": { + "External id": 10004, + "cbid": 211, "correlation": 76761 + } + }, + { + "ph": "s", "id": 76761, "pid": 435, "tid": 435, "ts": 1776649791962684, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791962713, "dur": 0, + "args": { + "External id": 10006, + "cbid": 210, "correlation": 76788 + } + }, + { + "ph": "f", "id": 76788, "pid": 435, "tid": 435, "ts": 1776649791962713, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791963004, "dur": 43, + "args": { + "External id": 10006, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76789, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76789, "pid": 0, "tid": 7, "ts": 1776649791963004, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791962714, "dur": 6, + "args": { + "External id": 10006, + "cbid": 211, "correlation": 76789 + } + }, + { + "ph": "s", "id": 76789, "pid": 435, "tid": 435, "ts": 1776649791962714, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791962743, "dur": 0, + "args": { + "External id": 10008, + "cbid": 210, "correlation": 76816 + } + }, + { + "ph": "f", "id": 76816, "pid": 435, "tid": 435, "ts": 1776649791962743, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791963049, "dur": 44, + "args": { + "External id": 10008, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76817, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76817, "pid": 0, "tid": 7, "ts": 1776649791963049, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791962744, "dur": 5, + "args": { + "External id": 10008, + "cbid": 211, "correlation": 76817 + } + }, + { + "ph": "s", "id": 76817, "pid": 435, "tid": 435, "ts": 1776649791962744, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791962772, "dur": 0, + "args": { + "External id": 10010, + "cbid": 210, "correlation": 76844 + } + }, + { + "ph": "f", "id": 76844, "pid": 435, "tid": 435, "ts": 1776649791962772, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791963095, "dur": 46, + "args": { + "External id": 10010, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76845, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76845, "pid": 0, "tid": 7, "ts": 1776649791963095, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791962774, "dur": 5, + "args": { + "External id": 10010, + "cbid": 211, "correlation": 76845 + } + }, + { + "ph": "s", "id": 76845, "pid": 435, "tid": 435, "ts": 1776649791962774, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791962802, "dur": 0, + "args": { + "External id": 10012, + "cbid": 210, "correlation": 76872 + } + }, + { + "ph": "f", "id": 76872, "pid": 435, "tid": 435, "ts": 1776649791962802, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791963143, "dur": 47, + "args": { + "External id": 10012, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76873, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76873, "pid": 0, "tid": 7, "ts": 1776649791963143, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791962804, "dur": 5, + "args": { + "External id": 10012, + "cbid": 211, "correlation": 76873 + } + }, + { + "ph": "s", "id": 76873, "pid": 435, "tid": 435, "ts": 1776649791962804, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791962832, "dur": 0, + "args": { + "External id": 10014, + "cbid": 210, "correlation": 76900 + } + }, + { + "ph": "f", "id": 76900, "pid": 435, "tid": 435, "ts": 1776649791962832, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791963192, "dur": 47, + "args": { + "External id": 10014, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76901, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76901, "pid": 0, "tid": 7, "ts": 1776649791963192, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791962833, "dur": 6, + "args": { + "External id": 10014, + "cbid": 211, "correlation": 76901 + } + }, + { + "ph": "s", "id": 76901, "pid": 435, "tid": 435, "ts": 1776649791962833, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791962862, "dur": 0, + "args": { + "External id": 10016, + "cbid": 210, "correlation": 76928 + } + }, + { + "ph": "f", "id": 76928, "pid": 435, "tid": 435, "ts": 1776649791962862, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791963241, "dur": 48, + "args": { + "External id": 10016, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76929, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76929, "pid": 0, "tid": 7, "ts": 1776649791963241, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791962863, "dur": 6, + "args": { + "External id": 10016, + "cbid": 211, "correlation": 76929 + } + }, + { + "ph": "s", "id": 76929, "pid": 435, "tid": 435, "ts": 1776649791962863, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791962892, "dur": 0, + "args": { + "External id": 10018, + "cbid": 210, "correlation": 76956 + } + }, + { + "ph": "f", "id": 76956, "pid": 435, "tid": 435, "ts": 1776649791962892, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791963291, "dur": 50, + "args": { + "External id": 10018, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76957, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76957, "pid": 0, "tid": 7, "ts": 1776649791963291, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791962893, "dur": 5, + "args": { + "External id": 10018, + "cbid": 211, "correlation": 76957 + } + }, + { + "ph": "s", "id": 76957, "pid": 435, "tid": 435, "ts": 1776649791962893, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791962922, "dur": 0, + "args": { + "External id": 10020, + "cbid": 210, "correlation": 76984 + } + }, + { + "ph": "f", "id": 76984, "pid": 435, "tid": 435, "ts": 1776649791962922, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791963343, "dur": 52, + "args": { + "External id": 10020, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76985, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 76985, "pid": 0, "tid": 7, "ts": 1776649791963343, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791962923, "dur": 5, + "args": { + "External id": 10020, + "cbid": 211, "correlation": 76985 + } + }, + { + "ph": "s", "id": 76985, "pid": 435, "tid": 435, "ts": 1776649791962923, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791962951, "dur": 0, + "args": { + "External id": 10022, + "cbid": 210, "correlation": 77012 + } + }, + { + "ph": "f", "id": 77012, "pid": 435, "tid": 435, "ts": 1776649791962951, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791963396, "dur": 50, + "args": { + "External id": 10022, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77013, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77013, "pid": 0, "tid": 7, "ts": 1776649791963396, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791962953, "dur": 6, + "args": { + "External id": 10022, + "cbid": 211, "correlation": 77013 + } + }, + { + "ph": "s", "id": 77013, "pid": 435, "tid": 435, "ts": 1776649791962953, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791962982, "dur": 0, + "args": { + "External id": 10024, + "cbid": 210, "correlation": 77040 + } + }, + { + "ph": "f", "id": 77040, "pid": 435, "tid": 435, "ts": 1776649791962982, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791963448, "dur": 44, + "args": { + "External id": 10024, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77041, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77041, "pid": 0, "tid": 7, "ts": 1776649791963448, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791962983, "dur": 6, + "args": { + "External id": 10024, + "cbid": 211, "correlation": 77041 + } + }, + { + "ph": "s", "id": 77041, "pid": 435, "tid": 435, "ts": 1776649791962983, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791963012, "dur": 0, + "args": { + "External id": 10026, + "cbid": 210, "correlation": 77068 + } + }, + { + "ph": "f", "id": 77068, "pid": 435, "tid": 435, "ts": 1776649791963012, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791963494, "dur": 47, + "args": { + "External id": 10026, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77069, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77069, "pid": 0, "tid": 7, "ts": 1776649791963494, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791963013, "dur": 6, + "args": { + "External id": 10026, + "cbid": 211, "correlation": 77069 + } + }, + { + "ph": "s", "id": 77069, "pid": 435, "tid": 435, "ts": 1776649791963013, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791963043, "dur": 0, + "args": { + "External id": 10028, + "cbid": 210, "correlation": 77096 + } + }, + { + "ph": "f", "id": 77096, "pid": 435, "tid": 435, "ts": 1776649791963043, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791963543, "dur": 47, + "args": { + "External id": 10028, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77097, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77097, "pid": 0, "tid": 7, "ts": 1776649791963543, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791963045, "dur": 6, + "args": { + "External id": 10028, + "cbid": 211, "correlation": 77097 + } + }, + { + "ph": "s", "id": 77097, "pid": 435, "tid": 435, "ts": 1776649791963045, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791963076, "dur": 0, + "args": { + "External id": 10030, + "cbid": 210, "correlation": 77124 + } + }, + { + "ph": "f", "id": 77124, "pid": 435, "tid": 435, "ts": 1776649791963076, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791963592, "dur": 46, + "args": { + "External id": 10030, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77125, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77125, "pid": 0, "tid": 7, "ts": 1776649791963592, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791963078, "dur": 6, + "args": { + "External id": 10030, + "cbid": 211, "correlation": 77125 + } + }, + { + "ph": "s", "id": 77125, "pid": 435, "tid": 435, "ts": 1776649791963078, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791963106, "dur": 0, + "args": { + "External id": 10032, + "cbid": 210, "correlation": 77152 + } + }, + { + "ph": "f", "id": 77152, "pid": 435, "tid": 435, "ts": 1776649791963106, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791963640, "dur": 47, + "args": { + "External id": 10032, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77153, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77153, "pid": 0, "tid": 7, "ts": 1776649791963640, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791963107, "dur": 5, + "args": { + "External id": 10032, + "cbid": 211, "correlation": 77153 + } + }, + { + "ph": "s", "id": 77153, "pid": 435, "tid": 435, "ts": 1776649791963107, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791963136, "dur": 0, + "args": { + "External id": 10034, + "cbid": 210, "correlation": 77180 + } + }, + { + "ph": "f", "id": 77180, "pid": 435, "tid": 435, "ts": 1776649791963136, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791963689, "dur": 41, + "args": { + "External id": 10034, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77181, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77181, "pid": 0, "tid": 7, "ts": 1776649791963689, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791963137, "dur": 5, + "args": { + "External id": 10034, + "cbid": 211, "correlation": 77181 + } + }, + { + "ph": "s", "id": 77181, "pid": 435, "tid": 435, "ts": 1776649791963137, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791963166, "dur": 0, + "args": { + "External id": 10036, + "cbid": 210, "correlation": 77208 + } + }, + { + "ph": "f", "id": 77208, "pid": 435, "tid": 435, "ts": 1776649791963166, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791963733, "dur": 47, + "args": { + "External id": 10036, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77209, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77209, "pid": 0, "tid": 7, "ts": 1776649791963733, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791963167, "dur": 6, + "args": { + "External id": 10036, + "cbid": 211, "correlation": 77209 + } + }, + { + "ph": "s", "id": 77209, "pid": 435, "tid": 435, "ts": 1776649791963167, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791963195, "dur": 0, + "args": { + "External id": 10038, + "cbid": 210, "correlation": 77236 + } + }, + { + "ph": "f", "id": 77236, "pid": 435, "tid": 435, "ts": 1776649791963195, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791963781, "dur": 44, + "args": { + "External id": 10038, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77237, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77237, "pid": 0, "tid": 7, "ts": 1776649791963781, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791963196, "dur": 6, + "args": { + "External id": 10038, + "cbid": 211, "correlation": 77237 + } + }, + { + "ph": "s", "id": 77237, "pid": 435, "tid": 435, "ts": 1776649791963196, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791963225, "dur": 0, + "args": { + "External id": 10040, + "cbid": 210, "correlation": 77264 + } + }, + { + "ph": "f", "id": 77264, "pid": 435, "tid": 435, "ts": 1776649791963225, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791963826, "dur": 44, + "args": { + "External id": 10040, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77265, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77265, "pid": 0, "tid": 7, "ts": 1776649791963826, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791963226, "dur": 5, + "args": { + "External id": 10040, + "cbid": 211, "correlation": 77265 + } + }, + { + "ph": "s", "id": 77265, "pid": 435, "tid": 435, "ts": 1776649791963226, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791963255, "dur": 0, + "args": { + "External id": 10042, + "cbid": 210, "correlation": 77292 + } + }, + { + "ph": "f", "id": 77292, "pid": 435, "tid": 435, "ts": 1776649791963255, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791963872, "dur": 46, + "args": { + "External id": 10042, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77293, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77293, "pid": 0, "tid": 7, "ts": 1776649791963872, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791963256, "dur": 5, + "args": { + "External id": 10042, + "cbid": 211, "correlation": 77293 + } + }, + { + "ph": "s", "id": 77293, "pid": 435, "tid": 435, "ts": 1776649791963256, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791963285, "dur": 0, + "args": { + "External id": 10044, + "cbid": 210, "correlation": 77320 + } + }, + { + "ph": "f", "id": 77320, "pid": 435, "tid": 435, "ts": 1776649791963285, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791963920, "dur": 48, + "args": { + "External id": 10044, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77321, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77321, "pid": 0, "tid": 7, "ts": 1776649791963920, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791963286, "dur": 5, + "args": { + "External id": 10044, + "cbid": 211, "correlation": 77321 + } + }, + { + "ph": "s", "id": 77321, "pid": 435, "tid": 435, "ts": 1776649791963286, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791963314, "dur": 0, + "args": { + "External id": 10046, + "cbid": 210, "correlation": 77348 + } + }, + { + "ph": "f", "id": 77348, "pid": 435, "tid": 435, "ts": 1776649791963314, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791963970, "dur": 49, + "args": { + "External id": 10046, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77349, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77349, "pid": 0, "tid": 7, "ts": 1776649791963970, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791963316, "dur": 6, + "args": { + "External id": 10046, + "cbid": 211, "correlation": 77349 + } + }, + { + "ph": "s", "id": 77349, "pid": 435, "tid": 435, "ts": 1776649791963316, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791963344, "dur": 0, + "args": { + "External id": 10048, + "cbid": 210, "correlation": 77376 + } + }, + { + "ph": "f", "id": 77376, "pid": 435, "tid": 435, "ts": 1776649791963344, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791964021, "dur": 46, + "args": { + "External id": 10048, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77377, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77377, "pid": 0, "tid": 7, "ts": 1776649791964021, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791963345, "dur": 6, + "args": { + "External id": 10048, + "cbid": 211, "correlation": 77377 + } + }, + { + "ph": "s", "id": 77377, "pid": 435, "tid": 435, "ts": 1776649791963345, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791963374, "dur": 0, + "args": { + "External id": 10050, + "cbid": 210, "correlation": 77404 + } + }, + { + "ph": "f", "id": 77404, "pid": 435, "tid": 435, "ts": 1776649791963374, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791964069, "dur": 51, + "args": { + "External id": 10050, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77405, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77405, "pid": 0, "tid": 7, "ts": 1776649791964069, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791963375, "dur": 6, + "args": { + "External id": 10050, + "cbid": 211, "correlation": 77405 + } + }, + { + "ph": "s", "id": 77405, "pid": 435, "tid": 435, "ts": 1776649791963375, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791963404, "dur": 0, + "args": { + "External id": 10052, + "cbid": 210, "correlation": 77432 + } + }, + { + "ph": "f", "id": 77432, "pid": 435, "tid": 435, "ts": 1776649791963404, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791964122, "dur": 50, + "args": { + "External id": 10052, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77433, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77433, "pid": 0, "tid": 7, "ts": 1776649791964122, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791963405, "dur": 5, + "args": { + "External id": 10052, + "cbid": 211, "correlation": 77433 + } + }, + { + "ph": "s", "id": 77433, "pid": 435, "tid": 435, "ts": 1776649791963405, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791963433, "dur": 0, + "args": { + "External id": 10054, + "cbid": 210, "correlation": 77460 + } + }, + { + "ph": "f", "id": 77460, "pid": 435, "tid": 435, "ts": 1776649791963433, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791964174, "dur": 47, + "args": { + "External id": 10054, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77461, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77461, "pid": 0, "tid": 7, "ts": 1776649791964174, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791963435, "dur": 5, + "args": { + "External id": 10054, + "cbid": 211, "correlation": 77461 + } + }, + { + "ph": "s", "id": 77461, "pid": 435, "tid": 435, "ts": 1776649791963435, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791963463, "dur": 0, + "args": { + "External id": 10056, + "cbid": 210, "correlation": 77488 + } + }, + { + "ph": "f", "id": 77488, "pid": 435, "tid": 435, "ts": 1776649791963463, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791964222, "dur": 44, + "args": { + "External id": 10056, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77489, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77489, "pid": 0, "tid": 7, "ts": 1776649791964222, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791963464, "dur": 5, + "args": { + "External id": 10056, + "cbid": 211, "correlation": 77489 + } + }, + { + "ph": "s", "id": 77489, "pid": 435, "tid": 435, "ts": 1776649791963464, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791963493, "dur": 0, + "args": { + "External id": 10058, + "cbid": 210, "correlation": 77516 + } + }, + { + "ph": "f", "id": 77516, "pid": 435, "tid": 435, "ts": 1776649791963493, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791964268, "dur": 49, + "args": { + "External id": 10058, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77517, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77517, "pid": 0, "tid": 7, "ts": 1776649791964268, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791963494, "dur": 5, + "args": { + "External id": 10058, + "cbid": 211, "correlation": 77517 + } + }, + { + "ph": "s", "id": 77517, "pid": 435, "tid": 435, "ts": 1776649791963494, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791963523, "dur": 0, + "args": { + "External id": 10060, + "cbid": 210, "correlation": 77544 + } + }, + { + "ph": "f", "id": 77544, "pid": 435, "tid": 435, "ts": 1776649791963523, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791964319, "dur": 46, + "args": { + "External id": 10060, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77545, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77545, "pid": 0, "tid": 7, "ts": 1776649791964319, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791963524, "dur": 5, + "args": { + "External id": 10060, + "cbid": 211, "correlation": 77545 + } + }, + { + "ph": "s", "id": 77545, "pid": 435, "tid": 435, "ts": 1776649791963524, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791963552, "dur": 0, + "args": { + "External id": 10062, + "cbid": 210, "correlation": 77572 + } + }, + { + "ph": "f", "id": 77572, "pid": 435, "tid": 435, "ts": 1776649791963552, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791964367, "dur": 46, + "args": { + "External id": 10062, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77573, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77573, "pid": 0, "tid": 7, "ts": 1776649791964367, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791963554, "dur": 5, + "args": { + "External id": 10062, + "cbid": 211, "correlation": 77573 + } + }, + { + "ph": "s", "id": 77573, "pid": 435, "tid": 435, "ts": 1776649791963554, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791963582, "dur": 0, + "args": { + "External id": 10064, + "cbid": 210, "correlation": 77600 + } + }, + { + "ph": "f", "id": 77600, "pid": 435, "tid": 435, "ts": 1776649791963582, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791964416, "dur": 49, + "args": { + "External id": 10064, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77601, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77601, "pid": 0, "tid": 7, "ts": 1776649791964416, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791963583, "dur": 5, + "args": { + "External id": 10064, + "cbid": 211, "correlation": 77601 + } + }, + { + "ph": "s", "id": 77601, "pid": 435, "tid": 435, "ts": 1776649791963583, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791963612, "dur": 0, + "args": { + "External id": 10066, + "cbid": 210, "correlation": 77628 + } + }, + { + "ph": "f", "id": 77628, "pid": 435, "tid": 435, "ts": 1776649791963612, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791964467, "dur": 47, + "args": { + "External id": 10066, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77629, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77629, "pid": 0, "tid": 7, "ts": 1776649791964467, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791963613, "dur": 6, + "args": { + "External id": 10066, + "cbid": 211, "correlation": 77629 + } + }, + { + "ph": "s", "id": 77629, "pid": 435, "tid": 435, "ts": 1776649791963613, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791963641, "dur": 0, + "args": { + "External id": 10068, + "cbid": 210, "correlation": 77656 + } + }, + { + "ph": "f", "id": 77656, "pid": 435, "tid": 435, "ts": 1776649791963641, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791964515, "dur": 48, + "args": { + "External id": 10068, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77657, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77657, "pid": 0, "tid": 7, "ts": 1776649791964515, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791963643, "dur": 5, + "args": { + "External id": 10068, + "cbid": 211, "correlation": 77657 + } + }, + { + "ph": "s", "id": 77657, "pid": 435, "tid": 435, "ts": 1776649791963643, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791963671, "dur": 0, + "args": { + "External id": 10070, + "cbid": 210, "correlation": 77684 + } + }, + { + "ph": "f", "id": 77684, "pid": 435, "tid": 435, "ts": 1776649791963671, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791964565, "dur": 44, + "args": { + "External id": 10070, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77685, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77685, "pid": 0, "tid": 7, "ts": 1776649791964565, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791963672, "dur": 5, + "args": { + "External id": 10070, + "cbid": 211, "correlation": 77685 + } + }, + { + "ph": "s", "id": 77685, "pid": 435, "tid": 435, "ts": 1776649791963672, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791963701, "dur": 0, + "args": { + "External id": 10072, + "cbid": 210, "correlation": 77712 + } + }, + { + "ph": "f", "id": 77712, "pid": 435, "tid": 435, "ts": 1776649791963701, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791964611, "dur": 49, + "args": { + "External id": 10072, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77713, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77713, "pid": 0, "tid": 7, "ts": 1776649791964611, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791963702, "dur": 6, + "args": { + "External id": 10072, + "cbid": 211, "correlation": 77713 + } + }, + { + "ph": "s", "id": 77713, "pid": 435, "tid": 435, "ts": 1776649791963702, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791963734, "dur": 4, + "args": { + "External id": 77724, + "cbid": 135, "correlation": 77724 + } + }, + { + "ph": "f", "id": 77724, "pid": 435, "tid": 435, "ts": 1776649791963734, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791963749, "dur": 920, + "args": { + "External id": 77729, + "cbid": 165, "correlation": 77729 + } + }, + { + "ph": "s", "id": 77729, "pid": 435, "tid": 435, "ts": 1776649791963749, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649791964673, "dur": 4, + "args": { + "External id": 77732, + "cbid": 139, "correlation": 77732 + } + }, + { + "ph": "f", "id": 77732, "pid": 435, "tid": 435, "ts": 1776649791964673, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791964713, "dur": 1, + "args": { + "External id": 10074, + "cbid": 210, "correlation": 77766 + } + }, + { + "ph": "f", "id": 77766, "pid": 435, "tid": 435, "ts": 1776649791964713, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791964727, "dur": 41, + "args": { + "External id": 10074, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77767, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77767, "pid": 0, "tid": 7, "ts": 1776649791964727, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791964715, "dur": 8, + "args": { + "External id": 10074, + "cbid": 211, "correlation": 77767 + } + }, + { + "ph": "s", "id": 77767, "pid": 435, "tid": 435, "ts": 1776649791964715, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791964750, "dur": 0, + "args": { + "External id": 10076, + "cbid": 210, "correlation": 77794 + } + }, + { + "ph": "f", "id": 77794, "pid": 435, "tid": 435, "ts": 1776649791964750, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791964769, "dur": 41, + "args": { + "External id": 10076, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77795, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77795, "pid": 0, "tid": 7, "ts": 1776649791964769, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791964751, "dur": 6, + "args": { + "External id": 10076, + "cbid": 211, "correlation": 77795 + } + }, + { + "ph": "s", "id": 77795, "pid": 435, "tid": 435, "ts": 1776649791964751, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791964781, "dur": 0, + "args": { + "External id": 10078, + "cbid": 210, "correlation": 77822 + } + }, + { + "ph": "f", "id": 77822, "pid": 435, "tid": 435, "ts": 1776649791964781, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791964811, "dur": 49, + "args": { + "External id": 10078, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77823, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77823, "pid": 0, "tid": 7, "ts": 1776649791964811, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791964782, "dur": 6, + "args": { + "External id": 10078, + "cbid": 211, "correlation": 77823 + } + }, + { + "ph": "s", "id": 77823, "pid": 435, "tid": 435, "ts": 1776649791964782, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791964812, "dur": 0, + "args": { + "External id": 10080, + "cbid": 210, "correlation": 77850 + } + }, + { + "ph": "f", "id": 77850, "pid": 435, "tid": 435, "ts": 1776649791964812, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791964863, "dur": 48, + "args": { + "External id": 10080, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77851, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77851, "pid": 0, "tid": 7, "ts": 1776649791964863, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791964813, "dur": 6, + "args": { + "External id": 10080, + "cbid": 211, "correlation": 77851 + } + }, + { + "ph": "s", "id": 77851, "pid": 435, "tid": 435, "ts": 1776649791964813, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791964842, "dur": 0, + "args": { + "External id": 10082, + "cbid": 210, "correlation": 77878 + } + }, + { + "ph": "f", "id": 77878, "pid": 435, "tid": 435, "ts": 1776649791964842, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791964913, "dur": 47, + "args": { + "External id": 10082, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77879, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77879, "pid": 0, "tid": 7, "ts": 1776649791964913, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791964843, "dur": 6, + "args": { + "External id": 10082, + "cbid": 211, "correlation": 77879 + } + }, + { + "ph": "s", "id": 77879, "pid": 435, "tid": 435, "ts": 1776649791964843, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791964872, "dur": 0, + "args": { + "External id": 10084, + "cbid": 210, "correlation": 77906 + } + }, + { + "ph": "f", "id": 77906, "pid": 435, "tid": 435, "ts": 1776649791964872, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791964961, "dur": 46, + "args": { + "External id": 10084, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77907, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77907, "pid": 0, "tid": 7, "ts": 1776649791964961, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791964873, "dur": 5, + "args": { + "External id": 10084, + "cbid": 211, "correlation": 77907 + } + }, + { + "ph": "s", "id": 77907, "pid": 435, "tid": 435, "ts": 1776649791964873, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791964902, "dur": 0, + "args": { + "External id": 10086, + "cbid": 210, "correlation": 77934 + } + }, + { + "ph": "f", "id": 77934, "pid": 435, "tid": 435, "ts": 1776649791964902, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791965009, "dur": 47, + "args": { + "External id": 10086, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77935, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77935, "pid": 0, "tid": 7, "ts": 1776649791965009, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791964904, "dur": 6, + "args": { + "External id": 10086, + "cbid": 211, "correlation": 77935 + } + }, + { + "ph": "s", "id": 77935, "pid": 435, "tid": 435, "ts": 1776649791964904, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791964932, "dur": 0, + "args": { + "External id": 10088, + "cbid": 210, "correlation": 77962 + } + }, + { + "ph": "f", "id": 77962, "pid": 435, "tid": 435, "ts": 1776649791964932, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791965058, "dur": 46, + "args": { + "External id": 10088, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77963, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77963, "pid": 0, "tid": 7, "ts": 1776649791965058, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791964934, "dur": 5, + "args": { + "External id": 10088, + "cbid": 211, "correlation": 77963 + } + }, + { + "ph": "s", "id": 77963, "pid": 435, "tid": 435, "ts": 1776649791964934, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791964962, "dur": 0, + "args": { + "External id": 10090, + "cbid": 210, "correlation": 77990 + } + }, + { + "ph": "f", "id": 77990, "pid": 435, "tid": 435, "ts": 1776649791964962, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791965106, "dur": 43, + "args": { + "External id": 10090, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 77991, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 77991, "pid": 0, "tid": 7, "ts": 1776649791965106, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791964964, "dur": 6, + "args": { + "External id": 10090, + "cbid": 211, "correlation": 77991 + } + }, + { + "ph": "s", "id": 77991, "pid": 435, "tid": 435, "ts": 1776649791964964, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791964992, "dur": 0, + "args": { + "External id": 10092, + "cbid": 210, "correlation": 78018 + } + }, + { + "ph": "f", "id": 78018, "pid": 435, "tid": 435, "ts": 1776649791964992, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791965151, "dur": 47, + "args": { + "External id": 10092, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78019, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78019, "pid": 0, "tid": 7, "ts": 1776649791965151, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791964994, "dur": 5, + "args": { + "External id": 10092, + "cbid": 211, "correlation": 78019 + } + }, + { + "ph": "s", "id": 78019, "pid": 435, "tid": 435, "ts": 1776649791964994, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791965019, "dur": 183, + "args": { + "External id": 78025, + "cbid": 165, "correlation": 78025 + } + }, + { + "ph": "s", "id": 78025, "pid": 435, "tid": 435, "ts": 1776649791965019, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791965222, "dur": 3, + "args": { + "External id": 78034, + "cbid": 135, "correlation": 78034 + } + }, + { + "ph": "f", "id": 78034, "pid": 435, "tid": 435, "ts": 1776649791965222, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791965252, "dur": 1, + "args": { + "External id": 10094, + "cbid": 210, "correlation": 78060 + } + }, + { + "ph": "f", "id": 78060, "pid": 435, "tid": 435, "ts": 1776649791965252, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791965264, "dur": 40, + "args": { + "External id": 10094, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78061, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78061, "pid": 0, "tid": 7, "ts": 1776649791965264, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791965254, "dur": 7, + "args": { + "External id": 10094, + "cbid": 211, "correlation": 78061 + } + }, + { + "ph": "s", "id": 78061, "pid": 435, "tid": 435, "ts": 1776649791965254, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791965287, "dur": 0, + "args": { + "External id": 10096, + "cbid": 210, "correlation": 78088 + } + }, + { + "ph": "f", "id": 78088, "pid": 435, "tid": 435, "ts": 1776649791965287, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791965306, "dur": 45, + "args": { + "External id": 10096, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78089, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78089, "pid": 0, "tid": 7, "ts": 1776649791965306, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791965288, "dur": 6, + "args": { + "External id": 10096, + "cbid": 211, "correlation": 78089 + } + }, + { + "ph": "s", "id": 78089, "pid": 435, "tid": 435, "ts": 1776649791965288, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791965318, "dur": 0, + "args": { + "External id": 10098, + "cbid": 210, "correlation": 78116 + } + }, + { + "ph": "f", "id": 78116, "pid": 435, "tid": 435, "ts": 1776649791965318, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791965352, "dur": 49, + "args": { + "External id": 10098, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78117, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78117, "pid": 0, "tid": 7, "ts": 1776649791965352, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791965319, "dur": 6, + "args": { + "External id": 10098, + "cbid": 211, "correlation": 78117 + } + }, + { + "ph": "s", "id": 78117, "pid": 435, "tid": 435, "ts": 1776649791965319, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791965348, "dur": 0, + "args": { + "External id": 10100, + "cbid": 210, "correlation": 78144 + } + }, + { + "ph": "f", "id": 78144, "pid": 435, "tid": 435, "ts": 1776649791965348, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791965403, "dur": 47, + "args": { + "External id": 10100, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78145, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78145, "pid": 0, "tid": 7, "ts": 1776649791965403, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791965349, "dur": 6, + "args": { + "External id": 10100, + "cbid": 211, "correlation": 78145 + } + }, + { + "ph": "s", "id": 78145, "pid": 435, "tid": 435, "ts": 1776649791965349, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791965378, "dur": 0, + "args": { + "External id": 10102, + "cbid": 210, "correlation": 78172 + } + }, + { + "ph": "f", "id": 78172, "pid": 435, "tid": 435, "ts": 1776649791965378, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791965452, "dur": 49, + "args": { + "External id": 10102, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78173, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78173, "pid": 0, "tid": 7, "ts": 1776649791965452, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791965379, "dur": 6, + "args": { + "External id": 10102, + "cbid": 211, "correlation": 78173 + } + }, + { + "ph": "s", "id": 78173, "pid": 435, "tid": 435, "ts": 1776649791965379, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791965408, "dur": 0, + "args": { + "External id": 10104, + "cbid": 210, "correlation": 78200 + } + }, + { + "ph": "f", "id": 78200, "pid": 435, "tid": 435, "ts": 1776649791965408, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791965502, "dur": 47, + "args": { + "External id": 10104, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78201, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78201, "pid": 0, "tid": 7, "ts": 1776649791965502, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791965409, "dur": 5, + "args": { + "External id": 10104, + "cbid": 211, "correlation": 78201 + } + }, + { + "ph": "s", "id": 78201, "pid": 435, "tid": 435, "ts": 1776649791965409, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791965438, "dur": 0, + "args": { + "External id": 10106, + "cbid": 210, "correlation": 78228 + } + }, + { + "ph": "f", "id": 78228, "pid": 435, "tid": 435, "ts": 1776649791965438, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791965551, "dur": 49, + "args": { + "External id": 10106, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78229, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78229, "pid": 0, "tid": 7, "ts": 1776649791965551, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791965439, "dur": 5, + "args": { + "External id": 10106, + "cbid": 211, "correlation": 78229 + } + }, + { + "ph": "s", "id": 78229, "pid": 435, "tid": 435, "ts": 1776649791965439, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791965468, "dur": 0, + "args": { + "External id": 10108, + "cbid": 210, "correlation": 78256 + } + }, + { + "ph": "f", "id": 78256, "pid": 435, "tid": 435, "ts": 1776649791965468, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791965602, "dur": 49, + "args": { + "External id": 10108, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78257, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78257, "pid": 0, "tid": 7, "ts": 1776649791965602, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791965469, "dur": 6, + "args": { + "External id": 10108, + "cbid": 211, "correlation": 78257 + } + }, + { + "ph": "s", "id": 78257, "pid": 435, "tid": 435, "ts": 1776649791965469, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791965497, "dur": 0, + "args": { + "External id": 10110, + "cbid": 210, "correlation": 78284 + } + }, + { + "ph": "f", "id": 78284, "pid": 435, "tid": 435, "ts": 1776649791965497, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791965652, "dur": 44, + "args": { + "External id": 10110, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78285, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78285, "pid": 0, "tid": 7, "ts": 1776649791965652, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791965499, "dur": 5, + "args": { + "External id": 10110, + "cbid": 211, "correlation": 78285 + } + }, + { + "ph": "s", "id": 78285, "pid": 435, "tid": 435, "ts": 1776649791965499, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791965527, "dur": 0, + "args": { + "External id": 10112, + "cbid": 210, "correlation": 78312 + } + }, + { + "ph": "f", "id": 78312, "pid": 435, "tid": 435, "ts": 1776649791965527, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791965698, "dur": 45, + "args": { + "External id": 10112, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78313, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78313, "pid": 0, "tid": 7, "ts": 1776649791965698, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791965529, "dur": 6, + "args": { + "External id": 10112, + "cbid": 211, "correlation": 78313 + } + }, + { + "ph": "s", "id": 78313, "pid": 435, "tid": 435, "ts": 1776649791965529, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791965557, "dur": 0, + "args": { + "External id": 10114, + "cbid": 210, "correlation": 78340 + } + }, + { + "ph": "f", "id": 78340, "pid": 435, "tid": 435, "ts": 1776649791965557, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791965745, "dur": 46, + "args": { + "External id": 10114, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78341, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78341, "pid": 0, "tid": 7, "ts": 1776649791965745, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791965558, "dur": 5, + "args": { + "External id": 10114, + "cbid": 211, "correlation": 78341 + } + }, + { + "ph": "s", "id": 78341, "pid": 435, "tid": 435, "ts": 1776649791965558, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791965587, "dur": 0, + "args": { + "External id": 10116, + "cbid": 210, "correlation": 78368 + } + }, + { + "ph": "f", "id": 78368, "pid": 435, "tid": 435, "ts": 1776649791965587, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791965793, "dur": 46, + "args": { + "External id": 10116, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78369, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78369, "pid": 0, "tid": 7, "ts": 1776649791965793, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791965588, "dur": 5, + "args": { + "External id": 10116, + "cbid": 211, "correlation": 78369 + } + }, + { + "ph": "s", "id": 78369, "pid": 435, "tid": 435, "ts": 1776649791965588, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791965616, "dur": 0, + "args": { + "External id": 10118, + "cbid": 210, "correlation": 78396 + } + }, + { + "ph": "f", "id": 78396, "pid": 435, "tid": 435, "ts": 1776649791965616, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791965840, "dur": 46, + "args": { + "External id": 10118, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78397, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78397, "pid": 0, "tid": 7, "ts": 1776649791965840, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791965618, "dur": 6, + "args": { + "External id": 10118, + "cbid": 211, "correlation": 78397 + } + }, + { + "ph": "s", "id": 78397, "pid": 435, "tid": 435, "ts": 1776649791965618, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791965646, "dur": 0, + "args": { + "External id": 10120, + "cbid": 210, "correlation": 78424 + } + }, + { + "ph": "f", "id": 78424, "pid": 435, "tid": 435, "ts": 1776649791965646, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791965888, "dur": 46, + "args": { + "External id": 10120, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78425, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78425, "pid": 0, "tid": 7, "ts": 1776649791965888, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791965648, "dur": 5, + "args": { + "External id": 10120, + "cbid": 211, "correlation": 78425 + } + }, + { + "ph": "s", "id": 78425, "pid": 435, "tid": 435, "ts": 1776649791965648, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791965676, "dur": 0, + "args": { + "External id": 10122, + "cbid": 210, "correlation": 78452 + } + }, + { + "ph": "f", "id": 78452, "pid": 435, "tid": 435, "ts": 1776649791965676, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791965936, "dur": 46, + "args": { + "External id": 10122, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78453, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78453, "pid": 0, "tid": 7, "ts": 1776649791965936, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791965678, "dur": 6, + "args": { + "External id": 10122, + "cbid": 211, "correlation": 78453 + } + }, + { + "ph": "s", "id": 78453, "pid": 435, "tid": 435, "ts": 1776649791965678, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791965706, "dur": 0, + "args": { + "External id": 10124, + "cbid": 210, "correlation": 78480 + } + }, + { + "ph": "f", "id": 78480, "pid": 435, "tid": 435, "ts": 1776649791965706, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791965984, "dur": 51, + "args": { + "External id": 10124, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78481, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78481, "pid": 0, "tid": 7, "ts": 1776649791965984, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791965708, "dur": 6, + "args": { + "External id": 10124, + "cbid": 211, "correlation": 78481 + } + }, + { + "ph": "s", "id": 78481, "pid": 435, "tid": 435, "ts": 1776649791965708, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791965736, "dur": 0, + "args": { + "External id": 10126, + "cbid": 210, "correlation": 78508 + } + }, + { + "ph": "f", "id": 78508, "pid": 435, "tid": 435, "ts": 1776649791965736, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791966037, "dur": 51, + "args": { + "External id": 10126, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78509, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78509, "pid": 0, "tid": 7, "ts": 1776649791966037, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791965737, "dur": 5, + "args": { + "External id": 10126, + "cbid": 211, "correlation": 78509 + } + }, + { + "ph": "s", "id": 78509, "pid": 435, "tid": 435, "ts": 1776649791965737, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791965766, "dur": 0, + "args": { + "External id": 10128, + "cbid": 210, "correlation": 78536 + } + }, + { + "ph": "f", "id": 78536, "pid": 435, "tid": 435, "ts": 1776649791965766, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791966089, "dur": 48, + "args": { + "External id": 10128, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78537, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78537, "pid": 0, "tid": 7, "ts": 1776649791966089, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791965767, "dur": 6, + "args": { + "External id": 10128, + "cbid": 211, "correlation": 78537 + } + }, + { + "ph": "s", "id": 78537, "pid": 435, "tid": 435, "ts": 1776649791965767, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791965796, "dur": 0, + "args": { + "External id": 10130, + "cbid": 210, "correlation": 78564 + } + }, + { + "ph": "f", "id": 78564, "pid": 435, "tid": 435, "ts": 1776649791965796, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791966139, "dur": 44, + "args": { + "External id": 10130, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78565, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78565, "pid": 0, "tid": 7, "ts": 1776649791966139, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791965798, "dur": 5, + "args": { + "External id": 10130, + "cbid": 211, "correlation": 78565 + } + }, + { + "ph": "s", "id": 78565, "pid": 435, "tid": 435, "ts": 1776649791965798, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791965827, "dur": 0, + "args": { + "External id": 10132, + "cbid": 210, "correlation": 78592 + } + }, + { + "ph": "f", "id": 78592, "pid": 435, "tid": 435, "ts": 1776649791965827, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791966185, "dur": 49, + "args": { + "External id": 10132, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78593, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78593, "pid": 0, "tid": 7, "ts": 1776649791966185, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791965828, "dur": 5, + "args": { + "External id": 10132, + "cbid": 211, "correlation": 78593 + } + }, + { + "ph": "s", "id": 78593, "pid": 435, "tid": 435, "ts": 1776649791965828, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791965856, "dur": 0, + "args": { + "External id": 10134, + "cbid": 210, "correlation": 78620 + } + }, + { + "ph": "f", "id": 78620, "pid": 435, "tid": 435, "ts": 1776649791965856, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791966235, "dur": 44, + "args": { + "External id": 10134, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78621, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78621, "pid": 0, "tid": 7, "ts": 1776649791966235, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791965858, "dur": 6, + "args": { + "External id": 10134, + "cbid": 211, "correlation": 78621 + } + }, + { + "ph": "s", "id": 78621, "pid": 435, "tid": 435, "ts": 1776649791965858, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791965886, "dur": 0, + "args": { + "External id": 10136, + "cbid": 210, "correlation": 78648 + } + }, + { + "ph": "f", "id": 78648, "pid": 435, "tid": 435, "ts": 1776649791965886, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791966281, "dur": 44, + "args": { + "External id": 10136, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78649, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78649, "pid": 0, "tid": 7, "ts": 1776649791966281, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791965888, "dur": 6, + "args": { + "External id": 10136, + "cbid": 211, "correlation": 78649 + } + }, + { + "ph": "s", "id": 78649, "pid": 435, "tid": 435, "ts": 1776649791965888, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791965916, "dur": 0, + "args": { + "External id": 10138, + "cbid": 210, "correlation": 78676 + } + }, + { + "ph": "f", "id": 78676, "pid": 435, "tid": 435, "ts": 1776649791965916, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791966326, "dur": 46, + "args": { + "External id": 10138, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78677, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78677, "pid": 0, "tid": 7, "ts": 1776649791966326, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791965918, "dur": 5, + "args": { + "External id": 10138, + "cbid": 211, "correlation": 78677 + } + }, + { + "ph": "s", "id": 78677, "pid": 435, "tid": 435, "ts": 1776649791965918, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791965946, "dur": 0, + "args": { + "External id": 10140, + "cbid": 210, "correlation": 78704 + } + }, + { + "ph": "f", "id": 78704, "pid": 435, "tid": 435, "ts": 1776649791965946, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791966374, "dur": 50, + "args": { + "External id": 10140, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78705, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78705, "pid": 0, "tid": 7, "ts": 1776649791966374, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791965947, "dur": 6, + "args": { + "External id": 10140, + "cbid": 211, "correlation": 78705 + } + }, + { + "ph": "s", "id": 78705, "pid": 435, "tid": 435, "ts": 1776649791965947, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791965976, "dur": 0, + "args": { + "External id": 10142, + "cbid": 210, "correlation": 78732 + } + }, + { + "ph": "f", "id": 78732, "pid": 435, "tid": 435, "ts": 1776649791965976, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791966426, "dur": 49, + "args": { + "External id": 10142, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78733, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78733, "pid": 0, "tid": 7, "ts": 1776649791966426, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791965977, "dur": 5, + "args": { + "External id": 10142, + "cbid": 211, "correlation": 78733 + } + }, + { + "ph": "s", "id": 78733, "pid": 435, "tid": 435, "ts": 1776649791965977, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791966005, "dur": 0, + "args": { + "External id": 10144, + "cbid": 210, "correlation": 78760 + } + }, + { + "ph": "f", "id": 78760, "pid": 435, "tid": 435, "ts": 1776649791966005, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791966477, "dur": 41, + "args": { + "External id": 10144, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78761, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78761, "pid": 0, "tid": 7, "ts": 1776649791966477, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791966007, "dur": 5, + "args": { + "External id": 10144, + "cbid": 211, "correlation": 78761 + } + }, + { + "ph": "s", "id": 78761, "pid": 435, "tid": 435, "ts": 1776649791966007, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791966036, "dur": 0, + "args": { + "External id": 10146, + "cbid": 210, "correlation": 78788 + } + }, + { + "ph": "f", "id": 78788, "pid": 435, "tid": 435, "ts": 1776649791966036, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791966519, "dur": 49, + "args": { + "External id": 10146, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78789, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78789, "pid": 0, "tid": 7, "ts": 1776649791966519, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791966037, "dur": 6, + "args": { + "External id": 10146, + "cbid": 211, "correlation": 78789 + } + }, + { + "ph": "s", "id": 78789, "pid": 435, "tid": 435, "ts": 1776649791966037, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791966066, "dur": 0, + "args": { + "External id": 10148, + "cbid": 210, "correlation": 78816 + } + }, + { + "ph": "f", "id": 78816, "pid": 435, "tid": 435, "ts": 1776649791966066, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791966570, "dur": 46, + "args": { + "External id": 10148, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78817, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78817, "pid": 0, "tid": 7, "ts": 1776649791966570, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791966067, "dur": 5, + "args": { + "External id": 10148, + "cbid": 211, "correlation": 78817 + } + }, + { + "ph": "s", "id": 78817, "pid": 435, "tid": 435, "ts": 1776649791966067, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791966095, "dur": 0, + "args": { + "External id": 10150, + "cbid": 210, "correlation": 78844 + } + }, + { + "ph": "f", "id": 78844, "pid": 435, "tid": 435, "ts": 1776649791966095, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791966618, "dur": 44, + "args": { + "External id": 10150, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78845, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78845, "pid": 0, "tid": 7, "ts": 1776649791966618, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791966097, "dur": 5, + "args": { + "External id": 10150, + "cbid": 211, "correlation": 78845 + } + }, + { + "ph": "s", "id": 78845, "pid": 435, "tid": 435, "ts": 1776649791966097, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791966126, "dur": 0, + "args": { + "External id": 10152, + "cbid": 210, "correlation": 78872 + } + }, + { + "ph": "f", "id": 78872, "pid": 435, "tid": 435, "ts": 1776649791966126, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791966664, "dur": 49, + "args": { + "External id": 10152, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78873, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78873, "pid": 0, "tid": 7, "ts": 1776649791966664, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791966127, "dur": 5, + "args": { + "External id": 10152, + "cbid": 211, "correlation": 78873 + } + }, + { + "ph": "s", "id": 78873, "pid": 435, "tid": 435, "ts": 1776649791966127, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791966155, "dur": 0, + "args": { + "External id": 10154, + "cbid": 210, "correlation": 78900 + } + }, + { + "ph": "f", "id": 78900, "pid": 435, "tid": 435, "ts": 1776649791966155, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791966714, "dur": 52, + "args": { + "External id": 10154, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78901, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78901, "pid": 0, "tid": 7, "ts": 1776649791966714, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791966156, "dur": 6, + "args": { + "External id": 10154, + "cbid": 211, "correlation": 78901 + } + }, + { + "ph": "s", "id": 78901, "pid": 435, "tid": 435, "ts": 1776649791966156, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791966185, "dur": 0, + "args": { + "External id": 10156, + "cbid": 210, "correlation": 78928 + } + }, + { + "ph": "f", "id": 78928, "pid": 435, "tid": 435, "ts": 1776649791966185, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791966768, "dur": 46, + "args": { + "External id": 10156, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78929, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78929, "pid": 0, "tid": 7, "ts": 1776649791966768, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791966186, "dur": 5, + "args": { + "External id": 10156, + "cbid": 211, "correlation": 78929 + } + }, + { + "ph": "s", "id": 78929, "pid": 435, "tid": 435, "ts": 1776649791966186, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791966215, "dur": 0, + "args": { + "External id": 10158, + "cbid": 210, "correlation": 78956 + } + }, + { + "ph": "f", "id": 78956, "pid": 435, "tid": 435, "ts": 1776649791966215, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791966816, "dur": 44, + "args": { + "External id": 10158, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78957, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78957, "pid": 0, "tid": 7, "ts": 1776649791966816, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791966216, "dur": 5, + "args": { + "External id": 10158, + "cbid": 211, "correlation": 78957 + } + }, + { + "ph": "s", "id": 78957, "pid": 435, "tid": 435, "ts": 1776649791966216, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791966245, "dur": 0, + "args": { + "External id": 10160, + "cbid": 210, "correlation": 78984 + } + }, + { + "ph": "f", "id": 78984, "pid": 435, "tid": 435, "ts": 1776649791966245, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791966862, "dur": 52, + "args": { + "External id": 10160, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 78985, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 78985, "pid": 0, "tid": 7, "ts": 1776649791966862, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791966246, "dur": 5, + "args": { + "External id": 10160, + "cbid": 211, "correlation": 78985 + } + }, + { + "ph": "s", "id": 78985, "pid": 435, "tid": 435, "ts": 1776649791966246, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791966275, "dur": 0, + "args": { + "External id": 10162, + "cbid": 210, "correlation": 79012 + } + }, + { + "ph": "f", "id": 79012, "pid": 435, "tid": 435, "ts": 1776649791966275, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791966915, "dur": 43, + "args": { + "External id": 10162, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79013, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79013, "pid": 0, "tid": 7, "ts": 1776649791966915, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791966276, "dur": 5, + "args": { + "External id": 10162, + "cbid": 211, "correlation": 79013 + } + }, + { + "ph": "s", "id": 79013, "pid": 435, "tid": 435, "ts": 1776649791966276, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791966305, "dur": 0, + "args": { + "External id": 10164, + "cbid": 210, "correlation": 79040 + } + }, + { + "ph": "f", "id": 79040, "pid": 435, "tid": 435, "ts": 1776649791966305, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791966960, "dur": 49, + "args": { + "External id": 10164, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79041, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79041, "pid": 0, "tid": 7, "ts": 1776649791966960, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791966306, "dur": 5, + "args": { + "External id": 10164, + "cbid": 211, "correlation": 79041 + } + }, + { + "ph": "s", "id": 79041, "pid": 435, "tid": 435, "ts": 1776649791966306, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791966335, "dur": 0, + "args": { + "External id": 10166, + "cbid": 210, "correlation": 79068 + } + }, + { + "ph": "f", "id": 79068, "pid": 435, "tid": 435, "ts": 1776649791966335, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791967011, "dur": 47, + "args": { + "External id": 10166, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79069, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79069, "pid": 0, "tid": 7, "ts": 1776649791967011, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791966336, "dur": 6, + "args": { + "External id": 10166, + "cbid": 211, "correlation": 79069 + } + }, + { + "ph": "s", "id": 79069, "pid": 435, "tid": 435, "ts": 1776649791966336, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791966365, "dur": 0, + "args": { + "External id": 10168, + "cbid": 210, "correlation": 79096 + } + }, + { + "ph": "f", "id": 79096, "pid": 435, "tid": 435, "ts": 1776649791966365, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791967059, "dur": 50, + "args": { + "External id": 10168, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79097, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79097, "pid": 0, "tid": 7, "ts": 1776649791967059, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791966366, "dur": 5, + "args": { + "External id": 10168, + "cbid": 211, "correlation": 79097 + } + }, + { + "ph": "s", "id": 79097, "pid": 435, "tid": 435, "ts": 1776649791966366, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791966395, "dur": 0, + "args": { + "External id": 10170, + "cbid": 210, "correlation": 79124 + } + }, + { + "ph": "f", "id": 79124, "pid": 435, "tid": 435, "ts": 1776649791966395, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791967111, "dur": 43, + "args": { + "External id": 10170, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79125, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79125, "pid": 0, "tid": 7, "ts": 1776649791967111, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791966396, "dur": 6, + "args": { + "External id": 10170, + "cbid": 211, "correlation": 79125 + } + }, + { + "ph": "s", "id": 79125, "pid": 435, "tid": 435, "ts": 1776649791966396, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791966425, "dur": 0, + "args": { + "External id": 10172, + "cbid": 210, "correlation": 79152 + } + }, + { + "ph": "f", "id": 79152, "pid": 435, "tid": 435, "ts": 1776649791966425, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791967156, "dur": 47, + "args": { + "External id": 10172, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79153, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79153, "pid": 0, "tid": 7, "ts": 1776649791967156, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791966426, "dur": 5, + "args": { + "External id": 10172, + "cbid": 211, "correlation": 79153 + } + }, + { + "ph": "s", "id": 79153, "pid": 435, "tid": 435, "ts": 1776649791966426, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791966454, "dur": 0, + "args": { + "External id": 10174, + "cbid": 210, "correlation": 79180 + } + }, + { + "ph": "f", "id": 79180, "pid": 435, "tid": 435, "ts": 1776649791966454, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791967205, "dur": 46, + "args": { + "External id": 10174, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79181, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79181, "pid": 0, "tid": 7, "ts": 1776649791967205, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791966455, "dur": 5, + "args": { + "External id": 10174, + "cbid": 211, "correlation": 79181 + } + }, + { + "ph": "s", "id": 79181, "pid": 435, "tid": 435, "ts": 1776649791966455, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791966484, "dur": 0, + "args": { + "External id": 10176, + "cbid": 210, "correlation": 79208 + } + }, + { + "ph": "f", "id": 79208, "pid": 435, "tid": 435, "ts": 1776649791966484, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791967252, "dur": 44, + "args": { + "External id": 10176, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79209, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79209, "pid": 0, "tid": 7, "ts": 1776649791967252, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791966485, "dur": 5, + "args": { + "External id": 10176, + "cbid": 211, "correlation": 79209 + } + }, + { + "ph": "s", "id": 79209, "pid": 435, "tid": 435, "ts": 1776649791966485, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791966514, "dur": 0, + "args": { + "External id": 10178, + "cbid": 210, "correlation": 79236 + } + }, + { + "ph": "f", "id": 79236, "pid": 435, "tid": 435, "ts": 1776649791966514, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791967298, "dur": 49, + "args": { + "External id": 10178, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79237, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79237, "pid": 0, "tid": 7, "ts": 1776649791967298, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791966515, "dur": 5, + "args": { + "External id": 10178, + "cbid": 211, "correlation": 79237 + } + }, + { + "ph": "s", "id": 79237, "pid": 435, "tid": 435, "ts": 1776649791966515, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791966543, "dur": 0, + "args": { + "External id": 10180, + "cbid": 210, "correlation": 79264 + } + }, + { + "ph": "f", "id": 79264, "pid": 435, "tid": 435, "ts": 1776649791966543, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791967349, "dur": 49, + "args": { + "External id": 10180, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79265, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79265, "pid": 0, "tid": 7, "ts": 1776649791967349, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791966545, "dur": 6, + "args": { + "External id": 10180, + "cbid": 211, "correlation": 79265 + } + }, + { + "ph": "s", "id": 79265, "pid": 435, "tid": 435, "ts": 1776649791966545, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791966573, "dur": 0, + "args": { + "External id": 10182, + "cbid": 210, "correlation": 79292 + } + }, + { + "ph": "f", "id": 79292, "pid": 435, "tid": 435, "ts": 1776649791966573, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791967400, "dur": 46, + "args": { + "External id": 10182, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79293, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79293, "pid": 0, "tid": 7, "ts": 1776649791967400, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791966574, "dur": 5, + "args": { + "External id": 10182, + "cbid": 211, "correlation": 79293 + } + }, + { + "ph": "s", "id": 79293, "pid": 435, "tid": 435, "ts": 1776649791966574, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791966603, "dur": 0, + "args": { + "External id": 10184, + "cbid": 210, "correlation": 79320 + } + }, + { + "ph": "f", "id": 79320, "pid": 435, "tid": 435, "ts": 1776649791966603, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791967448, "dur": 44, + "args": { + "External id": 10184, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79321, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79321, "pid": 0, "tid": 7, "ts": 1776649791967448, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791966604, "dur": 5, + "args": { + "External id": 10184, + "cbid": 211, "correlation": 79321 + } + }, + { + "ph": "s", "id": 79321, "pid": 435, "tid": 435, "ts": 1776649791966604, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791966633, "dur": 0, + "args": { + "External id": 10186, + "cbid": 210, "correlation": 79348 + } + }, + { + "ph": "f", "id": 79348, "pid": 435, "tid": 435, "ts": 1776649791966633, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791967493, "dur": 46, + "args": { + "External id": 10186, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79349, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79349, "pid": 0, "tid": 7, "ts": 1776649791967493, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791966634, "dur": 6, + "args": { + "External id": 10186, + "cbid": 211, "correlation": 79349 + } + }, + { + "ph": "s", "id": 79349, "pid": 435, "tid": 435, "ts": 1776649791966634, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791966663, "dur": 0, + "args": { + "External id": 10188, + "cbid": 210, "correlation": 79376 + } + }, + { + "ph": "f", "id": 79376, "pid": 435, "tid": 435, "ts": 1776649791966663, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791967541, "dur": 49, + "args": { + "External id": 10188, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79377, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79377, "pid": 0, "tid": 7, "ts": 1776649791967541, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791966665, "dur": 5, + "args": { + "External id": 10188, + "cbid": 211, "correlation": 79377 + } + }, + { + "ph": "s", "id": 79377, "pid": 435, "tid": 435, "ts": 1776649791966665, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791966693, "dur": 0, + "args": { + "External id": 10190, + "cbid": 210, "correlation": 79404 + } + }, + { + "ph": "f", "id": 79404, "pid": 435, "tid": 435, "ts": 1776649791966693, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791967594, "dur": 49, + "args": { + "External id": 10190, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79405, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79405, "pid": 0, "tid": 7, "ts": 1776649791967594, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791966695, "dur": 6, + "args": { + "External id": 10190, + "cbid": 211, "correlation": 79405 + } + }, + { + "ph": "s", "id": 79405, "pid": 435, "tid": 435, "ts": 1776649791966695, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791966726, "dur": 0, + "args": { + "External id": 10192, + "cbid": 210, "correlation": 79432 + } + }, + { + "ph": "f", "id": 79432, "pid": 435, "tid": 435, "ts": 1776649791966726, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791967645, "dur": 47, + "args": { + "External id": 10192, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79433, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79433, "pid": 0, "tid": 7, "ts": 1776649791967645, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791966727, "dur": 6, + "args": { + "External id": 10192, + "cbid": 211, "correlation": 79433 + } + }, + { + "ph": "s", "id": 79433, "pid": 435, "tid": 435, "ts": 1776649791966727, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791966761, "dur": 4, + "args": { + "External id": 79444, + "cbid": 135, "correlation": 79444 + } + }, + { + "ph": "f", "id": 79444, "pid": 435, "tid": 435, "ts": 1776649791966761, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791966776, "dur": 925, + "args": { + "External id": 79449, + "cbid": 165, "correlation": 79449 + } + }, + { + "ph": "s", "id": 79449, "pid": 435, "tid": 435, "ts": 1776649791966776, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649791967705, "dur": 4, + "args": { + "External id": 79452, + "cbid": 139, "correlation": 79452 + } + }, + { + "ph": "f", "id": 79452, "pid": 435, "tid": 435, "ts": 1776649791967705, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor >, at::detail::Array >(int, at::native::AUnaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791967804, "dur": 33, + "args": { + "External id": 10195, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79469, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 79469, "pid": 0, "tid": 7, "ts": 1776649791967804, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791967791, "dur": 11, + "args": { + "External id": 10195, + "cbid": 211, "correlation": 79469 + } + }, + { + "ph": "s", "id": 79469, "pid": 435, "tid": 435, "ts": 1776649791967791, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649791967844, "dur": 57, + "args": { + "External id": 10199, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79482, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 79482, "pid": 0, "tid": 7, "ts": 1776649791967844, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791967832, "dur": 7, + "args": { + "External id": 10199, + "cbid": 211, "correlation": 79482 + } + }, + { + "ph": "s", "id": 79482, "pid": 435, "tid": 435, "ts": 1776649791967832, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BUnaryFunctor, at::detail::Array >(int, at::native::BUnaryFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791967903, "dur": 33, + "args": { + "External id": 10200, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79496, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 79496, "pid": 0, "tid": 7, "ts": 1776649791967903, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791967866, "dur": 7, + "args": { + "External id": 10200, + "cbid": 211, "correlation": 79496 + } + }, + { + "ph": "s", "id": 79496, "pid": 435, "tid": 435, "ts": 1776649791967866, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649791967938, "dur": 57, + "args": { + "External id": 10205, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79509, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 79509, "pid": 0, "tid": 7, "ts": 1776649791967938, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791967893, "dur": 6, + "args": { + "External id": 10205, + "cbid": 211, "correlation": 79509 + } + }, + { + "ph": "s", "id": 79509, "pid": 435, "tid": 435, "ts": 1776649791967893, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy_contig, unsigned int, 3, 128, 1>(at::native::(anonymous namespace)::OpaqueType<1u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata, unsigned int, 128, 1>, at::native::(anonymous namespace)::TensorSizeStride, int, unsigned int)", "pid": 0, "tid": 7, + "ts": 1776649791967997, "dur": 126, + "args": { + "External id": 10211, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79521, + "registers per thread": 21, + "shared memory": 0, + "blocks per SM": 4.000000, + "warps per SM": 64.000000, + "grid": [216, 2, 1], + "block": [512, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 79521, "pid": 0, "tid": 7, "ts": 1776649791967997, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791967953, "dur": 8, + "args": { + "External id": 10211, + "cbid": 211, "correlation": 79521 + } + }, + { + "ph": "s", "id": 79521, "pid": 435, "tid": 435, "ts": 1776649791967953, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649791968125, "dur": 135, + "args": { + "External id": 10222, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79534, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 79534, "pid": 0, "tid": 7, "ts": 1776649791968125, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791968014, "dur": 9, + "args": { + "External id": 10222, + "cbid": 211, "correlation": 79534 + } + }, + { + "ph": "s", "id": 79534, "pid": 435, "tid": 435, "ts": 1776649791968014, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, at::detail::Array >(int, at::native::CUDAFunctorOnSelf_add, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791968262, "dur": 63, + "args": { + "External id": 10225, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79545, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 79545, "pid": 0, "tid": 7, "ts": 1776649791968262, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791968050, "dur": 7, + "args": { + "External id": 10225, + "cbid": 211, "correlation": 79545 + } + }, + { + "ph": "s", "id": 79545, "pid": 435, "tid": 435, "ts": 1776649791968050, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649791968326, "dur": 179, + "args": { + "External id": 10229, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79558, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 79558, "pid": 0, "tid": 7, "ts": 1776649791968326, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791968079, "dur": 7, + "args": { + "External id": 10229, + "cbid": 211, "correlation": 79558 + } + }, + { + "ph": "s", "id": 79558, "pid": 435, "tid": 435, "ts": 1776649791968079, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor >, at::detail::Array >(int, at::native::BinaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791968507, "dur": 115, + "args": { + "External id": 10230, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79569, + "registers per thread": 22, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 79569, "pid": 0, "tid": 7, "ts": 1776649791968507, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791968104, "dur": 6, + "args": { + "External id": 10230, + "cbid": 211, "correlation": 79569 + } + }, + { + "ph": "s", "id": 79569, "pid": 435, "tid": 435, "ts": 1776649791968104, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649791968623, "dur": 514, + "args": { + "External id": 10238, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79583, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 79583, "pid": 0, "tid": 7, "ts": 1776649791968623, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791968171, "dur": 8, + "args": { + "External id": 10238, + "cbid": 211, "correlation": 79583 + } + }, + { + "ph": "s", "id": 79583, "pid": 435, "tid": 435, "ts": 1776649791968171, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649792277157, "dur": 7067, + "args": { + "External id": 10254, + "device": 0, "context": 1, + "stream": 7, "correlation": 79599, + "bytes": 67108864, "memory bandwidth (GB/s)": 9.495131451998688 + } + }, + { + "ph": "f", "id": 79599, "pid": 0, "tid": 7, "ts": 1776649792277157, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649792277010, "dur": 7183, + "args": { + "External id": 10254, + "cbid": 41, "correlation": 79599 + } + }, + { + "ph": "s", "id": 79599, "pid": 435, "tid": 435, "ts": 1776649792277010, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792284201, "dur": 57, + "args": { + "External id": 10254, + "cbid": 131, "correlation": 79600 + } + }, + { + "ph": "s", "id": 79600, "pid": 435, "tid": 435, "ts": 1776649792284201, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7, + "ts": 1776649792397280, "dur": 76, + "args": { + "External id": 10257, + "device": 0, "context": 1, + "stream": 7, "correlation": 79605, + "bytes": 67108864, "memory bandwidth (GB/s)": 876.7472401134003 + } + }, + { + "ph": "f", "id": 79605, "pid": 0, "tid": 7, "ts": 1776649792397280, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649792397201, "dur": 80, + "args": { + "External id": 10257, + "cbid": 41, "correlation": 79605 + } + }, + { + "ph": "s", "id": 79605, "pid": 435, "tid": 435, "ts": 1776649792397201, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649792397894, "dur": 4449, + "args": { + "External id": 10266, + "device": 0, "context": 1, + "stream": 7, "correlation": 79620, + "bytes": 16777216, "memory bandwidth (GB/s)": 3.7707328639737487 + } + }, + { + "ph": "f", "id": 79620, "pid": 0, "tid": 7, "ts": 1776649792397894, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649792397596, "dur": 4705, + "args": { + "External id": 10266, + "cbid": 41, "correlation": 79620 + } + }, + { + "ph": "s", "id": 79620, "pid": 435, "tid": 435, "ts": 1776649792397596, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792402303, "dur": 51, + "args": { + "External id": 10266, + "cbid": 131, "correlation": 79621 + } + }, + { + "ph": "s", "id": 79621, "pid": 435, "tid": 435, "ts": 1776649792402303, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649792402438, "dur": 24, + "args": { + "External id": 10270, + "device": 0, "context": 1, + "stream": 7, "correlation": 79633, + "bytes": 524288, "memory bandwidth (GB/s)": 21.78723404255319 + } + }, + { + "ph": "f", "id": 79633, "pid": 0, "tid": 7, "ts": 1776649792402438, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649792402394, "dur": 42, + "args": { + "External id": 10270, + "cbid": 41, "correlation": 79633 + } + }, + { + "ph": "s", "id": 79633, "pid": 435, "tid": 435, "ts": 1776649792402394, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792402437, "dur": 32, + "args": { + "External id": 10270, + "cbid": 131, "correlation": 79634 + } + }, + { + "ph": "s", "id": 79634, "pid": 435, "tid": 435, "ts": 1776649792402437, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649792402502, "dur": 1, + "args": { + "External id": 10274, + "device": 0, "context": 1, + "stream": 7, "correlation": 79646, + "bytes": 4096, "memory bandwidth (GB/s)": 2.6122448979591835 + } + }, + { + "ph": "f", "id": 79646, "pid": 0, "tid": 7, "ts": 1776649792402502, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649792402492, "dur": 6, + "args": { + "External id": 10274, + "cbid": 41, "correlation": 79646 + } + }, + { + "ph": "s", "id": 79646, "pid": 435, "tid": 435, "ts": 1776649792402492, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792402498, "dur": 10, + "args": { + "External id": 10274, + "cbid": 131, "correlation": 79647 + } + }, + { + "ph": "s", "id": 79647, "pid": 435, "tid": 435, "ts": 1776649792402498, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649792402690, "dur": 1039, + "args": { + "External id": 10289, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79659, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 79659, "pid": 0, "tid": 7, "ts": 1776649792402690, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792402669, "dur": 21, + "args": { + "External id": 10289, + "cbid": 211, "correlation": 79659 + } + }, + { + "ph": "s", "id": 79659, "pid": 435, "tid": 435, "ts": 1776649792402669, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649792403731, "dur": 7, + "args": { + "External id": 10295, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79673, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 4.740741, + "warps per SM": 18.962963, + "grid": [512, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 30 + } + }, + { + "ph": "f", "id": 79673, "pid": 0, "tid": 7, "ts": 1776649792403731, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792402729, "dur": 8, + "args": { + "External id": 10295, + "cbid": 211, "correlation": 79673 + } + }, + { + "ph": "s", "id": 79673, "pid": 435, "tid": 435, "ts": 1776649792402729, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649792403739, "dur": 247, + "args": { + "External id": 10297, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79684, + "registers per thread": 20, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 79684, "pid": 0, "tid": 7, "ts": 1776649792403739, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792402775, "dur": 8, + "args": { + "External id": 10297, + "cbid": 211, "correlation": 79684 + } + }, + { + "ph": "s", "id": 79684, "pid": 435, "tid": 435, "ts": 1776649792402775, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792403988, "dur": 76, + "args": { + "External id": 10298, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79694, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 79694, "pid": 0, "tid": 7, "ts": 1776649792403988, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792402808, "dur": 7, + "args": { + "External id": 10298, + "cbid": 211, "correlation": 79694 + } + }, + { + "ph": "s", "id": 79694, "pid": 435, "tid": 435, "ts": 1776649792402808, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649792404065, "dur": 192, + "args": { + "External id": 10302, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79707, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 79707, "pid": 0, "tid": 7, "ts": 1776649792404065, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792402841, "dur": 9, + "args": { + "External id": 10302, + "cbid": 211, "correlation": 79707 + } + }, + { + "ph": "s", "id": 79707, "pid": 435, "tid": 435, "ts": 1776649792402841, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, at::detail::Array >(int, at::native::CUDAFunctorOnSelf_add, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792404259, "dur": 152, + "args": { + "External id": 10303, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79714, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 79714, "pid": 0, "tid": 7, "ts": 1776649792404259, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792402889, "dur": 9, + "args": { + "External id": 10303, + "cbid": 211, "correlation": 79714 + } + }, + { + "ph": "s", "id": 79714, "pid": 435, "tid": 435, "ts": 1776649792402889, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1}, at::detail::Array >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792404413, "dur": 157, + "args": { + "External id": 10304, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79724, + "registers per thread": 22, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 79724, "pid": 0, "tid": 7, "ts": 1776649792404413, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792402936, "dur": 8, + "args": { + "External id": 10304, + "cbid": 211, "correlation": 79724 + } + }, + { + "ph": "s", "id": 79724, "pid": 435, "tid": 435, "ts": 1776649792402936, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649792404572, "dur": 198, + "args": { + "External id": 10314, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79737, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 1213.629639, + "warps per SM": 4854.518555, + "grid": [131072, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 79737, "pid": 0, "tid": 7, "ts": 1776649792404572, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792402996, "dur": 9, + "args": { + "External id": 10314, + "cbid": 211, "correlation": 79737 + } + }, + { + "ph": "s", "id": 79737, "pid": 435, "tid": 435, "ts": 1776649792402996, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649792404773, "dur": 1, + "args": { + "External id": 10324, + "device": 0, "context": 1, + "stream": 7, "correlation": 79752, + "bytes": 512, "memory bandwidth (GB/s)": 0.38095238095238093 + } + }, + { + "ph": "f", "id": 79752, "pid": 0, "tid": 7, "ts": 1776649792404773, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649792403096, "dur": 10, + "args": { + "External id": 10324, + "cbid": 41, "correlation": 79752 + } + }, + { + "ph": "s", "id": 79752, "pid": 435, "tid": 435, "ts": 1776649792403096, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792403107, "dur": 1673, + "args": { + "External id": 10324, + "cbid": 131, "correlation": 79753 + } + }, + { + "ph": "s", "id": 79753, "pid": 435, "tid": 435, "ts": 1776649792403107, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649792404844, "dur": 9, + "args": { + "External id": 10326, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79762, + "registers per thread": 40, + "shared memory": 0, + "blocks per SM": 4.740741, + "warps per SM": 18.962963, + "grid": [512, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 30 + } + }, + { + "ph": "f", "id": 79762, "pid": 0, "tid": 7, "ts": 1776649792404844, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792404831, "dur": 12, + "args": { + "External id": 10326, + "cbid": 211, "correlation": 79762 + } + }, + { + "ph": "s", "id": 79762, "pid": 435, "tid": 435, "ts": 1776649792404831, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649792404911, "dur": 216, + "args": { + "External id": 10340, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79777, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 1213.629639, + "warps per SM": 4854.518555, + "grid": [131072, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 79777, "pid": 0, "tid": 7, "ts": 1776649792404911, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792404901, "dur": 8, + "args": { + "External id": 10340, + "cbid": 211, "correlation": 79777 + } + }, + { + "ph": "s", "id": 79777, "pid": 435, "tid": 435, "ts": 1776649792404901, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649792405131, "dur": 1, + "args": { + "External id": 10350, + "device": 0, "context": 1, + "stream": 7, "correlation": 79792, + "bytes": 8192, "memory bandwidth (GB/s)": 4.654545454545454 + } + }, + { + "ph": "f", "id": 79792, "pid": 0, "tid": 7, "ts": 1776649792405131, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649792404960, "dur": 9, + "args": { + "External id": 10350, + "cbid": 41, "correlation": 79792 + } + }, + { + "ph": "s", "id": 79792, "pid": 435, "tid": 435, "ts": 1776649792404960, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792404970, "dur": 167, + "args": { + "External id": 10350, + "cbid": 131, "correlation": 79793 + } + }, + { + "ph": "s", "id": 79793, "pid": 435, "tid": 435, "ts": 1776649792404970, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649792405168, "dur": 245, + "args": { + "External id": 10346, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79802, + "registers per thread": 40, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 75 + } + }, + { + "ph": "f", "id": 79802, "pid": 0, "tid": 7, "ts": 1776649792405168, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792405156, "dur": 10, + "args": { + "External id": 10346, + "cbid": 211, "correlation": 79802 + } + }, + { + "ph": "s", "id": 79802, "pid": 435, "tid": 435, "ts": 1776649792405156, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoH (Device -> Pageable)", "pid": 0, "tid": 7, + "ts": 1776649792427924, "dur": 91047, + "args": { + "External id": 10359, + "device": 0, "context": 1, + "stream": 7, "correlation": 79810, + "bytes": 134217728, "memory bandwidth (GB/s)": 1.4741541595294867 + } + }, + { + "ph": "f", "id": 79810, "pid": 0, "tid": 7, "ts": 1776649792427924, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649792427877, "dur": 92365, + "args": { + "External id": 10359, + "cbid": 41, "correlation": 79810 + } + }, + { + "ph": "s", "id": 79810, "pid": 435, "tid": 435, "ts": 1776649792427877, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792520249, "dur": 17, + "args": { + "External id": 10359, + "cbid": 131, "correlation": 79811 + } + }, + { + "ph": "s", "id": 79811, "pid": 435, "tid": 435, "ts": 1776649792520249, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649792641813, "dur": 6541, + "args": { + "External id": 10369, + "device": 0, "context": 1, + "stream": 7, "correlation": 79823, + "bytes": 16777216, "memory bandwidth (GB/s)": 2.5646694390482674 + } + }, + { + "ph": "f", "id": 79823, "pid": 0, "tid": 7, "ts": 1776649792641813, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649792641656, "dur": 6664, + "args": { + "External id": 10369, + "cbid": 41, "correlation": 79823 + } + }, + { + "ph": "s", "id": 79823, "pid": 435, "tid": 435, "ts": 1776649792641656, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792648328, "dur": 36, + "args": { + "External id": 10369, + "cbid": 131, "correlation": 79824 + } + }, + { + "ph": "s", "id": 79824, "pid": 435, "tid": 435, "ts": 1776649792648328, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7, + "ts": 1776649792648655, "dur": 15, + "args": { + "External id": 10375, + "device": 0, "context": 1, + "stream": 7, "correlation": 79829, + "bytes": 16777216, "memory bandwidth (GB/s)": 1050.6773547094188 + } + }, + { + "ph": "f", "id": 79829, "pid": 0, "tid": 7, "ts": 1776649792648655, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649792648605, "dur": 47, + "args": { + "External id": 10375, + "cbid": 41, "correlation": 79829 + } + }, + { + "ph": "s", "id": 79829, "pid": 435, "tid": 435, "ts": 1776649792648605, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7, + "ts": 1776649792648708, "dur": 3, + "args": { + "External id": 10381, + "device": 0, "context": 1, + "stream": 7, "correlation": 79837, + "bytes": 524288, "memory bandwidth (GB/s)": 141.24137931034483 + } + }, + { + "ph": "f", "id": 79837, "pid": 0, "tid": 7, "ts": 1776649792648708, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649792648689, "dur": 16, + "args": { + "External id": 10381, + "cbid": 41, "correlation": 79837 + } + }, + { + "ph": "s", "id": 79837, "pid": 435, "tid": 435, "ts": 1776649792648689, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792649914, "dur": 1, + "args": { + "External id": 10386, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79860, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79860, "pid": 0, "tid": 7, "ts": 1776649792649914, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792649891, "dur": 24, + "args": { + "External id": 10386, + "cbid": 211, "correlation": 79860 + } + }, + { + "ph": "s", "id": 79860, "pid": 435, "tid": 435, "ts": 1776649792649891, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792649933, "dur": 2, + "args": { + "External id": 79866, + "cbid": 200, "correlation": 79866 + } + }, + { + "ph": "f", "id": 79866, "pid": 435, "tid": 435, "ts": 1776649792649933, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792649936, "dur": 3, + "args": { + "External id": 79867, + "cbid": 273, "correlation": 79867 + } + }, + { + "ph": "f", "id": 79867, "pid": 435, "tid": 435, "ts": 1776649792649936, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792649951, "dur": 19, + "args": { + "External id": 79868, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79868, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79868, "pid": 0, "tid": 7, "ts": 1776649792649951, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792649941, "dur": 7, + "args": { + "External id": 79868, + "cbid": 211, "correlation": 79868 + } + }, + { + "ph": "s", "id": 79868, "pid": 435, "tid": 435, "ts": 1776649792649941, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792649995, "dur": 1, + "args": { + "External id": 10391, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79886, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79886, "pid": 0, "tid": 7, "ts": 1776649792649995, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792649985, "dur": 7, + "args": { + "External id": 10391, + "cbid": 211, "correlation": 79886 + } + }, + { + "ph": "s", "id": 79886, "pid": 435, "tid": 435, "ts": 1776649792649985, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792649999, "dur": 0, + "args": { + "External id": 79892, + "cbid": 200, "correlation": 79892 + } + }, + { + "ph": "f", "id": 79892, "pid": 435, "tid": 435, "ts": 1776649792649999, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650000, "dur": 0, + "args": { + "External id": 79893, + "cbid": 273, "correlation": 79893 + } + }, + { + "ph": "f", "id": 79893, "pid": 435, "tid": 435, "ts": 1776649792650000, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792650009, "dur": 19, + "args": { + "External id": 79894, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79894, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79894, "pid": 0, "tid": 7, "ts": 1776649792650009, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650001, "dur": 5, + "args": { + "External id": 79894, + "cbid": 211, "correlation": 79894 + } + }, + { + "ph": "s", "id": 79894, "pid": 435, "tid": 435, "ts": 1776649792650001, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792650047, "dur": 1, + "args": { + "External id": 10396, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79912, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79912, "pid": 0, "tid": 7, "ts": 1776649792650047, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650037, "dur": 7, + "args": { + "External id": 10396, + "cbid": 211, "correlation": 79912 + } + }, + { + "ph": "s", "id": 79912, "pid": 435, "tid": 435, "ts": 1776649792650037, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650051, "dur": 0, + "args": { + "External id": 79918, + "cbid": 200, "correlation": 79918 + } + }, + { + "ph": "f", "id": 79918, "pid": 435, "tid": 435, "ts": 1776649792650051, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650052, "dur": 0, + "args": { + "External id": 79919, + "cbid": 273, "correlation": 79919 + } + }, + { + "ph": "f", "id": 79919, "pid": 435, "tid": 435, "ts": 1776649792650052, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792650061, "dur": 19, + "args": { + "External id": 79920, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79920, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79920, "pid": 0, "tid": 7, "ts": 1776649792650061, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650053, "dur": 5, + "args": { + "External id": 79920, + "cbid": 211, "correlation": 79920 + } + }, + { + "ph": "s", "id": 79920, "pid": 435, "tid": 435, "ts": 1776649792650053, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792650097, "dur": 1, + "args": { + "External id": 10401, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79938, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79938, "pid": 0, "tid": 7, "ts": 1776649792650097, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650088, "dur": 6, + "args": { + "External id": 10401, + "cbid": 211, "correlation": 79938 + } + }, + { + "ph": "s", "id": 79938, "pid": 435, "tid": 435, "ts": 1776649792650088, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650100, "dur": 0, + "args": { + "External id": 79944, + "cbid": 200, "correlation": 79944 + } + }, + { + "ph": "f", "id": 79944, "pid": 435, "tid": 435, "ts": 1776649792650100, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650101, "dur": 0, + "args": { + "External id": 79945, + "cbid": 273, "correlation": 79945 + } + }, + { + "ph": "f", "id": 79945, "pid": 435, "tid": 435, "ts": 1776649792650101, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792650110, "dur": 19, + "args": { + "External id": 79946, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79946, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79946, "pid": 0, "tid": 7, "ts": 1776649792650110, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650102, "dur": 5, + "args": { + "External id": 79946, + "cbid": 211, "correlation": 79946 + } + }, + { + "ph": "s", "id": 79946, "pid": 435, "tid": 435, "ts": 1776649792650102, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792650146, "dur": 1, + "args": { + "External id": 10406, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79964, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79964, "pid": 0, "tid": 7, "ts": 1776649792650146, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650136, "dur": 7, + "args": { + "External id": 10406, + "cbid": 211, "correlation": 79964 + } + }, + { + "ph": "s", "id": 79964, "pid": 435, "tid": 435, "ts": 1776649792650136, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650149, "dur": 0, + "args": { + "External id": 79970, + "cbid": 200, "correlation": 79970 + } + }, + { + "ph": "f", "id": 79970, "pid": 435, "tid": 435, "ts": 1776649792650149, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650150, "dur": 0, + "args": { + "External id": 79971, + "cbid": 273, "correlation": 79971 + } + }, + { + "ph": "f", "id": 79971, "pid": 435, "tid": 435, "ts": 1776649792650150, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792650159, "dur": 19, + "args": { + "External id": 79972, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79972, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79972, "pid": 0, "tid": 7, "ts": 1776649792650159, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650151, "dur": 5, + "args": { + "External id": 79972, + "cbid": 211, "correlation": 79972 + } + }, + { + "ph": "s", "id": 79972, "pid": 435, "tid": 435, "ts": 1776649792650151, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792650195, "dur": 1, + "args": { + "External id": 10411, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79990, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79990, "pid": 0, "tid": 7, "ts": 1776649792650195, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650186, "dur": 7, + "args": { + "External id": 10411, + "cbid": 211, "correlation": 79990 + } + }, + { + "ph": "s", "id": 79990, "pid": 435, "tid": 435, "ts": 1776649792650186, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650199, "dur": 0, + "args": { + "External id": 79996, + "cbid": 200, "correlation": 79996 + } + }, + { + "ph": "f", "id": 79996, "pid": 435, "tid": 435, "ts": 1776649792650199, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650199, "dur": 0, + "args": { + "External id": 79997, + "cbid": 273, "correlation": 79997 + } + }, + { + "ph": "f", "id": 79997, "pid": 435, "tid": 435, "ts": 1776649792650199, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792650209, "dur": 19, + "args": { + "External id": 79998, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 79998, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 79998, "pid": 0, "tid": 7, "ts": 1776649792650209, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650201, "dur": 4, + "args": { + "External id": 79998, + "cbid": 211, "correlation": 79998 + } + }, + { + "ph": "s", "id": 79998, "pid": 435, "tid": 435, "ts": 1776649792650201, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792650244, "dur": 1, + "args": { + "External id": 10416, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80016, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80016, "pid": 0, "tid": 7, "ts": 1776649792650244, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650235, "dur": 6, + "args": { + "External id": 10416, + "cbid": 211, "correlation": 80016 + } + }, + { + "ph": "s", "id": 80016, "pid": 435, "tid": 435, "ts": 1776649792650235, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650247, "dur": 0, + "args": { + "External id": 80022, + "cbid": 200, "correlation": 80022 + } + }, + { + "ph": "f", "id": 80022, "pid": 435, "tid": 435, "ts": 1776649792650247, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650248, "dur": 0, + "args": { + "External id": 80023, + "cbid": 273, "correlation": 80023 + } + }, + { + "ph": "f", "id": 80023, "pid": 435, "tid": 435, "ts": 1776649792650248, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792650257, "dur": 19, + "args": { + "External id": 80024, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80024, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80024, "pid": 0, "tid": 7, "ts": 1776649792650257, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650249, "dur": 4, + "args": { + "External id": 80024, + "cbid": 211, "correlation": 80024 + } + }, + { + "ph": "s", "id": 80024, "pid": 435, "tid": 435, "ts": 1776649792650249, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792650293, "dur": 1, + "args": { + "External id": 10421, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80042, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80042, "pid": 0, "tid": 7, "ts": 1776649792650293, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650283, "dur": 6, + "args": { + "External id": 10421, + "cbid": 211, "correlation": 80042 + } + }, + { + "ph": "s", "id": 80042, "pid": 435, "tid": 435, "ts": 1776649792650283, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650296, "dur": 0, + "args": { + "External id": 80048, + "cbid": 200, "correlation": 80048 + } + }, + { + "ph": "f", "id": 80048, "pid": 435, "tid": 435, "ts": 1776649792650296, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650297, "dur": 0, + "args": { + "External id": 80049, + "cbid": 273, "correlation": 80049 + } + }, + { + "ph": "f", "id": 80049, "pid": 435, "tid": 435, "ts": 1776649792650297, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792650306, "dur": 19, + "args": { + "External id": 80050, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80050, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80050, "pid": 0, "tid": 7, "ts": 1776649792650306, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650298, "dur": 4, + "args": { + "External id": 80050, + "cbid": 211, "correlation": 80050 + } + }, + { + "ph": "s", "id": 80050, "pid": 435, "tid": 435, "ts": 1776649792650298, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792650341, "dur": 1, + "args": { + "External id": 10426, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80068, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80068, "pid": 0, "tid": 7, "ts": 1776649792650341, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650332, "dur": 6, + "args": { + "External id": 10426, + "cbid": 211, "correlation": 80068 + } + }, + { + "ph": "s", "id": 80068, "pid": 435, "tid": 435, "ts": 1776649792650332, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650345, "dur": 0, + "args": { + "External id": 80074, + "cbid": 200, "correlation": 80074 + } + }, + { + "ph": "f", "id": 80074, "pid": 435, "tid": 435, "ts": 1776649792650345, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650345, "dur": 0, + "args": { + "External id": 80075, + "cbid": 273, "correlation": 80075 + } + }, + { + "ph": "f", "id": 80075, "pid": 435, "tid": 435, "ts": 1776649792650345, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792650354, "dur": 19, + "args": { + "External id": 80076, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80076, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80076, "pid": 0, "tid": 7, "ts": 1776649792650354, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650346, "dur": 4, + "args": { + "External id": 80076, + "cbid": 211, "correlation": 80076 + } + }, + { + "ph": "s", "id": 80076, "pid": 435, "tid": 435, "ts": 1776649792650346, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792650390, "dur": 1, + "args": { + "External id": 10431, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80094, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80094, "pid": 0, "tid": 7, "ts": 1776649792650390, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650381, "dur": 6, + "args": { + "External id": 10431, + "cbid": 211, "correlation": 80094 + } + }, + { + "ph": "s", "id": 80094, "pid": 435, "tid": 435, "ts": 1776649792650381, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650393, "dur": 0, + "args": { + "External id": 80100, + "cbid": 200, "correlation": 80100 + } + }, + { + "ph": "f", "id": 80100, "pid": 435, "tid": 435, "ts": 1776649792650393, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650394, "dur": 0, + "args": { + "External id": 80101, + "cbid": 273, "correlation": 80101 + } + }, + { + "ph": "f", "id": 80101, "pid": 435, "tid": 435, "ts": 1776649792650394, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792650403, "dur": 19, + "args": { + "External id": 80102, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80102, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80102, "pid": 0, "tid": 7, "ts": 1776649792650403, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650395, "dur": 5, + "args": { + "External id": 80102, + "cbid": 211, "correlation": 80102 + } + }, + { + "ph": "s", "id": 80102, "pid": 435, "tid": 435, "ts": 1776649792650395, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792650438, "dur": 10, + "args": { + "External id": 80105, + "cbid": 165, "correlation": 80105 + } + }, + { + "ph": "s", "id": 80105, "pid": 435, "tid": 435, "ts": 1776649792650438, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649792650494, "dur": 8, + "args": { + "External id": 80114, + "cbid": 135, "correlation": 80114 + } + }, + { + "ph": "f", "id": 80114, "pid": 435, "tid": 435, "ts": 1776649792650494, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792650553, "dur": 1, + "args": { + "External id": 10436, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80134, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80134, "pid": 0, "tid": 7, "ts": 1776649792650553, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650541, "dur": 9, + "args": { + "External id": 10436, + "cbid": 211, "correlation": 80134 + } + }, + { + "ph": "s", "id": 80134, "pid": 435, "tid": 435, "ts": 1776649792650541, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650558, "dur": 0, + "args": { + "External id": 80140, + "cbid": 200, "correlation": 80140 + } + }, + { + "ph": "f", "id": 80140, "pid": 435, "tid": 435, "ts": 1776649792650558, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650559, "dur": 0, + "args": { + "External id": 80141, + "cbid": 273, "correlation": 80141 + } + }, + { + "ph": "f", "id": 80141, "pid": 435, "tid": 435, "ts": 1776649792650559, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792650568, "dur": 19, + "args": { + "External id": 80142, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80142, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80142, "pid": 0, "tid": 7, "ts": 1776649792650568, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650560, "dur": 5, + "args": { + "External id": 80142, + "cbid": 211, "correlation": 80142 + } + }, + { + "ph": "s", "id": 80142, "pid": 435, "tid": 435, "ts": 1776649792650560, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792650606, "dur": 1, + "args": { + "External id": 10441, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80160, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80160, "pid": 0, "tid": 7, "ts": 1776649792650606, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650597, "dur": 6, + "args": { + "External id": 10441, + "cbid": 211, "correlation": 80160 + } + }, + { + "ph": "s", "id": 80160, "pid": 435, "tid": 435, "ts": 1776649792650597, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650610, "dur": 0, + "args": { + "External id": 80166, + "cbid": 200, "correlation": 80166 + } + }, + { + "ph": "f", "id": 80166, "pid": 435, "tid": 435, "ts": 1776649792650610, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650610, "dur": 0, + "args": { + "External id": 80167, + "cbid": 273, "correlation": 80167 + } + }, + { + "ph": "f", "id": 80167, "pid": 435, "tid": 435, "ts": 1776649792650610, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792650620, "dur": 19, + "args": { + "External id": 80168, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80168, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80168, "pid": 0, "tid": 7, "ts": 1776649792650620, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650612, "dur": 5, + "args": { + "External id": 80168, + "cbid": 211, "correlation": 80168 + } + }, + { + "ph": "s", "id": 80168, "pid": 435, "tid": 435, "ts": 1776649792650612, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792650656, "dur": 1, + "args": { + "External id": 10446, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80186, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80186, "pid": 0, "tid": 7, "ts": 1776649792650656, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650647, "dur": 6, + "args": { + "External id": 10446, + "cbid": 211, "correlation": 80186 + } + }, + { + "ph": "s", "id": 80186, "pid": 435, "tid": 435, "ts": 1776649792650647, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650659, "dur": 0, + "args": { + "External id": 80192, + "cbid": 200, "correlation": 80192 + } + }, + { + "ph": "f", "id": 80192, "pid": 435, "tid": 435, "ts": 1776649792650659, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650660, "dur": 0, + "args": { + "External id": 80193, + "cbid": 273, "correlation": 80193 + } + }, + { + "ph": "f", "id": 80193, "pid": 435, "tid": 435, "ts": 1776649792650660, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792650669, "dur": 19, + "args": { + "External id": 80194, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80194, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80194, "pid": 0, "tid": 7, "ts": 1776649792650669, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650661, "dur": 5, + "args": { + "External id": 80194, + "cbid": 211, "correlation": 80194 + } + }, + { + "ph": "s", "id": 80194, "pid": 435, "tid": 435, "ts": 1776649792650661, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792650707, "dur": 1, + "args": { + "External id": 10451, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80212, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80212, "pid": 0, "tid": 7, "ts": 1776649792650707, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650697, "dur": 6, + "args": { + "External id": 10451, + "cbid": 211, "correlation": 80212 + } + }, + { + "ph": "s", "id": 80212, "pid": 435, "tid": 435, "ts": 1776649792650697, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650710, "dur": 0, + "args": { + "External id": 80218, + "cbid": 200, "correlation": 80218 + } + }, + { + "ph": "f", "id": 80218, "pid": 435, "tid": 435, "ts": 1776649792650710, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650711, "dur": 0, + "args": { + "External id": 80219, + "cbid": 273, "correlation": 80219 + } + }, + { + "ph": "f", "id": 80219, "pid": 435, "tid": 435, "ts": 1776649792650711, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792650720, "dur": 19, + "args": { + "External id": 80220, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80220, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80220, "pid": 0, "tid": 7, "ts": 1776649792650720, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650712, "dur": 5, + "args": { + "External id": 80220, + "cbid": 211, "correlation": 80220 + } + }, + { + "ph": "s", "id": 80220, "pid": 435, "tid": 435, "ts": 1776649792650712, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792650755, "dur": 1, + "args": { + "External id": 10456, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80238, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80238, "pid": 0, "tid": 7, "ts": 1776649792650755, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650746, "dur": 7, + "args": { + "External id": 10456, + "cbid": 211, "correlation": 80238 + } + }, + { + "ph": "s", "id": 80238, "pid": 435, "tid": 435, "ts": 1776649792650746, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650759, "dur": 0, + "args": { + "External id": 80244, + "cbid": 200, "correlation": 80244 + } + }, + { + "ph": "f", "id": 80244, "pid": 435, "tid": 435, "ts": 1776649792650759, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650759, "dur": 0, + "args": { + "External id": 80245, + "cbid": 273, "correlation": 80245 + } + }, + { + "ph": "f", "id": 80245, "pid": 435, "tid": 435, "ts": 1776649792650759, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792650768, "dur": 19, + "args": { + "External id": 80246, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80246, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80246, "pid": 0, "tid": 7, "ts": 1776649792650768, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650760, "dur": 5, + "args": { + "External id": 80246, + "cbid": 211, "correlation": 80246 + } + }, + { + "ph": "s", "id": 80246, "pid": 435, "tid": 435, "ts": 1776649792650760, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792650804, "dur": 1, + "args": { + "External id": 10461, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80264, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80264, "pid": 0, "tid": 7, "ts": 1776649792650804, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650795, "dur": 6, + "args": { + "External id": 10461, + "cbid": 211, "correlation": 80264 + } + }, + { + "ph": "s", "id": 80264, "pid": 435, "tid": 435, "ts": 1776649792650795, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650808, "dur": 0, + "args": { + "External id": 80270, + "cbid": 200, "correlation": 80270 + } + }, + { + "ph": "f", "id": 80270, "pid": 435, "tid": 435, "ts": 1776649792650808, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650808, "dur": 0, + "args": { + "External id": 80271, + "cbid": 273, "correlation": 80271 + } + }, + { + "ph": "f", "id": 80271, "pid": 435, "tid": 435, "ts": 1776649792650808, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792650817, "dur": 19, + "args": { + "External id": 80272, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80272, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80272, "pid": 0, "tid": 7, "ts": 1776649792650817, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650809, "dur": 5, + "args": { + "External id": 80272, + "cbid": 211, "correlation": 80272 + } + }, + { + "ph": "s", "id": 80272, "pid": 435, "tid": 435, "ts": 1776649792650809, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792650853, "dur": 1, + "args": { + "External id": 10466, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80290, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80290, "pid": 0, "tid": 7, "ts": 1776649792650853, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650844, "dur": 6, + "args": { + "External id": 10466, + "cbid": 211, "correlation": 80290 + } + }, + { + "ph": "s", "id": 80290, "pid": 435, "tid": 435, "ts": 1776649792650844, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650856, "dur": 0, + "args": { + "External id": 80296, + "cbid": 200, "correlation": 80296 + } + }, + { + "ph": "f", "id": 80296, "pid": 435, "tid": 435, "ts": 1776649792650856, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650857, "dur": 0, + "args": { + "External id": 80297, + "cbid": 273, "correlation": 80297 + } + }, + { + "ph": "f", "id": 80297, "pid": 435, "tid": 435, "ts": 1776649792650857, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792650866, "dur": 19, + "args": { + "External id": 80298, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80298, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80298, "pid": 0, "tid": 7, "ts": 1776649792650866, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650858, "dur": 4, + "args": { + "External id": 80298, + "cbid": 211, "correlation": 80298 + } + }, + { + "ph": "s", "id": 80298, "pid": 435, "tid": 435, "ts": 1776649792650858, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792650902, "dur": 1, + "args": { + "External id": 10471, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80316, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80316, "pid": 0, "tid": 7, "ts": 1776649792650902, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650892, "dur": 6, + "args": { + "External id": 10471, + "cbid": 211, "correlation": 80316 + } + }, + { + "ph": "s", "id": 80316, "pid": 435, "tid": 435, "ts": 1776649792650892, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650905, "dur": 0, + "args": { + "External id": 80322, + "cbid": 200, "correlation": 80322 + } + }, + { + "ph": "f", "id": 80322, "pid": 435, "tid": 435, "ts": 1776649792650905, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650906, "dur": 0, + "args": { + "External id": 80323, + "cbid": 273, "correlation": 80323 + } + }, + { + "ph": "f", "id": 80323, "pid": 435, "tid": 435, "ts": 1776649792650906, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792650915, "dur": 19, + "args": { + "External id": 80324, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80324, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80324, "pid": 0, "tid": 7, "ts": 1776649792650915, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650907, "dur": 5, + "args": { + "External id": 80324, + "cbid": 211, "correlation": 80324 + } + }, + { + "ph": "s", "id": 80324, "pid": 435, "tid": 435, "ts": 1776649792650907, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792650951, "dur": 1, + "args": { + "External id": 10476, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80342, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80342, "pid": 0, "tid": 7, "ts": 1776649792650951, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650941, "dur": 7, + "args": { + "External id": 10476, + "cbid": 211, "correlation": 80342 + } + }, + { + "ph": "s", "id": 80342, "pid": 435, "tid": 435, "ts": 1776649792650941, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650954, "dur": 0, + "args": { + "External id": 80348, + "cbid": 200, "correlation": 80348 + } + }, + { + "ph": "f", "id": 80348, "pid": 435, "tid": 435, "ts": 1776649792650954, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792650955, "dur": 0, + "args": { + "External id": 80349, + "cbid": 273, "correlation": 80349 + } + }, + { + "ph": "f", "id": 80349, "pid": 435, "tid": 435, "ts": 1776649792650955, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792650964, "dur": 19, + "args": { + "External id": 80350, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80350, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80350, "pid": 0, "tid": 7, "ts": 1776649792650964, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650956, "dur": 5, + "args": { + "External id": 80350, + "cbid": 211, "correlation": 80350 + } + }, + { + "ph": "s", "id": 80350, "pid": 435, "tid": 435, "ts": 1776649792650956, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792650999, "dur": 1, + "args": { + "External id": 10481, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80368, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80368, "pid": 0, "tid": 7, "ts": 1776649792650999, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792650990, "dur": 6, + "args": { + "External id": 10481, + "cbid": 211, "correlation": 80368 + } + }, + { + "ph": "s", "id": 80368, "pid": 435, "tid": 435, "ts": 1776649792650990, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651002, "dur": 0, + "args": { + "External id": 80374, + "cbid": 200, "correlation": 80374 + } + }, + { + "ph": "f", "id": 80374, "pid": 435, "tid": 435, "ts": 1776649792651002, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651003, "dur": 0, + "args": { + "External id": 80375, + "cbid": 273, "correlation": 80375 + } + }, + { + "ph": "f", "id": 80375, "pid": 435, "tid": 435, "ts": 1776649792651003, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792651012, "dur": 19, + "args": { + "External id": 80376, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80376, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80376, "pid": 0, "tid": 7, "ts": 1776649792651012, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651004, "dur": 5, + "args": { + "External id": 80376, + "cbid": 211, "correlation": 80376 + } + }, + { + "ph": "s", "id": 80376, "pid": 435, "tid": 435, "ts": 1776649792651004, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792651047, "dur": 1, + "args": { + "External id": 10486, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80394, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80394, "pid": 0, "tid": 7, "ts": 1776649792651047, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651038, "dur": 6, + "args": { + "External id": 10486, + "cbid": 211, "correlation": 80394 + } + }, + { + "ph": "s", "id": 80394, "pid": 435, "tid": 435, "ts": 1776649792651038, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651051, "dur": 0, + "args": { + "External id": 80400, + "cbid": 200, "correlation": 80400 + } + }, + { + "ph": "f", "id": 80400, "pid": 435, "tid": 435, "ts": 1776649792651051, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651051, "dur": 0, + "args": { + "External id": 80401, + "cbid": 273, "correlation": 80401 + } + }, + { + "ph": "f", "id": 80401, "pid": 435, "tid": 435, "ts": 1776649792651051, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792651060, "dur": 19, + "args": { + "External id": 80402, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80402, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80402, "pid": 0, "tid": 7, "ts": 1776649792651060, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651052, "dur": 4, + "args": { + "External id": 80402, + "cbid": 211, "correlation": 80402 + } + }, + { + "ph": "s", "id": 80402, "pid": 435, "tid": 435, "ts": 1776649792651052, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792651096, "dur": 1, + "args": { + "External id": 10491, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80420, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80420, "pid": 0, "tid": 7, "ts": 1776649792651096, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651087, "dur": 6, + "args": { + "External id": 10491, + "cbid": 211, "correlation": 80420 + } + }, + { + "ph": "s", "id": 80420, "pid": 435, "tid": 435, "ts": 1776649792651087, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651099, "dur": 0, + "args": { + "External id": 80426, + "cbid": 200, "correlation": 80426 + } + }, + { + "ph": "f", "id": 80426, "pid": 435, "tid": 435, "ts": 1776649792651099, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651100, "dur": 0, + "args": { + "External id": 80427, + "cbid": 273, "correlation": 80427 + } + }, + { + "ph": "f", "id": 80427, "pid": 435, "tid": 435, "ts": 1776649792651100, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792651109, "dur": 19, + "args": { + "External id": 80428, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80428, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80428, "pid": 0, "tid": 7, "ts": 1776649792651109, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651101, "dur": 4, + "args": { + "External id": 80428, + "cbid": 211, "correlation": 80428 + } + }, + { + "ph": "s", "id": 80428, "pid": 435, "tid": 435, "ts": 1776649792651101, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792651144, "dur": 1, + "args": { + "External id": 10496, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80446, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80446, "pid": 0, "tid": 7, "ts": 1776649792651144, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651135, "dur": 6, + "args": { + "External id": 10496, + "cbid": 211, "correlation": 80446 + } + }, + { + "ph": "s", "id": 80446, "pid": 435, "tid": 435, "ts": 1776649792651135, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651147, "dur": 0, + "args": { + "External id": 80452, + "cbid": 200, "correlation": 80452 + } + }, + { + "ph": "f", "id": 80452, "pid": 435, "tid": 435, "ts": 1776649792651147, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651148, "dur": 0, + "args": { + "External id": 80453, + "cbid": 273, "correlation": 80453 + } + }, + { + "ph": "f", "id": 80453, "pid": 435, "tid": 435, "ts": 1776649792651148, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792651157, "dur": 19, + "args": { + "External id": 80454, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80454, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80454, "pid": 0, "tid": 7, "ts": 1776649792651157, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651149, "dur": 4, + "args": { + "External id": 80454, + "cbid": 211, "correlation": 80454 + } + }, + { + "ph": "s", "id": 80454, "pid": 435, "tid": 435, "ts": 1776649792651149, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792651193, "dur": 1, + "args": { + "External id": 10501, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80472, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80472, "pid": 0, "tid": 7, "ts": 1776649792651193, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651184, "dur": 6, + "args": { + "External id": 10501, + "cbid": 211, "correlation": 80472 + } + }, + { + "ph": "s", "id": 80472, "pid": 435, "tid": 435, "ts": 1776649792651184, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651196, "dur": 0, + "args": { + "External id": 80478, + "cbid": 200, "correlation": 80478 + } + }, + { + "ph": "f", "id": 80478, "pid": 435, "tid": 435, "ts": 1776649792651196, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651197, "dur": 0, + "args": { + "External id": 80479, + "cbid": 273, "correlation": 80479 + } + }, + { + "ph": "f", "id": 80479, "pid": 435, "tid": 435, "ts": 1776649792651197, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792651206, "dur": 19, + "args": { + "External id": 80480, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80480, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80480, "pid": 0, "tid": 7, "ts": 1776649792651206, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651198, "dur": 4, + "args": { + "External id": 80480, + "cbid": 211, "correlation": 80480 + } + }, + { + "ph": "s", "id": 80480, "pid": 435, "tid": 435, "ts": 1776649792651198, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792651241, "dur": 1, + "args": { + "External id": 10506, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80498, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80498, "pid": 0, "tid": 7, "ts": 1776649792651241, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651232, "dur": 6, + "args": { + "External id": 10506, + "cbid": 211, "correlation": 80498 + } + }, + { + "ph": "s", "id": 80498, "pid": 435, "tid": 435, "ts": 1776649792651232, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651244, "dur": 0, + "args": { + "External id": 80504, + "cbid": 200, "correlation": 80504 + } + }, + { + "ph": "f", "id": 80504, "pid": 435, "tid": 435, "ts": 1776649792651244, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651245, "dur": 0, + "args": { + "External id": 80505, + "cbid": 273, "correlation": 80505 + } + }, + { + "ph": "f", "id": 80505, "pid": 435, "tid": 435, "ts": 1776649792651245, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792651254, "dur": 19, + "args": { + "External id": 80506, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80506, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80506, "pid": 0, "tid": 7, "ts": 1776649792651254, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651246, "dur": 4, + "args": { + "External id": 80506, + "cbid": 211, "correlation": 80506 + } + }, + { + "ph": "s", "id": 80506, "pid": 435, "tid": 435, "ts": 1776649792651246, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792651289, "dur": 1, + "args": { + "External id": 10511, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80524, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80524, "pid": 0, "tid": 7, "ts": 1776649792651289, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651280, "dur": 6, + "args": { + "External id": 10511, + "cbid": 211, "correlation": 80524 + } + }, + { + "ph": "s", "id": 80524, "pid": 435, "tid": 435, "ts": 1776649792651280, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651293, "dur": 0, + "args": { + "External id": 80530, + "cbid": 200, "correlation": 80530 + } + }, + { + "ph": "f", "id": 80530, "pid": 435, "tid": 435, "ts": 1776649792651293, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651293, "dur": 0, + "args": { + "External id": 80531, + "cbid": 273, "correlation": 80531 + } + }, + { + "ph": "f", "id": 80531, "pid": 435, "tid": 435, "ts": 1776649792651293, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792651302, "dur": 19, + "args": { + "External id": 80532, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80532, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80532, "pid": 0, "tid": 7, "ts": 1776649792651302, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651294, "dur": 5, + "args": { + "External id": 80532, + "cbid": 211, "correlation": 80532 + } + }, + { + "ph": "s", "id": 80532, "pid": 435, "tid": 435, "ts": 1776649792651294, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792651338, "dur": 1, + "args": { + "External id": 10516, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80550, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80550, "pid": 0, "tid": 7, "ts": 1776649792651338, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651328, "dur": 7, + "args": { + "External id": 10516, + "cbid": 211, "correlation": 80550 + } + }, + { + "ph": "s", "id": 80550, "pid": 435, "tid": 435, "ts": 1776649792651328, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651343, "dur": 0, + "args": { + "External id": 80556, + "cbid": 200, "correlation": 80556 + } + }, + { + "ph": "f", "id": 80556, "pid": 435, "tid": 435, "ts": 1776649792651343, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651344, "dur": 0, + "args": { + "External id": 80557, + "cbid": 273, "correlation": 80557 + } + }, + { + "ph": "f", "id": 80557, "pid": 435, "tid": 435, "ts": 1776649792651344, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792651353, "dur": 19, + "args": { + "External id": 80558, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80558, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80558, "pid": 0, "tid": 7, "ts": 1776649792651353, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651345, "dur": 5, + "args": { + "External id": 80558, + "cbid": 211, "correlation": 80558 + } + }, + { + "ph": "s", "id": 80558, "pid": 435, "tid": 435, "ts": 1776649792651345, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792651390, "dur": 1, + "args": { + "External id": 10521, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80576, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80576, "pid": 0, "tid": 7, "ts": 1776649792651390, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651381, "dur": 6, + "args": { + "External id": 10521, + "cbid": 211, "correlation": 80576 + } + }, + { + "ph": "s", "id": 80576, "pid": 435, "tid": 435, "ts": 1776649792651381, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651393, "dur": 0, + "args": { + "External id": 80582, + "cbid": 200, "correlation": 80582 + } + }, + { + "ph": "f", "id": 80582, "pid": 435, "tid": 435, "ts": 1776649792651393, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651394, "dur": 0, + "args": { + "External id": 80583, + "cbid": 273, "correlation": 80583 + } + }, + { + "ph": "f", "id": 80583, "pid": 435, "tid": 435, "ts": 1776649792651394, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792651403, "dur": 19, + "args": { + "External id": 80584, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80584, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80584, "pid": 0, "tid": 7, "ts": 1776649792651403, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651395, "dur": 4, + "args": { + "External id": 80584, + "cbid": 211, "correlation": 80584 + } + }, + { + "ph": "s", "id": 80584, "pid": 435, "tid": 435, "ts": 1776649792651395, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792651438, "dur": 1, + "args": { + "External id": 10526, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80602, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80602, "pid": 0, "tid": 7, "ts": 1776649792651438, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651429, "dur": 6, + "args": { + "External id": 10526, + "cbid": 211, "correlation": 80602 + } + }, + { + "ph": "s", "id": 80602, "pid": 435, "tid": 435, "ts": 1776649792651429, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651441, "dur": 0, + "args": { + "External id": 80608, + "cbid": 200, "correlation": 80608 + } + }, + { + "ph": "f", "id": 80608, "pid": 435, "tid": 435, "ts": 1776649792651441, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651442, "dur": 0, + "args": { + "External id": 80609, + "cbid": 273, "correlation": 80609 + } + }, + { + "ph": "f", "id": 80609, "pid": 435, "tid": 435, "ts": 1776649792651442, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792651451, "dur": 19, + "args": { + "External id": 80610, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80610, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80610, "pid": 0, "tid": 7, "ts": 1776649792651451, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651443, "dur": 4, + "args": { + "External id": 80610, + "cbid": 211, "correlation": 80610 + } + }, + { + "ph": "s", "id": 80610, "pid": 435, "tid": 435, "ts": 1776649792651443, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792651486, "dur": 1, + "args": { + "External id": 10531, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80628, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80628, "pid": 0, "tid": 7, "ts": 1776649792651486, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651477, "dur": 6, + "args": { + "External id": 10531, + "cbid": 211, "correlation": 80628 + } + }, + { + "ph": "s", "id": 80628, "pid": 435, "tid": 435, "ts": 1776649792651477, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651489, "dur": 0, + "args": { + "External id": 80634, + "cbid": 200, "correlation": 80634 + } + }, + { + "ph": "f", "id": 80634, "pid": 435, "tid": 435, "ts": 1776649792651489, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651489, "dur": 0, + "args": { + "External id": 80635, + "cbid": 273, "correlation": 80635 + } + }, + { + "ph": "f", "id": 80635, "pid": 435, "tid": 435, "ts": 1776649792651489, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792651498, "dur": 19, + "args": { + "External id": 80636, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80636, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80636, "pid": 0, "tid": 7, "ts": 1776649792651498, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651491, "dur": 4, + "args": { + "External id": 80636, + "cbid": 211, "correlation": 80636 + } + }, + { + "ph": "s", "id": 80636, "pid": 435, "tid": 435, "ts": 1776649792651491, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792651534, "dur": 1, + "args": { + "External id": 10536, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80654, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80654, "pid": 0, "tid": 7, "ts": 1776649792651534, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651524, "dur": 6, + "args": { + "External id": 10536, + "cbid": 211, "correlation": 80654 + } + }, + { + "ph": "s", "id": 80654, "pid": 435, "tid": 435, "ts": 1776649792651524, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651537, "dur": 0, + "args": { + "External id": 80660, + "cbid": 200, "correlation": 80660 + } + }, + { + "ph": "f", "id": 80660, "pid": 435, "tid": 435, "ts": 1776649792651537, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651537, "dur": 0, + "args": { + "External id": 80661, + "cbid": 273, "correlation": 80661 + } + }, + { + "ph": "f", "id": 80661, "pid": 435, "tid": 435, "ts": 1776649792651537, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792651547, "dur": 19, + "args": { + "External id": 80662, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80662, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80662, "pid": 0, "tid": 7, "ts": 1776649792651547, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651538, "dur": 5, + "args": { + "External id": 80662, + "cbid": 211, "correlation": 80662 + } + }, + { + "ph": "s", "id": 80662, "pid": 435, "tid": 435, "ts": 1776649792651538, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792651582, "dur": 1, + "args": { + "External id": 10541, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80680, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80680, "pid": 0, "tid": 7, "ts": 1776649792651582, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651573, "dur": 6, + "args": { + "External id": 10541, + "cbid": 211, "correlation": 80680 + } + }, + { + "ph": "s", "id": 80680, "pid": 435, "tid": 435, "ts": 1776649792651573, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651585, "dur": 0, + "args": { + "External id": 80686, + "cbid": 200, "correlation": 80686 + } + }, + { + "ph": "f", "id": 80686, "pid": 435, "tid": 435, "ts": 1776649792651585, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651586, "dur": 0, + "args": { + "External id": 80687, + "cbid": 273, "correlation": 80687 + } + }, + { + "ph": "f", "id": 80687, "pid": 435, "tid": 435, "ts": 1776649792651586, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792651595, "dur": 19, + "args": { + "External id": 80688, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80688, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80688, "pid": 0, "tid": 7, "ts": 1776649792651595, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651587, "dur": 4, + "args": { + "External id": 80688, + "cbid": 211, "correlation": 80688 + } + }, + { + "ph": "s", "id": 80688, "pid": 435, "tid": 435, "ts": 1776649792651587, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792651630, "dur": 1, + "args": { + "External id": 10546, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80706, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80706, "pid": 0, "tid": 7, "ts": 1776649792651630, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651621, "dur": 6, + "args": { + "External id": 10546, + "cbid": 211, "correlation": 80706 + } + }, + { + "ph": "s", "id": 80706, "pid": 435, "tid": 435, "ts": 1776649792651621, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651634, "dur": 0, + "args": { + "External id": 80712, + "cbid": 200, "correlation": 80712 + } + }, + { + "ph": "f", "id": 80712, "pid": 435, "tid": 435, "ts": 1776649792651634, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651634, "dur": 0, + "args": { + "External id": 80713, + "cbid": 273, "correlation": 80713 + } + }, + { + "ph": "f", "id": 80713, "pid": 435, "tid": 435, "ts": 1776649792651634, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792651643, "dur": 19, + "args": { + "External id": 80714, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80714, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80714, "pid": 0, "tid": 7, "ts": 1776649792651643, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651635, "dur": 4, + "args": { + "External id": 80714, + "cbid": 211, "correlation": 80714 + } + }, + { + "ph": "s", "id": 80714, "pid": 435, "tid": 435, "ts": 1776649792651635, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792651679, "dur": 1, + "args": { + "External id": 10551, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80732, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80732, "pid": 0, "tid": 7, "ts": 1776649792651679, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651670, "dur": 6, + "args": { + "External id": 10551, + "cbid": 211, "correlation": 80732 + } + }, + { + "ph": "s", "id": 80732, "pid": 435, "tid": 435, "ts": 1776649792651670, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651682, "dur": 0, + "args": { + "External id": 80738, + "cbid": 200, "correlation": 80738 + } + }, + { + "ph": "f", "id": 80738, "pid": 435, "tid": 435, "ts": 1776649792651682, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651683, "dur": 0, + "args": { + "External id": 80739, + "cbid": 273, "correlation": 80739 + } + }, + { + "ph": "f", "id": 80739, "pid": 435, "tid": 435, "ts": 1776649792651683, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792651692, "dur": 19, + "args": { + "External id": 80740, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80740, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80740, "pid": 0, "tid": 7, "ts": 1776649792651692, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651684, "dur": 4, + "args": { + "External id": 80740, + "cbid": 211, "correlation": 80740 + } + }, + { + "ph": "s", "id": 80740, "pid": 435, "tid": 435, "ts": 1776649792651684, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792651727, "dur": 1, + "args": { + "External id": 10556, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80758, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80758, "pid": 0, "tid": 7, "ts": 1776649792651727, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651718, "dur": 6, + "args": { + "External id": 10556, + "cbid": 211, "correlation": 80758 + } + }, + { + "ph": "s", "id": 80758, "pid": 435, "tid": 435, "ts": 1776649792651718, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651730, "dur": 0, + "args": { + "External id": 80764, + "cbid": 200, "correlation": 80764 + } + }, + { + "ph": "f", "id": 80764, "pid": 435, "tid": 435, "ts": 1776649792651730, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651731, "dur": 0, + "args": { + "External id": 80765, + "cbid": 273, "correlation": 80765 + } + }, + { + "ph": "f", "id": 80765, "pid": 435, "tid": 435, "ts": 1776649792651731, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792651740, "dur": 19, + "args": { + "External id": 80766, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80766, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80766, "pid": 0, "tid": 7, "ts": 1776649792651740, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651732, "dur": 5, + "args": { + "External id": 80766, + "cbid": 211, "correlation": 80766 + } + }, + { + "ph": "s", "id": 80766, "pid": 435, "tid": 435, "ts": 1776649792651732, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792651775, "dur": 1, + "args": { + "External id": 10561, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80784, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80784, "pid": 0, "tid": 7, "ts": 1776649792651775, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651766, "dur": 6, + "args": { + "External id": 10561, + "cbid": 211, "correlation": 80784 + } + }, + { + "ph": "s", "id": 80784, "pid": 435, "tid": 435, "ts": 1776649792651766, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651778, "dur": 0, + "args": { + "External id": 80790, + "cbid": 200, "correlation": 80790 + } + }, + { + "ph": "f", "id": 80790, "pid": 435, "tid": 435, "ts": 1776649792651778, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651779, "dur": 0, + "args": { + "External id": 80791, + "cbid": 273, "correlation": 80791 + } + }, + { + "ph": "f", "id": 80791, "pid": 435, "tid": 435, "ts": 1776649792651779, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792651788, "dur": 19, + "args": { + "External id": 80792, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80792, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80792, "pid": 0, "tid": 7, "ts": 1776649792651788, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651780, "dur": 4, + "args": { + "External id": 80792, + "cbid": 211, "correlation": 80792 + } + }, + { + "ph": "s", "id": 80792, "pid": 435, "tid": 435, "ts": 1776649792651780, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792651823, "dur": 1, + "args": { + "External id": 10566, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80810, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80810, "pid": 0, "tid": 7, "ts": 1776649792651823, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651814, "dur": 6, + "args": { + "External id": 10566, + "cbid": 211, "correlation": 80810 + } + }, + { + "ph": "s", "id": 80810, "pid": 435, "tid": 435, "ts": 1776649792651814, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651827, "dur": 0, + "args": { + "External id": 80816, + "cbid": 200, "correlation": 80816 + } + }, + { + "ph": "f", "id": 80816, "pid": 435, "tid": 435, "ts": 1776649792651827, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651827, "dur": 0, + "args": { + "External id": 80817, + "cbid": 273, "correlation": 80817 + } + }, + { + "ph": "f", "id": 80817, "pid": 435, "tid": 435, "ts": 1776649792651827, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792651836, "dur": 19, + "args": { + "External id": 80818, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80818, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80818, "pid": 0, "tid": 7, "ts": 1776649792651836, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651828, "dur": 4, + "args": { + "External id": 80818, + "cbid": 211, "correlation": 80818 + } + }, + { + "ph": "s", "id": 80818, "pid": 435, "tid": 435, "ts": 1776649792651828, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792651871, "dur": 1, + "args": { + "External id": 10571, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80836, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80836, "pid": 0, "tid": 7, "ts": 1776649792651871, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651862, "dur": 6, + "args": { + "External id": 10571, + "cbid": 211, "correlation": 80836 + } + }, + { + "ph": "s", "id": 80836, "pid": 435, "tid": 435, "ts": 1776649792651862, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651875, "dur": 0, + "args": { + "External id": 80842, + "cbid": 200, "correlation": 80842 + } + }, + { + "ph": "f", "id": 80842, "pid": 435, "tid": 435, "ts": 1776649792651875, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651875, "dur": 0, + "args": { + "External id": 80843, + "cbid": 273, "correlation": 80843 + } + }, + { + "ph": "f", "id": 80843, "pid": 435, "tid": 435, "ts": 1776649792651875, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792651884, "dur": 19, + "args": { + "External id": 80844, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80844, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80844, "pid": 0, "tid": 7, "ts": 1776649792651884, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651876, "dur": 5, + "args": { + "External id": 80844, + "cbid": 211, "correlation": 80844 + } + }, + { + "ph": "s", "id": 80844, "pid": 435, "tid": 435, "ts": 1776649792651876, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792651920, "dur": 1, + "args": { + "External id": 10576, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80862, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80862, "pid": 0, "tid": 7, "ts": 1776649792651920, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651911, "dur": 6, + "args": { + "External id": 10576, + "cbid": 211, "correlation": 80862 + } + }, + { + "ph": "s", "id": 80862, "pid": 435, "tid": 435, "ts": 1776649792651911, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651923, "dur": 0, + "args": { + "External id": 80868, + "cbid": 200, "correlation": 80868 + } + }, + { + "ph": "f", "id": 80868, "pid": 435, "tid": 435, "ts": 1776649792651923, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651924, "dur": 0, + "args": { + "External id": 80869, + "cbid": 273, "correlation": 80869 + } + }, + { + "ph": "f", "id": 80869, "pid": 435, "tid": 435, "ts": 1776649792651924, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792651933, "dur": 19, + "args": { + "External id": 80870, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80870, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80870, "pid": 0, "tid": 7, "ts": 1776649792651933, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651925, "dur": 4, + "args": { + "External id": 80870, + "cbid": 211, "correlation": 80870 + } + }, + { + "ph": "s", "id": 80870, "pid": 435, "tid": 435, "ts": 1776649792651925, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792651969, "dur": 1, + "args": { + "External id": 10581, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80888, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80888, "pid": 0, "tid": 7, "ts": 1776649792651969, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651960, "dur": 6, + "args": { + "External id": 10581, + "cbid": 211, "correlation": 80888 + } + }, + { + "ph": "s", "id": 80888, "pid": 435, "tid": 435, "ts": 1776649792651960, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651972, "dur": 0, + "args": { + "External id": 80894, + "cbid": 200, "correlation": 80894 + } + }, + { + "ph": "f", "id": 80894, "pid": 435, "tid": 435, "ts": 1776649792651972, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792651973, "dur": 0, + "args": { + "External id": 80895, + "cbid": 273, "correlation": 80895 + } + }, + { + "ph": "f", "id": 80895, "pid": 435, "tid": 435, "ts": 1776649792651973, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792651982, "dur": 19, + "args": { + "External id": 80896, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80896, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80896, "pid": 0, "tid": 7, "ts": 1776649792651982, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792651974, "dur": 5, + "args": { + "External id": 80896, + "cbid": 211, "correlation": 80896 + } + }, + { + "ph": "s", "id": 80896, "pid": 435, "tid": 435, "ts": 1776649792651974, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792652017, "dur": 1, + "args": { + "External id": 10586, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80914, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80914, "pid": 0, "tid": 7, "ts": 1776649792652017, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792652008, "dur": 6, + "args": { + "External id": 10586, + "cbid": 211, "correlation": 80914 + } + }, + { + "ph": "s", "id": 80914, "pid": 435, "tid": 435, "ts": 1776649792652008, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792652020, "dur": 0, + "args": { + "External id": 80920, + "cbid": 200, "correlation": 80920 + } + }, + { + "ph": "f", "id": 80920, "pid": 435, "tid": 435, "ts": 1776649792652020, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792652021, "dur": 0, + "args": { + "External id": 80921, + "cbid": 273, "correlation": 80921 + } + }, + { + "ph": "f", "id": 80921, "pid": 435, "tid": 435, "ts": 1776649792652021, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792652030, "dur": 19, + "args": { + "External id": 80922, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80922, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80922, "pid": 0, "tid": 7, "ts": 1776649792652030, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792652022, "dur": 4, + "args": { + "External id": 80922, + "cbid": 211, "correlation": 80922 + } + }, + { + "ph": "s", "id": 80922, "pid": 435, "tid": 435, "ts": 1776649792652022, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792652066, "dur": 1, + "args": { + "External id": 10591, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80940, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80940, "pid": 0, "tid": 7, "ts": 1776649792652066, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792652056, "dur": 6, + "args": { + "External id": 10591, + "cbid": 211, "correlation": 80940 + } + }, + { + "ph": "s", "id": 80940, "pid": 435, "tid": 435, "ts": 1776649792652056, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792652069, "dur": 0, + "args": { + "External id": 80946, + "cbid": 200, "correlation": 80946 + } + }, + { + "ph": "f", "id": 80946, "pid": 435, "tid": 435, "ts": 1776649792652069, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792652069, "dur": 0, + "args": { + "External id": 80947, + "cbid": 273, "correlation": 80947 + } + }, + { + "ph": "f", "id": 80947, "pid": 435, "tid": 435, "ts": 1776649792652069, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792652078, "dur": 19, + "args": { + "External id": 80948, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80948, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80948, "pid": 0, "tid": 7, "ts": 1776649792652078, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792652071, "dur": 4, + "args": { + "External id": 80948, + "cbid": 211, "correlation": 80948 + } + }, + { + "ph": "s", "id": 80948, "pid": 435, "tid": 435, "ts": 1776649792652071, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792652114, "dur": 1, + "args": { + "External id": 10596, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80966, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80966, "pid": 0, "tid": 7, "ts": 1776649792652114, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792652104, "dur": 6, + "args": { + "External id": 10596, + "cbid": 211, "correlation": 80966 + } + }, + { + "ph": "s", "id": 80966, "pid": 435, "tid": 435, "ts": 1776649792652104, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792652117, "dur": 0, + "args": { + "External id": 80972, + "cbid": 200, "correlation": 80972 + } + }, + { + "ph": "f", "id": 80972, "pid": 435, "tid": 435, "ts": 1776649792652117, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792652117, "dur": 0, + "args": { + "External id": 80973, + "cbid": 273, "correlation": 80973 + } + }, + { + "ph": "f", "id": 80973, "pid": 435, "tid": 435, "ts": 1776649792652117, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792652127, "dur": 19, + "args": { + "External id": 80974, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80974, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80974, "pid": 0, "tid": 7, "ts": 1776649792652127, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792652119, "dur": 4, + "args": { + "External id": 80974, + "cbid": 211, "correlation": 80974 + } + }, + { + "ph": "s", "id": 80974, "pid": 435, "tid": 435, "ts": 1776649792652119, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792652162, "dur": 1, + "args": { + "External id": 10601, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 80992, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 80992, "pid": 0, "tid": 7, "ts": 1776649792652162, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792652152, "dur": 6, + "args": { + "External id": 10601, + "cbid": 211, "correlation": 80992 + } + }, + { + "ph": "s", "id": 80992, "pid": 435, "tid": 435, "ts": 1776649792652152, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792652165, "dur": 0, + "args": { + "External id": 80998, + "cbid": 200, "correlation": 80998 + } + }, + { + "ph": "f", "id": 80998, "pid": 435, "tid": 435, "ts": 1776649792652165, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792652165, "dur": 0, + "args": { + "External id": 80999, + "cbid": 273, "correlation": 80999 + } + }, + { + "ph": "f", "id": 80999, "pid": 435, "tid": 435, "ts": 1776649792652165, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792652174, "dur": 19, + "args": { + "External id": 81000, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81000, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81000, "pid": 0, "tid": 7, "ts": 1776649792652174, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792652167, "dur": 5, + "args": { + "External id": 81000, + "cbid": 211, "correlation": 81000 + } + }, + { + "ph": "s", "id": 81000, "pid": 435, "tid": 435, "ts": 1776649792652167, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792652209, "dur": 1, + "args": { + "External id": 10606, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81018, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81018, "pid": 0, "tid": 7, "ts": 1776649792652209, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792652200, "dur": 6, + "args": { + "External id": 10606, + "cbid": 211, "correlation": 81018 + } + }, + { + "ph": "s", "id": 81018, "pid": 435, "tid": 435, "ts": 1776649792652200, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792652213, "dur": 0, + "args": { + "External id": 81024, + "cbid": 200, "correlation": 81024 + } + }, + { + "ph": "f", "id": 81024, "pid": 435, "tid": 435, "ts": 1776649792652213, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792652213, "dur": 0, + "args": { + "External id": 81025, + "cbid": 273, "correlation": 81025 + } + }, + { + "ph": "f", "id": 81025, "pid": 435, "tid": 435, "ts": 1776649792652213, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792656321, "dur": 19, + "args": { + "External id": 81026, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81026, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81026, "pid": 0, "tid": 7, "ts": 1776649792656321, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792652214, "dur": 4108, + "args": { + "External id": 81026, + "cbid": 211, "correlation": 81026 + } + }, + { + "ph": "s", "id": 81026, "pid": 435, "tid": 435, "ts": 1776649792652214, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792656410, "dur": 1, + "args": { + "External id": 10611, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81044, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81044, "pid": 0, "tid": 7, "ts": 1776649792656410, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792656397, "dur": 10, + "args": { + "External id": 10611, + "cbid": 211, "correlation": 81044 + } + }, + { + "ph": "s", "id": 81044, "pid": 435, "tid": 435, "ts": 1776649792656397, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792656416, "dur": 0, + "args": { + "External id": 81050, + "cbid": 200, "correlation": 81050 + } + }, + { + "ph": "f", "id": 81050, "pid": 435, "tid": 435, "ts": 1776649792656416, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792656418, "dur": 1, + "args": { + "External id": 81051, + "cbid": 273, "correlation": 81051 + } + }, + { + "ph": "f", "id": 81051, "pid": 435, "tid": 435, "ts": 1776649792656418, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792656428, "dur": 19, + "args": { + "External id": 81052, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81052, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81052, "pid": 0, "tid": 7, "ts": 1776649792656428, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792656419, "dur": 6, + "args": { + "External id": 81052, + "cbid": 211, "correlation": 81052 + } + }, + { + "ph": "s", "id": 81052, "pid": 435, "tid": 435, "ts": 1776649792656419, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792656468, "dur": 1, + "args": { + "External id": 10616, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81070, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81070, "pid": 0, "tid": 7, "ts": 1776649792656468, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792656458, "dur": 7, + "args": { + "External id": 10616, + "cbid": 211, "correlation": 81070 + } + }, + { + "ph": "s", "id": 81070, "pid": 435, "tid": 435, "ts": 1776649792656458, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792656472, "dur": 0, + "args": { + "External id": 81076, + "cbid": 200, "correlation": 81076 + } + }, + { + "ph": "f", "id": 81076, "pid": 435, "tid": 435, "ts": 1776649792656472, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792656472, "dur": 0, + "args": { + "External id": 81077, + "cbid": 273, "correlation": 81077 + } + }, + { + "ph": "f", "id": 81077, "pid": 435, "tid": 435, "ts": 1776649792656472, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792656482, "dur": 19, + "args": { + "External id": 81078, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81078, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81078, "pid": 0, "tid": 7, "ts": 1776649792656482, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792656474, "dur": 5, + "args": { + "External id": 81078, + "cbid": 211, "correlation": 81078 + } + }, + { + "ph": "s", "id": 81078, "pid": 435, "tid": 435, "ts": 1776649792656474, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792656518, "dur": 1, + "args": { + "External id": 10621, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81096, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81096, "pid": 0, "tid": 7, "ts": 1776649792656518, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792656509, "dur": 7, + "args": { + "External id": 10621, + "cbid": 211, "correlation": 81096 + } + }, + { + "ph": "s", "id": 81096, "pid": 435, "tid": 435, "ts": 1776649792656509, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792656522, "dur": 0, + "args": { + "External id": 81102, + "cbid": 200, "correlation": 81102 + } + }, + { + "ph": "f", "id": 81102, "pid": 435, "tid": 435, "ts": 1776649792656522, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792656522, "dur": 0, + "args": { + "External id": 81103, + "cbid": 273, "correlation": 81103 + } + }, + { + "ph": "f", "id": 81103, "pid": 435, "tid": 435, "ts": 1776649792656522, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792656531, "dur": 19, + "args": { + "External id": 81104, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81104, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81104, "pid": 0, "tid": 7, "ts": 1776649792656531, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792656523, "dur": 5, + "args": { + "External id": 81104, + "cbid": 211, "correlation": 81104 + } + }, + { + "ph": "s", "id": 81104, "pid": 435, "tid": 435, "ts": 1776649792656523, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792656567, "dur": 1, + "args": { + "External id": 10626, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81122, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81122, "pid": 0, "tid": 7, "ts": 1776649792656567, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792656558, "dur": 7, + "args": { + "External id": 10626, + "cbid": 211, "correlation": 81122 + } + }, + { + "ph": "s", "id": 81122, "pid": 435, "tid": 435, "ts": 1776649792656558, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792656570, "dur": 0, + "args": { + "External id": 81128, + "cbid": 200, "correlation": 81128 + } + }, + { + "ph": "f", "id": 81128, "pid": 435, "tid": 435, "ts": 1776649792656570, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792656571, "dur": 0, + "args": { + "External id": 81129, + "cbid": 273, "correlation": 81129 + } + }, + { + "ph": "f", "id": 81129, "pid": 435, "tid": 435, "ts": 1776649792656571, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792656580, "dur": 19, + "args": { + "External id": 81130, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81130, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81130, "pid": 0, "tid": 7, "ts": 1776649792656580, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792656572, "dur": 5, + "args": { + "External id": 81130, + "cbid": 211, "correlation": 81130 + } + }, + { + "ph": "s", "id": 81130, "pid": 435, "tid": 435, "ts": 1776649792656572, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792656616, "dur": 1, + "args": { + "External id": 10631, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81148, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81148, "pid": 0, "tid": 7, "ts": 1776649792656616, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792656607, "dur": 6, + "args": { + "External id": 10631, + "cbid": 211, "correlation": 81148 + } + }, + { + "ph": "s", "id": 81148, "pid": 435, "tid": 435, "ts": 1776649792656607, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792656619, "dur": 0, + "args": { + "External id": 81154, + "cbid": 200, "correlation": 81154 + } + }, + { + "ph": "f", "id": 81154, "pid": 435, "tid": 435, "ts": 1776649792656619, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792656620, "dur": 0, + "args": { + "External id": 81155, + "cbid": 273, "correlation": 81155 + } + }, + { + "ph": "f", "id": 81155, "pid": 435, "tid": 435, "ts": 1776649792656620, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792656629, "dur": 19, + "args": { + "External id": 81156, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81156, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81156, "pid": 0, "tid": 7, "ts": 1776649792656629, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792656621, "dur": 4, + "args": { + "External id": 81156, + "cbid": 211, "correlation": 81156 + } + }, + { + "ph": "s", "id": 81156, "pid": 435, "tid": 435, "ts": 1776649792656621, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792656664, "dur": 1, + "args": { + "External id": 10636, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81174, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81174, "pid": 0, "tid": 7, "ts": 1776649792656664, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792656655, "dur": 6, + "args": { + "External id": 10636, + "cbid": 211, "correlation": 81174 + } + }, + { + "ph": "s", "id": 81174, "pid": 435, "tid": 435, "ts": 1776649792656655, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792656667, "dur": 0, + "args": { + "External id": 81180, + "cbid": 200, "correlation": 81180 + } + }, + { + "ph": "f", "id": 81180, "pid": 435, "tid": 435, "ts": 1776649792656667, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792656668, "dur": 0, + "args": { + "External id": 81181, + "cbid": 273, "correlation": 81181 + } + }, + { + "ph": "f", "id": 81181, "pid": 435, "tid": 435, "ts": 1776649792656668, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792656677, "dur": 19, + "args": { + "External id": 81182, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81182, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81182, "pid": 0, "tid": 7, "ts": 1776649792656677, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792656669, "dur": 4, + "args": { + "External id": 81182, + "cbid": 211, "correlation": 81182 + } + }, + { + "ph": "s", "id": 81182, "pid": 435, "tid": 435, "ts": 1776649792656669, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792656712, "dur": 1, + "args": { + "External id": 10641, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81200, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81200, "pid": 0, "tid": 7, "ts": 1776649792656712, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792656703, "dur": 6, + "args": { + "External id": 10641, + "cbid": 211, "correlation": 81200 + } + }, + { + "ph": "s", "id": 81200, "pid": 435, "tid": 435, "ts": 1776649792656703, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792656715, "dur": 0, + "args": { + "External id": 81206, + "cbid": 200, "correlation": 81206 + } + }, + { + "ph": "f", "id": 81206, "pid": 435, "tid": 435, "ts": 1776649792656715, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792656716, "dur": 0, + "args": { + "External id": 81207, + "cbid": 273, "correlation": 81207 + } + }, + { + "ph": "f", "id": 81207, "pid": 435, "tid": 435, "ts": 1776649792656716, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792656725, "dur": 19, + "args": { + "External id": 81208, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81208, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81208, "pid": 0, "tid": 7, "ts": 1776649792656725, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792656717, "dur": 4, + "args": { + "External id": 81208, + "cbid": 211, "correlation": 81208 + } + }, + { + "ph": "s", "id": 81208, "pid": 435, "tid": 435, "ts": 1776649792656717, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792656760, "dur": 1, + "args": { + "External id": 10646, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81226, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81226, "pid": 0, "tid": 7, "ts": 1776649792656760, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792656751, "dur": 6, + "args": { + "External id": 10646, + "cbid": 211, "correlation": 81226 + } + }, + { + "ph": "s", "id": 81226, "pid": 435, "tid": 435, "ts": 1776649792656751, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792656763, "dur": 0, + "args": { + "External id": 81232, + "cbid": 200, "correlation": 81232 + } + }, + { + "ph": "f", "id": 81232, "pid": 435, "tid": 435, "ts": 1776649792656763, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792656764, "dur": 0, + "args": { + "External id": 81233, + "cbid": 273, "correlation": 81233 + } + }, + { + "ph": "f", "id": 81233, "pid": 435, "tid": 435, "ts": 1776649792656764, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792656773, "dur": 19, + "args": { + "External id": 81234, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81234, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81234, "pid": 0, "tid": 7, "ts": 1776649792656773, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792656765, "dur": 4, + "args": { + "External id": 81234, + "cbid": 211, "correlation": 81234 + } + }, + { + "ph": "s", "id": 81234, "pid": 435, "tid": 435, "ts": 1776649792656765, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792656815, "dur": 1, + "args": { + "External id": 10651, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81252, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81252, "pid": 0, "tid": 7, "ts": 1776649792656815, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792656805, "dur": 7, + "args": { + "External id": 10651, + "cbid": 211, "correlation": 81252 + } + }, + { + "ph": "s", "id": 81252, "pid": 435, "tid": 435, "ts": 1776649792656805, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792656819, "dur": 0, + "args": { + "External id": 81258, + "cbid": 200, "correlation": 81258 + } + }, + { + "ph": "f", "id": 81258, "pid": 435, "tid": 435, "ts": 1776649792656819, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792656819, "dur": 0, + "args": { + "External id": 81259, + "cbid": 273, "correlation": 81259 + } + }, + { + "ph": "f", "id": 81259, "pid": 435, "tid": 435, "ts": 1776649792656819, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792656828, "dur": 19, + "args": { + "External id": 81260, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81260, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81260, "pid": 0, "tid": 7, "ts": 1776649792656828, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792656821, "dur": 5, + "args": { + "External id": 81260, + "cbid": 211, "correlation": 81260 + } + }, + { + "ph": "s", "id": 81260, "pid": 435, "tid": 435, "ts": 1776649792656821, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792656865, "dur": 1, + "args": { + "External id": 10656, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81278, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81278, "pid": 0, "tid": 7, "ts": 1776649792656865, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792656856, "dur": 6, + "args": { + "External id": 10656, + "cbid": 211, "correlation": 81278 + } + }, + { + "ph": "s", "id": 81278, "pid": 435, "tid": 435, "ts": 1776649792656856, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792656868, "dur": 0, + "args": { + "External id": 81284, + "cbid": 200, "correlation": 81284 + } + }, + { + "ph": "f", "id": 81284, "pid": 435, "tid": 435, "ts": 1776649792656868, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792656869, "dur": 0, + "args": { + "External id": 81285, + "cbid": 273, "correlation": 81285 + } + }, + { + "ph": "f", "id": 81285, "pid": 435, "tid": 435, "ts": 1776649792656869, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792656879, "dur": 19, + "args": { + "External id": 81286, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81286, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81286, "pid": 0, "tid": 7, "ts": 1776649792656879, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792656870, "dur": 4, + "args": { + "External id": 81286, + "cbid": 211, "correlation": 81286 + } + }, + { + "ph": "s", "id": 81286, "pid": 435, "tid": 435, "ts": 1776649792656870, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792656914, "dur": 1, + "args": { + "External id": 10661, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81304, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81304, "pid": 0, "tid": 7, "ts": 1776649792656914, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792656905, "dur": 6, + "args": { + "External id": 10661, + "cbid": 211, "correlation": 81304 + } + }, + { + "ph": "s", "id": 81304, "pid": 435, "tid": 435, "ts": 1776649792656905, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792656917, "dur": 0, + "args": { + "External id": 81310, + "cbid": 200, "correlation": 81310 + } + }, + { + "ph": "f", "id": 81310, "pid": 435, "tid": 435, "ts": 1776649792656917, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792656918, "dur": 0, + "args": { + "External id": 81311, + "cbid": 273, "correlation": 81311 + } + }, + { + "ph": "f", "id": 81311, "pid": 435, "tid": 435, "ts": 1776649792656918, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792656927, "dur": 19, + "args": { + "External id": 81312, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81312, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81312, "pid": 0, "tid": 7, "ts": 1776649792656927, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792656919, "dur": 5, + "args": { + "External id": 81312, + "cbid": 211, "correlation": 81312 + } + }, + { + "ph": "s", "id": 81312, "pid": 435, "tid": 435, "ts": 1776649792656919, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792656963, "dur": 1, + "args": { + "External id": 10666, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81330, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81330, "pid": 0, "tid": 7, "ts": 1776649792656963, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792656953, "dur": 6, + "args": { + "External id": 10666, + "cbid": 211, "correlation": 81330 + } + }, + { + "ph": "s", "id": 81330, "pid": 435, "tid": 435, "ts": 1776649792656953, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792656966, "dur": 0, + "args": { + "External id": 81336, + "cbid": 200, "correlation": 81336 + } + }, + { + "ph": "f", "id": 81336, "pid": 435, "tid": 435, "ts": 1776649792656966, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792656966, "dur": 0, + "args": { + "External id": 81337, + "cbid": 273, "correlation": 81337 + } + }, + { + "ph": "f", "id": 81337, "pid": 435, "tid": 435, "ts": 1776649792656966, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792656975, "dur": 19, + "args": { + "External id": 81338, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81338, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81338, "pid": 0, "tid": 7, "ts": 1776649792656975, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792656967, "dur": 5, + "args": { + "External id": 81338, + "cbid": 211, "correlation": 81338 + } + }, + { + "ph": "s", "id": 81338, "pid": 435, "tid": 435, "ts": 1776649792656967, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792657011, "dur": 1, + "args": { + "External id": 10671, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81356, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81356, "pid": 0, "tid": 7, "ts": 1776649792657011, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657001, "dur": 6, + "args": { + "External id": 10671, + "cbid": 211, "correlation": 81356 + } + }, + { + "ph": "s", "id": 81356, "pid": 435, "tid": 435, "ts": 1776649792657001, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657014, "dur": 0, + "args": { + "External id": 81362, + "cbid": 200, "correlation": 81362 + } + }, + { + "ph": "f", "id": 81362, "pid": 435, "tid": 435, "ts": 1776649792657014, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657015, "dur": 0, + "args": { + "External id": 81363, + "cbid": 273, "correlation": 81363 + } + }, + { + "ph": "f", "id": 81363, "pid": 435, "tid": 435, "ts": 1776649792657015, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792657023, "dur": 19, + "args": { + "External id": 81364, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81364, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81364, "pid": 0, "tid": 7, "ts": 1776649792657023, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657016, "dur": 5, + "args": { + "External id": 81364, + "cbid": 211, "correlation": 81364 + } + }, + { + "ph": "s", "id": 81364, "pid": 435, "tid": 435, "ts": 1776649792657016, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792657059, "dur": 1, + "args": { + "External id": 10676, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81382, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81382, "pid": 0, "tid": 7, "ts": 1776649792657059, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657050, "dur": 6, + "args": { + "External id": 10676, + "cbid": 211, "correlation": 81382 + } + }, + { + "ph": "s", "id": 81382, "pid": 435, "tid": 435, "ts": 1776649792657050, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657062, "dur": 0, + "args": { + "External id": 81388, + "cbid": 200, "correlation": 81388 + } + }, + { + "ph": "f", "id": 81388, "pid": 435, "tid": 435, "ts": 1776649792657062, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657063, "dur": 0, + "args": { + "External id": 81389, + "cbid": 273, "correlation": 81389 + } + }, + { + "ph": "f", "id": 81389, "pid": 435, "tid": 435, "ts": 1776649792657063, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792657072, "dur": 19, + "args": { + "External id": 81390, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81390, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81390, "pid": 0, "tid": 7, "ts": 1776649792657072, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657064, "dur": 4, + "args": { + "External id": 81390, + "cbid": 211, "correlation": 81390 + } + }, + { + "ph": "s", "id": 81390, "pid": 435, "tid": 435, "ts": 1776649792657064, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792657107, "dur": 1, + "args": { + "External id": 10681, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81408, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81408, "pid": 0, "tid": 7, "ts": 1776649792657107, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657098, "dur": 6, + "args": { + "External id": 10681, + "cbid": 211, "correlation": 81408 + } + }, + { + "ph": "s", "id": 81408, "pid": 435, "tid": 435, "ts": 1776649792657098, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657110, "dur": 0, + "args": { + "External id": 81414, + "cbid": 200, "correlation": 81414 + } + }, + { + "ph": "f", "id": 81414, "pid": 435, "tid": 435, "ts": 1776649792657110, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657111, "dur": 0, + "args": { + "External id": 81415, + "cbid": 273, "correlation": 81415 + } + }, + { + "ph": "f", "id": 81415, "pid": 435, "tid": 435, "ts": 1776649792657111, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792657120, "dur": 19, + "args": { + "External id": 81416, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81416, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81416, "pid": 0, "tid": 7, "ts": 1776649792657120, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657112, "dur": 5, + "args": { + "External id": 81416, + "cbid": 211, "correlation": 81416 + } + }, + { + "ph": "s", "id": 81416, "pid": 435, "tid": 435, "ts": 1776649792657112, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649792657150, "dur": 4, + "args": { + "External id": 81424, + "cbid": 135, "correlation": 81424 + } + }, + { + "ph": "f", "id": 81424, "pid": 435, "tid": 435, "ts": 1776649792657150, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792657167, "dur": 6, + "args": { + "External id": 81429, + "cbid": 165, "correlation": 81429 + } + }, + { + "ph": "s", "id": 81429, "pid": 435, "tid": 435, "ts": 1776649792657167, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649792657178, "dur": 7, + "args": { + "External id": 81432, + "cbid": 139, "correlation": 81432 + } + }, + { + "ph": "f", "id": 81432, "pid": 435, "tid": 435, "ts": 1776649792657178, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792657243, "dur": 1, + "args": { + "External id": 10686, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81460, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81460, "pid": 0, "tid": 7, "ts": 1776649792657243, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657231, "dur": 9, + "args": { + "External id": 10686, + "cbid": 211, "correlation": 81460 + } + }, + { + "ph": "s", "id": 81460, "pid": 435, "tid": 435, "ts": 1776649792657231, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657247, "dur": 0, + "args": { + "External id": 81466, + "cbid": 200, "correlation": 81466 + } + }, + { + "ph": "f", "id": 81466, "pid": 435, "tid": 435, "ts": 1776649792657247, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657248, "dur": 1, + "args": { + "External id": 81467, + "cbid": 273, "correlation": 81467 + } + }, + { + "ph": "f", "id": 81467, "pid": 435, "tid": 435, "ts": 1776649792657248, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792657258, "dur": 19, + "args": { + "External id": 81468, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81468, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81468, "pid": 0, "tid": 7, "ts": 1776649792657258, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657250, "dur": 5, + "args": { + "External id": 81468, + "cbid": 211, "correlation": 81468 + } + }, + { + "ph": "s", "id": 81468, "pid": 435, "tid": 435, "ts": 1776649792657250, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792657297, "dur": 1, + "args": { + "External id": 10691, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81486, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81486, "pid": 0, "tid": 7, "ts": 1776649792657297, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657287, "dur": 7, + "args": { + "External id": 10691, + "cbid": 211, "correlation": 81486 + } + }, + { + "ph": "s", "id": 81486, "pid": 435, "tid": 435, "ts": 1776649792657287, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657300, "dur": 0, + "args": { + "External id": 81492, + "cbid": 200, "correlation": 81492 + } + }, + { + "ph": "f", "id": 81492, "pid": 435, "tid": 435, "ts": 1776649792657300, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657301, "dur": 0, + "args": { + "External id": 81493, + "cbid": 273, "correlation": 81493 + } + }, + { + "ph": "f", "id": 81493, "pid": 435, "tid": 435, "ts": 1776649792657301, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792657310, "dur": 19, + "args": { + "External id": 81494, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81494, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81494, "pid": 0, "tid": 7, "ts": 1776649792657310, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657302, "dur": 5, + "args": { + "External id": 81494, + "cbid": 211, "correlation": 81494 + } + }, + { + "ph": "s", "id": 81494, "pid": 435, "tid": 435, "ts": 1776649792657302, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792657346, "dur": 1, + "args": { + "External id": 10696, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81512, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81512, "pid": 0, "tid": 7, "ts": 1776649792657346, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657337, "dur": 7, + "args": { + "External id": 10696, + "cbid": 211, "correlation": 81512 + } + }, + { + "ph": "s", "id": 81512, "pid": 435, "tid": 435, "ts": 1776649792657337, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657349, "dur": 0, + "args": { + "External id": 81518, + "cbid": 200, "correlation": 81518 + } + }, + { + "ph": "f", "id": 81518, "pid": 435, "tid": 435, "ts": 1776649792657349, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657350, "dur": 0, + "args": { + "External id": 81519, + "cbid": 273, "correlation": 81519 + } + }, + { + "ph": "f", "id": 81519, "pid": 435, "tid": 435, "ts": 1776649792657350, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792657359, "dur": 19, + "args": { + "External id": 81520, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81520, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81520, "pid": 0, "tid": 7, "ts": 1776649792657359, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657351, "dur": 4, + "args": { + "External id": 81520, + "cbid": 211, "correlation": 81520 + } + }, + { + "ph": "s", "id": 81520, "pid": 435, "tid": 435, "ts": 1776649792657351, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792657395, "dur": 1, + "args": { + "External id": 10701, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81538, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81538, "pid": 0, "tid": 7, "ts": 1776649792657395, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657386, "dur": 6, + "args": { + "External id": 10701, + "cbid": 211, "correlation": 81538 + } + }, + { + "ph": "s", "id": 81538, "pid": 435, "tid": 435, "ts": 1776649792657386, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657398, "dur": 0, + "args": { + "External id": 81544, + "cbid": 200, "correlation": 81544 + } + }, + { + "ph": "f", "id": 81544, "pid": 435, "tid": 435, "ts": 1776649792657398, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657399, "dur": 0, + "args": { + "External id": 81545, + "cbid": 273, "correlation": 81545 + } + }, + { + "ph": "f", "id": 81545, "pid": 435, "tid": 435, "ts": 1776649792657399, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792657408, "dur": 19, + "args": { + "External id": 81546, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81546, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81546, "pid": 0, "tid": 7, "ts": 1776649792657408, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657400, "dur": 5, + "args": { + "External id": 81546, + "cbid": 211, "correlation": 81546 + } + }, + { + "ph": "s", "id": 81546, "pid": 435, "tid": 435, "ts": 1776649792657400, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792657446, "dur": 1, + "args": { + "External id": 10706, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81564, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81564, "pid": 0, "tid": 7, "ts": 1776649792657446, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657437, "dur": 6, + "args": { + "External id": 10706, + "cbid": 211, "correlation": 81564 + } + }, + { + "ph": "s", "id": 81564, "pid": 435, "tid": 435, "ts": 1776649792657437, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657449, "dur": 0, + "args": { + "External id": 81570, + "cbid": 200, "correlation": 81570 + } + }, + { + "ph": "f", "id": 81570, "pid": 435, "tid": 435, "ts": 1776649792657449, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657450, "dur": 0, + "args": { + "External id": 81571, + "cbid": 273, "correlation": 81571 + } + }, + { + "ph": "f", "id": 81571, "pid": 435, "tid": 435, "ts": 1776649792657450, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792657459, "dur": 19, + "args": { + "External id": 81572, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81572, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81572, "pid": 0, "tid": 7, "ts": 1776649792657459, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657451, "dur": 5, + "args": { + "External id": 81572, + "cbid": 211, "correlation": 81572 + } + }, + { + "ph": "s", "id": 81572, "pid": 435, "tid": 435, "ts": 1776649792657451, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792657494, "dur": 1, + "args": { + "External id": 10711, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81590, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81590, "pid": 0, "tid": 7, "ts": 1776649792657494, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657485, "dur": 6, + "args": { + "External id": 10711, + "cbid": 211, "correlation": 81590 + } + }, + { + "ph": "s", "id": 81590, "pid": 435, "tid": 435, "ts": 1776649792657485, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657498, "dur": 0, + "args": { + "External id": 81596, + "cbid": 200, "correlation": 81596 + } + }, + { + "ph": "f", "id": 81596, "pid": 435, "tid": 435, "ts": 1776649792657498, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657498, "dur": 0, + "args": { + "External id": 81597, + "cbid": 273, "correlation": 81597 + } + }, + { + "ph": "f", "id": 81597, "pid": 435, "tid": 435, "ts": 1776649792657498, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792657507, "dur": 19, + "args": { + "External id": 81598, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81598, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81598, "pid": 0, "tid": 7, "ts": 1776649792657507, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657499, "dur": 4, + "args": { + "External id": 81598, + "cbid": 211, "correlation": 81598 + } + }, + { + "ph": "s", "id": 81598, "pid": 435, "tid": 435, "ts": 1776649792657499, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792657543, "dur": 1, + "args": { + "External id": 10716, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81616, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81616, "pid": 0, "tid": 7, "ts": 1776649792657543, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657533, "dur": 6, + "args": { + "External id": 10716, + "cbid": 211, "correlation": 81616 + } + }, + { + "ph": "s", "id": 81616, "pid": 435, "tid": 435, "ts": 1776649792657533, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657545, "dur": 0, + "args": { + "External id": 81622, + "cbid": 200, "correlation": 81622 + } + }, + { + "ph": "f", "id": 81622, "pid": 435, "tid": 435, "ts": 1776649792657545, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657546, "dur": 0, + "args": { + "External id": 81623, + "cbid": 273, "correlation": 81623 + } + }, + { + "ph": "f", "id": 81623, "pid": 435, "tid": 435, "ts": 1776649792657546, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792657555, "dur": 19, + "args": { + "External id": 81624, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81624, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81624, "pid": 0, "tid": 7, "ts": 1776649792657555, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657547, "dur": 4, + "args": { + "External id": 81624, + "cbid": 211, "correlation": 81624 + } + }, + { + "ph": "s", "id": 81624, "pid": 435, "tid": 435, "ts": 1776649792657547, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792657590, "dur": 1, + "args": { + "External id": 10721, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81642, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81642, "pid": 0, "tid": 7, "ts": 1776649792657590, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657581, "dur": 6, + "args": { + "External id": 10721, + "cbid": 211, "correlation": 81642 + } + }, + { + "ph": "s", "id": 81642, "pid": 435, "tid": 435, "ts": 1776649792657581, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657593, "dur": 0, + "args": { + "External id": 81648, + "cbid": 200, "correlation": 81648 + } + }, + { + "ph": "f", "id": 81648, "pid": 435, "tid": 435, "ts": 1776649792657593, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657594, "dur": 0, + "args": { + "External id": 81649, + "cbid": 273, "correlation": 81649 + } + }, + { + "ph": "f", "id": 81649, "pid": 435, "tid": 435, "ts": 1776649792657594, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792657603, "dur": 19, + "args": { + "External id": 81650, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81650, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81650, "pid": 0, "tid": 7, "ts": 1776649792657603, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657595, "dur": 4, + "args": { + "External id": 81650, + "cbid": 211, "correlation": 81650 + } + }, + { + "ph": "s", "id": 81650, "pid": 435, "tid": 435, "ts": 1776649792657595, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792657639, "dur": 1, + "args": { + "External id": 10726, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81668, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81668, "pid": 0, "tid": 7, "ts": 1776649792657639, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657629, "dur": 7, + "args": { + "External id": 10726, + "cbid": 211, "correlation": 81668 + } + }, + { + "ph": "s", "id": 81668, "pid": 435, "tid": 435, "ts": 1776649792657629, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657642, "dur": 0, + "args": { + "External id": 81674, + "cbid": 200, "correlation": 81674 + } + }, + { + "ph": "f", "id": 81674, "pid": 435, "tid": 435, "ts": 1776649792657642, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657643, "dur": 0, + "args": { + "External id": 81675, + "cbid": 273, "correlation": 81675 + } + }, + { + "ph": "f", "id": 81675, "pid": 435, "tid": 435, "ts": 1776649792657643, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792657652, "dur": 19, + "args": { + "External id": 81676, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81676, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81676, "pid": 0, "tid": 7, "ts": 1776649792657652, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657644, "dur": 5, + "args": { + "External id": 81676, + "cbid": 211, "correlation": 81676 + } + }, + { + "ph": "s", "id": 81676, "pid": 435, "tid": 435, "ts": 1776649792657644, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792657687, "dur": 1, + "args": { + "External id": 10731, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81694, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81694, "pid": 0, "tid": 7, "ts": 1776649792657687, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657677, "dur": 6, + "args": { + "External id": 10731, + "cbid": 211, "correlation": 81694 + } + }, + { + "ph": "s", "id": 81694, "pid": 435, "tid": 435, "ts": 1776649792657677, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657690, "dur": 0, + "args": { + "External id": 81700, + "cbid": 200, "correlation": 81700 + } + }, + { + "ph": "f", "id": 81700, "pid": 435, "tid": 435, "ts": 1776649792657690, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657690, "dur": 0, + "args": { + "External id": 81701, + "cbid": 273, "correlation": 81701 + } + }, + { + "ph": "f", "id": 81701, "pid": 435, "tid": 435, "ts": 1776649792657690, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792657699, "dur": 19, + "args": { + "External id": 81702, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81702, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81702, "pid": 0, "tid": 7, "ts": 1776649792657699, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657691, "dur": 4, + "args": { + "External id": 81702, + "cbid": 211, "correlation": 81702 + } + }, + { + "ph": "s", "id": 81702, "pid": 435, "tid": 435, "ts": 1776649792657691, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792657715, "dur": 8, + "args": { + "External id": 81705, + "cbid": 165, "correlation": 81705 + } + }, + { + "ph": "s", "id": 81705, "pid": 435, "tid": 435, "ts": 1776649792657715, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649792657746, "dur": 3, + "args": { + "External id": 81714, + "cbid": 135, "correlation": 81714 + } + }, + { + "ph": "f", "id": 81714, "pid": 435, "tid": 435, "ts": 1776649792657746, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792657795, "dur": 1, + "args": { + "External id": 10736, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81734, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81734, "pid": 0, "tid": 7, "ts": 1776649792657795, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657784, "dur": 8, + "args": { + "External id": 10736, + "cbid": 211, "correlation": 81734 + } + }, + { + "ph": "s", "id": 81734, "pid": 435, "tid": 435, "ts": 1776649792657784, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657799, "dur": 0, + "args": { + "External id": 81740, + "cbid": 200, "correlation": 81740 + } + }, + { + "ph": "f", "id": 81740, "pid": 435, "tid": 435, "ts": 1776649792657799, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657800, "dur": 0, + "args": { + "External id": 81741, + "cbid": 273, "correlation": 81741 + } + }, + { + "ph": "f", "id": 81741, "pid": 435, "tid": 435, "ts": 1776649792657800, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792657810, "dur": 19, + "args": { + "External id": 81742, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81742, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81742, "pid": 0, "tid": 7, "ts": 1776649792657810, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657801, "dur": 5, + "args": { + "External id": 81742, + "cbid": 211, "correlation": 81742 + } + }, + { + "ph": "s", "id": 81742, "pid": 435, "tid": 435, "ts": 1776649792657801, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792657848, "dur": 1, + "args": { + "External id": 10741, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81760, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81760, "pid": 0, "tid": 7, "ts": 1776649792657848, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657838, "dur": 7, + "args": { + "External id": 10741, + "cbid": 211, "correlation": 81760 + } + }, + { + "ph": "s", "id": 81760, "pid": 435, "tid": 435, "ts": 1776649792657838, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657851, "dur": 0, + "args": { + "External id": 81766, + "cbid": 200, "correlation": 81766 + } + }, + { + "ph": "f", "id": 81766, "pid": 435, "tid": 435, "ts": 1776649792657851, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657852, "dur": 0, + "args": { + "External id": 81767, + "cbid": 273, "correlation": 81767 + } + }, + { + "ph": "f", "id": 81767, "pid": 435, "tid": 435, "ts": 1776649792657852, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792657861, "dur": 19, + "args": { + "External id": 81768, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81768, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81768, "pid": 0, "tid": 7, "ts": 1776649792657861, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657853, "dur": 4, + "args": { + "External id": 81768, + "cbid": 211, "correlation": 81768 + } + }, + { + "ph": "s", "id": 81768, "pid": 435, "tid": 435, "ts": 1776649792657853, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792657897, "dur": 1, + "args": { + "External id": 10746, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81786, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81786, "pid": 0, "tid": 7, "ts": 1776649792657897, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657887, "dur": 6, + "args": { + "External id": 10746, + "cbid": 211, "correlation": 81786 + } + }, + { + "ph": "s", "id": 81786, "pid": 435, "tid": 435, "ts": 1776649792657887, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657900, "dur": 0, + "args": { + "External id": 81792, + "cbid": 200, "correlation": 81792 + } + }, + { + "ph": "f", "id": 81792, "pid": 435, "tid": 435, "ts": 1776649792657900, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657900, "dur": 0, + "args": { + "External id": 81793, + "cbid": 273, "correlation": 81793 + } + }, + { + "ph": "f", "id": 81793, "pid": 435, "tid": 435, "ts": 1776649792657900, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792657909, "dur": 19, + "args": { + "External id": 81794, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81794, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81794, "pid": 0, "tid": 7, "ts": 1776649792657909, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657901, "dur": 5, + "args": { + "External id": 81794, + "cbid": 211, "correlation": 81794 + } + }, + { + "ph": "s", "id": 81794, "pid": 435, "tid": 435, "ts": 1776649792657901, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792657945, "dur": 1, + "args": { + "External id": 10751, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81812, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81812, "pid": 0, "tid": 7, "ts": 1776649792657945, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657936, "dur": 6, + "args": { + "External id": 10751, + "cbid": 211, "correlation": 81812 + } + }, + { + "ph": "s", "id": 81812, "pid": 435, "tid": 435, "ts": 1776649792657936, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657948, "dur": 0, + "args": { + "External id": 81818, + "cbid": 200, "correlation": 81818 + } + }, + { + "ph": "f", "id": 81818, "pid": 435, "tid": 435, "ts": 1776649792657948, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792657949, "dur": 0, + "args": { + "External id": 81819, + "cbid": 273, "correlation": 81819 + } + }, + { + "ph": "f", "id": 81819, "pid": 435, "tid": 435, "ts": 1776649792657949, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792657958, "dur": 19, + "args": { + "External id": 81820, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81820, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81820, "pid": 0, "tid": 7, "ts": 1776649792657958, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657950, "dur": 4, + "args": { + "External id": 81820, + "cbid": 211, "correlation": 81820 + } + }, + { + "ph": "s", "id": 81820, "pid": 435, "tid": 435, "ts": 1776649792657950, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792658003, "dur": 1, + "args": { + "External id": 10756, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81838, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81838, "pid": 0, "tid": 7, "ts": 1776649792658003, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792657994, "dur": 7, + "args": { + "External id": 10756, + "cbid": 211, "correlation": 81838 + } + }, + { + "ph": "s", "id": 81838, "pid": 435, "tid": 435, "ts": 1776649792657994, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658006, "dur": 0, + "args": { + "External id": 81844, + "cbid": 200, "correlation": 81844 + } + }, + { + "ph": "f", "id": 81844, "pid": 435, "tid": 435, "ts": 1776649792658006, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658007, "dur": 0, + "args": { + "External id": 81845, + "cbid": 273, "correlation": 81845 + } + }, + { + "ph": "f", "id": 81845, "pid": 435, "tid": 435, "ts": 1776649792658007, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792658016, "dur": 19, + "args": { + "External id": 81846, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81846, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81846, "pid": 0, "tid": 7, "ts": 1776649792658016, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658008, "dur": 5, + "args": { + "External id": 81846, + "cbid": 211, "correlation": 81846 + } + }, + { + "ph": "s", "id": 81846, "pid": 435, "tid": 435, "ts": 1776649792658008, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792658052, "dur": 1, + "args": { + "External id": 10761, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81864, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81864, "pid": 0, "tid": 7, "ts": 1776649792658052, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658043, "dur": 6, + "args": { + "External id": 10761, + "cbid": 211, "correlation": 81864 + } + }, + { + "ph": "s", "id": 81864, "pid": 435, "tid": 435, "ts": 1776649792658043, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658055, "dur": 0, + "args": { + "External id": 81870, + "cbid": 200, "correlation": 81870 + } + }, + { + "ph": "f", "id": 81870, "pid": 435, "tid": 435, "ts": 1776649792658055, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658056, "dur": 0, + "args": { + "External id": 81871, + "cbid": 273, "correlation": 81871 + } + }, + { + "ph": "f", "id": 81871, "pid": 435, "tid": 435, "ts": 1776649792658056, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792658065, "dur": 19, + "args": { + "External id": 81872, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81872, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81872, "pid": 0, "tid": 7, "ts": 1776649792658065, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658057, "dur": 4, + "args": { + "External id": 81872, + "cbid": 211, "correlation": 81872 + } + }, + { + "ph": "s", "id": 81872, "pid": 435, "tid": 435, "ts": 1776649792658057, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792658100, "dur": 1, + "args": { + "External id": 10766, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81890, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81890, "pid": 0, "tid": 7, "ts": 1776649792658100, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658091, "dur": 6, + "args": { + "External id": 10766, + "cbid": 211, "correlation": 81890 + } + }, + { + "ph": "s", "id": 81890, "pid": 435, "tid": 435, "ts": 1776649792658091, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658103, "dur": 0, + "args": { + "External id": 81896, + "cbid": 200, "correlation": 81896 + } + }, + { + "ph": "f", "id": 81896, "pid": 435, "tid": 435, "ts": 1776649792658103, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658104, "dur": 0, + "args": { + "External id": 81897, + "cbid": 273, "correlation": 81897 + } + }, + { + "ph": "f", "id": 81897, "pid": 435, "tid": 435, "ts": 1776649792658104, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792658113, "dur": 19, + "args": { + "External id": 81898, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81898, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81898, "pid": 0, "tid": 7, "ts": 1776649792658113, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658105, "dur": 4, + "args": { + "External id": 81898, + "cbid": 211, "correlation": 81898 + } + }, + { + "ph": "s", "id": 81898, "pid": 435, "tid": 435, "ts": 1776649792658105, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792658148, "dur": 1, + "args": { + "External id": 10771, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81916, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81916, "pid": 0, "tid": 7, "ts": 1776649792658148, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658139, "dur": 6, + "args": { + "External id": 10771, + "cbid": 211, "correlation": 81916 + } + }, + { + "ph": "s", "id": 81916, "pid": 435, "tid": 435, "ts": 1776649792658139, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658151, "dur": 0, + "args": { + "External id": 81922, + "cbid": 200, "correlation": 81922 + } + }, + { + "ph": "f", "id": 81922, "pid": 435, "tid": 435, "ts": 1776649792658151, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658152, "dur": 0, + "args": { + "External id": 81923, + "cbid": 273, "correlation": 81923 + } + }, + { + "ph": "f", "id": 81923, "pid": 435, "tid": 435, "ts": 1776649792658152, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792658161, "dur": 19, + "args": { + "External id": 81924, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81924, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81924, "pid": 0, "tid": 7, "ts": 1776649792658161, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658153, "dur": 4, + "args": { + "External id": 81924, + "cbid": 211, "correlation": 81924 + } + }, + { + "ph": "s", "id": 81924, "pid": 435, "tid": 435, "ts": 1776649792658153, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792658196, "dur": 1, + "args": { + "External id": 10776, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81942, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81942, "pid": 0, "tid": 7, "ts": 1776649792658196, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658187, "dur": 6, + "args": { + "External id": 10776, + "cbid": 211, "correlation": 81942 + } + }, + { + "ph": "s", "id": 81942, "pid": 435, "tid": 435, "ts": 1776649792658187, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658200, "dur": 0, + "args": { + "External id": 81948, + "cbid": 200, "correlation": 81948 + } + }, + { + "ph": "f", "id": 81948, "pid": 435, "tid": 435, "ts": 1776649792658200, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658200, "dur": 0, + "args": { + "External id": 81949, + "cbid": 273, "correlation": 81949 + } + }, + { + "ph": "f", "id": 81949, "pid": 435, "tid": 435, "ts": 1776649792658200, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792658209, "dur": 19, + "args": { + "External id": 81950, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81950, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81950, "pid": 0, "tid": 7, "ts": 1776649792658209, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658201, "dur": 4, + "args": { + "External id": 81950, + "cbid": 211, "correlation": 81950 + } + }, + { + "ph": "s", "id": 81950, "pid": 435, "tid": 435, "ts": 1776649792658201, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792658245, "dur": 1, + "args": { + "External id": 10781, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81968, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81968, "pid": 0, "tid": 7, "ts": 1776649792658245, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658236, "dur": 6, + "args": { + "External id": 10781, + "cbid": 211, "correlation": 81968 + } + }, + { + "ph": "s", "id": 81968, "pid": 435, "tid": 435, "ts": 1776649792658236, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658248, "dur": 0, + "args": { + "External id": 81974, + "cbid": 200, "correlation": 81974 + } + }, + { + "ph": "f", "id": 81974, "pid": 435, "tid": 435, "ts": 1776649792658248, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658249, "dur": 0, + "args": { + "External id": 81975, + "cbid": 273, "correlation": 81975 + } + }, + { + "ph": "f", "id": 81975, "pid": 435, "tid": 435, "ts": 1776649792658249, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792658258, "dur": 19, + "args": { + "External id": 81976, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81976, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81976, "pid": 0, "tid": 7, "ts": 1776649792658258, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658250, "dur": 4, + "args": { + "External id": 81976, + "cbid": 211, "correlation": 81976 + } + }, + { + "ph": "s", "id": 81976, "pid": 435, "tid": 435, "ts": 1776649792658250, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792658293, "dur": 1, + "args": { + "External id": 10786, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 81994, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 81994, "pid": 0, "tid": 7, "ts": 1776649792658293, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658284, "dur": 6, + "args": { + "External id": 10786, + "cbid": 211, "correlation": 81994 + } + }, + { + "ph": "s", "id": 81994, "pid": 435, "tid": 435, "ts": 1776649792658284, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658296, "dur": 0, + "args": { + "External id": 82000, + "cbid": 200, "correlation": 82000 + } + }, + { + "ph": "f", "id": 82000, "pid": 435, "tid": 435, "ts": 1776649792658296, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658297, "dur": 0, + "args": { + "External id": 82001, + "cbid": 273, "correlation": 82001 + } + }, + { + "ph": "f", "id": 82001, "pid": 435, "tid": 435, "ts": 1776649792658297, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792658306, "dur": 19, + "args": { + "External id": 82002, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82002, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82002, "pid": 0, "tid": 7, "ts": 1776649792658306, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658298, "dur": 5, + "args": { + "External id": 82002, + "cbid": 211, "correlation": 82002 + } + }, + { + "ph": "s", "id": 82002, "pid": 435, "tid": 435, "ts": 1776649792658298, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792658341, "dur": 1, + "args": { + "External id": 10791, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82020, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82020, "pid": 0, "tid": 7, "ts": 1776649792658341, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658332, "dur": 6, + "args": { + "External id": 10791, + "cbid": 211, "correlation": 82020 + } + }, + { + "ph": "s", "id": 82020, "pid": 435, "tid": 435, "ts": 1776649792658332, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658344, "dur": 0, + "args": { + "External id": 82026, + "cbid": 200, "correlation": 82026 + } + }, + { + "ph": "f", "id": 82026, "pid": 435, "tid": 435, "ts": 1776649792658344, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658345, "dur": 0, + "args": { + "External id": 82027, + "cbid": 273, "correlation": 82027 + } + }, + { + "ph": "f", "id": 82027, "pid": 435, "tid": 435, "ts": 1776649792658345, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792658354, "dur": 19, + "args": { + "External id": 82028, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82028, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82028, "pid": 0, "tid": 7, "ts": 1776649792658354, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658346, "dur": 4, + "args": { + "External id": 82028, + "cbid": 211, "correlation": 82028 + } + }, + { + "ph": "s", "id": 82028, "pid": 435, "tid": 435, "ts": 1776649792658346, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792658389, "dur": 1, + "args": { + "External id": 10796, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82046, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82046, "pid": 0, "tid": 7, "ts": 1776649792658389, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658380, "dur": 7, + "args": { + "External id": 10796, + "cbid": 211, "correlation": 82046 + } + }, + { + "ph": "s", "id": 82046, "pid": 435, "tid": 435, "ts": 1776649792658380, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658392, "dur": 0, + "args": { + "External id": 82052, + "cbid": 200, "correlation": 82052 + } + }, + { + "ph": "f", "id": 82052, "pid": 435, "tid": 435, "ts": 1776649792658392, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658393, "dur": 0, + "args": { + "External id": 82053, + "cbid": 273, "correlation": 82053 + } + }, + { + "ph": "f", "id": 82053, "pid": 435, "tid": 435, "ts": 1776649792658393, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792658402, "dur": 19, + "args": { + "External id": 82054, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82054, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82054, "pid": 0, "tid": 7, "ts": 1776649792658402, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658394, "dur": 5, + "args": { + "External id": 82054, + "cbid": 211, "correlation": 82054 + } + }, + { + "ph": "s", "id": 82054, "pid": 435, "tid": 435, "ts": 1776649792658394, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792658437, "dur": 1, + "args": { + "External id": 10801, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82072, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82072, "pid": 0, "tid": 7, "ts": 1776649792658437, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658428, "dur": 6, + "args": { + "External id": 10801, + "cbid": 211, "correlation": 82072 + } + }, + { + "ph": "s", "id": 82072, "pid": 435, "tid": 435, "ts": 1776649792658428, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658440, "dur": 0, + "args": { + "External id": 82078, + "cbid": 200, "correlation": 82078 + } + }, + { + "ph": "f", "id": 82078, "pid": 435, "tid": 435, "ts": 1776649792658440, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658441, "dur": 0, + "args": { + "External id": 82079, + "cbid": 273, "correlation": 82079 + } + }, + { + "ph": "f", "id": 82079, "pid": 435, "tid": 435, "ts": 1776649792658441, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792658450, "dur": 19, + "args": { + "External id": 82080, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82080, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82080, "pid": 0, "tid": 7, "ts": 1776649792658450, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658442, "dur": 4, + "args": { + "External id": 82080, + "cbid": 211, "correlation": 82080 + } + }, + { + "ph": "s", "id": 82080, "pid": 435, "tid": 435, "ts": 1776649792658442, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792658485, "dur": 1, + "args": { + "External id": 10806, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82098, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82098, "pid": 0, "tid": 7, "ts": 1776649792658485, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658476, "dur": 6, + "args": { + "External id": 10806, + "cbid": 211, "correlation": 82098 + } + }, + { + "ph": "s", "id": 82098, "pid": 435, "tid": 435, "ts": 1776649792658476, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658488, "dur": 0, + "args": { + "External id": 82104, + "cbid": 200, "correlation": 82104 + } + }, + { + "ph": "f", "id": 82104, "pid": 435, "tid": 435, "ts": 1776649792658488, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658489, "dur": 0, + "args": { + "External id": 82105, + "cbid": 273, "correlation": 82105 + } + }, + { + "ph": "f", "id": 82105, "pid": 435, "tid": 435, "ts": 1776649792658489, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792658498, "dur": 19, + "args": { + "External id": 82106, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82106, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82106, "pid": 0, "tid": 7, "ts": 1776649792658498, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658490, "dur": 5, + "args": { + "External id": 82106, + "cbid": 211, "correlation": 82106 + } + }, + { + "ph": "s", "id": 82106, "pid": 435, "tid": 435, "ts": 1776649792658490, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792658533, "dur": 1, + "args": { + "External id": 10811, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82124, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82124, "pid": 0, "tid": 7, "ts": 1776649792658533, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658524, "dur": 6, + "args": { + "External id": 10811, + "cbid": 211, "correlation": 82124 + } + }, + { + "ph": "s", "id": 82124, "pid": 435, "tid": 435, "ts": 1776649792658524, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658536, "dur": 0, + "args": { + "External id": 82130, + "cbid": 200, "correlation": 82130 + } + }, + { + "ph": "f", "id": 82130, "pid": 435, "tid": 435, "ts": 1776649792658536, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658537, "dur": 0, + "args": { + "External id": 82131, + "cbid": 273, "correlation": 82131 + } + }, + { + "ph": "f", "id": 82131, "pid": 435, "tid": 435, "ts": 1776649792658537, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792658546, "dur": 19, + "args": { + "External id": 82132, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82132, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82132, "pid": 0, "tid": 7, "ts": 1776649792658546, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658538, "dur": 4, + "args": { + "External id": 82132, + "cbid": 211, "correlation": 82132 + } + }, + { + "ph": "s", "id": 82132, "pid": 435, "tid": 435, "ts": 1776649792658538, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792658580, "dur": 1, + "args": { + "External id": 10816, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82150, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82150, "pid": 0, "tid": 7, "ts": 1776649792658580, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658572, "dur": 6, + "args": { + "External id": 10816, + "cbid": 211, "correlation": 82150 + } + }, + { + "ph": "s", "id": 82150, "pid": 435, "tid": 435, "ts": 1776649792658572, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658584, "dur": 0, + "args": { + "External id": 82156, + "cbid": 200, "correlation": 82156 + } + }, + { + "ph": "f", "id": 82156, "pid": 435, "tid": 435, "ts": 1776649792658584, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658584, "dur": 0, + "args": { + "External id": 82157, + "cbid": 273, "correlation": 82157 + } + }, + { + "ph": "f", "id": 82157, "pid": 435, "tid": 435, "ts": 1776649792658584, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792658593, "dur": 19, + "args": { + "External id": 82158, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82158, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82158, "pid": 0, "tid": 7, "ts": 1776649792658593, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658585, "dur": 5, + "args": { + "External id": 82158, + "cbid": 211, "correlation": 82158 + } + }, + { + "ph": "s", "id": 82158, "pid": 435, "tid": 435, "ts": 1776649792658585, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792658629, "dur": 1, + "args": { + "External id": 10821, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82176, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82176, "pid": 0, "tid": 7, "ts": 1776649792658629, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658620, "dur": 6, + "args": { + "External id": 10821, + "cbid": 211, "correlation": 82176 + } + }, + { + "ph": "s", "id": 82176, "pid": 435, "tid": 435, "ts": 1776649792658620, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658632, "dur": 0, + "args": { + "External id": 82182, + "cbid": 200, "correlation": 82182 + } + }, + { + "ph": "f", "id": 82182, "pid": 435, "tid": 435, "ts": 1776649792658632, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658633, "dur": 0, + "args": { + "External id": 82183, + "cbid": 273, "correlation": 82183 + } + }, + { + "ph": "f", "id": 82183, "pid": 435, "tid": 435, "ts": 1776649792658633, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792658642, "dur": 19, + "args": { + "External id": 82184, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82184, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82184, "pid": 0, "tid": 7, "ts": 1776649792658642, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658634, "dur": 4, + "args": { + "External id": 82184, + "cbid": 211, "correlation": 82184 + } + }, + { + "ph": "s", "id": 82184, "pid": 435, "tid": 435, "ts": 1776649792658634, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792658677, "dur": 1, + "args": { + "External id": 10826, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82202, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82202, "pid": 0, "tid": 7, "ts": 1776649792658677, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658667, "dur": 6, + "args": { + "External id": 10826, + "cbid": 211, "correlation": 82202 + } + }, + { + "ph": "s", "id": 82202, "pid": 435, "tid": 435, "ts": 1776649792658667, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658680, "dur": 0, + "args": { + "External id": 82208, + "cbid": 200, "correlation": 82208 + } + }, + { + "ph": "f", "id": 82208, "pid": 435, "tid": 435, "ts": 1776649792658680, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658680, "dur": 0, + "args": { + "External id": 82209, + "cbid": 273, "correlation": 82209 + } + }, + { + "ph": "f", "id": 82209, "pid": 435, "tid": 435, "ts": 1776649792658680, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792658689, "dur": 19, + "args": { + "External id": 82210, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82210, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82210, "pid": 0, "tid": 7, "ts": 1776649792658689, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658682, "dur": 5, + "args": { + "External id": 82210, + "cbid": 211, "correlation": 82210 + } + }, + { + "ph": "s", "id": 82210, "pid": 435, "tid": 435, "ts": 1776649792658682, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792658724, "dur": 1, + "args": { + "External id": 10831, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82228, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82228, "pid": 0, "tid": 7, "ts": 1776649792658724, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658715, "dur": 6, + "args": { + "External id": 10831, + "cbid": 211, "correlation": 82228 + } + }, + { + "ph": "s", "id": 82228, "pid": 435, "tid": 435, "ts": 1776649792658715, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658727, "dur": 0, + "args": { + "External id": 82234, + "cbid": 200, "correlation": 82234 + } + }, + { + "ph": "f", "id": 82234, "pid": 435, "tid": 435, "ts": 1776649792658727, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658728, "dur": 0, + "args": { + "External id": 82235, + "cbid": 273, "correlation": 82235 + } + }, + { + "ph": "f", "id": 82235, "pid": 435, "tid": 435, "ts": 1776649792658728, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792658737, "dur": 19, + "args": { + "External id": 82236, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82236, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82236, "pid": 0, "tid": 7, "ts": 1776649792658737, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658729, "dur": 4, + "args": { + "External id": 82236, + "cbid": 211, "correlation": 82236 + } + }, + { + "ph": "s", "id": 82236, "pid": 435, "tid": 435, "ts": 1776649792658729, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792658772, "dur": 1, + "args": { + "External id": 10836, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82254, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82254, "pid": 0, "tid": 7, "ts": 1776649792658772, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658763, "dur": 6, + "args": { + "External id": 10836, + "cbid": 211, "correlation": 82254 + } + }, + { + "ph": "s", "id": 82254, "pid": 435, "tid": 435, "ts": 1776649792658763, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658775, "dur": 0, + "args": { + "External id": 82260, + "cbid": 200, "correlation": 82260 + } + }, + { + "ph": "f", "id": 82260, "pid": 435, "tid": 435, "ts": 1776649792658775, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658776, "dur": 0, + "args": { + "External id": 82261, + "cbid": 273, "correlation": 82261 + } + }, + { + "ph": "f", "id": 82261, "pid": 435, "tid": 435, "ts": 1776649792658776, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792658784, "dur": 19, + "args": { + "External id": 82262, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82262, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82262, "pid": 0, "tid": 7, "ts": 1776649792658784, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658777, "dur": 4, + "args": { + "External id": 82262, + "cbid": 211, "correlation": 82262 + } + }, + { + "ph": "s", "id": 82262, "pid": 435, "tid": 435, "ts": 1776649792658777, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792658820, "dur": 1, + "args": { + "External id": 10841, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82280, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82280, "pid": 0, "tid": 7, "ts": 1776649792658820, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658811, "dur": 6, + "args": { + "External id": 10841, + "cbid": 211, "correlation": 82280 + } + }, + { + "ph": "s", "id": 82280, "pid": 435, "tid": 435, "ts": 1776649792658811, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658823, "dur": 0, + "args": { + "External id": 82286, + "cbid": 200, "correlation": 82286 + } + }, + { + "ph": "f", "id": 82286, "pid": 435, "tid": 435, "ts": 1776649792658823, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658824, "dur": 0, + "args": { + "External id": 82287, + "cbid": 273, "correlation": 82287 + } + }, + { + "ph": "f", "id": 82287, "pid": 435, "tid": 435, "ts": 1776649792658824, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792658833, "dur": 19, + "args": { + "External id": 82288, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82288, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82288, "pid": 0, "tid": 7, "ts": 1776649792658833, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658825, "dur": 4, + "args": { + "External id": 82288, + "cbid": 211, "correlation": 82288 + } + }, + { + "ph": "s", "id": 82288, "pid": 435, "tid": 435, "ts": 1776649792658825, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792658868, "dur": 1, + "args": { + "External id": 10846, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82306, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82306, "pid": 0, "tid": 7, "ts": 1776649792658868, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658859, "dur": 6, + "args": { + "External id": 10846, + "cbid": 211, "correlation": 82306 + } + }, + { + "ph": "s", "id": 82306, "pid": 435, "tid": 435, "ts": 1776649792658859, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658871, "dur": 0, + "args": { + "External id": 82312, + "cbid": 200, "correlation": 82312 + } + }, + { + "ph": "f", "id": 82312, "pid": 435, "tid": 435, "ts": 1776649792658871, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658872, "dur": 0, + "args": { + "External id": 82313, + "cbid": 273, "correlation": 82313 + } + }, + { + "ph": "f", "id": 82313, "pid": 435, "tid": 435, "ts": 1776649792658872, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792658881, "dur": 19, + "args": { + "External id": 82314, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82314, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82314, "pid": 0, "tid": 7, "ts": 1776649792658881, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658873, "dur": 5, + "args": { + "External id": 82314, + "cbid": 211, "correlation": 82314 + } + }, + { + "ph": "s", "id": 82314, "pid": 435, "tid": 435, "ts": 1776649792658873, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792658917, "dur": 1, + "args": { + "External id": 10851, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82332, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82332, "pid": 0, "tid": 7, "ts": 1776649792658917, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658907, "dur": 6, + "args": { + "External id": 10851, + "cbid": 211, "correlation": 82332 + } + }, + { + "ph": "s", "id": 82332, "pid": 435, "tid": 435, "ts": 1776649792658907, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658919, "dur": 0, + "args": { + "External id": 82338, + "cbid": 200, "correlation": 82338 + } + }, + { + "ph": "f", "id": 82338, "pid": 435, "tid": 435, "ts": 1776649792658919, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658920, "dur": 0, + "args": { + "External id": 82339, + "cbid": 273, "correlation": 82339 + } + }, + { + "ph": "f", "id": 82339, "pid": 435, "tid": 435, "ts": 1776649792658920, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792658929, "dur": 19, + "args": { + "External id": 82340, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82340, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82340, "pid": 0, "tid": 7, "ts": 1776649792658929, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658921, "dur": 4, + "args": { + "External id": 82340, + "cbid": 211, "correlation": 82340 + } + }, + { + "ph": "s", "id": 82340, "pid": 435, "tid": 435, "ts": 1776649792658921, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792658964, "dur": 1, + "args": { + "External id": 10856, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82358, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82358, "pid": 0, "tid": 7, "ts": 1776649792658964, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658955, "dur": 6, + "args": { + "External id": 10856, + "cbid": 211, "correlation": 82358 + } + }, + { + "ph": "s", "id": 82358, "pid": 435, "tid": 435, "ts": 1776649792658955, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658967, "dur": 0, + "args": { + "External id": 82364, + "cbid": 200, "correlation": 82364 + } + }, + { + "ph": "f", "id": 82364, "pid": 435, "tid": 435, "ts": 1776649792658967, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792658968, "dur": 0, + "args": { + "External id": 82365, + "cbid": 273, "correlation": 82365 + } + }, + { + "ph": "f", "id": 82365, "pid": 435, "tid": 435, "ts": 1776649792658968, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792658977, "dur": 19, + "args": { + "External id": 82366, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82366, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82366, "pid": 0, "tid": 7, "ts": 1776649792658977, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792658969, "dur": 5, + "args": { + "External id": 82366, + "cbid": 211, "correlation": 82366 + } + }, + { + "ph": "s", "id": 82366, "pid": 435, "tid": 435, "ts": 1776649792658969, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792659014, "dur": 1, + "args": { + "External id": 10861, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82384, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82384, "pid": 0, "tid": 7, "ts": 1776649792659014, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659004, "dur": 6, + "args": { + "External id": 10861, + "cbid": 211, "correlation": 82384 + } + }, + { + "ph": "s", "id": 82384, "pid": 435, "tid": 435, "ts": 1776649792659004, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659017, "dur": 0, + "args": { + "External id": 82390, + "cbid": 200, "correlation": 82390 + } + }, + { + "ph": "f", "id": 82390, "pid": 435, "tid": 435, "ts": 1776649792659017, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659017, "dur": 0, + "args": { + "External id": 82391, + "cbid": 273, "correlation": 82391 + } + }, + { + "ph": "f", "id": 82391, "pid": 435, "tid": 435, "ts": 1776649792659017, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792659026, "dur": 19, + "args": { + "External id": 82392, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82392, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82392, "pid": 0, "tid": 7, "ts": 1776649792659026, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659018, "dur": 4, + "args": { + "External id": 82392, + "cbid": 211, "correlation": 82392 + } + }, + { + "ph": "s", "id": 82392, "pid": 435, "tid": 435, "ts": 1776649792659018, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792659061, "dur": 1, + "args": { + "External id": 10866, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82410, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82410, "pid": 0, "tid": 7, "ts": 1776649792659061, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659052, "dur": 6, + "args": { + "External id": 10866, + "cbid": 211, "correlation": 82410 + } + }, + { + "ph": "s", "id": 82410, "pid": 435, "tid": 435, "ts": 1776649792659052, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659064, "dur": 0, + "args": { + "External id": 82416, + "cbid": 200, "correlation": 82416 + } + }, + { + "ph": "f", "id": 82416, "pid": 435, "tid": 435, "ts": 1776649792659064, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659065, "dur": 1, + "args": { + "External id": 82417, + "cbid": 273, "correlation": 82417 + } + }, + { + "ph": "f", "id": 82417, "pid": 435, "tid": 435, "ts": 1776649792659065, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792659074, "dur": 19, + "args": { + "External id": 82418, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82418, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82418, "pid": 0, "tid": 7, "ts": 1776649792659074, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659066, "dur": 4, + "args": { + "External id": 82418, + "cbid": 211, "correlation": 82418 + } + }, + { + "ph": "s", "id": 82418, "pid": 435, "tid": 435, "ts": 1776649792659066, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792659109, "dur": 1, + "args": { + "External id": 10871, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82436, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82436, "pid": 0, "tid": 7, "ts": 1776649792659109, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659100, "dur": 6, + "args": { + "External id": 10871, + "cbid": 211, "correlation": 82436 + } + }, + { + "ph": "s", "id": 82436, "pid": 435, "tid": 435, "ts": 1776649792659100, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659112, "dur": 0, + "args": { + "External id": 82442, + "cbid": 200, "correlation": 82442 + } + }, + { + "ph": "f", "id": 82442, "pid": 435, "tid": 435, "ts": 1776649792659112, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659113, "dur": 0, + "args": { + "External id": 82443, + "cbid": 273, "correlation": 82443 + } + }, + { + "ph": "f", "id": 82443, "pid": 435, "tid": 435, "ts": 1776649792659113, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792659122, "dur": 19, + "args": { + "External id": 82444, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82444, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82444, "pid": 0, "tid": 7, "ts": 1776649792659122, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659114, "dur": 4, + "args": { + "External id": 82444, + "cbid": 211, "correlation": 82444 + } + }, + { + "ph": "s", "id": 82444, "pid": 435, "tid": 435, "ts": 1776649792659114, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792659157, "dur": 1, + "args": { + "External id": 10876, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82462, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82462, "pid": 0, "tid": 7, "ts": 1776649792659157, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659148, "dur": 6, + "args": { + "External id": 10876, + "cbid": 211, "correlation": 82462 + } + }, + { + "ph": "s", "id": 82462, "pid": 435, "tid": 435, "ts": 1776649792659148, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659160, "dur": 0, + "args": { + "External id": 82468, + "cbid": 200, "correlation": 82468 + } + }, + { + "ph": "f", "id": 82468, "pid": 435, "tid": 435, "ts": 1776649792659160, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659161, "dur": 0, + "args": { + "External id": 82469, + "cbid": 273, "correlation": 82469 + } + }, + { + "ph": "f", "id": 82469, "pid": 435, "tid": 435, "ts": 1776649792659161, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792659170, "dur": 19, + "args": { + "External id": 82470, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82470, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82470, "pid": 0, "tid": 7, "ts": 1776649792659170, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659162, "dur": 5, + "args": { + "External id": 82470, + "cbid": 211, "correlation": 82470 + } + }, + { + "ph": "s", "id": 82470, "pid": 435, "tid": 435, "ts": 1776649792659162, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792659205, "dur": 1, + "args": { + "External id": 10881, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82488, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82488, "pid": 0, "tid": 7, "ts": 1776649792659205, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659196, "dur": 6, + "args": { + "External id": 10881, + "cbid": 211, "correlation": 82488 + } + }, + { + "ph": "s", "id": 82488, "pid": 435, "tid": 435, "ts": 1776649792659196, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659208, "dur": 0, + "args": { + "External id": 82494, + "cbid": 200, "correlation": 82494 + } + }, + { + "ph": "f", "id": 82494, "pid": 435, "tid": 435, "ts": 1776649792659208, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659209, "dur": 0, + "args": { + "External id": 82495, + "cbid": 273, "correlation": 82495 + } + }, + { + "ph": "f", "id": 82495, "pid": 435, "tid": 435, "ts": 1776649792659209, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792659218, "dur": 19, + "args": { + "External id": 82496, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82496, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82496, "pid": 0, "tid": 7, "ts": 1776649792659218, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659210, "dur": 5, + "args": { + "External id": 82496, + "cbid": 211, "correlation": 82496 + } + }, + { + "ph": "s", "id": 82496, "pid": 435, "tid": 435, "ts": 1776649792659210, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792659253, "dur": 1, + "args": { + "External id": 10886, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82514, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82514, "pid": 0, "tid": 7, "ts": 1776649792659253, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659244, "dur": 6, + "args": { + "External id": 10886, + "cbid": 211, "correlation": 82514 + } + }, + { + "ph": "s", "id": 82514, "pid": 435, "tid": 435, "ts": 1776649792659244, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659256, "dur": 0, + "args": { + "External id": 82520, + "cbid": 200, "correlation": 82520 + } + }, + { + "ph": "f", "id": 82520, "pid": 435, "tid": 435, "ts": 1776649792659256, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659257, "dur": 0, + "args": { + "External id": 82521, + "cbid": 273, "correlation": 82521 + } + }, + { + "ph": "f", "id": 82521, "pid": 435, "tid": 435, "ts": 1776649792659257, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792659266, "dur": 19, + "args": { + "External id": 82522, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82522, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82522, "pid": 0, "tid": 7, "ts": 1776649792659266, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659258, "dur": 5, + "args": { + "External id": 82522, + "cbid": 211, "correlation": 82522 + } + }, + { + "ph": "s", "id": 82522, "pid": 435, "tid": 435, "ts": 1776649792659258, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792659302, "dur": 1, + "args": { + "External id": 10891, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82540, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82540, "pid": 0, "tid": 7, "ts": 1776649792659302, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659293, "dur": 6, + "args": { + "External id": 10891, + "cbid": 211, "correlation": 82540 + } + }, + { + "ph": "s", "id": 82540, "pid": 435, "tid": 435, "ts": 1776649792659293, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659305, "dur": 0, + "args": { + "External id": 82546, + "cbid": 200, "correlation": 82546 + } + }, + { + "ph": "f", "id": 82546, "pid": 435, "tid": 435, "ts": 1776649792659305, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659306, "dur": 0, + "args": { + "External id": 82547, + "cbid": 273, "correlation": 82547 + } + }, + { + "ph": "f", "id": 82547, "pid": 435, "tid": 435, "ts": 1776649792659306, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792659315, "dur": 19, + "args": { + "External id": 82548, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82548, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82548, "pid": 0, "tid": 7, "ts": 1776649792659315, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659307, "dur": 4, + "args": { + "External id": 82548, + "cbid": 211, "correlation": 82548 + } + }, + { + "ph": "s", "id": 82548, "pid": 435, "tid": 435, "ts": 1776649792659307, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792659350, "dur": 1, + "args": { + "External id": 10896, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82566, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82566, "pid": 0, "tid": 7, "ts": 1776649792659350, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659340, "dur": 7, + "args": { + "External id": 10896, + "cbid": 211, "correlation": 82566 + } + }, + { + "ph": "s", "id": 82566, "pid": 435, "tid": 435, "ts": 1776649792659340, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659353, "dur": 0, + "args": { + "External id": 82572, + "cbid": 200, "correlation": 82572 + } + }, + { + "ph": "f", "id": 82572, "pid": 435, "tid": 435, "ts": 1776649792659353, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659354, "dur": 0, + "args": { + "External id": 82573, + "cbid": 273, "correlation": 82573 + } + }, + { + "ph": "f", "id": 82573, "pid": 435, "tid": 435, "ts": 1776649792659354, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792659363, "dur": 19, + "args": { + "External id": 82574, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82574, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82574, "pid": 0, "tid": 7, "ts": 1776649792659363, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659355, "dur": 5, + "args": { + "External id": 82574, + "cbid": 211, "correlation": 82574 + } + }, + { + "ph": "s", "id": 82574, "pid": 435, "tid": 435, "ts": 1776649792659355, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792659398, "dur": 1, + "args": { + "External id": 10901, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82592, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82592, "pid": 0, "tid": 7, "ts": 1776649792659398, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659389, "dur": 6, + "args": { + "External id": 10901, + "cbid": 211, "correlation": 82592 + } + }, + { + "ph": "s", "id": 82592, "pid": 435, "tid": 435, "ts": 1776649792659389, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659401, "dur": 0, + "args": { + "External id": 82598, + "cbid": 200, "correlation": 82598 + } + }, + { + "ph": "f", "id": 82598, "pid": 435, "tid": 435, "ts": 1776649792659401, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659401, "dur": 0, + "args": { + "External id": 82599, + "cbid": 273, "correlation": 82599 + } + }, + { + "ph": "f", "id": 82599, "pid": 435, "tid": 435, "ts": 1776649792659401, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792659411, "dur": 19, + "args": { + "External id": 82600, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82600, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82600, "pid": 0, "tid": 7, "ts": 1776649792659411, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659403, "dur": 4, + "args": { + "External id": 82600, + "cbid": 211, "correlation": 82600 + } + }, + { + "ph": "s", "id": 82600, "pid": 435, "tid": 435, "ts": 1776649792659403, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792659445, "dur": 1, + "args": { + "External id": 10906, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82618, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82618, "pid": 0, "tid": 7, "ts": 1776649792659445, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659436, "dur": 6, + "args": { + "External id": 10906, + "cbid": 211, "correlation": 82618 + } + }, + { + "ph": "s", "id": 82618, "pid": 435, "tid": 435, "ts": 1776649792659436, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659449, "dur": 0, + "args": { + "External id": 82624, + "cbid": 200, "correlation": 82624 + } + }, + { + "ph": "f", "id": 82624, "pid": 435, "tid": 435, "ts": 1776649792659449, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659449, "dur": 0, + "args": { + "External id": 82625, + "cbid": 273, "correlation": 82625 + } + }, + { + "ph": "f", "id": 82625, "pid": 435, "tid": 435, "ts": 1776649792659449, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792659458, "dur": 19, + "args": { + "External id": 82626, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82626, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82626, "pid": 0, "tid": 7, "ts": 1776649792659458, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659450, "dur": 5, + "args": { + "External id": 82626, + "cbid": 211, "correlation": 82626 + } + }, + { + "ph": "s", "id": 82626, "pid": 435, "tid": 435, "ts": 1776649792659450, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792659493, "dur": 1, + "args": { + "External id": 10911, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82644, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82644, "pid": 0, "tid": 7, "ts": 1776649792659493, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659484, "dur": 6, + "args": { + "External id": 10911, + "cbid": 211, "correlation": 82644 + } + }, + { + "ph": "s", "id": 82644, "pid": 435, "tid": 435, "ts": 1776649792659484, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659496, "dur": 0, + "args": { + "External id": 82650, + "cbid": 200, "correlation": 82650 + } + }, + { + "ph": "f", "id": 82650, "pid": 435, "tid": 435, "ts": 1776649792659496, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659497, "dur": 0, + "args": { + "External id": 82651, + "cbid": 273, "correlation": 82651 + } + }, + { + "ph": "f", "id": 82651, "pid": 435, "tid": 435, "ts": 1776649792659497, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792659506, "dur": 19, + "args": { + "External id": 82652, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82652, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82652, "pid": 0, "tid": 7, "ts": 1776649792659506, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659498, "dur": 5, + "args": { + "External id": 82652, + "cbid": 211, "correlation": 82652 + } + }, + { + "ph": "s", "id": 82652, "pid": 435, "tid": 435, "ts": 1776649792659498, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792659541, "dur": 1, + "args": { + "External id": 10916, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82670, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82670, "pid": 0, "tid": 7, "ts": 1776649792659541, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659532, "dur": 6, + "args": { + "External id": 10916, + "cbid": 211, "correlation": 82670 + } + }, + { + "ph": "s", "id": 82670, "pid": 435, "tid": 435, "ts": 1776649792659532, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659544, "dur": 0, + "args": { + "External id": 82676, + "cbid": 200, "correlation": 82676 + } + }, + { + "ph": "f", "id": 82676, "pid": 435, "tid": 435, "ts": 1776649792659544, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659545, "dur": 0, + "args": { + "External id": 82677, + "cbid": 273, "correlation": 82677 + } + }, + { + "ph": "f", "id": 82677, "pid": 435, "tid": 435, "ts": 1776649792659545, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792659554, "dur": 19, + "args": { + "External id": 82678, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82678, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82678, "pid": 0, "tid": 7, "ts": 1776649792659554, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659546, "dur": 5, + "args": { + "External id": 82678, + "cbid": 211, "correlation": 82678 + } + }, + { + "ph": "s", "id": 82678, "pid": 435, "tid": 435, "ts": 1776649792659546, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792659590, "dur": 1, + "args": { + "External id": 10921, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82696, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82696, "pid": 0, "tid": 7, "ts": 1776649792659590, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659580, "dur": 6, + "args": { + "External id": 10921, + "cbid": 211, "correlation": 82696 + } + }, + { + "ph": "s", "id": 82696, "pid": 435, "tid": 435, "ts": 1776649792659580, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659593, "dur": 0, + "args": { + "External id": 82702, + "cbid": 200, "correlation": 82702 + } + }, + { + "ph": "f", "id": 82702, "pid": 435, "tid": 435, "ts": 1776649792659593, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659593, "dur": 0, + "args": { + "External id": 82703, + "cbid": 273, "correlation": 82703 + } + }, + { + "ph": "f", "id": 82703, "pid": 435, "tid": 435, "ts": 1776649792659593, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792659603, "dur": 19, + "args": { + "External id": 82704, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82704, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82704, "pid": 0, "tid": 7, "ts": 1776649792659603, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659594, "dur": 4, + "args": { + "External id": 82704, + "cbid": 211, "correlation": 82704 + } + }, + { + "ph": "s", "id": 82704, "pid": 435, "tid": 435, "ts": 1776649792659594, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792659637, "dur": 1, + "args": { + "External id": 10926, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82722, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82722, "pid": 0, "tid": 7, "ts": 1776649792659637, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659628, "dur": 6, + "args": { + "External id": 10926, + "cbid": 211, "correlation": 82722 + } + }, + { + "ph": "s", "id": 82722, "pid": 435, "tid": 435, "ts": 1776649792659628, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659640, "dur": 0, + "args": { + "External id": 82728, + "cbid": 200, "correlation": 82728 + } + }, + { + "ph": "f", "id": 82728, "pid": 435, "tid": 435, "ts": 1776649792659640, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659641, "dur": 0, + "args": { + "External id": 82729, + "cbid": 273, "correlation": 82729 + } + }, + { + "ph": "f", "id": 82729, "pid": 435, "tid": 435, "ts": 1776649792659641, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792659650, "dur": 19, + "args": { + "External id": 82730, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82730, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82730, "pid": 0, "tid": 7, "ts": 1776649792659650, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659642, "dur": 5, + "args": { + "External id": 82730, + "cbid": 211, "correlation": 82730 + } + }, + { + "ph": "s", "id": 82730, "pid": 435, "tid": 435, "ts": 1776649792659642, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792659685, "dur": 1, + "args": { + "External id": 10931, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82748, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82748, "pid": 0, "tid": 7, "ts": 1776649792659685, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659676, "dur": 6, + "args": { + "External id": 10931, + "cbid": 211, "correlation": 82748 + } + }, + { + "ph": "s", "id": 82748, "pid": 435, "tid": 435, "ts": 1776649792659676, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659688, "dur": 0, + "args": { + "External id": 82754, + "cbid": 200, "correlation": 82754 + } + }, + { + "ph": "f", "id": 82754, "pid": 435, "tid": 435, "ts": 1776649792659688, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659689, "dur": 0, + "args": { + "External id": 82755, + "cbid": 273, "correlation": 82755 + } + }, + { + "ph": "f", "id": 82755, "pid": 435, "tid": 435, "ts": 1776649792659689, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792659698, "dur": 19, + "args": { + "External id": 82756, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82756, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82756, "pid": 0, "tid": 7, "ts": 1776649792659698, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659690, "dur": 4, + "args": { + "External id": 82756, + "cbid": 211, "correlation": 82756 + } + }, + { + "ph": "s", "id": 82756, "pid": 435, "tid": 435, "ts": 1776649792659690, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792659733, "dur": 1, + "args": { + "External id": 10936, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82774, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82774, "pid": 0, "tid": 7, "ts": 1776649792659733, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659723, "dur": 6, + "args": { + "External id": 10936, + "cbid": 211, "correlation": 82774 + } + }, + { + "ph": "s", "id": 82774, "pid": 435, "tid": 435, "ts": 1776649792659723, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659736, "dur": 0, + "args": { + "External id": 82780, + "cbid": 200, "correlation": 82780 + } + }, + { + "ph": "f", "id": 82780, "pid": 435, "tid": 435, "ts": 1776649792659736, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659736, "dur": 0, + "args": { + "External id": 82781, + "cbid": 273, "correlation": 82781 + } + }, + { + "ph": "f", "id": 82781, "pid": 435, "tid": 435, "ts": 1776649792659736, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792659745, "dur": 19, + "args": { + "External id": 82782, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82782, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82782, "pid": 0, "tid": 7, "ts": 1776649792659745, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659737, "dur": 5, + "args": { + "External id": 82782, + "cbid": 211, "correlation": 82782 + } + }, + { + "ph": "s", "id": 82782, "pid": 435, "tid": 435, "ts": 1776649792659737, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792659780, "dur": 1, + "args": { + "External id": 10941, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82800, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82800, "pid": 0, "tid": 7, "ts": 1776649792659780, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659771, "dur": 6, + "args": { + "External id": 10941, + "cbid": 211, "correlation": 82800 + } + }, + { + "ph": "s", "id": 82800, "pid": 435, "tid": 435, "ts": 1776649792659771, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659783, "dur": 0, + "args": { + "External id": 82806, + "cbid": 200, "correlation": 82806 + } + }, + { + "ph": "f", "id": 82806, "pid": 435, "tid": 435, "ts": 1776649792659783, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659784, "dur": 0, + "args": { + "External id": 82807, + "cbid": 273, "correlation": 82807 + } + }, + { + "ph": "f", "id": 82807, "pid": 435, "tid": 435, "ts": 1776649792659784, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792659793, "dur": 19, + "args": { + "External id": 82808, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82808, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82808, "pid": 0, "tid": 7, "ts": 1776649792659793, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659785, "dur": 5, + "args": { + "External id": 82808, + "cbid": 211, "correlation": 82808 + } + }, + { + "ph": "s", "id": 82808, "pid": 435, "tid": 435, "ts": 1776649792659785, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792659828, "dur": 1, + "args": { + "External id": 10946, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82826, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82826, "pid": 0, "tid": 7, "ts": 1776649792659828, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659819, "dur": 6, + "args": { + "External id": 10946, + "cbid": 211, "correlation": 82826 + } + }, + { + "ph": "s", "id": 82826, "pid": 435, "tid": 435, "ts": 1776649792659819, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659831, "dur": 0, + "args": { + "External id": 82832, + "cbid": 200, "correlation": 82832 + } + }, + { + "ph": "f", "id": 82832, "pid": 435, "tid": 435, "ts": 1776649792659831, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659831, "dur": 0, + "args": { + "External id": 82833, + "cbid": 273, "correlation": 82833 + } + }, + { + "ph": "f", "id": 82833, "pid": 435, "tid": 435, "ts": 1776649792659831, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792659840, "dur": 19, + "args": { + "External id": 82834, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82834, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82834, "pid": 0, "tid": 7, "ts": 1776649792659840, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659832, "dur": 5, + "args": { + "External id": 82834, + "cbid": 211, "correlation": 82834 + } + }, + { + "ph": "s", "id": 82834, "pid": 435, "tid": 435, "ts": 1776649792659832, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792659876, "dur": 1, + "args": { + "External id": 10951, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82852, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82852, "pid": 0, "tid": 7, "ts": 1776649792659876, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659866, "dur": 6, + "args": { + "External id": 10951, + "cbid": 211, "correlation": 82852 + } + }, + { + "ph": "s", "id": 82852, "pid": 435, "tid": 435, "ts": 1776649792659866, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659879, "dur": 0, + "args": { + "External id": 82858, + "cbid": 200, "correlation": 82858 + } + }, + { + "ph": "f", "id": 82858, "pid": 435, "tid": 435, "ts": 1776649792659879, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659880, "dur": 0, + "args": { + "External id": 82859, + "cbid": 273, "correlation": 82859 + } + }, + { + "ph": "f", "id": 82859, "pid": 435, "tid": 435, "ts": 1776649792659880, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792659889, "dur": 19, + "args": { + "External id": 82860, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82860, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82860, "pid": 0, "tid": 7, "ts": 1776649792659889, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659881, "dur": 5, + "args": { + "External id": 82860, + "cbid": 211, "correlation": 82860 + } + }, + { + "ph": "s", "id": 82860, "pid": 435, "tid": 435, "ts": 1776649792659881, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792659924, "dur": 1, + "args": { + "External id": 10956, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82878, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82878, "pid": 0, "tid": 7, "ts": 1776649792659924, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659914, "dur": 6, + "args": { + "External id": 10956, + "cbid": 211, "correlation": 82878 + } + }, + { + "ph": "s", "id": 82878, "pid": 435, "tid": 435, "ts": 1776649792659914, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659927, "dur": 0, + "args": { + "External id": 82884, + "cbid": 200, "correlation": 82884 + } + }, + { + "ph": "f", "id": 82884, "pid": 435, "tid": 435, "ts": 1776649792659927, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659927, "dur": 0, + "args": { + "External id": 82885, + "cbid": 273, "correlation": 82885 + } + }, + { + "ph": "f", "id": 82885, "pid": 435, "tid": 435, "ts": 1776649792659927, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792659937, "dur": 19, + "args": { + "External id": 82886, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82886, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82886, "pid": 0, "tid": 7, "ts": 1776649792659937, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659929, "dur": 5, + "args": { + "External id": 82886, + "cbid": 211, "correlation": 82886 + } + }, + { + "ph": "s", "id": 82886, "pid": 435, "tid": 435, "ts": 1776649792659929, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792659972, "dur": 1, + "args": { + "External id": 10961, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82904, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82904, "pid": 0, "tid": 7, "ts": 1776649792659972, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659962, "dur": 6, + "args": { + "External id": 10961, + "cbid": 211, "correlation": 82904 + } + }, + { + "ph": "s", "id": 82904, "pid": 435, "tid": 435, "ts": 1776649792659962, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659975, "dur": 0, + "args": { + "External id": 82910, + "cbid": 200, "correlation": 82910 + } + }, + { + "ph": "f", "id": 82910, "pid": 435, "tid": 435, "ts": 1776649792659975, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792659975, "dur": 0, + "args": { + "External id": 82911, + "cbid": 273, "correlation": 82911 + } + }, + { + "ph": "f", "id": 82911, "pid": 435, "tid": 435, "ts": 1776649792659975, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792659984, "dur": 19, + "args": { + "External id": 82912, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82912, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82912, "pid": 0, "tid": 7, "ts": 1776649792659984, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792659976, "dur": 5, + "args": { + "External id": 82912, + "cbid": 211, "correlation": 82912 + } + }, + { + "ph": "s", "id": 82912, "pid": 435, "tid": 435, "ts": 1776649792659976, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792660020, "dur": 1, + "args": { + "External id": 10966, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82930, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82930, "pid": 0, "tid": 7, "ts": 1776649792660020, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660010, "dur": 6, + "args": { + "External id": 10966, + "cbid": 211, "correlation": 82930 + } + }, + { + "ph": "s", "id": 82930, "pid": 435, "tid": 435, "ts": 1776649792660010, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660023, "dur": 0, + "args": { + "External id": 82936, + "cbid": 200, "correlation": 82936 + } + }, + { + "ph": "f", "id": 82936, "pid": 435, "tid": 435, "ts": 1776649792660023, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660023, "dur": 0, + "args": { + "External id": 82937, + "cbid": 273, "correlation": 82937 + } + }, + { + "ph": "f", "id": 82937, "pid": 435, "tid": 435, "ts": 1776649792660023, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792660035, "dur": 19, + "args": { + "External id": 82938, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82938, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82938, "pid": 0, "tid": 7, "ts": 1776649792660035, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660024, "dur": 4, + "args": { + "External id": 82938, + "cbid": 211, "correlation": 82938 + } + }, + { + "ph": "s", "id": 82938, "pid": 435, "tid": 435, "ts": 1776649792660024, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792660067, "dur": 1, + "args": { + "External id": 10971, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82956, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82956, "pid": 0, "tid": 7, "ts": 1776649792660067, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660058, "dur": 6, + "args": { + "External id": 10971, + "cbid": 211, "correlation": 82956 + } + }, + { + "ph": "s", "id": 82956, "pid": 435, "tid": 435, "ts": 1776649792660058, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660070, "dur": 0, + "args": { + "External id": 82962, + "cbid": 200, "correlation": 82962 + } + }, + { + "ph": "f", "id": 82962, "pid": 435, "tid": 435, "ts": 1776649792660070, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660071, "dur": 0, + "args": { + "External id": 82963, + "cbid": 273, "correlation": 82963 + } + }, + { + "ph": "f", "id": 82963, "pid": 435, "tid": 435, "ts": 1776649792660071, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792660080, "dur": 19, + "args": { + "External id": 82964, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82964, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82964, "pid": 0, "tid": 7, "ts": 1776649792660080, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660072, "dur": 4, + "args": { + "External id": 82964, + "cbid": 211, "correlation": 82964 + } + }, + { + "ph": "s", "id": 82964, "pid": 435, "tid": 435, "ts": 1776649792660072, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792660116, "dur": 1, + "args": { + "External id": 10976, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82982, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82982, "pid": 0, "tid": 7, "ts": 1776649792660116, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660106, "dur": 6, + "args": { + "External id": 10976, + "cbid": 211, "correlation": 82982 + } + }, + { + "ph": "s", "id": 82982, "pid": 435, "tid": 435, "ts": 1776649792660106, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660119, "dur": 0, + "args": { + "External id": 82988, + "cbid": 200, "correlation": 82988 + } + }, + { + "ph": "f", "id": 82988, "pid": 435, "tid": 435, "ts": 1776649792660119, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660119, "dur": 0, + "args": { + "External id": 82989, + "cbid": 273, "correlation": 82989 + } + }, + { + "ph": "f", "id": 82989, "pid": 435, "tid": 435, "ts": 1776649792660119, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792660128, "dur": 19, + "args": { + "External id": 82990, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 82990, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 82990, "pid": 0, "tid": 7, "ts": 1776649792660128, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660120, "dur": 5, + "args": { + "External id": 82990, + "cbid": 211, "correlation": 82990 + } + }, + { + "ph": "s", "id": 82990, "pid": 435, "tid": 435, "ts": 1776649792660120, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792660164, "dur": 1, + "args": { + "External id": 10981, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83008, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83008, "pid": 0, "tid": 7, "ts": 1776649792660164, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660154, "dur": 6, + "args": { + "External id": 10981, + "cbid": 211, "correlation": 83008 + } + }, + { + "ph": "s", "id": 83008, "pid": 435, "tid": 435, "ts": 1776649792660154, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660166, "dur": 0, + "args": { + "External id": 83014, + "cbid": 200, "correlation": 83014 + } + }, + { + "ph": "f", "id": 83014, "pid": 435, "tid": 435, "ts": 1776649792660166, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660167, "dur": 0, + "args": { + "External id": 83015, + "cbid": 273, "correlation": 83015 + } + }, + { + "ph": "f", "id": 83015, "pid": 435, "tid": 435, "ts": 1776649792660167, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792660176, "dur": 19, + "args": { + "External id": 83016, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83016, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83016, "pid": 0, "tid": 7, "ts": 1776649792660176, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660168, "dur": 4, + "args": { + "External id": 83016, + "cbid": 211, "correlation": 83016 + } + }, + { + "ph": "s", "id": 83016, "pid": 435, "tid": 435, "ts": 1776649792660168, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649792660198, "dur": 3, + "args": { + "External id": 83024, + "cbid": 135, "correlation": 83024 + } + }, + { + "ph": "f", "id": 83024, "pid": 435, "tid": 435, "ts": 1776649792660198, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792660213, "dur": 25, + "args": { + "External id": 83029, + "cbid": 165, "correlation": 83029 + } + }, + { + "ph": "s", "id": 83029, "pid": 435, "tid": 435, "ts": 1776649792660213, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649792660242, "dur": 5, + "args": { + "External id": 83032, + "cbid": 139, "correlation": 83032 + } + }, + { + "ph": "f", "id": 83032, "pid": 435, "tid": 435, "ts": 1776649792660242, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792660302, "dur": 1, + "args": { + "External id": 10986, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83060, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83060, "pid": 0, "tid": 7, "ts": 1776649792660302, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660290, "dur": 9, + "args": { + "External id": 10986, + "cbid": 211, "correlation": 83060 + } + }, + { + "ph": "s", "id": 83060, "pid": 435, "tid": 435, "ts": 1776649792660290, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660307, "dur": 0, + "args": { + "External id": 83066, + "cbid": 200, "correlation": 83066 + } + }, + { + "ph": "f", "id": 83066, "pid": 435, "tid": 435, "ts": 1776649792660307, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660308, "dur": 1, + "args": { + "External id": 83067, + "cbid": 273, "correlation": 83067 + } + }, + { + "ph": "f", "id": 83067, "pid": 435, "tid": 435, "ts": 1776649792660308, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792660317, "dur": 19, + "args": { + "External id": 83068, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83068, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83068, "pid": 0, "tid": 7, "ts": 1776649792660317, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660309, "dur": 5, + "args": { + "External id": 83068, + "cbid": 211, "correlation": 83068 + } + }, + { + "ph": "s", "id": 83068, "pid": 435, "tid": 435, "ts": 1776649792660309, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792660356, "dur": 1, + "args": { + "External id": 10991, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83086, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83086, "pid": 0, "tid": 7, "ts": 1776649792660356, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660346, "dur": 6, + "args": { + "External id": 10991, + "cbid": 211, "correlation": 83086 + } + }, + { + "ph": "s", "id": 83086, "pid": 435, "tid": 435, "ts": 1776649792660346, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660359, "dur": 0, + "args": { + "External id": 83092, + "cbid": 200, "correlation": 83092 + } + }, + { + "ph": "f", "id": 83092, "pid": 435, "tid": 435, "ts": 1776649792660359, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660360, "dur": 0, + "args": { + "External id": 83093, + "cbid": 273, "correlation": 83093 + } + }, + { + "ph": "f", "id": 83093, "pid": 435, "tid": 435, "ts": 1776649792660360, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792660369, "dur": 19, + "args": { + "External id": 83094, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83094, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83094, "pid": 0, "tid": 7, "ts": 1776649792660369, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660361, "dur": 5, + "args": { + "External id": 83094, + "cbid": 211, "correlation": 83094 + } + }, + { + "ph": "s", "id": 83094, "pid": 435, "tid": 435, "ts": 1776649792660361, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792660405, "dur": 1, + "args": { + "External id": 10996, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83112, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83112, "pid": 0, "tid": 7, "ts": 1776649792660405, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660396, "dur": 6, + "args": { + "External id": 10996, + "cbid": 211, "correlation": 83112 + } + }, + { + "ph": "s", "id": 83112, "pid": 435, "tid": 435, "ts": 1776649792660396, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660408, "dur": 0, + "args": { + "External id": 83118, + "cbid": 200, "correlation": 83118 + } + }, + { + "ph": "f", "id": 83118, "pid": 435, "tid": 435, "ts": 1776649792660408, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660409, "dur": 0, + "args": { + "External id": 83119, + "cbid": 273, "correlation": 83119 + } + }, + { + "ph": "f", "id": 83119, "pid": 435, "tid": 435, "ts": 1776649792660409, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792660418, "dur": 19, + "args": { + "External id": 83120, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83120, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83120, "pid": 0, "tid": 7, "ts": 1776649792660418, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660410, "dur": 4, + "args": { + "External id": 83120, + "cbid": 211, "correlation": 83120 + } + }, + { + "ph": "s", "id": 83120, "pid": 435, "tid": 435, "ts": 1776649792660410, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792660454, "dur": 1, + "args": { + "External id": 11001, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83138, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83138, "pid": 0, "tid": 7, "ts": 1776649792660454, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660444, "dur": 6, + "args": { + "External id": 11001, + "cbid": 211, "correlation": 83138 + } + }, + { + "ph": "s", "id": 83138, "pid": 435, "tid": 435, "ts": 1776649792660444, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660457, "dur": 0, + "args": { + "External id": 83144, + "cbid": 200, "correlation": 83144 + } + }, + { + "ph": "f", "id": 83144, "pid": 435, "tid": 435, "ts": 1776649792660457, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660457, "dur": 0, + "args": { + "External id": 83145, + "cbid": 273, "correlation": 83145 + } + }, + { + "ph": "f", "id": 83145, "pid": 435, "tid": 435, "ts": 1776649792660457, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792660466, "dur": 19, + "args": { + "External id": 83146, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83146, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83146, "pid": 0, "tid": 7, "ts": 1776649792660466, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660459, "dur": 5, + "args": { + "External id": 83146, + "cbid": 211, "correlation": 83146 + } + }, + { + "ph": "s", "id": 83146, "pid": 435, "tid": 435, "ts": 1776649792660459, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792660502, "dur": 1, + "args": { + "External id": 11006, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83164, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83164, "pid": 0, "tid": 7, "ts": 1776649792660502, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660493, "dur": 6, + "args": { + "External id": 11006, + "cbid": 211, "correlation": 83164 + } + }, + { + "ph": "s", "id": 83164, "pid": 435, "tid": 435, "ts": 1776649792660493, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660505, "dur": 0, + "args": { + "External id": 83170, + "cbid": 200, "correlation": 83170 + } + }, + { + "ph": "f", "id": 83170, "pid": 435, "tid": 435, "ts": 1776649792660505, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660506, "dur": 0, + "args": { + "External id": 83171, + "cbid": 273, "correlation": 83171 + } + }, + { + "ph": "f", "id": 83171, "pid": 435, "tid": 435, "ts": 1776649792660506, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792660515, "dur": 19, + "args": { + "External id": 83172, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83172, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83172, "pid": 0, "tid": 7, "ts": 1776649792660515, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660507, "dur": 5, + "args": { + "External id": 83172, + "cbid": 211, "correlation": 83172 + } + }, + { + "ph": "s", "id": 83172, "pid": 435, "tid": 435, "ts": 1776649792660507, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792660551, "dur": 1, + "args": { + "External id": 11011, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83190, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83190, "pid": 0, "tid": 7, "ts": 1776649792660551, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660541, "dur": 6, + "args": { + "External id": 11011, + "cbid": 211, "correlation": 83190 + } + }, + { + "ph": "s", "id": 83190, "pid": 435, "tid": 435, "ts": 1776649792660541, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660553, "dur": 0, + "args": { + "External id": 83196, + "cbid": 200, "correlation": 83196 + } + }, + { + "ph": "f", "id": 83196, "pid": 435, "tid": 435, "ts": 1776649792660553, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660554, "dur": 0, + "args": { + "External id": 83197, + "cbid": 273, "correlation": 83197 + } + }, + { + "ph": "f", "id": 83197, "pid": 435, "tid": 435, "ts": 1776649792660554, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792660563, "dur": 19, + "args": { + "External id": 83198, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83198, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83198, "pid": 0, "tid": 7, "ts": 1776649792660563, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660555, "dur": 5, + "args": { + "External id": 83198, + "cbid": 211, "correlation": 83198 + } + }, + { + "ph": "s", "id": 83198, "pid": 435, "tid": 435, "ts": 1776649792660555, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792660599, "dur": 1, + "args": { + "External id": 11016, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83216, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83216, "pid": 0, "tid": 7, "ts": 1776649792660599, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660589, "dur": 6, + "args": { + "External id": 11016, + "cbid": 211, "correlation": 83216 + } + }, + { + "ph": "s", "id": 83216, "pid": 435, "tid": 435, "ts": 1776649792660589, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660602, "dur": 0, + "args": { + "External id": 83222, + "cbid": 200, "correlation": 83222 + } + }, + { + "ph": "f", "id": 83222, "pid": 435, "tid": 435, "ts": 1776649792660602, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660603, "dur": 0, + "args": { + "External id": 83223, + "cbid": 273, "correlation": 83223 + } + }, + { + "ph": "f", "id": 83223, "pid": 435, "tid": 435, "ts": 1776649792660603, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792660611, "dur": 19, + "args": { + "External id": 83224, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83224, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83224, "pid": 0, "tid": 7, "ts": 1776649792660611, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660604, "dur": 4, + "args": { + "External id": 83224, + "cbid": 211, "correlation": 83224 + } + }, + { + "ph": "s", "id": 83224, "pid": 435, "tid": 435, "ts": 1776649792660604, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792660647, "dur": 1, + "args": { + "External id": 11021, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83242, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83242, "pid": 0, "tid": 7, "ts": 1776649792660647, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660638, "dur": 6, + "args": { + "External id": 11021, + "cbid": 211, "correlation": 83242 + } + }, + { + "ph": "s", "id": 83242, "pid": 435, "tid": 435, "ts": 1776649792660638, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660650, "dur": 0, + "args": { + "External id": 83248, + "cbid": 200, "correlation": 83248 + } + }, + { + "ph": "f", "id": 83248, "pid": 435, "tid": 435, "ts": 1776649792660650, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660651, "dur": 0, + "args": { + "External id": 83249, + "cbid": 273, "correlation": 83249 + } + }, + { + "ph": "f", "id": 83249, "pid": 435, "tid": 435, "ts": 1776649792660651, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792660660, "dur": 19, + "args": { + "External id": 83250, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83250, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83250, "pid": 0, "tid": 7, "ts": 1776649792660660, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660652, "dur": 4, + "args": { + "External id": 83250, + "cbid": 211, "correlation": 83250 + } + }, + { + "ph": "s", "id": 83250, "pid": 435, "tid": 435, "ts": 1776649792660652, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792660694, "dur": 1, + "args": { + "External id": 11026, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83268, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83268, "pid": 0, "tid": 7, "ts": 1776649792660694, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660685, "dur": 6, + "args": { + "External id": 11026, + "cbid": 211, "correlation": 83268 + } + }, + { + "ph": "s", "id": 83268, "pid": 435, "tid": 435, "ts": 1776649792660685, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660697, "dur": 0, + "args": { + "External id": 83274, + "cbid": 200, "correlation": 83274 + } + }, + { + "ph": "f", "id": 83274, "pid": 435, "tid": 435, "ts": 1776649792660697, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660698, "dur": 0, + "args": { + "External id": 83275, + "cbid": 273, "correlation": 83275 + } + }, + { + "ph": "f", "id": 83275, "pid": 435, "tid": 435, "ts": 1776649792660698, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792660707, "dur": 19, + "args": { + "External id": 83276, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83276, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83276, "pid": 0, "tid": 7, "ts": 1776649792660707, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660699, "dur": 4, + "args": { + "External id": 83276, + "cbid": 211, "correlation": 83276 + } + }, + { + "ph": "s", "id": 83276, "pid": 435, "tid": 435, "ts": 1776649792660699, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792660742, "dur": 1, + "args": { + "External id": 11031, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83294, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83294, "pid": 0, "tid": 7, "ts": 1776649792660742, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660733, "dur": 6, + "args": { + "External id": 11031, + "cbid": 211, "correlation": 83294 + } + }, + { + "ph": "s", "id": 83294, "pid": 435, "tid": 435, "ts": 1776649792660733, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660746, "dur": 0, + "args": { + "External id": 83300, + "cbid": 200, "correlation": 83300 + } + }, + { + "ph": "f", "id": 83300, "pid": 435, "tid": 435, "ts": 1776649792660746, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660746, "dur": 0, + "args": { + "External id": 83301, + "cbid": 273, "correlation": 83301 + } + }, + { + "ph": "f", "id": 83301, "pid": 435, "tid": 435, "ts": 1776649792660746, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792660755, "dur": 19, + "args": { + "External id": 83302, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83302, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83302, "pid": 0, "tid": 7, "ts": 1776649792660755, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660747, "dur": 4, + "args": { + "External id": 83302, + "cbid": 211, "correlation": 83302 + } + }, + { + "ph": "s", "id": 83302, "pid": 435, "tid": 435, "ts": 1776649792660747, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792660769, "dur": 9, + "args": { + "External id": 83305, + "cbid": 165, "correlation": 83305 + } + }, + { + "ph": "s", "id": 83305, "pid": 435, "tid": 435, "ts": 1776649792660769, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649792660801, "dur": 3, + "args": { + "External id": 83314, + "cbid": 135, "correlation": 83314 + } + }, + { + "ph": "f", "id": 83314, "pid": 435, "tid": 435, "ts": 1776649792660801, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792660849, "dur": 1, + "args": { + "External id": 11036, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83334, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83334, "pid": 0, "tid": 7, "ts": 1776649792660849, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660839, "dur": 8, + "args": { + "External id": 11036, + "cbid": 211, "correlation": 83334 + } + }, + { + "ph": "s", "id": 83334, "pid": 435, "tid": 435, "ts": 1776649792660839, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660854, "dur": 0, + "args": { + "External id": 83340, + "cbid": 200, "correlation": 83340 + } + }, + { + "ph": "f", "id": 83340, "pid": 435, "tid": 435, "ts": 1776649792660854, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660854, "dur": 0, + "args": { + "External id": 83341, + "cbid": 273, "correlation": 83341 + } + }, + { + "ph": "f", "id": 83341, "pid": 435, "tid": 435, "ts": 1776649792660854, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792660864, "dur": 19, + "args": { + "External id": 83342, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83342, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83342, "pid": 0, "tid": 7, "ts": 1776649792660864, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660856, "dur": 5, + "args": { + "External id": 83342, + "cbid": 211, "correlation": 83342 + } + }, + { + "ph": "s", "id": 83342, "pid": 435, "tid": 435, "ts": 1776649792660856, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792660901, "dur": 1, + "args": { + "External id": 11041, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83360, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83360, "pid": 0, "tid": 7, "ts": 1776649792660901, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660892, "dur": 7, + "args": { + "External id": 11041, + "cbid": 211, "correlation": 83360 + } + }, + { + "ph": "s", "id": 83360, "pid": 435, "tid": 435, "ts": 1776649792660892, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660905, "dur": 0, + "args": { + "External id": 83366, + "cbid": 200, "correlation": 83366 + } + }, + { + "ph": "f", "id": 83366, "pid": 435, "tid": 435, "ts": 1776649792660905, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660905, "dur": 0, + "args": { + "External id": 83367, + "cbid": 273, "correlation": 83367 + } + }, + { + "ph": "f", "id": 83367, "pid": 435, "tid": 435, "ts": 1776649792660905, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792660914, "dur": 19, + "args": { + "External id": 83368, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83368, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83368, "pid": 0, "tid": 7, "ts": 1776649792660914, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660907, "dur": 5, + "args": { + "External id": 83368, + "cbid": 211, "correlation": 83368 + } + }, + { + "ph": "s", "id": 83368, "pid": 435, "tid": 435, "ts": 1776649792660907, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792660950, "dur": 1, + "args": { + "External id": 11046, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83386, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83386, "pid": 0, "tid": 7, "ts": 1776649792660950, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660941, "dur": 6, + "args": { + "External id": 11046, + "cbid": 211, "correlation": 83386 + } + }, + { + "ph": "s", "id": 83386, "pid": 435, "tid": 435, "ts": 1776649792660941, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660953, "dur": 0, + "args": { + "External id": 83392, + "cbid": 200, "correlation": 83392 + } + }, + { + "ph": "f", "id": 83392, "pid": 435, "tid": 435, "ts": 1776649792660953, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792660954, "dur": 0, + "args": { + "External id": 83393, + "cbid": 273, "correlation": 83393 + } + }, + { + "ph": "f", "id": 83393, "pid": 435, "tid": 435, "ts": 1776649792660954, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792660963, "dur": 19, + "args": { + "External id": 83394, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83394, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83394, "pid": 0, "tid": 7, "ts": 1776649792660963, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660955, "dur": 4, + "args": { + "External id": 83394, + "cbid": 211, "correlation": 83394 + } + }, + { + "ph": "s", "id": 83394, "pid": 435, "tid": 435, "ts": 1776649792660955, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792660998, "dur": 1, + "args": { + "External id": 11051, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83412, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83412, "pid": 0, "tid": 7, "ts": 1776649792660998, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792660989, "dur": 7, + "args": { + "External id": 11051, + "cbid": 211, "correlation": 83412 + } + }, + { + "ph": "s", "id": 83412, "pid": 435, "tid": 435, "ts": 1776649792660989, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661002, "dur": 0, + "args": { + "External id": 83418, + "cbid": 200, "correlation": 83418 + } + }, + { + "ph": "f", "id": 83418, "pid": 435, "tid": 435, "ts": 1776649792661002, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661002, "dur": 0, + "args": { + "External id": 83419, + "cbid": 273, "correlation": 83419 + } + }, + { + "ph": "f", "id": 83419, "pid": 435, "tid": 435, "ts": 1776649792661002, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792661011, "dur": 19, + "args": { + "External id": 83420, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83420, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83420, "pid": 0, "tid": 7, "ts": 1776649792661011, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661003, "dur": 5, + "args": { + "External id": 83420, + "cbid": 211, "correlation": 83420 + } + }, + { + "ph": "s", "id": 83420, "pid": 435, "tid": 435, "ts": 1776649792661003, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792661047, "dur": 1, + "args": { + "External id": 11056, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83438, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83438, "pid": 0, "tid": 7, "ts": 1776649792661047, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661038, "dur": 6, + "args": { + "External id": 11056, + "cbid": 211, "correlation": 83438 + } + }, + { + "ph": "s", "id": 83438, "pid": 435, "tid": 435, "ts": 1776649792661038, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661050, "dur": 0, + "args": { + "External id": 83444, + "cbid": 200, "correlation": 83444 + } + }, + { + "ph": "f", "id": 83444, "pid": 435, "tid": 435, "ts": 1776649792661050, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661051, "dur": 0, + "args": { + "External id": 83445, + "cbid": 273, "correlation": 83445 + } + }, + { + "ph": "f", "id": 83445, "pid": 435, "tid": 435, "ts": 1776649792661051, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792661060, "dur": 19, + "args": { + "External id": 83446, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83446, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83446, "pid": 0, "tid": 7, "ts": 1776649792661060, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661052, "dur": 4, + "args": { + "External id": 83446, + "cbid": 211, "correlation": 83446 + } + }, + { + "ph": "s", "id": 83446, "pid": 435, "tid": 435, "ts": 1776649792661052, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792661095, "dur": 1, + "args": { + "External id": 11061, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83464, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83464, "pid": 0, "tid": 7, "ts": 1776649792661095, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661086, "dur": 6, + "args": { + "External id": 11061, + "cbid": 211, "correlation": 83464 + } + }, + { + "ph": "s", "id": 83464, "pid": 435, "tid": 435, "ts": 1776649792661086, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661098, "dur": 0, + "args": { + "External id": 83470, + "cbid": 200, "correlation": 83470 + } + }, + { + "ph": "f", "id": 83470, "pid": 435, "tid": 435, "ts": 1776649792661098, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661099, "dur": 0, + "args": { + "External id": 83471, + "cbid": 273, "correlation": 83471 + } + }, + { + "ph": "f", "id": 83471, "pid": 435, "tid": 435, "ts": 1776649792661099, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792661108, "dur": 19, + "args": { + "External id": 83472, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83472, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83472, "pid": 0, "tid": 7, "ts": 1776649792661108, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661100, "dur": 5, + "args": { + "External id": 83472, + "cbid": 211, "correlation": 83472 + } + }, + { + "ph": "s", "id": 83472, "pid": 435, "tid": 435, "ts": 1776649792661100, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792661144, "dur": 1, + "args": { + "External id": 11066, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83490, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83490, "pid": 0, "tid": 7, "ts": 1776649792661144, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661135, "dur": 6, + "args": { + "External id": 11066, + "cbid": 211, "correlation": 83490 + } + }, + { + "ph": "s", "id": 83490, "pid": 435, "tid": 435, "ts": 1776649792661135, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661147, "dur": 0, + "args": { + "External id": 83496, + "cbid": 200, "correlation": 83496 + } + }, + { + "ph": "f", "id": 83496, "pid": 435, "tid": 435, "ts": 1776649792661147, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661147, "dur": 0, + "args": { + "External id": 83497, + "cbid": 273, "correlation": 83497 + } + }, + { + "ph": "f", "id": 83497, "pid": 435, "tid": 435, "ts": 1776649792661147, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792661157, "dur": 19, + "args": { + "External id": 83498, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83498, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83498, "pid": 0, "tid": 7, "ts": 1776649792661157, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661149, "dur": 4, + "args": { + "External id": 83498, + "cbid": 211, "correlation": 83498 + } + }, + { + "ph": "s", "id": 83498, "pid": 435, "tid": 435, "ts": 1776649792661149, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792661192, "dur": 1, + "args": { + "External id": 11071, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83516, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83516, "pid": 0, "tid": 7, "ts": 1776649792661192, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661183, "dur": 6, + "args": { + "External id": 11071, + "cbid": 211, "correlation": 83516 + } + }, + { + "ph": "s", "id": 83516, "pid": 435, "tid": 435, "ts": 1776649792661183, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661195, "dur": 0, + "args": { + "External id": 83522, + "cbid": 200, "correlation": 83522 + } + }, + { + "ph": "f", "id": 83522, "pid": 435, "tid": 435, "ts": 1776649792661195, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661195, "dur": 0, + "args": { + "External id": 83523, + "cbid": 273, "correlation": 83523 + } + }, + { + "ph": "f", "id": 83523, "pid": 435, "tid": 435, "ts": 1776649792661195, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792661204, "dur": 19, + "args": { + "External id": 83524, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83524, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83524, "pid": 0, "tid": 7, "ts": 1776649792661204, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661197, "dur": 5, + "args": { + "External id": 83524, + "cbid": 211, "correlation": 83524 + } + }, + { + "ph": "s", "id": 83524, "pid": 435, "tid": 435, "ts": 1776649792661197, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792661240, "dur": 1, + "args": { + "External id": 11076, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83542, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83542, "pid": 0, "tid": 7, "ts": 1776649792661240, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661231, "dur": 6, + "args": { + "External id": 11076, + "cbid": 211, "correlation": 83542 + } + }, + { + "ph": "s", "id": 83542, "pid": 435, "tid": 435, "ts": 1776649792661231, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661243, "dur": 0, + "args": { + "External id": 83548, + "cbid": 200, "correlation": 83548 + } + }, + { + "ph": "f", "id": 83548, "pid": 435, "tid": 435, "ts": 1776649792661243, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661244, "dur": 0, + "args": { + "External id": 83549, + "cbid": 273, "correlation": 83549 + } + }, + { + "ph": "f", "id": 83549, "pid": 435, "tid": 435, "ts": 1776649792661244, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792661253, "dur": 19, + "args": { + "External id": 83550, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83550, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83550, "pid": 0, "tid": 7, "ts": 1776649792661253, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661245, "dur": 4, + "args": { + "External id": 83550, + "cbid": 211, "correlation": 83550 + } + }, + { + "ph": "s", "id": 83550, "pid": 435, "tid": 435, "ts": 1776649792661245, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792661288, "dur": 1, + "args": { + "External id": 11081, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83568, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83568, "pid": 0, "tid": 7, "ts": 1776649792661288, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661279, "dur": 6, + "args": { + "External id": 11081, + "cbid": 211, "correlation": 83568 + } + }, + { + "ph": "s", "id": 83568, "pid": 435, "tid": 435, "ts": 1776649792661279, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661291, "dur": 0, + "args": { + "External id": 83574, + "cbid": 200, "correlation": 83574 + } + }, + { + "ph": "f", "id": 83574, "pid": 435, "tid": 435, "ts": 1776649792661291, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661292, "dur": 0, + "args": { + "External id": 83575, + "cbid": 273, "correlation": 83575 + } + }, + { + "ph": "f", "id": 83575, "pid": 435, "tid": 435, "ts": 1776649792661292, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792661301, "dur": 19, + "args": { + "External id": 83576, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83576, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83576, "pid": 0, "tid": 7, "ts": 1776649792661301, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661293, "dur": 5, + "args": { + "External id": 83576, + "cbid": 211, "correlation": 83576 + } + }, + { + "ph": "s", "id": 83576, "pid": 435, "tid": 435, "ts": 1776649792661293, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792661336, "dur": 1, + "args": { + "External id": 11086, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83594, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83594, "pid": 0, "tid": 7, "ts": 1776649792661336, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661327, "dur": 6, + "args": { + "External id": 11086, + "cbid": 211, "correlation": 83594 + } + }, + { + "ph": "s", "id": 83594, "pid": 435, "tid": 435, "ts": 1776649792661327, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661339, "dur": 0, + "args": { + "External id": 83600, + "cbid": 200, "correlation": 83600 + } + }, + { + "ph": "f", "id": 83600, "pid": 435, "tid": 435, "ts": 1776649792661339, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661340, "dur": 0, + "args": { + "External id": 83601, + "cbid": 273, "correlation": 83601 + } + }, + { + "ph": "f", "id": 83601, "pid": 435, "tid": 435, "ts": 1776649792661340, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792661349, "dur": 19, + "args": { + "External id": 83602, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83602, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83602, "pid": 0, "tid": 7, "ts": 1776649792661349, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661341, "dur": 4, + "args": { + "External id": 83602, + "cbid": 211, "correlation": 83602 + } + }, + { + "ph": "s", "id": 83602, "pid": 435, "tid": 435, "ts": 1776649792661341, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792661385, "dur": 1, + "args": { + "External id": 11091, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83620, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83620, "pid": 0, "tid": 7, "ts": 1776649792661385, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661375, "dur": 6, + "args": { + "External id": 11091, + "cbid": 211, "correlation": 83620 + } + }, + { + "ph": "s", "id": 83620, "pid": 435, "tid": 435, "ts": 1776649792661375, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661387, "dur": 0, + "args": { + "External id": 83626, + "cbid": 200, "correlation": 83626 + } + }, + { + "ph": "f", "id": 83626, "pid": 435, "tid": 435, "ts": 1776649792661387, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661388, "dur": 0, + "args": { + "External id": 83627, + "cbid": 273, "correlation": 83627 + } + }, + { + "ph": "f", "id": 83627, "pid": 435, "tid": 435, "ts": 1776649792661388, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792661396, "dur": 19, + "args": { + "External id": 83628, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83628, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83628, "pid": 0, "tid": 7, "ts": 1776649792661396, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661389, "dur": 4, + "args": { + "External id": 83628, + "cbid": 211, "correlation": 83628 + } + }, + { + "ph": "s", "id": 83628, "pid": 435, "tid": 435, "ts": 1776649792661389, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792661432, "dur": 1, + "args": { + "External id": 11096, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83646, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83646, "pid": 0, "tid": 7, "ts": 1776649792661432, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661423, "dur": 6, + "args": { + "External id": 11096, + "cbid": 211, "correlation": 83646 + } + }, + { + "ph": "s", "id": 83646, "pid": 435, "tid": 435, "ts": 1776649792661423, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661435, "dur": 0, + "args": { + "External id": 83652, + "cbid": 200, "correlation": 83652 + } + }, + { + "ph": "f", "id": 83652, "pid": 435, "tid": 435, "ts": 1776649792661435, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661436, "dur": 0, + "args": { + "External id": 83653, + "cbid": 273, "correlation": 83653 + } + }, + { + "ph": "f", "id": 83653, "pid": 435, "tid": 435, "ts": 1776649792661436, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792661445, "dur": 19, + "args": { + "External id": 83654, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83654, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83654, "pid": 0, "tid": 7, "ts": 1776649792661445, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661437, "dur": 5, + "args": { + "External id": 83654, + "cbid": 211, "correlation": 83654 + } + }, + { + "ph": "s", "id": 83654, "pid": 435, "tid": 435, "ts": 1776649792661437, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792661480, "dur": 1, + "args": { + "External id": 11101, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83672, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83672, "pid": 0, "tid": 7, "ts": 1776649792661480, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661471, "dur": 6, + "args": { + "External id": 11101, + "cbid": 211, "correlation": 83672 + } + }, + { + "ph": "s", "id": 83672, "pid": 435, "tid": 435, "ts": 1776649792661471, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661483, "dur": 0, + "args": { + "External id": 83678, + "cbid": 200, "correlation": 83678 + } + }, + { + "ph": "f", "id": 83678, "pid": 435, "tid": 435, "ts": 1776649792661483, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661484, "dur": 0, + "args": { + "External id": 83679, + "cbid": 273, "correlation": 83679 + } + }, + { + "ph": "f", "id": 83679, "pid": 435, "tid": 435, "ts": 1776649792661484, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792661493, "dur": 19, + "args": { + "External id": 83680, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83680, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83680, "pid": 0, "tid": 7, "ts": 1776649792661493, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661485, "dur": 4, + "args": { + "External id": 83680, + "cbid": 211, "correlation": 83680 + } + }, + { + "ph": "s", "id": 83680, "pid": 435, "tid": 435, "ts": 1776649792661485, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792661528, "dur": 1, + "args": { + "External id": 11106, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83698, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83698, "pid": 0, "tid": 7, "ts": 1776649792661528, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661519, "dur": 6, + "args": { + "External id": 11106, + "cbid": 211, "correlation": 83698 + } + }, + { + "ph": "s", "id": 83698, "pid": 435, "tid": 435, "ts": 1776649792661519, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661531, "dur": 0, + "args": { + "External id": 83704, + "cbid": 200, "correlation": 83704 + } + }, + { + "ph": "f", "id": 83704, "pid": 435, "tid": 435, "ts": 1776649792661531, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661531, "dur": 0, + "args": { + "External id": 83705, + "cbid": 273, "correlation": 83705 + } + }, + { + "ph": "f", "id": 83705, "pid": 435, "tid": 435, "ts": 1776649792661531, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792661541, "dur": 19, + "args": { + "External id": 83706, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83706, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83706, "pid": 0, "tid": 7, "ts": 1776649792661541, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661533, "dur": 4, + "args": { + "External id": 83706, + "cbid": 211, "correlation": 83706 + } + }, + { + "ph": "s", "id": 83706, "pid": 435, "tid": 435, "ts": 1776649792661533, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792661575, "dur": 1, + "args": { + "External id": 11111, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83724, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83724, "pid": 0, "tid": 7, "ts": 1776649792661575, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661566, "dur": 6, + "args": { + "External id": 11111, + "cbid": 211, "correlation": 83724 + } + }, + { + "ph": "s", "id": 83724, "pid": 435, "tid": 435, "ts": 1776649792661566, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661579, "dur": 0, + "args": { + "External id": 83730, + "cbid": 200, "correlation": 83730 + } + }, + { + "ph": "f", "id": 83730, "pid": 435, "tid": 435, "ts": 1776649792661579, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661579, "dur": 0, + "args": { + "External id": 83731, + "cbid": 273, "correlation": 83731 + } + }, + { + "ph": "f", "id": 83731, "pid": 435, "tid": 435, "ts": 1776649792661579, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792661588, "dur": 19, + "args": { + "External id": 83732, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83732, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83732, "pid": 0, "tid": 7, "ts": 1776649792661588, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661580, "dur": 5, + "args": { + "External id": 83732, + "cbid": 211, "correlation": 83732 + } + }, + { + "ph": "s", "id": 83732, "pid": 435, "tid": 435, "ts": 1776649792661580, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792661623, "dur": 1, + "args": { + "External id": 11116, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83750, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83750, "pid": 0, "tid": 7, "ts": 1776649792661623, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661614, "dur": 6, + "args": { + "External id": 11116, + "cbid": 211, "correlation": 83750 + } + }, + { + "ph": "s", "id": 83750, "pid": 435, "tid": 435, "ts": 1776649792661614, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661627, "dur": 0, + "args": { + "External id": 83756, + "cbid": 200, "correlation": 83756 + } + }, + { + "ph": "f", "id": 83756, "pid": 435, "tid": 435, "ts": 1776649792661627, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661627, "dur": 0, + "args": { + "External id": 83757, + "cbid": 273, "correlation": 83757 + } + }, + { + "ph": "f", "id": 83757, "pid": 435, "tid": 435, "ts": 1776649792661627, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792661636, "dur": 19, + "args": { + "External id": 83758, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83758, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83758, "pid": 0, "tid": 7, "ts": 1776649792661636, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661628, "dur": 4, + "args": { + "External id": 83758, + "cbid": 211, "correlation": 83758 + } + }, + { + "ph": "s", "id": 83758, "pid": 435, "tid": 435, "ts": 1776649792661628, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792661671, "dur": 1, + "args": { + "External id": 11121, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83776, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83776, "pid": 0, "tid": 7, "ts": 1776649792661671, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661662, "dur": 6, + "args": { + "External id": 11121, + "cbid": 211, "correlation": 83776 + } + }, + { + "ph": "s", "id": 83776, "pid": 435, "tid": 435, "ts": 1776649792661662, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661674, "dur": 0, + "args": { + "External id": 83782, + "cbid": 200, "correlation": 83782 + } + }, + { + "ph": "f", "id": 83782, "pid": 435, "tid": 435, "ts": 1776649792661674, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661675, "dur": 0, + "args": { + "External id": 83783, + "cbid": 273, "correlation": 83783 + } + }, + { + "ph": "f", "id": 83783, "pid": 435, "tid": 435, "ts": 1776649792661675, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792661684, "dur": 19, + "args": { + "External id": 83784, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83784, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83784, "pid": 0, "tid": 7, "ts": 1776649792661684, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661676, "dur": 4, + "args": { + "External id": 83784, + "cbid": 211, "correlation": 83784 + } + }, + { + "ph": "s", "id": 83784, "pid": 435, "tid": 435, "ts": 1776649792661676, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792661719, "dur": 1, + "args": { + "External id": 11126, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83802, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83802, "pid": 0, "tid": 7, "ts": 1776649792661719, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661710, "dur": 6, + "args": { + "External id": 11126, + "cbid": 211, "correlation": 83802 + } + }, + { + "ph": "s", "id": 83802, "pid": 435, "tid": 435, "ts": 1776649792661710, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661722, "dur": 0, + "args": { + "External id": 83808, + "cbid": 200, "correlation": 83808 + } + }, + { + "ph": "f", "id": 83808, "pid": 435, "tid": 435, "ts": 1776649792661722, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661723, "dur": 0, + "args": { + "External id": 83809, + "cbid": 273, "correlation": 83809 + } + }, + { + "ph": "f", "id": 83809, "pid": 435, "tid": 435, "ts": 1776649792661723, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792661732, "dur": 19, + "args": { + "External id": 83810, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83810, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83810, "pid": 0, "tid": 7, "ts": 1776649792661732, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661724, "dur": 4, + "args": { + "External id": 83810, + "cbid": 211, "correlation": 83810 + } + }, + { + "ph": "s", "id": 83810, "pid": 435, "tid": 435, "ts": 1776649792661724, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792661767, "dur": 1, + "args": { + "External id": 11131, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83828, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83828, "pid": 0, "tid": 7, "ts": 1776649792661767, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661758, "dur": 6, + "args": { + "External id": 11131, + "cbid": 211, "correlation": 83828 + } + }, + { + "ph": "s", "id": 83828, "pid": 435, "tid": 435, "ts": 1776649792661758, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661770, "dur": 0, + "args": { + "External id": 83834, + "cbid": 200, "correlation": 83834 + } + }, + { + "ph": "f", "id": 83834, "pid": 435, "tid": 435, "ts": 1776649792661770, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661771, "dur": 0, + "args": { + "External id": 83835, + "cbid": 273, "correlation": 83835 + } + }, + { + "ph": "f", "id": 83835, "pid": 435, "tid": 435, "ts": 1776649792661771, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792661780, "dur": 19, + "args": { + "External id": 83836, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83836, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83836, "pid": 0, "tid": 7, "ts": 1776649792661780, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661772, "dur": 4, + "args": { + "External id": 83836, + "cbid": 211, "correlation": 83836 + } + }, + { + "ph": "s", "id": 83836, "pid": 435, "tid": 435, "ts": 1776649792661772, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792661815, "dur": 1, + "args": { + "External id": 11136, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83854, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83854, "pid": 0, "tid": 7, "ts": 1776649792661815, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661806, "dur": 6, + "args": { + "External id": 11136, + "cbid": 211, "correlation": 83854 + } + }, + { + "ph": "s", "id": 83854, "pid": 435, "tid": 435, "ts": 1776649792661806, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661818, "dur": 0, + "args": { + "External id": 83860, + "cbid": 200, "correlation": 83860 + } + }, + { + "ph": "f", "id": 83860, "pid": 435, "tid": 435, "ts": 1776649792661818, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661819, "dur": 0, + "args": { + "External id": 83861, + "cbid": 273, "correlation": 83861 + } + }, + { + "ph": "f", "id": 83861, "pid": 435, "tid": 435, "ts": 1776649792661819, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792661828, "dur": 19, + "args": { + "External id": 83862, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83862, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83862, "pid": 0, "tid": 7, "ts": 1776649792661828, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661820, "dur": 5, + "args": { + "External id": 83862, + "cbid": 211, "correlation": 83862 + } + }, + { + "ph": "s", "id": 83862, "pid": 435, "tid": 435, "ts": 1776649792661820, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792661863, "dur": 1, + "args": { + "External id": 11141, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83880, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83880, "pid": 0, "tid": 7, "ts": 1776649792661863, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661854, "dur": 6, + "args": { + "External id": 11141, + "cbid": 211, "correlation": 83880 + } + }, + { + "ph": "s", "id": 83880, "pid": 435, "tid": 435, "ts": 1776649792661854, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661866, "dur": 0, + "args": { + "External id": 83886, + "cbid": 200, "correlation": 83886 + } + }, + { + "ph": "f", "id": 83886, "pid": 435, "tid": 435, "ts": 1776649792661866, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661867, "dur": 0, + "args": { + "External id": 83887, + "cbid": 273, "correlation": 83887 + } + }, + { + "ph": "f", "id": 83887, "pid": 435, "tid": 435, "ts": 1776649792661867, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792661876, "dur": 19, + "args": { + "External id": 83888, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83888, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83888, "pid": 0, "tid": 7, "ts": 1776649792661876, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661868, "dur": 4, + "args": { + "External id": 83888, + "cbid": 211, "correlation": 83888 + } + }, + { + "ph": "s", "id": 83888, "pid": 435, "tid": 435, "ts": 1776649792661868, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792661911, "dur": 1, + "args": { + "External id": 11146, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83906, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83906, "pid": 0, "tid": 7, "ts": 1776649792661911, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661902, "dur": 6, + "args": { + "External id": 11146, + "cbid": 211, "correlation": 83906 + } + }, + { + "ph": "s", "id": 83906, "pid": 435, "tid": 435, "ts": 1776649792661902, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661914, "dur": 0, + "args": { + "External id": 83912, + "cbid": 200, "correlation": 83912 + } + }, + { + "ph": "f", "id": 83912, "pid": 435, "tid": 435, "ts": 1776649792661914, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661914, "dur": 0, + "args": { + "External id": 83913, + "cbid": 273, "correlation": 83913 + } + }, + { + "ph": "f", "id": 83913, "pid": 435, "tid": 435, "ts": 1776649792661914, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792661923, "dur": 19, + "args": { + "External id": 83914, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83914, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83914, "pid": 0, "tid": 7, "ts": 1776649792661923, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661915, "dur": 4, + "args": { + "External id": 83914, + "cbid": 211, "correlation": 83914 + } + }, + { + "ph": "s", "id": 83914, "pid": 435, "tid": 435, "ts": 1776649792661915, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792661958, "dur": 1, + "args": { + "External id": 11151, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83932, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83932, "pid": 0, "tid": 7, "ts": 1776649792661958, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661949, "dur": 6, + "args": { + "External id": 11151, + "cbid": 211, "correlation": 83932 + } + }, + { + "ph": "s", "id": 83932, "pid": 435, "tid": 435, "ts": 1776649792661949, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661961, "dur": 0, + "args": { + "External id": 83938, + "cbid": 200, "correlation": 83938 + } + }, + { + "ph": "f", "id": 83938, "pid": 435, "tid": 435, "ts": 1776649792661961, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792661962, "dur": 0, + "args": { + "External id": 83939, + "cbid": 273, "correlation": 83939 + } + }, + { + "ph": "f", "id": 83939, "pid": 435, "tid": 435, "ts": 1776649792661962, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792661971, "dur": 19, + "args": { + "External id": 83940, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83940, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83940, "pid": 0, "tid": 7, "ts": 1776649792661971, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661963, "dur": 4, + "args": { + "External id": 83940, + "cbid": 211, "correlation": 83940 + } + }, + { + "ph": "s", "id": 83940, "pid": 435, "tid": 435, "ts": 1776649792661963, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792662006, "dur": 1, + "args": { + "External id": 11156, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83958, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83958, "pid": 0, "tid": 7, "ts": 1776649792662006, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792661997, "dur": 6, + "args": { + "External id": 11156, + "cbid": 211, "correlation": 83958 + } + }, + { + "ph": "s", "id": 83958, "pid": 435, "tid": 435, "ts": 1776649792661997, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662009, "dur": 0, + "args": { + "External id": 83964, + "cbid": 200, "correlation": 83964 + } + }, + { + "ph": "f", "id": 83964, "pid": 435, "tid": 435, "ts": 1776649792662009, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662010, "dur": 0, + "args": { + "External id": 83965, + "cbid": 273, "correlation": 83965 + } + }, + { + "ph": "f", "id": 83965, "pid": 435, "tid": 435, "ts": 1776649792662010, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792662019, "dur": 19, + "args": { + "External id": 83966, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83966, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83966, "pid": 0, "tid": 7, "ts": 1776649792662019, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662011, "dur": 4, + "args": { + "External id": 83966, + "cbid": 211, "correlation": 83966 + } + }, + { + "ph": "s", "id": 83966, "pid": 435, "tid": 435, "ts": 1776649792662011, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792662054, "dur": 1, + "args": { + "External id": 11161, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83984, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83984, "pid": 0, "tid": 7, "ts": 1776649792662054, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662044, "dur": 6, + "args": { + "External id": 11161, + "cbid": 211, "correlation": 83984 + } + }, + { + "ph": "s", "id": 83984, "pid": 435, "tid": 435, "ts": 1776649792662044, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662056, "dur": 0, + "args": { + "External id": 83990, + "cbid": 200, "correlation": 83990 + } + }, + { + "ph": "f", "id": 83990, "pid": 435, "tid": 435, "ts": 1776649792662056, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662057, "dur": 0, + "args": { + "External id": 83991, + "cbid": 273, "correlation": 83991 + } + }, + { + "ph": "f", "id": 83991, "pid": 435, "tid": 435, "ts": 1776649792662057, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792662066, "dur": 19, + "args": { + "External id": 83992, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 83992, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 83992, "pid": 0, "tid": 7, "ts": 1776649792662066, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662058, "dur": 5, + "args": { + "External id": 83992, + "cbid": 211, "correlation": 83992 + } + }, + { + "ph": "s", "id": 83992, "pid": 435, "tid": 435, "ts": 1776649792662058, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792662101, "dur": 1, + "args": { + "External id": 11166, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84010, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84010, "pid": 0, "tid": 7, "ts": 1776649792662101, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662092, "dur": 6, + "args": { + "External id": 11166, + "cbid": 211, "correlation": 84010 + } + }, + { + "ph": "s", "id": 84010, "pid": 435, "tid": 435, "ts": 1776649792662092, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662104, "dur": 0, + "args": { + "External id": 84016, + "cbid": 200, "correlation": 84016 + } + }, + { + "ph": "f", "id": 84016, "pid": 435, "tid": 435, "ts": 1776649792662104, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662105, "dur": 0, + "args": { + "External id": 84017, + "cbid": 273, "correlation": 84017 + } + }, + { + "ph": "f", "id": 84017, "pid": 435, "tid": 435, "ts": 1776649792662105, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792662114, "dur": 19, + "args": { + "External id": 84018, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84018, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84018, "pid": 0, "tid": 7, "ts": 1776649792662114, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662106, "dur": 4, + "args": { + "External id": 84018, + "cbid": 211, "correlation": 84018 + } + }, + { + "ph": "s", "id": 84018, "pid": 435, "tid": 435, "ts": 1776649792662106, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792662149, "dur": 1, + "args": { + "External id": 11171, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84036, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84036, "pid": 0, "tid": 7, "ts": 1776649792662149, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662139, "dur": 6, + "args": { + "External id": 11171, + "cbid": 211, "correlation": 84036 + } + }, + { + "ph": "s", "id": 84036, "pid": 435, "tid": 435, "ts": 1776649792662139, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662152, "dur": 0, + "args": { + "External id": 84042, + "cbid": 200, "correlation": 84042 + } + }, + { + "ph": "f", "id": 84042, "pid": 435, "tid": 435, "ts": 1776649792662152, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662152, "dur": 0, + "args": { + "External id": 84043, + "cbid": 273, "correlation": 84043 + } + }, + { + "ph": "f", "id": 84043, "pid": 435, "tid": 435, "ts": 1776649792662152, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792662162, "dur": 19, + "args": { + "External id": 84044, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84044, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84044, "pid": 0, "tid": 7, "ts": 1776649792662162, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662153, "dur": 5, + "args": { + "External id": 84044, + "cbid": 211, "correlation": 84044 + } + }, + { + "ph": "s", "id": 84044, "pid": 435, "tid": 435, "ts": 1776649792662153, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792662196, "dur": 1, + "args": { + "External id": 11176, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84062, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84062, "pid": 0, "tid": 7, "ts": 1776649792662196, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662188, "dur": 6, + "args": { + "External id": 11176, + "cbid": 211, "correlation": 84062 + } + }, + { + "ph": "s", "id": 84062, "pid": 435, "tid": 435, "ts": 1776649792662188, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662200, "dur": 0, + "args": { + "External id": 84068, + "cbid": 200, "correlation": 84068 + } + }, + { + "ph": "f", "id": 84068, "pid": 435, "tid": 435, "ts": 1776649792662200, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662201, "dur": 0, + "args": { + "External id": 84069, + "cbid": 273, "correlation": 84069 + } + }, + { + "ph": "f", "id": 84069, "pid": 435, "tid": 435, "ts": 1776649792662201, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792662210, "dur": 19, + "args": { + "External id": 84070, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84070, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84070, "pid": 0, "tid": 7, "ts": 1776649792662210, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662202, "dur": 4, + "args": { + "External id": 84070, + "cbid": 211, "correlation": 84070 + } + }, + { + "ph": "s", "id": 84070, "pid": 435, "tid": 435, "ts": 1776649792662202, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792662245, "dur": 1, + "args": { + "External id": 11181, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84088, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84088, "pid": 0, "tid": 7, "ts": 1776649792662245, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662236, "dur": 6, + "args": { + "External id": 11181, + "cbid": 211, "correlation": 84088 + } + }, + { + "ph": "s", "id": 84088, "pid": 435, "tid": 435, "ts": 1776649792662236, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662248, "dur": 0, + "args": { + "External id": 84094, + "cbid": 200, "correlation": 84094 + } + }, + { + "ph": "f", "id": 84094, "pid": 435, "tid": 435, "ts": 1776649792662248, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662248, "dur": 0, + "args": { + "External id": 84095, + "cbid": 273, "correlation": 84095 + } + }, + { + "ph": "f", "id": 84095, "pid": 435, "tid": 435, "ts": 1776649792662248, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792662257, "dur": 19, + "args": { + "External id": 84096, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84096, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84096, "pid": 0, "tid": 7, "ts": 1776649792662257, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662250, "dur": 5, + "args": { + "External id": 84096, + "cbid": 211, "correlation": 84096 + } + }, + { + "ph": "s", "id": 84096, "pid": 435, "tid": 435, "ts": 1776649792662250, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792662293, "dur": 1, + "args": { + "External id": 11186, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84114, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84114, "pid": 0, "tid": 7, "ts": 1776649792662293, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662284, "dur": 6, + "args": { + "External id": 11186, + "cbid": 211, "correlation": 84114 + } + }, + { + "ph": "s", "id": 84114, "pid": 435, "tid": 435, "ts": 1776649792662284, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662296, "dur": 0, + "args": { + "External id": 84120, + "cbid": 200, "correlation": 84120 + } + }, + { + "ph": "f", "id": 84120, "pid": 435, "tid": 435, "ts": 1776649792662296, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662297, "dur": 0, + "args": { + "External id": 84121, + "cbid": 273, "correlation": 84121 + } + }, + { + "ph": "f", "id": 84121, "pid": 435, "tid": 435, "ts": 1776649792662297, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792662306, "dur": 19, + "args": { + "External id": 84122, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84122, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84122, "pid": 0, "tid": 7, "ts": 1776649792662306, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662298, "dur": 4, + "args": { + "External id": 84122, + "cbid": 211, "correlation": 84122 + } + }, + { + "ph": "s", "id": 84122, "pid": 435, "tid": 435, "ts": 1776649792662298, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792662341, "dur": 1, + "args": { + "External id": 11191, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84140, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84140, "pid": 0, "tid": 7, "ts": 1776649792662341, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662332, "dur": 7, + "args": { + "External id": 11191, + "cbid": 211, "correlation": 84140 + } + }, + { + "ph": "s", "id": 84140, "pid": 435, "tid": 435, "ts": 1776649792662332, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662345, "dur": 0, + "args": { + "External id": 84146, + "cbid": 200, "correlation": 84146 + } + }, + { + "ph": "f", "id": 84146, "pid": 435, "tid": 435, "ts": 1776649792662345, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662345, "dur": 0, + "args": { + "External id": 84147, + "cbid": 273, "correlation": 84147 + } + }, + { + "ph": "f", "id": 84147, "pid": 435, "tid": 435, "ts": 1776649792662345, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792662354, "dur": 19, + "args": { + "External id": 84148, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84148, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84148, "pid": 0, "tid": 7, "ts": 1776649792662354, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662346, "dur": 5, + "args": { + "External id": 84148, + "cbid": 211, "correlation": 84148 + } + }, + { + "ph": "s", "id": 84148, "pid": 435, "tid": 435, "ts": 1776649792662346, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792662390, "dur": 1, + "args": { + "External id": 11196, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84166, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84166, "pid": 0, "tid": 7, "ts": 1776649792662390, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662381, "dur": 6, + "args": { + "External id": 11196, + "cbid": 211, "correlation": 84166 + } + }, + { + "ph": "s", "id": 84166, "pid": 435, "tid": 435, "ts": 1776649792662381, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662393, "dur": 0, + "args": { + "External id": 84172, + "cbid": 200, "correlation": 84172 + } + }, + { + "ph": "f", "id": 84172, "pid": 435, "tid": 435, "ts": 1776649792662393, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662393, "dur": 0, + "args": { + "External id": 84173, + "cbid": 273, "correlation": 84173 + } + }, + { + "ph": "f", "id": 84173, "pid": 435, "tid": 435, "ts": 1776649792662393, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792662402, "dur": 19, + "args": { + "External id": 84174, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84174, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84174, "pid": 0, "tid": 7, "ts": 1776649792662402, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662394, "dur": 4, + "args": { + "External id": 84174, + "cbid": 211, "correlation": 84174 + } + }, + { + "ph": "s", "id": 84174, "pid": 435, "tid": 435, "ts": 1776649792662394, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792662437, "dur": 1, + "args": { + "External id": 11201, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84192, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84192, "pid": 0, "tid": 7, "ts": 1776649792662437, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662428, "dur": 6, + "args": { + "External id": 11201, + "cbid": 211, "correlation": 84192 + } + }, + { + "ph": "s", "id": 84192, "pid": 435, "tid": 435, "ts": 1776649792662428, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662440, "dur": 0, + "args": { + "External id": 84198, + "cbid": 200, "correlation": 84198 + } + }, + { + "ph": "f", "id": 84198, "pid": 435, "tid": 435, "ts": 1776649792662440, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662441, "dur": 0, + "args": { + "External id": 84199, + "cbid": 273, "correlation": 84199 + } + }, + { + "ph": "f", "id": 84199, "pid": 435, "tid": 435, "ts": 1776649792662441, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792662450, "dur": 19, + "args": { + "External id": 84200, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84200, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84200, "pid": 0, "tid": 7, "ts": 1776649792662450, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662442, "dur": 5, + "args": { + "External id": 84200, + "cbid": 211, "correlation": 84200 + } + }, + { + "ph": "s", "id": 84200, "pid": 435, "tid": 435, "ts": 1776649792662442, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792662485, "dur": 1, + "args": { + "External id": 11206, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84218, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84218, "pid": 0, "tid": 7, "ts": 1776649792662485, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662476, "dur": 6, + "args": { + "External id": 11206, + "cbid": 211, "correlation": 84218 + } + }, + { + "ph": "s", "id": 84218, "pid": 435, "tid": 435, "ts": 1776649792662476, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662489, "dur": 0, + "args": { + "External id": 84224, + "cbid": 200, "correlation": 84224 + } + }, + { + "ph": "f", "id": 84224, "pid": 435, "tid": 435, "ts": 1776649792662489, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662489, "dur": 0, + "args": { + "External id": 84225, + "cbid": 273, "correlation": 84225 + } + }, + { + "ph": "f", "id": 84225, "pid": 435, "tid": 435, "ts": 1776649792662489, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792662498, "dur": 19, + "args": { + "External id": 84226, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84226, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84226, "pid": 0, "tid": 7, "ts": 1776649792662498, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662490, "dur": 5, + "args": { + "External id": 84226, + "cbid": 211, "correlation": 84226 + } + }, + { + "ph": "s", "id": 84226, "pid": 435, "tid": 435, "ts": 1776649792662490, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792662533, "dur": 1, + "args": { + "External id": 11211, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84244, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84244, "pid": 0, "tid": 7, "ts": 1776649792662533, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662524, "dur": 6, + "args": { + "External id": 11211, + "cbid": 211, "correlation": 84244 + } + }, + { + "ph": "s", "id": 84244, "pid": 435, "tid": 435, "ts": 1776649792662524, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662537, "dur": 0, + "args": { + "External id": 84250, + "cbid": 200, "correlation": 84250 + } + }, + { + "ph": "f", "id": 84250, "pid": 435, "tid": 435, "ts": 1776649792662537, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662537, "dur": 0, + "args": { + "External id": 84251, + "cbid": 273, "correlation": 84251 + } + }, + { + "ph": "f", "id": 84251, "pid": 435, "tid": 435, "ts": 1776649792662537, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792662547, "dur": 19, + "args": { + "External id": 84252, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84252, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84252, "pid": 0, "tid": 7, "ts": 1776649792662547, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662539, "dur": 5, + "args": { + "External id": 84252, + "cbid": 211, "correlation": 84252 + } + }, + { + "ph": "s", "id": 84252, "pid": 435, "tid": 435, "ts": 1776649792662539, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792662584, "dur": 1, + "args": { + "External id": 11216, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84270, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84270, "pid": 0, "tid": 7, "ts": 1776649792662584, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662574, "dur": 6, + "args": { + "External id": 11216, + "cbid": 211, "correlation": 84270 + } + }, + { + "ph": "s", "id": 84270, "pid": 435, "tid": 435, "ts": 1776649792662574, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662587, "dur": 0, + "args": { + "External id": 84276, + "cbid": 200, "correlation": 84276 + } + }, + { + "ph": "f", "id": 84276, "pid": 435, "tid": 435, "ts": 1776649792662587, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662588, "dur": 0, + "args": { + "External id": 84277, + "cbid": 273, "correlation": 84277 + } + }, + { + "ph": "f", "id": 84277, "pid": 435, "tid": 435, "ts": 1776649792662588, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792662597, "dur": 19, + "args": { + "External id": 84278, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84278, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84278, "pid": 0, "tid": 7, "ts": 1776649792662597, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662589, "dur": 5, + "args": { + "External id": 84278, + "cbid": 211, "correlation": 84278 + } + }, + { + "ph": "s", "id": 84278, "pid": 435, "tid": 435, "ts": 1776649792662589, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792662632, "dur": 1, + "args": { + "External id": 11221, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84296, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84296, "pid": 0, "tid": 7, "ts": 1776649792662632, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662623, "dur": 6, + "args": { + "External id": 11221, + "cbid": 211, "correlation": 84296 + } + }, + { + "ph": "s", "id": 84296, "pid": 435, "tid": 435, "ts": 1776649792662623, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662635, "dur": 0, + "args": { + "External id": 84302, + "cbid": 200, "correlation": 84302 + } + }, + { + "ph": "f", "id": 84302, "pid": 435, "tid": 435, "ts": 1776649792662635, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662635, "dur": 0, + "args": { + "External id": 84303, + "cbid": 273, "correlation": 84303 + } + }, + { + "ph": "f", "id": 84303, "pid": 435, "tid": 435, "ts": 1776649792662635, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792662644, "dur": 19, + "args": { + "External id": 84304, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84304, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84304, "pid": 0, "tid": 7, "ts": 1776649792662644, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662636, "dur": 4, + "args": { + "External id": 84304, + "cbid": 211, "correlation": 84304 + } + }, + { + "ph": "s", "id": 84304, "pid": 435, "tid": 435, "ts": 1776649792662636, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792662679, "dur": 1, + "args": { + "External id": 11226, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84322, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84322, "pid": 0, "tid": 7, "ts": 1776649792662679, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662670, "dur": 6, + "args": { + "External id": 11226, + "cbid": 211, "correlation": 84322 + } + }, + { + "ph": "s", "id": 84322, "pid": 435, "tid": 435, "ts": 1776649792662670, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662682, "dur": 0, + "args": { + "External id": 84328, + "cbid": 200, "correlation": 84328 + } + }, + { + "ph": "f", "id": 84328, "pid": 435, "tid": 435, "ts": 1776649792662682, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662683, "dur": 0, + "args": { + "External id": 84329, + "cbid": 273, "correlation": 84329 + } + }, + { + "ph": "f", "id": 84329, "pid": 435, "tid": 435, "ts": 1776649792662683, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792662692, "dur": 19, + "args": { + "External id": 84330, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84330, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84330, "pid": 0, "tid": 7, "ts": 1776649792662692, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662684, "dur": 5, + "args": { + "External id": 84330, + "cbid": 211, "correlation": 84330 + } + }, + { + "ph": "s", "id": 84330, "pid": 435, "tid": 435, "ts": 1776649792662684, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792662728, "dur": 1, + "args": { + "External id": 11231, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84348, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84348, "pid": 0, "tid": 7, "ts": 1776649792662728, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662719, "dur": 6, + "args": { + "External id": 11231, + "cbid": 211, "correlation": 84348 + } + }, + { + "ph": "s", "id": 84348, "pid": 435, "tid": 435, "ts": 1776649792662719, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662731, "dur": 0, + "args": { + "External id": 84354, + "cbid": 200, "correlation": 84354 + } + }, + { + "ph": "f", "id": 84354, "pid": 435, "tid": 435, "ts": 1776649792662731, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662732, "dur": 0, + "args": { + "External id": 84355, + "cbid": 273, "correlation": 84355 + } + }, + { + "ph": "f", "id": 84355, "pid": 435, "tid": 435, "ts": 1776649792662732, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792662741, "dur": 19, + "args": { + "External id": 84356, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84356, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84356, "pid": 0, "tid": 7, "ts": 1776649792662741, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662733, "dur": 5, + "args": { + "External id": 84356, + "cbid": 211, "correlation": 84356 + } + }, + { + "ph": "s", "id": 84356, "pid": 435, "tid": 435, "ts": 1776649792662733, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792662776, "dur": 1, + "args": { + "External id": 11236, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84374, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84374, "pid": 0, "tid": 7, "ts": 1776649792662776, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662767, "dur": 6, + "args": { + "External id": 11236, + "cbid": 211, "correlation": 84374 + } + }, + { + "ph": "s", "id": 84374, "pid": 435, "tid": 435, "ts": 1776649792662767, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662779, "dur": 0, + "args": { + "External id": 84380, + "cbid": 200, "correlation": 84380 + } + }, + { + "ph": "f", "id": 84380, "pid": 435, "tid": 435, "ts": 1776649792662779, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662779, "dur": 0, + "args": { + "External id": 84381, + "cbid": 273, "correlation": 84381 + } + }, + { + "ph": "f", "id": 84381, "pid": 435, "tid": 435, "ts": 1776649792662779, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792662788, "dur": 19, + "args": { + "External id": 84382, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84382, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84382, "pid": 0, "tid": 7, "ts": 1776649792662788, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662780, "dur": 4, + "args": { + "External id": 84382, + "cbid": 211, "correlation": 84382 + } + }, + { + "ph": "s", "id": 84382, "pid": 435, "tid": 435, "ts": 1776649792662780, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792662823, "dur": 1, + "args": { + "External id": 11241, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84400, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84400, "pid": 0, "tid": 7, "ts": 1776649792662823, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662814, "dur": 6, + "args": { + "External id": 11241, + "cbid": 211, "correlation": 84400 + } + }, + { + "ph": "s", "id": 84400, "pid": 435, "tid": 435, "ts": 1776649792662814, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662826, "dur": 0, + "args": { + "External id": 84406, + "cbid": 200, "correlation": 84406 + } + }, + { + "ph": "f", "id": 84406, "pid": 435, "tid": 435, "ts": 1776649792662826, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662827, "dur": 0, + "args": { + "External id": 84407, + "cbid": 273, "correlation": 84407 + } + }, + { + "ph": "f", "id": 84407, "pid": 435, "tid": 435, "ts": 1776649792662827, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792662836, "dur": 19, + "args": { + "External id": 84408, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84408, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84408, "pid": 0, "tid": 7, "ts": 1776649792662836, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662828, "dur": 4, + "args": { + "External id": 84408, + "cbid": 211, "correlation": 84408 + } + }, + { + "ph": "s", "id": 84408, "pid": 435, "tid": 435, "ts": 1776649792662828, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792662870, "dur": 1, + "args": { + "External id": 11246, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84426, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84426, "pid": 0, "tid": 7, "ts": 1776649792662870, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662861, "dur": 6, + "args": { + "External id": 11246, + "cbid": 211, "correlation": 84426 + } + }, + { + "ph": "s", "id": 84426, "pid": 435, "tid": 435, "ts": 1776649792662861, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662873, "dur": 0, + "args": { + "External id": 84432, + "cbid": 200, "correlation": 84432 + } + }, + { + "ph": "f", "id": 84432, "pid": 435, "tid": 435, "ts": 1776649792662873, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662874, "dur": 0, + "args": { + "External id": 84433, + "cbid": 273, "correlation": 84433 + } + }, + { + "ph": "f", "id": 84433, "pid": 435, "tid": 435, "ts": 1776649792662874, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792662883, "dur": 19, + "args": { + "External id": 84434, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84434, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84434, "pid": 0, "tid": 7, "ts": 1776649792662883, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662875, "dur": 4, + "args": { + "External id": 84434, + "cbid": 211, "correlation": 84434 + } + }, + { + "ph": "s", "id": 84434, "pid": 435, "tid": 435, "ts": 1776649792662875, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792662918, "dur": 1, + "args": { + "External id": 11251, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84452, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84452, "pid": 0, "tid": 7, "ts": 1776649792662918, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662909, "dur": 6, + "args": { + "External id": 11251, + "cbid": 211, "correlation": 84452 + } + }, + { + "ph": "s", "id": 84452, "pid": 435, "tid": 435, "ts": 1776649792662909, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662921, "dur": 0, + "args": { + "External id": 84458, + "cbid": 200, "correlation": 84458 + } + }, + { + "ph": "f", "id": 84458, "pid": 435, "tid": 435, "ts": 1776649792662921, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662922, "dur": 0, + "args": { + "External id": 84459, + "cbid": 273, "correlation": 84459 + } + }, + { + "ph": "f", "id": 84459, "pid": 435, "tid": 435, "ts": 1776649792662922, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792662931, "dur": 19, + "args": { + "External id": 84460, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84460, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84460, "pid": 0, "tid": 7, "ts": 1776649792662931, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662923, "dur": 5, + "args": { + "External id": 84460, + "cbid": 211, "correlation": 84460 + } + }, + { + "ph": "s", "id": 84460, "pid": 435, "tid": 435, "ts": 1776649792662923, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792662966, "dur": 1, + "args": { + "External id": 11256, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84478, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84478, "pid": 0, "tid": 7, "ts": 1776649792662966, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662957, "dur": 6, + "args": { + "External id": 11256, + "cbid": 211, "correlation": 84478 + } + }, + { + "ph": "s", "id": 84478, "pid": 435, "tid": 435, "ts": 1776649792662957, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662969, "dur": 0, + "args": { + "External id": 84484, + "cbid": 200, "correlation": 84484 + } + }, + { + "ph": "f", "id": 84484, "pid": 435, "tid": 435, "ts": 1776649792662969, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792662969, "dur": 0, + "args": { + "External id": 84485, + "cbid": 273, "correlation": 84485 + } + }, + { + "ph": "f", "id": 84485, "pid": 435, "tid": 435, "ts": 1776649792662969, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792662979, "dur": 19, + "args": { + "External id": 84486, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84486, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84486, "pid": 0, "tid": 7, "ts": 1776649792662979, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792662971, "dur": 4, + "args": { + "External id": 84486, + "cbid": 211, "correlation": 84486 + } + }, + { + "ph": "s", "id": 84486, "pid": 435, "tid": 435, "ts": 1776649792662971, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792663013, "dur": 1, + "args": { + "External id": 11261, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84504, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84504, "pid": 0, "tid": 7, "ts": 1776649792663013, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663004, "dur": 6, + "args": { + "External id": 11261, + "cbid": 211, "correlation": 84504 + } + }, + { + "ph": "s", "id": 84504, "pid": 435, "tid": 435, "ts": 1776649792663004, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663017, "dur": 0, + "args": { + "External id": 84510, + "cbid": 200, "correlation": 84510 + } + }, + { + "ph": "f", "id": 84510, "pid": 435, "tid": 435, "ts": 1776649792663017, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663017, "dur": 0, + "args": { + "External id": 84511, + "cbid": 273, "correlation": 84511 + } + }, + { + "ph": "f", "id": 84511, "pid": 435, "tid": 435, "ts": 1776649792663017, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792663026, "dur": 19, + "args": { + "External id": 84512, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84512, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84512, "pid": 0, "tid": 7, "ts": 1776649792663026, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663018, "dur": 4, + "args": { + "External id": 84512, + "cbid": 211, "correlation": 84512 + } + }, + { + "ph": "s", "id": 84512, "pid": 435, "tid": 435, "ts": 1776649792663018, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792663072, "dur": 1, + "args": { + "External id": 11266, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84530, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84530, "pid": 0, "tid": 7, "ts": 1776649792663072, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663062, "dur": 6, + "args": { + "External id": 11266, + "cbid": 211, "correlation": 84530 + } + }, + { + "ph": "s", "id": 84530, "pid": 435, "tid": 435, "ts": 1776649792663062, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663075, "dur": 0, + "args": { + "External id": 84536, + "cbid": 200, "correlation": 84536 + } + }, + { + "ph": "f", "id": 84536, "pid": 435, "tid": 435, "ts": 1776649792663075, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663076, "dur": 0, + "args": { + "External id": 84537, + "cbid": 273, "correlation": 84537 + } + }, + { + "ph": "f", "id": 84537, "pid": 435, "tid": 435, "ts": 1776649792663076, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792663085, "dur": 19, + "args": { + "External id": 84538, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84538, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84538, "pid": 0, "tid": 7, "ts": 1776649792663085, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663077, "dur": 4, + "args": { + "External id": 84538, + "cbid": 211, "correlation": 84538 + } + }, + { + "ph": "s", "id": 84538, "pid": 435, "tid": 435, "ts": 1776649792663077, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792663121, "dur": 1, + "args": { + "External id": 11271, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84556, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84556, "pid": 0, "tid": 7, "ts": 1776649792663121, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663112, "dur": 6, + "args": { + "External id": 11271, + "cbid": 211, "correlation": 84556 + } + }, + { + "ph": "s", "id": 84556, "pid": 435, "tid": 435, "ts": 1776649792663112, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663124, "dur": 0, + "args": { + "External id": 84562, + "cbid": 200, "correlation": 84562 + } + }, + { + "ph": "f", "id": 84562, "pid": 435, "tid": 435, "ts": 1776649792663124, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663125, "dur": 0, + "args": { + "External id": 84563, + "cbid": 273, "correlation": 84563 + } + }, + { + "ph": "f", "id": 84563, "pid": 435, "tid": 435, "ts": 1776649792663125, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792663133, "dur": 19, + "args": { + "External id": 84564, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84564, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84564, "pid": 0, "tid": 7, "ts": 1776649792663133, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663126, "dur": 5, + "args": { + "External id": 84564, + "cbid": 211, "correlation": 84564 + } + }, + { + "ph": "s", "id": 84564, "pid": 435, "tid": 435, "ts": 1776649792663126, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792663169, "dur": 1, + "args": { + "External id": 11276, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84582, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84582, "pid": 0, "tid": 7, "ts": 1776649792663169, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663159, "dur": 6, + "args": { + "External id": 11276, + "cbid": 211, "correlation": 84582 + } + }, + { + "ph": "s", "id": 84582, "pid": 435, "tid": 435, "ts": 1776649792663159, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663172, "dur": 0, + "args": { + "External id": 84588, + "cbid": 200, "correlation": 84588 + } + }, + { + "ph": "f", "id": 84588, "pid": 435, "tid": 435, "ts": 1776649792663172, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663173, "dur": 0, + "args": { + "External id": 84589, + "cbid": 273, "correlation": 84589 + } + }, + { + "ph": "f", "id": 84589, "pid": 435, "tid": 435, "ts": 1776649792663173, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792663182, "dur": 19, + "args": { + "External id": 84590, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84590, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84590, "pid": 0, "tid": 7, "ts": 1776649792663182, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663174, "dur": 5, + "args": { + "External id": 84590, + "cbid": 211, "correlation": 84590 + } + }, + { + "ph": "s", "id": 84590, "pid": 435, "tid": 435, "ts": 1776649792663174, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792663217, "dur": 1, + "args": { + "External id": 11281, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84608, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84608, "pid": 0, "tid": 7, "ts": 1776649792663217, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663208, "dur": 6, + "args": { + "External id": 11281, + "cbid": 211, "correlation": 84608 + } + }, + { + "ph": "s", "id": 84608, "pid": 435, "tid": 435, "ts": 1776649792663208, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663220, "dur": 0, + "args": { + "External id": 84614, + "cbid": 200, "correlation": 84614 + } + }, + { + "ph": "f", "id": 84614, "pid": 435, "tid": 435, "ts": 1776649792663220, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663221, "dur": 0, + "args": { + "External id": 84615, + "cbid": 273, "correlation": 84615 + } + }, + { + "ph": "f", "id": 84615, "pid": 435, "tid": 435, "ts": 1776649792663221, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792663230, "dur": 19, + "args": { + "External id": 84616, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84616, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84616, "pid": 0, "tid": 7, "ts": 1776649792663230, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663222, "dur": 5, + "args": { + "External id": 84616, + "cbid": 211, "correlation": 84616 + } + }, + { + "ph": "s", "id": 84616, "pid": 435, "tid": 435, "ts": 1776649792663222, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649792663252, "dur": 3, + "args": { + "External id": 84624, + "cbid": 135, "correlation": 84624 + } + }, + { + "ph": "f", "id": 84624, "pid": 435, "tid": 435, "ts": 1776649792663252, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792663267, "dur": 5, + "args": { + "External id": 84629, + "cbid": 165, "correlation": 84629 + } + }, + { + "ph": "s", "id": 84629, "pid": 435, "tid": 435, "ts": 1776649792663267, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649792663275, "dur": 5, + "args": { + "External id": 84632, + "cbid": 139, "correlation": 84632 + } + }, + { + "ph": "f", "id": 84632, "pid": 435, "tid": 435, "ts": 1776649792663275, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792663332, "dur": 1, + "args": { + "External id": 11286, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84660, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84660, "pid": 0, "tid": 7, "ts": 1776649792663332, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663320, "dur": 9, + "args": { + "External id": 11286, + "cbid": 211, "correlation": 84660 + } + }, + { + "ph": "s", "id": 84660, "pid": 435, "tid": 435, "ts": 1776649792663320, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663336, "dur": 0, + "args": { + "External id": 84666, + "cbid": 200, "correlation": 84666 + } + }, + { + "ph": "f", "id": 84666, "pid": 435, "tid": 435, "ts": 1776649792663336, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663337, "dur": 0, + "args": { + "External id": 84667, + "cbid": 273, "correlation": 84667 + } + }, + { + "ph": "f", "id": 84667, "pid": 435, "tid": 435, "ts": 1776649792663337, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792663346, "dur": 19, + "args": { + "External id": 84668, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84668, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84668, "pid": 0, "tid": 7, "ts": 1776649792663346, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663338, "dur": 5, + "args": { + "External id": 84668, + "cbid": 211, "correlation": 84668 + } + }, + { + "ph": "s", "id": 84668, "pid": 435, "tid": 435, "ts": 1776649792663338, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792663385, "dur": 1, + "args": { + "External id": 11291, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84686, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84686, "pid": 0, "tid": 7, "ts": 1776649792663385, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663375, "dur": 7, + "args": { + "External id": 11291, + "cbid": 211, "correlation": 84686 + } + }, + { + "ph": "s", "id": 84686, "pid": 435, "tid": 435, "ts": 1776649792663375, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663388, "dur": 0, + "args": { + "External id": 84692, + "cbid": 200, "correlation": 84692 + } + }, + { + "ph": "f", "id": 84692, "pid": 435, "tid": 435, "ts": 1776649792663388, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663389, "dur": 0, + "args": { + "External id": 84693, + "cbid": 273, "correlation": 84693 + } + }, + { + "ph": "f", "id": 84693, "pid": 435, "tid": 435, "ts": 1776649792663389, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792663398, "dur": 19, + "args": { + "External id": 84694, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84694, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84694, "pid": 0, "tid": 7, "ts": 1776649792663398, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663390, "dur": 5, + "args": { + "External id": 84694, + "cbid": 211, "correlation": 84694 + } + }, + { + "ph": "s", "id": 84694, "pid": 435, "tid": 435, "ts": 1776649792663390, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792663434, "dur": 1, + "args": { + "External id": 11296, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84712, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84712, "pid": 0, "tid": 7, "ts": 1776649792663434, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663424, "dur": 6, + "args": { + "External id": 11296, + "cbid": 211, "correlation": 84712 + } + }, + { + "ph": "s", "id": 84712, "pid": 435, "tid": 435, "ts": 1776649792663424, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663437, "dur": 0, + "args": { + "External id": 84718, + "cbid": 200, "correlation": 84718 + } + }, + { + "ph": "f", "id": 84718, "pid": 435, "tid": 435, "ts": 1776649792663437, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663437, "dur": 0, + "args": { + "External id": 84719, + "cbid": 273, "correlation": 84719 + } + }, + { + "ph": "f", "id": 84719, "pid": 435, "tid": 435, "ts": 1776649792663437, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792663446, "dur": 19, + "args": { + "External id": 84720, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84720, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84720, "pid": 0, "tid": 7, "ts": 1776649792663446, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663438, "dur": 4, + "args": { + "External id": 84720, + "cbid": 211, "correlation": 84720 + } + }, + { + "ph": "s", "id": 84720, "pid": 435, "tid": 435, "ts": 1776649792663438, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792663482, "dur": 1, + "args": { + "External id": 11301, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84738, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84738, "pid": 0, "tid": 7, "ts": 1776649792663482, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663472, "dur": 7, + "args": { + "External id": 11301, + "cbid": 211, "correlation": 84738 + } + }, + { + "ph": "s", "id": 84738, "pid": 435, "tid": 435, "ts": 1776649792663472, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663485, "dur": 0, + "args": { + "External id": 84744, + "cbid": 200, "correlation": 84744 + } + }, + { + "ph": "f", "id": 84744, "pid": 435, "tid": 435, "ts": 1776649792663485, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663486, "dur": 0, + "args": { + "External id": 84745, + "cbid": 273, "correlation": 84745 + } + }, + { + "ph": "f", "id": 84745, "pid": 435, "tid": 435, "ts": 1776649792663486, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792663495, "dur": 19, + "args": { + "External id": 84746, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84746, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84746, "pid": 0, "tid": 7, "ts": 1776649792663495, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663487, "dur": 5, + "args": { + "External id": 84746, + "cbid": 211, "correlation": 84746 + } + }, + { + "ph": "s", "id": 84746, "pid": 435, "tid": 435, "ts": 1776649792663487, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792663530, "dur": 1, + "args": { + "External id": 11306, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84764, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84764, "pid": 0, "tid": 7, "ts": 1776649792663530, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663521, "dur": 6, + "args": { + "External id": 11306, + "cbid": 211, "correlation": 84764 + } + }, + { + "ph": "s", "id": 84764, "pid": 435, "tid": 435, "ts": 1776649792663521, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663533, "dur": 0, + "args": { + "External id": 84770, + "cbid": 200, "correlation": 84770 + } + }, + { + "ph": "f", "id": 84770, "pid": 435, "tid": 435, "ts": 1776649792663533, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663534, "dur": 0, + "args": { + "External id": 84771, + "cbid": 273, "correlation": 84771 + } + }, + { + "ph": "f", "id": 84771, "pid": 435, "tid": 435, "ts": 1776649792663534, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792663543, "dur": 19, + "args": { + "External id": 84772, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84772, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84772, "pid": 0, "tid": 7, "ts": 1776649792663543, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663535, "dur": 4, + "args": { + "External id": 84772, + "cbid": 211, "correlation": 84772 + } + }, + { + "ph": "s", "id": 84772, "pid": 435, "tid": 435, "ts": 1776649792663535, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792663578, "dur": 1, + "args": { + "External id": 11311, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84790, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84790, "pid": 0, "tid": 7, "ts": 1776649792663578, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663569, "dur": 7, + "args": { + "External id": 11311, + "cbid": 211, "correlation": 84790 + } + }, + { + "ph": "s", "id": 84790, "pid": 435, "tid": 435, "ts": 1776649792663569, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663581, "dur": 0, + "args": { + "External id": 84796, + "cbid": 200, "correlation": 84796 + } + }, + { + "ph": "f", "id": 84796, "pid": 435, "tid": 435, "ts": 1776649792663581, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663582, "dur": 0, + "args": { + "External id": 84797, + "cbid": 273, "correlation": 84797 + } + }, + { + "ph": "f", "id": 84797, "pid": 435, "tid": 435, "ts": 1776649792663582, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792663591, "dur": 19, + "args": { + "External id": 84798, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84798, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84798, "pid": 0, "tid": 7, "ts": 1776649792663591, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663583, "dur": 4, + "args": { + "External id": 84798, + "cbid": 211, "correlation": 84798 + } + }, + { + "ph": "s", "id": 84798, "pid": 435, "tid": 435, "ts": 1776649792663583, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792663627, "dur": 1, + "args": { + "External id": 11316, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84816, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84816, "pid": 0, "tid": 7, "ts": 1776649792663627, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663617, "dur": 6, + "args": { + "External id": 11316, + "cbid": 211, "correlation": 84816 + } + }, + { + "ph": "s", "id": 84816, "pid": 435, "tid": 435, "ts": 1776649792663617, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663629, "dur": 0, + "args": { + "External id": 84822, + "cbid": 200, "correlation": 84822 + } + }, + { + "ph": "f", "id": 84822, "pid": 435, "tid": 435, "ts": 1776649792663629, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663630, "dur": 0, + "args": { + "External id": 84823, + "cbid": 273, "correlation": 84823 + } + }, + { + "ph": "f", "id": 84823, "pid": 435, "tid": 435, "ts": 1776649792663630, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792663639, "dur": 19, + "args": { + "External id": 84824, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84824, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84824, "pid": 0, "tid": 7, "ts": 1776649792663639, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663631, "dur": 5, + "args": { + "External id": 84824, + "cbid": 211, "correlation": 84824 + } + }, + { + "ph": "s", "id": 84824, "pid": 435, "tid": 435, "ts": 1776649792663631, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792663675, "dur": 1, + "args": { + "External id": 11321, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84842, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84842, "pid": 0, "tid": 7, "ts": 1776649792663675, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663666, "dur": 6, + "args": { + "External id": 11321, + "cbid": 211, "correlation": 84842 + } + }, + { + "ph": "s", "id": 84842, "pid": 435, "tid": 435, "ts": 1776649792663666, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663678, "dur": 0, + "args": { + "External id": 84848, + "cbid": 200, "correlation": 84848 + } + }, + { + "ph": "f", "id": 84848, "pid": 435, "tid": 435, "ts": 1776649792663678, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663678, "dur": 0, + "args": { + "External id": 84849, + "cbid": 273, "correlation": 84849 + } + }, + { + "ph": "f", "id": 84849, "pid": 435, "tid": 435, "ts": 1776649792663678, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792663688, "dur": 19, + "args": { + "External id": 84850, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84850, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84850, "pid": 0, "tid": 7, "ts": 1776649792663688, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663679, "dur": 5, + "args": { + "External id": 84850, + "cbid": 211, "correlation": 84850 + } + }, + { + "ph": "s", "id": 84850, "pid": 435, "tid": 435, "ts": 1776649792663679, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792663723, "dur": 1, + "args": { + "External id": 11326, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84868, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84868, "pid": 0, "tid": 7, "ts": 1776649792663723, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663713, "dur": 6, + "args": { + "External id": 11326, + "cbid": 211, "correlation": 84868 + } + }, + { + "ph": "s", "id": 84868, "pid": 435, "tid": 435, "ts": 1776649792663713, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663726, "dur": 0, + "args": { + "External id": 84874, + "cbid": 200, "correlation": 84874 + } + }, + { + "ph": "f", "id": 84874, "pid": 435, "tid": 435, "ts": 1776649792663726, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663726, "dur": 0, + "args": { + "External id": 84875, + "cbid": 273, "correlation": 84875 + } + }, + { + "ph": "f", "id": 84875, "pid": 435, "tid": 435, "ts": 1776649792663726, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792663735, "dur": 19, + "args": { + "External id": 84876, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84876, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84876, "pid": 0, "tid": 7, "ts": 1776649792663735, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663727, "dur": 4, + "args": { + "External id": 84876, + "cbid": 211, "correlation": 84876 + } + }, + { + "ph": "s", "id": 84876, "pid": 435, "tid": 435, "ts": 1776649792663727, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792663771, "dur": 1, + "args": { + "External id": 11331, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84894, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84894, "pid": 0, "tid": 7, "ts": 1776649792663771, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663761, "dur": 6, + "args": { + "External id": 11331, + "cbid": 211, "correlation": 84894 + } + }, + { + "ph": "s", "id": 84894, "pid": 435, "tid": 435, "ts": 1776649792663761, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663774, "dur": 0, + "args": { + "External id": 84900, + "cbid": 200, "correlation": 84900 + } + }, + { + "ph": "f", "id": 84900, "pid": 435, "tid": 435, "ts": 1776649792663774, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663774, "dur": 0, + "args": { + "External id": 84901, + "cbid": 273, "correlation": 84901 + } + }, + { + "ph": "f", "id": 84901, "pid": 435, "tid": 435, "ts": 1776649792663774, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792663783, "dur": 19, + "args": { + "External id": 84902, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84902, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84902, "pid": 0, "tid": 7, "ts": 1776649792663783, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663776, "dur": 4, + "args": { + "External id": 84902, + "cbid": 211, "correlation": 84902 + } + }, + { + "ph": "s", "id": 84902, "pid": 435, "tid": 435, "ts": 1776649792663776, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792663798, "dur": 9, + "args": { + "External id": 84905, + "cbid": 165, "correlation": 84905 + } + }, + { + "ph": "s", "id": 84905, "pid": 435, "tid": 435, "ts": 1776649792663798, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649792663828, "dur": 3, + "args": { + "External id": 84914, + "cbid": 135, "correlation": 84914 + } + }, + { + "ph": "f", "id": 84914, "pid": 435, "tid": 435, "ts": 1776649792663828, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792663877, "dur": 1, + "args": { + "External id": 11336, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84934, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84934, "pid": 0, "tid": 7, "ts": 1776649792663877, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663866, "dur": 8, + "args": { + "External id": 11336, + "cbid": 211, "correlation": 84934 + } + }, + { + "ph": "s", "id": 84934, "pid": 435, "tid": 435, "ts": 1776649792663866, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663881, "dur": 0, + "args": { + "External id": 84940, + "cbid": 200, "correlation": 84940 + } + }, + { + "ph": "f", "id": 84940, "pid": 435, "tid": 435, "ts": 1776649792663881, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663882, "dur": 0, + "args": { + "External id": 84941, + "cbid": 273, "correlation": 84941 + } + }, + { + "ph": "f", "id": 84941, "pid": 435, "tid": 435, "ts": 1776649792663882, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792663891, "dur": 19, + "args": { + "External id": 84942, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84942, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84942, "pid": 0, "tid": 7, "ts": 1776649792663891, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663883, "dur": 5, + "args": { + "External id": 84942, + "cbid": 211, "correlation": 84942 + } + }, + { + "ph": "s", "id": 84942, "pid": 435, "tid": 435, "ts": 1776649792663883, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792663929, "dur": 1, + "args": { + "External id": 11341, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84960, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84960, "pid": 0, "tid": 7, "ts": 1776649792663929, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663919, "dur": 7, + "args": { + "External id": 11341, + "cbid": 211, "correlation": 84960 + } + }, + { + "ph": "s", "id": 84960, "pid": 435, "tid": 435, "ts": 1776649792663919, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663932, "dur": 0, + "args": { + "External id": 84966, + "cbid": 200, "correlation": 84966 + } + }, + { + "ph": "f", "id": 84966, "pid": 435, "tid": 435, "ts": 1776649792663932, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663933, "dur": 0, + "args": { + "External id": 84967, + "cbid": 273, "correlation": 84967 + } + }, + { + "ph": "f", "id": 84967, "pid": 435, "tid": 435, "ts": 1776649792663933, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792663942, "dur": 19, + "args": { + "External id": 84968, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84968, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84968, "pid": 0, "tid": 7, "ts": 1776649792663942, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663934, "dur": 5, + "args": { + "External id": 84968, + "cbid": 211, "correlation": 84968 + } + }, + { + "ph": "s", "id": 84968, "pid": 435, "tid": 435, "ts": 1776649792663934, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792663978, "dur": 1, + "args": { + "External id": 11346, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84986, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84986, "pid": 0, "tid": 7, "ts": 1776649792663978, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663968, "dur": 6, + "args": { + "External id": 11346, + "cbid": 211, "correlation": 84986 + } + }, + { + "ph": "s", "id": 84986, "pid": 435, "tid": 435, "ts": 1776649792663968, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663981, "dur": 0, + "args": { + "External id": 84992, + "cbid": 200, "correlation": 84992 + } + }, + { + "ph": "f", "id": 84992, "pid": 435, "tid": 435, "ts": 1776649792663981, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792663982, "dur": 0, + "args": { + "External id": 84993, + "cbid": 273, "correlation": 84993 + } + }, + { + "ph": "f", "id": 84993, "pid": 435, "tid": 435, "ts": 1776649792663982, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792663991, "dur": 19, + "args": { + "External id": 84994, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 84994, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 84994, "pid": 0, "tid": 7, "ts": 1776649792663991, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792663983, "dur": 5, + "args": { + "External id": 84994, + "cbid": 211, "correlation": 84994 + } + }, + { + "ph": "s", "id": 84994, "pid": 435, "tid": 435, "ts": 1776649792663983, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792664027, "dur": 1, + "args": { + "External id": 11351, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85012, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85012, "pid": 0, "tid": 7, "ts": 1776649792664027, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664017, "dur": 6, + "args": { + "External id": 11351, + "cbid": 211, "correlation": 85012 + } + }, + { + "ph": "s", "id": 85012, "pid": 435, "tid": 435, "ts": 1776649792664017, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664030, "dur": 0, + "args": { + "External id": 85018, + "cbid": 200, "correlation": 85018 + } + }, + { + "ph": "f", "id": 85018, "pid": 435, "tid": 435, "ts": 1776649792664030, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664030, "dur": 0, + "args": { + "External id": 85019, + "cbid": 273, "correlation": 85019 + } + }, + { + "ph": "f", "id": 85019, "pid": 435, "tid": 435, "ts": 1776649792664030, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792664040, "dur": 19, + "args": { + "External id": 85020, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85020, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85020, "pid": 0, "tid": 7, "ts": 1776649792664040, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664032, "dur": 5, + "args": { + "External id": 85020, + "cbid": 211, "correlation": 85020 + } + }, + { + "ph": "s", "id": 85020, "pid": 435, "tid": 435, "ts": 1776649792664032, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792664075, "dur": 1, + "args": { + "External id": 11356, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85038, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85038, "pid": 0, "tid": 7, "ts": 1776649792664075, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664065, "dur": 6, + "args": { + "External id": 11356, + "cbid": 211, "correlation": 85038 + } + }, + { + "ph": "s", "id": 85038, "pid": 435, "tid": 435, "ts": 1776649792664065, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664078, "dur": 0, + "args": { + "External id": 85044, + "cbid": 200, "correlation": 85044 + } + }, + { + "ph": "f", "id": 85044, "pid": 435, "tid": 435, "ts": 1776649792664078, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664079, "dur": 0, + "args": { + "External id": 85045, + "cbid": 273, "correlation": 85045 + } + }, + { + "ph": "f", "id": 85045, "pid": 435, "tid": 435, "ts": 1776649792664079, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792664088, "dur": 19, + "args": { + "External id": 85046, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85046, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85046, "pid": 0, "tid": 7, "ts": 1776649792664088, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664080, "dur": 5, + "args": { + "External id": 85046, + "cbid": 211, "correlation": 85046 + } + }, + { + "ph": "s", "id": 85046, "pid": 435, "tid": 435, "ts": 1776649792664080, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792664123, "dur": 1, + "args": { + "External id": 11361, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85064, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85064, "pid": 0, "tid": 7, "ts": 1776649792664123, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664114, "dur": 6, + "args": { + "External id": 11361, + "cbid": 211, "correlation": 85064 + } + }, + { + "ph": "s", "id": 85064, "pid": 435, "tid": 435, "ts": 1776649792664114, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664126, "dur": 0, + "args": { + "External id": 85070, + "cbid": 200, "correlation": 85070 + } + }, + { + "ph": "f", "id": 85070, "pid": 435, "tid": 435, "ts": 1776649792664126, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664127, "dur": 0, + "args": { + "External id": 85071, + "cbid": 273, "correlation": 85071 + } + }, + { + "ph": "f", "id": 85071, "pid": 435, "tid": 435, "ts": 1776649792664127, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792664136, "dur": 19, + "args": { + "External id": 85072, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85072, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85072, "pid": 0, "tid": 7, "ts": 1776649792664136, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664128, "dur": 4, + "args": { + "External id": 85072, + "cbid": 211, "correlation": 85072 + } + }, + { + "ph": "s", "id": 85072, "pid": 435, "tid": 435, "ts": 1776649792664128, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792664171, "dur": 1, + "args": { + "External id": 11366, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85090, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85090, "pid": 0, "tid": 7, "ts": 1776649792664171, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664162, "dur": 6, + "args": { + "External id": 11366, + "cbid": 211, "correlation": 85090 + } + }, + { + "ph": "s", "id": 85090, "pid": 435, "tid": 435, "ts": 1776649792664162, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664175, "dur": 0, + "args": { + "External id": 85096, + "cbid": 200, "correlation": 85096 + } + }, + { + "ph": "f", "id": 85096, "pid": 435, "tid": 435, "ts": 1776649792664175, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664175, "dur": 0, + "args": { + "External id": 85097, + "cbid": 273, "correlation": 85097 + } + }, + { + "ph": "f", "id": 85097, "pid": 435, "tid": 435, "ts": 1776649792664175, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792664184, "dur": 19, + "args": { + "External id": 85098, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85098, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85098, "pid": 0, "tid": 7, "ts": 1776649792664184, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664176, "dur": 4, + "args": { + "External id": 85098, + "cbid": 211, "correlation": 85098 + } + }, + { + "ph": "s", "id": 85098, "pid": 435, "tid": 435, "ts": 1776649792664176, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792664237, "dur": 1, + "args": { + "External id": 11371, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85116, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85116, "pid": 0, "tid": 7, "ts": 1776649792664237, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664210, "dur": 24, + "args": { + "External id": 11371, + "cbid": 211, "correlation": 85116 + } + }, + { + "ph": "s", "id": 85116, "pid": 435, "tid": 435, "ts": 1776649792664210, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664242, "dur": 0, + "args": { + "External id": 85122, + "cbid": 200, "correlation": 85122 + } + }, + { + "ph": "f", "id": 85122, "pid": 435, "tid": 435, "ts": 1776649792664242, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664243, "dur": 1, + "args": { + "External id": 85123, + "cbid": 273, "correlation": 85123 + } + }, + { + "ph": "f", "id": 85123, "pid": 435, "tid": 435, "ts": 1776649792664243, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792664253, "dur": 19, + "args": { + "External id": 85124, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85124, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85124, "pid": 0, "tid": 7, "ts": 1776649792664253, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664244, "dur": 5, + "args": { + "External id": 85124, + "cbid": 211, "correlation": 85124 + } + }, + { + "ph": "s", "id": 85124, "pid": 435, "tid": 435, "ts": 1776649792664244, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792664292, "dur": 1, + "args": { + "External id": 11376, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85142, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85142, "pid": 0, "tid": 7, "ts": 1776649792664292, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664282, "dur": 6, + "args": { + "External id": 11376, + "cbid": 211, "correlation": 85142 + } + }, + { + "ph": "s", "id": 85142, "pid": 435, "tid": 435, "ts": 1776649792664282, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664295, "dur": 0, + "args": { + "External id": 85148, + "cbid": 200, "correlation": 85148 + } + }, + { + "ph": "f", "id": 85148, "pid": 435, "tid": 435, "ts": 1776649792664295, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664296, "dur": 0, + "args": { + "External id": 85149, + "cbid": 273, "correlation": 85149 + } + }, + { + "ph": "f", "id": 85149, "pid": 435, "tid": 435, "ts": 1776649792664296, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792664305, "dur": 19, + "args": { + "External id": 85150, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85150, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85150, "pid": 0, "tid": 7, "ts": 1776649792664305, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664297, "dur": 5, + "args": { + "External id": 85150, + "cbid": 211, "correlation": 85150 + } + }, + { + "ph": "s", "id": 85150, "pid": 435, "tid": 435, "ts": 1776649792664297, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792664344, "dur": 1, + "args": { + "External id": 11381, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85168, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85168, "pid": 0, "tid": 7, "ts": 1776649792664344, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664334, "dur": 7, + "args": { + "External id": 11381, + "cbid": 211, "correlation": 85168 + } + }, + { + "ph": "s", "id": 85168, "pid": 435, "tid": 435, "ts": 1776649792664334, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664347, "dur": 0, + "args": { + "External id": 85174, + "cbid": 200, "correlation": 85174 + } + }, + { + "ph": "f", "id": 85174, "pid": 435, "tid": 435, "ts": 1776649792664347, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664348, "dur": 0, + "args": { + "External id": 85175, + "cbid": 273, "correlation": 85175 + } + }, + { + "ph": "f", "id": 85175, "pid": 435, "tid": 435, "ts": 1776649792664348, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792664357, "dur": 19, + "args": { + "External id": 85176, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85176, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85176, "pid": 0, "tid": 7, "ts": 1776649792664357, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664349, "dur": 5, + "args": { + "External id": 85176, + "cbid": 211, "correlation": 85176 + } + }, + { + "ph": "s", "id": 85176, "pid": 435, "tid": 435, "ts": 1776649792664349, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792664397, "dur": 1, + "args": { + "External id": 11386, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85194, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85194, "pid": 0, "tid": 7, "ts": 1776649792664397, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664387, "dur": 7, + "args": { + "External id": 11386, + "cbid": 211, "correlation": 85194 + } + }, + { + "ph": "s", "id": 85194, "pid": 435, "tid": 435, "ts": 1776649792664387, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664400, "dur": 0, + "args": { + "External id": 85200, + "cbid": 200, "correlation": 85200 + } + }, + { + "ph": "f", "id": 85200, "pid": 435, "tid": 435, "ts": 1776649792664400, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664401, "dur": 0, + "args": { + "External id": 85201, + "cbid": 273, "correlation": 85201 + } + }, + { + "ph": "f", "id": 85201, "pid": 435, "tid": 435, "ts": 1776649792664401, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792664410, "dur": 19, + "args": { + "External id": 85202, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85202, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85202, "pid": 0, "tid": 7, "ts": 1776649792664410, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664402, "dur": 5, + "args": { + "External id": 85202, + "cbid": 211, "correlation": 85202 + } + }, + { + "ph": "s", "id": 85202, "pid": 435, "tid": 435, "ts": 1776649792664402, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792664446, "dur": 1, + "args": { + "External id": 11391, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85220, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85220, "pid": 0, "tid": 7, "ts": 1776649792664446, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664436, "dur": 7, + "args": { + "External id": 11391, + "cbid": 211, "correlation": 85220 + } + }, + { + "ph": "s", "id": 85220, "pid": 435, "tid": 435, "ts": 1776649792664436, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664449, "dur": 0, + "args": { + "External id": 85226, + "cbid": 200, "correlation": 85226 + } + }, + { + "ph": "f", "id": 85226, "pid": 435, "tid": 435, "ts": 1776649792664449, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664450, "dur": 0, + "args": { + "External id": 85227, + "cbid": 273, "correlation": 85227 + } + }, + { + "ph": "f", "id": 85227, "pid": 435, "tid": 435, "ts": 1776649792664450, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792664459, "dur": 19, + "args": { + "External id": 85228, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85228, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85228, "pid": 0, "tid": 7, "ts": 1776649792664459, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664451, "dur": 4, + "args": { + "External id": 85228, + "cbid": 211, "correlation": 85228 + } + }, + { + "ph": "s", "id": 85228, "pid": 435, "tid": 435, "ts": 1776649792664451, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792664494, "dur": 1, + "args": { + "External id": 11396, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85246, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85246, "pid": 0, "tid": 7, "ts": 1776649792664494, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664485, "dur": 6, + "args": { + "External id": 11396, + "cbid": 211, "correlation": 85246 + } + }, + { + "ph": "s", "id": 85246, "pid": 435, "tid": 435, "ts": 1776649792664485, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664497, "dur": 0, + "args": { + "External id": 85252, + "cbid": 200, "correlation": 85252 + } + }, + { + "ph": "f", "id": 85252, "pid": 435, "tid": 435, "ts": 1776649792664497, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664498, "dur": 0, + "args": { + "External id": 85253, + "cbid": 273, "correlation": 85253 + } + }, + { + "ph": "f", "id": 85253, "pid": 435, "tid": 435, "ts": 1776649792664498, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792664507, "dur": 19, + "args": { + "External id": 85254, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85254, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85254, "pid": 0, "tid": 7, "ts": 1776649792664507, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664499, "dur": 4, + "args": { + "External id": 85254, + "cbid": 211, "correlation": 85254 + } + }, + { + "ph": "s", "id": 85254, "pid": 435, "tid": 435, "ts": 1776649792664499, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792664542, "dur": 1, + "args": { + "External id": 11401, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85272, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85272, "pid": 0, "tid": 7, "ts": 1776649792664542, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664533, "dur": 6, + "args": { + "External id": 11401, + "cbid": 211, "correlation": 85272 + } + }, + { + "ph": "s", "id": 85272, "pid": 435, "tid": 435, "ts": 1776649792664533, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664545, "dur": 0, + "args": { + "External id": 85278, + "cbid": 200, "correlation": 85278 + } + }, + { + "ph": "f", "id": 85278, "pid": 435, "tid": 435, "ts": 1776649792664545, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664546, "dur": 0, + "args": { + "External id": 85279, + "cbid": 273, "correlation": 85279 + } + }, + { + "ph": "f", "id": 85279, "pid": 435, "tid": 435, "ts": 1776649792664546, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792664555, "dur": 19, + "args": { + "External id": 85280, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85280, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85280, "pid": 0, "tid": 7, "ts": 1776649792664555, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664547, "dur": 4, + "args": { + "External id": 85280, + "cbid": 211, "correlation": 85280 + } + }, + { + "ph": "s", "id": 85280, "pid": 435, "tid": 435, "ts": 1776649792664547, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792664590, "dur": 1, + "args": { + "External id": 11406, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85298, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85298, "pid": 0, "tid": 7, "ts": 1776649792664590, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664581, "dur": 6, + "args": { + "External id": 11406, + "cbid": 211, "correlation": 85298 + } + }, + { + "ph": "s", "id": 85298, "pid": 435, "tid": 435, "ts": 1776649792664581, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664594, "dur": 0, + "args": { + "External id": 85304, + "cbid": 200, "correlation": 85304 + } + }, + { + "ph": "f", "id": 85304, "pid": 435, "tid": 435, "ts": 1776649792664594, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664594, "dur": 0, + "args": { + "External id": 85305, + "cbid": 273, "correlation": 85305 + } + }, + { + "ph": "f", "id": 85305, "pid": 435, "tid": 435, "ts": 1776649792664594, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792664603, "dur": 19, + "args": { + "External id": 85306, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85306, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85306, "pid": 0, "tid": 7, "ts": 1776649792664603, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664595, "dur": 5, + "args": { + "External id": 85306, + "cbid": 211, "correlation": 85306 + } + }, + { + "ph": "s", "id": 85306, "pid": 435, "tid": 435, "ts": 1776649792664595, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792664638, "dur": 1, + "args": { + "External id": 11411, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85324, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85324, "pid": 0, "tid": 7, "ts": 1776649792664638, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664629, "dur": 6, + "args": { + "External id": 11411, + "cbid": 211, "correlation": 85324 + } + }, + { + "ph": "s", "id": 85324, "pid": 435, "tid": 435, "ts": 1776649792664629, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664642, "dur": 0, + "args": { + "External id": 85330, + "cbid": 200, "correlation": 85330 + } + }, + { + "ph": "f", "id": 85330, "pid": 435, "tid": 435, "ts": 1776649792664642, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664642, "dur": 0, + "args": { + "External id": 85331, + "cbid": 273, "correlation": 85331 + } + }, + { + "ph": "f", "id": 85331, "pid": 435, "tid": 435, "ts": 1776649792664642, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792664651, "dur": 19, + "args": { + "External id": 85332, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85332, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85332, "pid": 0, "tid": 7, "ts": 1776649792664651, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664643, "dur": 4, + "args": { + "External id": 85332, + "cbid": 211, "correlation": 85332 + } + }, + { + "ph": "s", "id": 85332, "pid": 435, "tid": 435, "ts": 1776649792664643, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792664686, "dur": 1, + "args": { + "External id": 11416, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85350, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85350, "pid": 0, "tid": 7, "ts": 1776649792664686, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664677, "dur": 6, + "args": { + "External id": 11416, + "cbid": 211, "correlation": 85350 + } + }, + { + "ph": "s", "id": 85350, "pid": 435, "tid": 435, "ts": 1776649792664677, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664690, "dur": 0, + "args": { + "External id": 85356, + "cbid": 200, "correlation": 85356 + } + }, + { + "ph": "f", "id": 85356, "pid": 435, "tid": 435, "ts": 1776649792664690, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664690, "dur": 0, + "args": { + "External id": 85357, + "cbid": 273, "correlation": 85357 + } + }, + { + "ph": "f", "id": 85357, "pid": 435, "tid": 435, "ts": 1776649792664690, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792664699, "dur": 19, + "args": { + "External id": 85358, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85358, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85358, "pid": 0, "tid": 7, "ts": 1776649792664699, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664691, "dur": 4, + "args": { + "External id": 85358, + "cbid": 211, "correlation": 85358 + } + }, + { + "ph": "s", "id": 85358, "pid": 435, "tid": 435, "ts": 1776649792664691, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792664735, "dur": 1, + "args": { + "External id": 11421, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85376, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85376, "pid": 0, "tid": 7, "ts": 1776649792664735, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664725, "dur": 7, + "args": { + "External id": 11421, + "cbid": 211, "correlation": 85376 + } + }, + { + "ph": "s", "id": 85376, "pid": 435, "tid": 435, "ts": 1776649792664725, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664738, "dur": 0, + "args": { + "External id": 85382, + "cbid": 200, "correlation": 85382 + } + }, + { + "ph": "f", "id": 85382, "pid": 435, "tid": 435, "ts": 1776649792664738, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664739, "dur": 0, + "args": { + "External id": 85383, + "cbid": 273, "correlation": 85383 + } + }, + { + "ph": "f", "id": 85383, "pid": 435, "tid": 435, "ts": 1776649792664739, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792664748, "dur": 19, + "args": { + "External id": 85384, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85384, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85384, "pid": 0, "tid": 7, "ts": 1776649792664748, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664740, "dur": 4, + "args": { + "External id": 85384, + "cbid": 211, "correlation": 85384 + } + }, + { + "ph": "s", "id": 85384, "pid": 435, "tid": 435, "ts": 1776649792664740, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792664783, "dur": 1, + "args": { + "External id": 11426, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85402, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85402, "pid": 0, "tid": 7, "ts": 1776649792664783, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664774, "dur": 6, + "args": { + "External id": 11426, + "cbid": 211, "correlation": 85402 + } + }, + { + "ph": "s", "id": 85402, "pid": 435, "tid": 435, "ts": 1776649792664774, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664786, "dur": 0, + "args": { + "External id": 85408, + "cbid": 200, "correlation": 85408 + } + }, + { + "ph": "f", "id": 85408, "pid": 435, "tid": 435, "ts": 1776649792664786, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664787, "dur": 0, + "args": { + "External id": 85409, + "cbid": 273, "correlation": 85409 + } + }, + { + "ph": "f", "id": 85409, "pid": 435, "tid": 435, "ts": 1776649792664787, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792664796, "dur": 19, + "args": { + "External id": 85410, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85410, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85410, "pid": 0, "tid": 7, "ts": 1776649792664796, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664788, "dur": 5, + "args": { + "External id": 85410, + "cbid": 211, "correlation": 85410 + } + }, + { + "ph": "s", "id": 85410, "pid": 435, "tid": 435, "ts": 1776649792664788, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792664831, "dur": 1, + "args": { + "External id": 11431, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85428, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85428, "pid": 0, "tid": 7, "ts": 1776649792664831, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664822, "dur": 6, + "args": { + "External id": 11431, + "cbid": 211, "correlation": 85428 + } + }, + { + "ph": "s", "id": 85428, "pid": 435, "tid": 435, "ts": 1776649792664822, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664834, "dur": 0, + "args": { + "External id": 85434, + "cbid": 200, "correlation": 85434 + } + }, + { + "ph": "f", "id": 85434, "pid": 435, "tid": 435, "ts": 1776649792664834, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664835, "dur": 0, + "args": { + "External id": 85435, + "cbid": 273, "correlation": 85435 + } + }, + { + "ph": "f", "id": 85435, "pid": 435, "tid": 435, "ts": 1776649792664835, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792664844, "dur": 19, + "args": { + "External id": 85436, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85436, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85436, "pid": 0, "tid": 7, "ts": 1776649792664844, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664836, "dur": 4, + "args": { + "External id": 85436, + "cbid": 211, "correlation": 85436 + } + }, + { + "ph": "s", "id": 85436, "pid": 435, "tid": 435, "ts": 1776649792664836, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792664879, "dur": 1, + "args": { + "External id": 11436, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85454, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85454, "pid": 0, "tid": 7, "ts": 1776649792664879, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664869, "dur": 6, + "args": { + "External id": 11436, + "cbid": 211, "correlation": 85454 + } + }, + { + "ph": "s", "id": 85454, "pid": 435, "tid": 435, "ts": 1776649792664869, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664882, "dur": 0, + "args": { + "External id": 85460, + "cbid": 200, "correlation": 85460 + } + }, + { + "ph": "f", "id": 85460, "pid": 435, "tid": 435, "ts": 1776649792664882, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664882, "dur": 0, + "args": { + "External id": 85461, + "cbid": 273, "correlation": 85461 + } + }, + { + "ph": "f", "id": 85461, "pid": 435, "tid": 435, "ts": 1776649792664882, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792664891, "dur": 19, + "args": { + "External id": 85462, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85462, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85462, "pid": 0, "tid": 7, "ts": 1776649792664891, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664884, "dur": 4, + "args": { + "External id": 85462, + "cbid": 211, "correlation": 85462 + } + }, + { + "ph": "s", "id": 85462, "pid": 435, "tid": 435, "ts": 1776649792664884, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792664930, "dur": 1, + "args": { + "External id": 11441, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85480, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85480, "pid": 0, "tid": 7, "ts": 1776649792664930, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664917, "dur": 7, + "args": { + "External id": 11441, + "cbid": 211, "correlation": 85480 + } + }, + { + "ph": "s", "id": 85480, "pid": 435, "tid": 435, "ts": 1776649792664917, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664931, "dur": 0, + "args": { + "External id": 85486, + "cbid": 200, "correlation": 85486 + } + }, + { + "ph": "f", "id": 85486, "pid": 435, "tid": 435, "ts": 1776649792664931, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664932, "dur": 0, + "args": { + "External id": 85487, + "cbid": 273, "correlation": 85487 + } + }, + { + "ph": "f", "id": 85487, "pid": 435, "tid": 435, "ts": 1776649792664932, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792664941, "dur": 19, + "args": { + "External id": 85488, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85488, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85488, "pid": 0, "tid": 7, "ts": 1776649792664941, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664933, "dur": 5, + "args": { + "External id": 85488, + "cbid": 211, "correlation": 85488 + } + }, + { + "ph": "s", "id": 85488, "pid": 435, "tid": 435, "ts": 1776649792664933, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792664976, "dur": 1, + "args": { + "External id": 11446, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85506, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85506, "pid": 0, "tid": 7, "ts": 1776649792664976, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664967, "dur": 6, + "args": { + "External id": 11446, + "cbid": 211, "correlation": 85506 + } + }, + { + "ph": "s", "id": 85506, "pid": 435, "tid": 435, "ts": 1776649792664967, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664979, "dur": 0, + "args": { + "External id": 85512, + "cbid": 200, "correlation": 85512 + } + }, + { + "ph": "f", "id": 85512, "pid": 435, "tid": 435, "ts": 1776649792664979, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792664980, "dur": 0, + "args": { + "External id": 85513, + "cbid": 273, "correlation": 85513 + } + }, + { + "ph": "f", "id": 85513, "pid": 435, "tid": 435, "ts": 1776649792664980, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792664989, "dur": 19, + "args": { + "External id": 85514, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85514, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85514, "pid": 0, "tid": 7, "ts": 1776649792664989, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792664981, "dur": 4, + "args": { + "External id": 85514, + "cbid": 211, "correlation": 85514 + } + }, + { + "ph": "s", "id": 85514, "pid": 435, "tid": 435, "ts": 1776649792664981, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792665024, "dur": 1, + "args": { + "External id": 11451, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85532, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85532, "pid": 0, "tid": 7, "ts": 1776649792665024, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665015, "dur": 6, + "args": { + "External id": 11451, + "cbid": 211, "correlation": 85532 + } + }, + { + "ph": "s", "id": 85532, "pid": 435, "tid": 435, "ts": 1776649792665015, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665027, "dur": 0, + "args": { + "External id": 85538, + "cbid": 200, "correlation": 85538 + } + }, + { + "ph": "f", "id": 85538, "pid": 435, "tid": 435, "ts": 1776649792665027, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665028, "dur": 0, + "args": { + "External id": 85539, + "cbid": 273, "correlation": 85539 + } + }, + { + "ph": "f", "id": 85539, "pid": 435, "tid": 435, "ts": 1776649792665028, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792665037, "dur": 19, + "args": { + "External id": 85540, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85540, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85540, "pid": 0, "tid": 7, "ts": 1776649792665037, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665029, "dur": 4, + "args": { + "External id": 85540, + "cbid": 211, "correlation": 85540 + } + }, + { + "ph": "s", "id": 85540, "pid": 435, "tid": 435, "ts": 1776649792665029, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792665071, "dur": 1, + "args": { + "External id": 11456, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85558, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85558, "pid": 0, "tid": 7, "ts": 1776649792665071, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665062, "dur": 6, + "args": { + "External id": 11456, + "cbid": 211, "correlation": 85558 + } + }, + { + "ph": "s", "id": 85558, "pid": 435, "tid": 435, "ts": 1776649792665062, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665075, "dur": 0, + "args": { + "External id": 85564, + "cbid": 200, "correlation": 85564 + } + }, + { + "ph": "f", "id": 85564, "pid": 435, "tid": 435, "ts": 1776649792665075, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665075, "dur": 0, + "args": { + "External id": 85565, + "cbid": 273, "correlation": 85565 + } + }, + { + "ph": "f", "id": 85565, "pid": 435, "tid": 435, "ts": 1776649792665075, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792665085, "dur": 19, + "args": { + "External id": 85566, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85566, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85566, "pid": 0, "tid": 7, "ts": 1776649792665085, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665076, "dur": 5, + "args": { + "External id": 85566, + "cbid": 211, "correlation": 85566 + } + }, + { + "ph": "s", "id": 85566, "pid": 435, "tid": 435, "ts": 1776649792665076, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792665120, "dur": 1, + "args": { + "External id": 11461, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85584, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85584, "pid": 0, "tid": 7, "ts": 1776649792665120, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665111, "dur": 6, + "args": { + "External id": 11461, + "cbid": 211, "correlation": 85584 + } + }, + { + "ph": "s", "id": 85584, "pid": 435, "tid": 435, "ts": 1776649792665111, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665123, "dur": 0, + "args": { + "External id": 85590, + "cbid": 200, "correlation": 85590 + } + }, + { + "ph": "f", "id": 85590, "pid": 435, "tid": 435, "ts": 1776649792665123, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665124, "dur": 0, + "args": { + "External id": 85591, + "cbid": 273, "correlation": 85591 + } + }, + { + "ph": "f", "id": 85591, "pid": 435, "tid": 435, "ts": 1776649792665124, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792665133, "dur": 19, + "args": { + "External id": 85592, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85592, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85592, "pid": 0, "tid": 7, "ts": 1776649792665133, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665125, "dur": 4, + "args": { + "External id": 85592, + "cbid": 211, "correlation": 85592 + } + }, + { + "ph": "s", "id": 85592, "pid": 435, "tid": 435, "ts": 1776649792665125, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792665168, "dur": 1, + "args": { + "External id": 11466, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85610, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85610, "pid": 0, "tid": 7, "ts": 1776649792665168, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665159, "dur": 6, + "args": { + "External id": 11466, + "cbid": 211, "correlation": 85610 + } + }, + { + "ph": "s", "id": 85610, "pid": 435, "tid": 435, "ts": 1776649792665159, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665171, "dur": 0, + "args": { + "External id": 85616, + "cbid": 200, "correlation": 85616 + } + }, + { + "ph": "f", "id": 85616, "pid": 435, "tid": 435, "ts": 1776649792665171, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665172, "dur": 0, + "args": { + "External id": 85617, + "cbid": 273, "correlation": 85617 + } + }, + { + "ph": "f", "id": 85617, "pid": 435, "tid": 435, "ts": 1776649792665172, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792665180, "dur": 19, + "args": { + "External id": 85618, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85618, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85618, "pid": 0, "tid": 7, "ts": 1776649792665180, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665173, "dur": 4, + "args": { + "External id": 85618, + "cbid": 211, "correlation": 85618 + } + }, + { + "ph": "s", "id": 85618, "pid": 435, "tid": 435, "ts": 1776649792665173, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792665216, "dur": 1, + "args": { + "External id": 11471, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85636, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85636, "pid": 0, "tid": 7, "ts": 1776649792665216, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665206, "dur": 6, + "args": { + "External id": 11471, + "cbid": 211, "correlation": 85636 + } + }, + { + "ph": "s", "id": 85636, "pid": 435, "tid": 435, "ts": 1776649792665206, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665219, "dur": 0, + "args": { + "External id": 85642, + "cbid": 200, "correlation": 85642 + } + }, + { + "ph": "f", "id": 85642, "pid": 435, "tid": 435, "ts": 1776649792665219, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665219, "dur": 0, + "args": { + "External id": 85643, + "cbid": 273, "correlation": 85643 + } + }, + { + "ph": "f", "id": 85643, "pid": 435, "tid": 435, "ts": 1776649792665219, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792665229, "dur": 19, + "args": { + "External id": 85644, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85644, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85644, "pid": 0, "tid": 7, "ts": 1776649792665229, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665221, "dur": 4, + "args": { + "External id": 85644, + "cbid": 211, "correlation": 85644 + } + }, + { + "ph": "s", "id": 85644, "pid": 435, "tid": 435, "ts": 1776649792665221, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792665264, "dur": 1, + "args": { + "External id": 11476, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85662, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85662, "pid": 0, "tid": 7, "ts": 1776649792665264, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665255, "dur": 6, + "args": { + "External id": 11476, + "cbid": 211, "correlation": 85662 + } + }, + { + "ph": "s", "id": 85662, "pid": 435, "tid": 435, "ts": 1776649792665255, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665268, "dur": 0, + "args": { + "External id": 85668, + "cbid": 200, "correlation": 85668 + } + }, + { + "ph": "f", "id": 85668, "pid": 435, "tid": 435, "ts": 1776649792665268, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665268, "dur": 0, + "args": { + "External id": 85669, + "cbid": 273, "correlation": 85669 + } + }, + { + "ph": "f", "id": 85669, "pid": 435, "tid": 435, "ts": 1776649792665268, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792665277, "dur": 19, + "args": { + "External id": 85670, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85670, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85670, "pid": 0, "tid": 7, "ts": 1776649792665277, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665269, "dur": 4, + "args": { + "External id": 85670, + "cbid": 211, "correlation": 85670 + } + }, + { + "ph": "s", "id": 85670, "pid": 435, "tid": 435, "ts": 1776649792665269, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792665312, "dur": 1, + "args": { + "External id": 11481, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85688, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85688, "pid": 0, "tid": 7, "ts": 1776649792665312, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665303, "dur": 6, + "args": { + "External id": 11481, + "cbid": 211, "correlation": 85688 + } + }, + { + "ph": "s", "id": 85688, "pid": 435, "tid": 435, "ts": 1776649792665303, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665315, "dur": 0, + "args": { + "External id": 85694, + "cbid": 200, "correlation": 85694 + } + }, + { + "ph": "f", "id": 85694, "pid": 435, "tid": 435, "ts": 1776649792665315, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665316, "dur": 0, + "args": { + "External id": 85695, + "cbid": 273, "correlation": 85695 + } + }, + { + "ph": "f", "id": 85695, "pid": 435, "tid": 435, "ts": 1776649792665316, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792665325, "dur": 19, + "args": { + "External id": 85696, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85696, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85696, "pid": 0, "tid": 7, "ts": 1776649792665325, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665317, "dur": 4, + "args": { + "External id": 85696, + "cbid": 211, "correlation": 85696 + } + }, + { + "ph": "s", "id": 85696, "pid": 435, "tid": 435, "ts": 1776649792665317, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792665361, "dur": 1, + "args": { + "External id": 11486, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85714, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85714, "pid": 0, "tid": 7, "ts": 1776649792665361, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665351, "dur": 7, + "args": { + "External id": 11486, + "cbid": 211, "correlation": 85714 + } + }, + { + "ph": "s", "id": 85714, "pid": 435, "tid": 435, "ts": 1776649792665351, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665366, "dur": 0, + "args": { + "External id": 85720, + "cbid": 200, "correlation": 85720 + } + }, + { + "ph": "f", "id": 85720, "pid": 435, "tid": 435, "ts": 1776649792665366, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665366, "dur": 0, + "args": { + "External id": 85721, + "cbid": 273, "correlation": 85721 + } + }, + { + "ph": "f", "id": 85721, "pid": 435, "tid": 435, "ts": 1776649792665366, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792665375, "dur": 19, + "args": { + "External id": 85722, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85722, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85722, "pid": 0, "tid": 7, "ts": 1776649792665375, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665367, "dur": 5, + "args": { + "External id": 85722, + "cbid": 211, "correlation": 85722 + } + }, + { + "ph": "s", "id": 85722, "pid": 435, "tid": 435, "ts": 1776649792665367, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792665412, "dur": 1, + "args": { + "External id": 11491, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85740, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85740, "pid": 0, "tid": 7, "ts": 1776649792665412, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665403, "dur": 6, + "args": { + "External id": 11491, + "cbid": 211, "correlation": 85740 + } + }, + { + "ph": "s", "id": 85740, "pid": 435, "tid": 435, "ts": 1776649792665403, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665415, "dur": 0, + "args": { + "External id": 85746, + "cbid": 200, "correlation": 85746 + } + }, + { + "ph": "f", "id": 85746, "pid": 435, "tid": 435, "ts": 1776649792665415, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665416, "dur": 0, + "args": { + "External id": 85747, + "cbid": 273, "correlation": 85747 + } + }, + { + "ph": "f", "id": 85747, "pid": 435, "tid": 435, "ts": 1776649792665416, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792665425, "dur": 19, + "args": { + "External id": 85748, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85748, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85748, "pid": 0, "tid": 7, "ts": 1776649792665425, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665417, "dur": 4, + "args": { + "External id": 85748, + "cbid": 211, "correlation": 85748 + } + }, + { + "ph": "s", "id": 85748, "pid": 435, "tid": 435, "ts": 1776649792665417, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792665461, "dur": 1, + "args": { + "External id": 11496, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85766, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85766, "pid": 0, "tid": 7, "ts": 1776649792665461, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665452, "dur": 6, + "args": { + "External id": 11496, + "cbid": 211, "correlation": 85766 + } + }, + { + "ph": "s", "id": 85766, "pid": 435, "tid": 435, "ts": 1776649792665452, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665464, "dur": 0, + "args": { + "External id": 85772, + "cbid": 200, "correlation": 85772 + } + }, + { + "ph": "f", "id": 85772, "pid": 435, "tid": 435, "ts": 1776649792665464, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665465, "dur": 0, + "args": { + "External id": 85773, + "cbid": 273, "correlation": 85773 + } + }, + { + "ph": "f", "id": 85773, "pid": 435, "tid": 435, "ts": 1776649792665465, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792665474, "dur": 19, + "args": { + "External id": 85774, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85774, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85774, "pid": 0, "tid": 7, "ts": 1776649792665474, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665466, "dur": 5, + "args": { + "External id": 85774, + "cbid": 211, "correlation": 85774 + } + }, + { + "ph": "s", "id": 85774, "pid": 435, "tid": 435, "ts": 1776649792665466, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792665510, "dur": 1, + "args": { + "External id": 11501, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85792, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85792, "pid": 0, "tid": 7, "ts": 1776649792665510, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665500, "dur": 6, + "args": { + "External id": 11501, + "cbid": 211, "correlation": 85792 + } + }, + { + "ph": "s", "id": 85792, "pid": 435, "tid": 435, "ts": 1776649792665500, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665513, "dur": 0, + "args": { + "External id": 85798, + "cbid": 200, "correlation": 85798 + } + }, + { + "ph": "f", "id": 85798, "pid": 435, "tid": 435, "ts": 1776649792665513, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665514, "dur": 0, + "args": { + "External id": 85799, + "cbid": 273, "correlation": 85799 + } + }, + { + "ph": "f", "id": 85799, "pid": 435, "tid": 435, "ts": 1776649792665514, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792665523, "dur": 19, + "args": { + "External id": 85800, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85800, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85800, "pid": 0, "tid": 7, "ts": 1776649792665523, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665515, "dur": 4, + "args": { + "External id": 85800, + "cbid": 211, "correlation": 85800 + } + }, + { + "ph": "s", "id": 85800, "pid": 435, "tid": 435, "ts": 1776649792665515, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792665558, "dur": 1, + "args": { + "External id": 11506, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85818, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85818, "pid": 0, "tid": 7, "ts": 1776649792665558, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665549, "dur": 6, + "args": { + "External id": 11506, + "cbid": 211, "correlation": 85818 + } + }, + { + "ph": "s", "id": 85818, "pid": 435, "tid": 435, "ts": 1776649792665549, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665561, "dur": 0, + "args": { + "External id": 85824, + "cbid": 200, "correlation": 85824 + } + }, + { + "ph": "f", "id": 85824, "pid": 435, "tid": 435, "ts": 1776649792665561, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665562, "dur": 0, + "args": { + "External id": 85825, + "cbid": 273, "correlation": 85825 + } + }, + { + "ph": "f", "id": 85825, "pid": 435, "tid": 435, "ts": 1776649792665562, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792665571, "dur": 19, + "args": { + "External id": 85826, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85826, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85826, "pid": 0, "tid": 7, "ts": 1776649792665571, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665563, "dur": 4, + "args": { + "External id": 85826, + "cbid": 211, "correlation": 85826 + } + }, + { + "ph": "s", "id": 85826, "pid": 435, "tid": 435, "ts": 1776649792665563, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792665606, "dur": 1, + "args": { + "External id": 11511, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85844, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85844, "pid": 0, "tid": 7, "ts": 1776649792665606, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665597, "dur": 6, + "args": { + "External id": 11511, + "cbid": 211, "correlation": 85844 + } + }, + { + "ph": "s", "id": 85844, "pid": 435, "tid": 435, "ts": 1776649792665597, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665609, "dur": 0, + "args": { + "External id": 85850, + "cbid": 200, "correlation": 85850 + } + }, + { + "ph": "f", "id": 85850, "pid": 435, "tid": 435, "ts": 1776649792665609, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665610, "dur": 0, + "args": { + "External id": 85851, + "cbid": 273, "correlation": 85851 + } + }, + { + "ph": "f", "id": 85851, "pid": 435, "tid": 435, "ts": 1776649792665610, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792665619, "dur": 19, + "args": { + "External id": 85852, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85852, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85852, "pid": 0, "tid": 7, "ts": 1776649792665619, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665611, "dur": 5, + "args": { + "External id": 85852, + "cbid": 211, "correlation": 85852 + } + }, + { + "ph": "s", "id": 85852, "pid": 435, "tid": 435, "ts": 1776649792665611, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792665654, "dur": 1, + "args": { + "External id": 11516, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85870, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85870, "pid": 0, "tid": 7, "ts": 1776649792665654, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665645, "dur": 6, + "args": { + "External id": 11516, + "cbid": 211, "correlation": 85870 + } + }, + { + "ph": "s", "id": 85870, "pid": 435, "tid": 435, "ts": 1776649792665645, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665657, "dur": 0, + "args": { + "External id": 85876, + "cbid": 200, "correlation": 85876 + } + }, + { + "ph": "f", "id": 85876, "pid": 435, "tid": 435, "ts": 1776649792665657, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665658, "dur": 0, + "args": { + "External id": 85877, + "cbid": 273, "correlation": 85877 + } + }, + { + "ph": "f", "id": 85877, "pid": 435, "tid": 435, "ts": 1776649792665658, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792665667, "dur": 19, + "args": { + "External id": 85878, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85878, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85878, "pid": 0, "tid": 7, "ts": 1776649792665667, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665659, "dur": 5, + "args": { + "External id": 85878, + "cbid": 211, "correlation": 85878 + } + }, + { + "ph": "s", "id": 85878, "pid": 435, "tid": 435, "ts": 1776649792665659, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792665702, "dur": 1, + "args": { + "External id": 11521, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85896, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85896, "pid": 0, "tid": 7, "ts": 1776649792665702, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665693, "dur": 6, + "args": { + "External id": 11521, + "cbid": 211, "correlation": 85896 + } + }, + { + "ph": "s", "id": 85896, "pid": 435, "tid": 435, "ts": 1776649792665693, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665705, "dur": 0, + "args": { + "External id": 85902, + "cbid": 200, "correlation": 85902 + } + }, + { + "ph": "f", "id": 85902, "pid": 435, "tid": 435, "ts": 1776649792665705, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665706, "dur": 0, + "args": { + "External id": 85903, + "cbid": 273, "correlation": 85903 + } + }, + { + "ph": "f", "id": 85903, "pid": 435, "tid": 435, "ts": 1776649792665706, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792665715, "dur": 19, + "args": { + "External id": 85904, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85904, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85904, "pid": 0, "tid": 7, "ts": 1776649792665715, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665707, "dur": 4, + "args": { + "External id": 85904, + "cbid": 211, "correlation": 85904 + } + }, + { + "ph": "s", "id": 85904, "pid": 435, "tid": 435, "ts": 1776649792665707, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792665750, "dur": 1, + "args": { + "External id": 11526, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85922, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85922, "pid": 0, "tid": 7, "ts": 1776649792665750, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665741, "dur": 6, + "args": { + "External id": 11526, + "cbid": 211, "correlation": 85922 + } + }, + { + "ph": "s", "id": 85922, "pid": 435, "tid": 435, "ts": 1776649792665741, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665753, "dur": 0, + "args": { + "External id": 85928, + "cbid": 200, "correlation": 85928 + } + }, + { + "ph": "f", "id": 85928, "pid": 435, "tid": 435, "ts": 1776649792665753, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665753, "dur": 0, + "args": { + "External id": 85929, + "cbid": 273, "correlation": 85929 + } + }, + { + "ph": "f", "id": 85929, "pid": 435, "tid": 435, "ts": 1776649792665753, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792665762, "dur": 19, + "args": { + "External id": 85930, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85930, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85930, "pid": 0, "tid": 7, "ts": 1776649792665762, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665755, "dur": 4, + "args": { + "External id": 85930, + "cbid": 211, "correlation": 85930 + } + }, + { + "ph": "s", "id": 85930, "pid": 435, "tid": 435, "ts": 1776649792665755, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792665798, "dur": 1, + "args": { + "External id": 11531, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85948, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85948, "pid": 0, "tid": 7, "ts": 1776649792665798, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665789, "dur": 6, + "args": { + "External id": 11531, + "cbid": 211, "correlation": 85948 + } + }, + { + "ph": "s", "id": 85948, "pid": 435, "tid": 435, "ts": 1776649792665789, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665801, "dur": 0, + "args": { + "External id": 85954, + "cbid": 200, "correlation": 85954 + } + }, + { + "ph": "f", "id": 85954, "pid": 435, "tid": 435, "ts": 1776649792665801, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665802, "dur": 0, + "args": { + "External id": 85955, + "cbid": 273, "correlation": 85955 + } + }, + { + "ph": "f", "id": 85955, "pid": 435, "tid": 435, "ts": 1776649792665802, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792665811, "dur": 19, + "args": { + "External id": 85956, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85956, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85956, "pid": 0, "tid": 7, "ts": 1776649792665811, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665803, "dur": 5, + "args": { + "External id": 85956, + "cbid": 211, "correlation": 85956 + } + }, + { + "ph": "s", "id": 85956, "pid": 435, "tid": 435, "ts": 1776649792665803, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792665846, "dur": 1, + "args": { + "External id": 11536, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85974, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85974, "pid": 0, "tid": 7, "ts": 1776649792665846, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665837, "dur": 6, + "args": { + "External id": 11536, + "cbid": 211, "correlation": 85974 + } + }, + { + "ph": "s", "id": 85974, "pid": 435, "tid": 435, "ts": 1776649792665837, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665849, "dur": 0, + "args": { + "External id": 85980, + "cbid": 200, "correlation": 85980 + } + }, + { + "ph": "f", "id": 85980, "pid": 435, "tid": 435, "ts": 1776649792665849, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665850, "dur": 0, + "args": { + "External id": 85981, + "cbid": 273, "correlation": 85981 + } + }, + { + "ph": "f", "id": 85981, "pid": 435, "tid": 435, "ts": 1776649792665850, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792665859, "dur": 19, + "args": { + "External id": 85982, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 85982, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 85982, "pid": 0, "tid": 7, "ts": 1776649792665859, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665851, "dur": 4, + "args": { + "External id": 85982, + "cbid": 211, "correlation": 85982 + } + }, + { + "ph": "s", "id": 85982, "pid": 435, "tid": 435, "ts": 1776649792665851, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792665894, "dur": 1, + "args": { + "External id": 11541, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86000, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86000, "pid": 0, "tid": 7, "ts": 1776649792665894, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665885, "dur": 6, + "args": { + "External id": 11541, + "cbid": 211, "correlation": 86000 + } + }, + { + "ph": "s", "id": 86000, "pid": 435, "tid": 435, "ts": 1776649792665885, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665897, "dur": 0, + "args": { + "External id": 86006, + "cbid": 200, "correlation": 86006 + } + }, + { + "ph": "f", "id": 86006, "pid": 435, "tid": 435, "ts": 1776649792665897, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665898, "dur": 0, + "args": { + "External id": 86007, + "cbid": 273, "correlation": 86007 + } + }, + { + "ph": "f", "id": 86007, "pid": 435, "tid": 435, "ts": 1776649792665898, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792665907, "dur": 19, + "args": { + "External id": 86008, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86008, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86008, "pid": 0, "tid": 7, "ts": 1776649792665907, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665899, "dur": 4, + "args": { + "External id": 86008, + "cbid": 211, "correlation": 86008 + } + }, + { + "ph": "s", "id": 86008, "pid": 435, "tid": 435, "ts": 1776649792665899, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792665942, "dur": 1, + "args": { + "External id": 11546, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86026, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86026, "pid": 0, "tid": 7, "ts": 1776649792665942, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665933, "dur": 6, + "args": { + "External id": 11546, + "cbid": 211, "correlation": 86026 + } + }, + { + "ph": "s", "id": 86026, "pid": 435, "tid": 435, "ts": 1776649792665933, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665945, "dur": 0, + "args": { + "External id": 86032, + "cbid": 200, "correlation": 86032 + } + }, + { + "ph": "f", "id": 86032, "pid": 435, "tid": 435, "ts": 1776649792665945, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665946, "dur": 0, + "args": { + "External id": 86033, + "cbid": 273, "correlation": 86033 + } + }, + { + "ph": "f", "id": 86033, "pid": 435, "tid": 435, "ts": 1776649792665946, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792665955, "dur": 19, + "args": { + "External id": 86034, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86034, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86034, "pid": 0, "tid": 7, "ts": 1776649792665955, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665947, "dur": 4, + "args": { + "External id": 86034, + "cbid": 211, "correlation": 86034 + } + }, + { + "ph": "s", "id": 86034, "pid": 435, "tid": 435, "ts": 1776649792665947, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792665990, "dur": 1, + "args": { + "External id": 11551, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86052, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86052, "pid": 0, "tid": 7, "ts": 1776649792665990, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665981, "dur": 6, + "args": { + "External id": 11551, + "cbid": 211, "correlation": 86052 + } + }, + { + "ph": "s", "id": 86052, "pid": 435, "tid": 435, "ts": 1776649792665981, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665994, "dur": 0, + "args": { + "External id": 86058, + "cbid": 200, "correlation": 86058 + } + }, + { + "ph": "f", "id": 86058, "pid": 435, "tid": 435, "ts": 1776649792665994, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792665994, "dur": 0, + "args": { + "External id": 86059, + "cbid": 273, "correlation": 86059 + } + }, + { + "ph": "f", "id": 86059, "pid": 435, "tid": 435, "ts": 1776649792665994, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792666003, "dur": 19, + "args": { + "External id": 86060, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86060, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86060, "pid": 0, "tid": 7, "ts": 1776649792666003, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792665995, "dur": 4, + "args": { + "External id": 86060, + "cbid": 211, "correlation": 86060 + } + }, + { + "ph": "s", "id": 86060, "pid": 435, "tid": 435, "ts": 1776649792665995, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792666038, "dur": 1, + "args": { + "External id": 11556, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86078, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86078, "pid": 0, "tid": 7, "ts": 1776649792666038, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666029, "dur": 6, + "args": { + "External id": 11556, + "cbid": 211, "correlation": 86078 + } + }, + { + "ph": "s", "id": 86078, "pid": 435, "tid": 435, "ts": 1776649792666029, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666042, "dur": 0, + "args": { + "External id": 86084, + "cbid": 200, "correlation": 86084 + } + }, + { + "ph": "f", "id": 86084, "pid": 435, "tid": 435, "ts": 1776649792666042, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666042, "dur": 0, + "args": { + "External id": 86085, + "cbid": 273, "correlation": 86085 + } + }, + { + "ph": "f", "id": 86085, "pid": 435, "tid": 435, "ts": 1776649792666042, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792666051, "dur": 19, + "args": { + "External id": 86086, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86086, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86086, "pid": 0, "tid": 7, "ts": 1776649792666051, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666043, "dur": 4, + "args": { + "External id": 86086, + "cbid": 211, "correlation": 86086 + } + }, + { + "ph": "s", "id": 86086, "pid": 435, "tid": 435, "ts": 1776649792666043, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792666086, "dur": 1, + "args": { + "External id": 11561, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86104, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86104, "pid": 0, "tid": 7, "ts": 1776649792666086, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666077, "dur": 6, + "args": { + "External id": 11561, + "cbid": 211, "correlation": 86104 + } + }, + { + "ph": "s", "id": 86104, "pid": 435, "tid": 435, "ts": 1776649792666077, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666089, "dur": 0, + "args": { + "External id": 86110, + "cbid": 200, "correlation": 86110 + } + }, + { + "ph": "f", "id": 86110, "pid": 435, "tid": 435, "ts": 1776649792666089, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666090, "dur": 0, + "args": { + "External id": 86111, + "cbid": 273, "correlation": 86111 + } + }, + { + "ph": "f", "id": 86111, "pid": 435, "tid": 435, "ts": 1776649792666090, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792666099, "dur": 19, + "args": { + "External id": 86112, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86112, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86112, "pid": 0, "tid": 7, "ts": 1776649792666099, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666091, "dur": 4, + "args": { + "External id": 86112, + "cbid": 211, "correlation": 86112 + } + }, + { + "ph": "s", "id": 86112, "pid": 435, "tid": 435, "ts": 1776649792666091, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792666134, "dur": 1, + "args": { + "External id": 11566, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86130, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86130, "pid": 0, "tid": 7, "ts": 1776649792666134, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666125, "dur": 6, + "args": { + "External id": 11566, + "cbid": 211, "correlation": 86130 + } + }, + { + "ph": "s", "id": 86130, "pid": 435, "tid": 435, "ts": 1776649792666125, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666137, "dur": 0, + "args": { + "External id": 86136, + "cbid": 200, "correlation": 86136 + } + }, + { + "ph": "f", "id": 86136, "pid": 435, "tid": 435, "ts": 1776649792666137, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666138, "dur": 0, + "args": { + "External id": 86137, + "cbid": 273, "correlation": 86137 + } + }, + { + "ph": "f", "id": 86137, "pid": 435, "tid": 435, "ts": 1776649792666138, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792666147, "dur": 19, + "args": { + "External id": 86138, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86138, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86138, "pid": 0, "tid": 7, "ts": 1776649792666147, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666139, "dur": 4, + "args": { + "External id": 86138, + "cbid": 211, "correlation": 86138 + } + }, + { + "ph": "s", "id": 86138, "pid": 435, "tid": 435, "ts": 1776649792666139, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792666182, "dur": 1, + "args": { + "External id": 11571, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86156, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86156, "pid": 0, "tid": 7, "ts": 1776649792666182, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666173, "dur": 6, + "args": { + "External id": 11571, + "cbid": 211, "correlation": 86156 + } + }, + { + "ph": "s", "id": 86156, "pid": 435, "tid": 435, "ts": 1776649792666173, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666185, "dur": 0, + "args": { + "External id": 86162, + "cbid": 200, "correlation": 86162 + } + }, + { + "ph": "f", "id": 86162, "pid": 435, "tid": 435, "ts": 1776649792666185, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666186, "dur": 0, + "args": { + "External id": 86163, + "cbid": 273, "correlation": 86163 + } + }, + { + "ph": "f", "id": 86163, "pid": 435, "tid": 435, "ts": 1776649792666186, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792666195, "dur": 19, + "args": { + "External id": 86164, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86164, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86164, "pid": 0, "tid": 7, "ts": 1776649792666195, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666187, "dur": 4, + "args": { + "External id": 86164, + "cbid": 211, "correlation": 86164 + } + }, + { + "ph": "s", "id": 86164, "pid": 435, "tid": 435, "ts": 1776649792666187, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792666230, "dur": 1, + "args": { + "External id": 11576, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86182, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86182, "pid": 0, "tid": 7, "ts": 1776649792666230, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666221, "dur": 6, + "args": { + "External id": 11576, + "cbid": 211, "correlation": 86182 + } + }, + { + "ph": "s", "id": 86182, "pid": 435, "tid": 435, "ts": 1776649792666221, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666233, "dur": 0, + "args": { + "External id": 86188, + "cbid": 200, "correlation": 86188 + } + }, + { + "ph": "f", "id": 86188, "pid": 435, "tid": 435, "ts": 1776649792666233, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666234, "dur": 0, + "args": { + "External id": 86189, + "cbid": 273, "correlation": 86189 + } + }, + { + "ph": "f", "id": 86189, "pid": 435, "tid": 435, "ts": 1776649792666234, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792666243, "dur": 19, + "args": { + "External id": 86190, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86190, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86190, "pid": 0, "tid": 7, "ts": 1776649792666243, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666235, "dur": 4, + "args": { + "External id": 86190, + "cbid": 211, "correlation": 86190 + } + }, + { + "ph": "s", "id": 86190, "pid": 435, "tid": 435, "ts": 1776649792666235, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792666278, "dur": 1, + "args": { + "External id": 11581, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86208, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86208, "pid": 0, "tid": 7, "ts": 1776649792666278, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666269, "dur": 6, + "args": { + "External id": 11581, + "cbid": 211, "correlation": 86208 + } + }, + { + "ph": "s", "id": 86208, "pid": 435, "tid": 435, "ts": 1776649792666269, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666281, "dur": 0, + "args": { + "External id": 86214, + "cbid": 200, "correlation": 86214 + } + }, + { + "ph": "f", "id": 86214, "pid": 435, "tid": 435, "ts": 1776649792666281, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666282, "dur": 0, + "args": { + "External id": 86215, + "cbid": 273, "correlation": 86215 + } + }, + { + "ph": "f", "id": 86215, "pid": 435, "tid": 435, "ts": 1776649792666282, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792666291, "dur": 19, + "args": { + "External id": 86216, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86216, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86216, "pid": 0, "tid": 7, "ts": 1776649792666291, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666283, "dur": 4, + "args": { + "External id": 86216, + "cbid": 211, "correlation": 86216 + } + }, + { + "ph": "s", "id": 86216, "pid": 435, "tid": 435, "ts": 1776649792666283, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649792666312, "dur": 4, + "args": { + "External id": 86224, + "cbid": 135, "correlation": 86224 + } + }, + { + "ph": "f", "id": 86224, "pid": 435, "tid": 435, "ts": 1776649792666312, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792666327, "dur": 5, + "args": { + "External id": 86229, + "cbid": 165, "correlation": 86229 + } + }, + { + "ph": "s", "id": 86229, "pid": 435, "tid": 435, "ts": 1776649792666327, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649792666335, "dur": 4, + "args": { + "External id": 86232, + "cbid": 139, "correlation": 86232 + } + }, + { + "ph": "f", "id": 86232, "pid": 435, "tid": 435, "ts": 1776649792666335, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792666390, "dur": 1, + "args": { + "External id": 11586, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86260, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86260, "pid": 0, "tid": 7, "ts": 1776649792666390, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666380, "dur": 8, + "args": { + "External id": 11586, + "cbid": 211, "correlation": 86260 + } + }, + { + "ph": "s", "id": 86260, "pid": 435, "tid": 435, "ts": 1776649792666380, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666395, "dur": 0, + "args": { + "External id": 86266, + "cbid": 200, "correlation": 86266 + } + }, + { + "ph": "f", "id": 86266, "pid": 435, "tid": 435, "ts": 1776649792666395, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666396, "dur": 1, + "args": { + "External id": 86267, + "cbid": 273, "correlation": 86267 + } + }, + { + "ph": "f", "id": 86267, "pid": 435, "tid": 435, "ts": 1776649792666396, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792666405, "dur": 19, + "args": { + "External id": 86268, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86268, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86268, "pid": 0, "tid": 7, "ts": 1776649792666405, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666397, "dur": 5, + "args": { + "External id": 86268, + "cbid": 211, "correlation": 86268 + } + }, + { + "ph": "s", "id": 86268, "pid": 435, "tid": 435, "ts": 1776649792666397, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792666443, "dur": 1, + "args": { + "External id": 11591, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86286, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86286, "pid": 0, "tid": 7, "ts": 1776649792666443, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666433, "dur": 6, + "args": { + "External id": 11591, + "cbid": 211, "correlation": 86286 + } + }, + { + "ph": "s", "id": 86286, "pid": 435, "tid": 435, "ts": 1776649792666433, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666446, "dur": 0, + "args": { + "External id": 86292, + "cbid": 200, "correlation": 86292 + } + }, + { + "ph": "f", "id": 86292, "pid": 435, "tid": 435, "ts": 1776649792666446, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666447, "dur": 0, + "args": { + "External id": 86293, + "cbid": 273, "correlation": 86293 + } + }, + { + "ph": "f", "id": 86293, "pid": 435, "tid": 435, "ts": 1776649792666447, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792666456, "dur": 19, + "args": { + "External id": 86294, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86294, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86294, "pid": 0, "tid": 7, "ts": 1776649792666456, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666448, "dur": 4, + "args": { + "External id": 86294, + "cbid": 211, "correlation": 86294 + } + }, + { + "ph": "s", "id": 86294, "pid": 435, "tid": 435, "ts": 1776649792666448, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792666491, "dur": 1, + "args": { + "External id": 11596, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86312, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86312, "pid": 0, "tid": 7, "ts": 1776649792666491, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666482, "dur": 6, + "args": { + "External id": 11596, + "cbid": 211, "correlation": 86312 + } + }, + { + "ph": "s", "id": 86312, "pid": 435, "tid": 435, "ts": 1776649792666482, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666494, "dur": 0, + "args": { + "External id": 86318, + "cbid": 200, "correlation": 86318 + } + }, + { + "ph": "f", "id": 86318, "pid": 435, "tid": 435, "ts": 1776649792666494, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666495, "dur": 0, + "args": { + "External id": 86319, + "cbid": 273, "correlation": 86319 + } + }, + { + "ph": "f", "id": 86319, "pid": 435, "tid": 435, "ts": 1776649792666495, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792666504, "dur": 19, + "args": { + "External id": 86320, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86320, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86320, "pid": 0, "tid": 7, "ts": 1776649792666504, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666496, "dur": 5, + "args": { + "External id": 86320, + "cbid": 211, "correlation": 86320 + } + }, + { + "ph": "s", "id": 86320, "pid": 435, "tid": 435, "ts": 1776649792666496, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792666540, "dur": 1, + "args": { + "External id": 11601, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86338, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86338, "pid": 0, "tid": 7, "ts": 1776649792666540, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666530, "dur": 7, + "args": { + "External id": 11601, + "cbid": 211, "correlation": 86338 + } + }, + { + "ph": "s", "id": 86338, "pid": 435, "tid": 435, "ts": 1776649792666530, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666543, "dur": 0, + "args": { + "External id": 86344, + "cbid": 200, "correlation": 86344 + } + }, + { + "ph": "f", "id": 86344, "pid": 435, "tid": 435, "ts": 1776649792666543, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666544, "dur": 0, + "args": { + "External id": 86345, + "cbid": 273, "correlation": 86345 + } + }, + { + "ph": "f", "id": 86345, "pid": 435, "tid": 435, "ts": 1776649792666544, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792666553, "dur": 19, + "args": { + "External id": 86346, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86346, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86346, "pid": 0, "tid": 7, "ts": 1776649792666553, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666545, "dur": 5, + "args": { + "External id": 86346, + "cbid": 211, "correlation": 86346 + } + }, + { + "ph": "s", "id": 86346, "pid": 435, "tid": 435, "ts": 1776649792666545, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792666588, "dur": 1, + "args": { + "External id": 11606, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86364, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86364, "pid": 0, "tid": 7, "ts": 1776649792666588, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666579, "dur": 6, + "args": { + "External id": 11606, + "cbid": 211, "correlation": 86364 + } + }, + { + "ph": "s", "id": 86364, "pid": 435, "tid": 435, "ts": 1776649792666579, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666591, "dur": 0, + "args": { + "External id": 86370, + "cbid": 200, "correlation": 86370 + } + }, + { + "ph": "f", "id": 86370, "pid": 435, "tid": 435, "ts": 1776649792666591, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666592, "dur": 0, + "args": { + "External id": 86371, + "cbid": 273, "correlation": 86371 + } + }, + { + "ph": "f", "id": 86371, "pid": 435, "tid": 435, "ts": 1776649792666592, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792666601, "dur": 19, + "args": { + "External id": 86372, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86372, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86372, "pid": 0, "tid": 7, "ts": 1776649792666601, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666593, "dur": 4, + "args": { + "External id": 86372, + "cbid": 211, "correlation": 86372 + } + }, + { + "ph": "s", "id": 86372, "pid": 435, "tid": 435, "ts": 1776649792666593, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792666637, "dur": 1, + "args": { + "External id": 11611, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86390, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86390, "pid": 0, "tid": 7, "ts": 1776649792666637, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666627, "dur": 6, + "args": { + "External id": 11611, + "cbid": 211, "correlation": 86390 + } + }, + { + "ph": "s", "id": 86390, "pid": 435, "tid": 435, "ts": 1776649792666627, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666639, "dur": 0, + "args": { + "External id": 86396, + "cbid": 200, "correlation": 86396 + } + }, + { + "ph": "f", "id": 86396, "pid": 435, "tid": 435, "ts": 1776649792666639, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666640, "dur": 0, + "args": { + "External id": 86397, + "cbid": 273, "correlation": 86397 + } + }, + { + "ph": "f", "id": 86397, "pid": 435, "tid": 435, "ts": 1776649792666640, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792666649, "dur": 19, + "args": { + "External id": 86398, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86398, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86398, "pid": 0, "tid": 7, "ts": 1776649792666649, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666641, "dur": 4, + "args": { + "External id": 86398, + "cbid": 211, "correlation": 86398 + } + }, + { + "ph": "s", "id": 86398, "pid": 435, "tid": 435, "ts": 1776649792666641, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792666685, "dur": 1, + "args": { + "External id": 11616, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86416, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86416, "pid": 0, "tid": 7, "ts": 1776649792666685, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666676, "dur": 6, + "args": { + "External id": 11616, + "cbid": 211, "correlation": 86416 + } + }, + { + "ph": "s", "id": 86416, "pid": 435, "tid": 435, "ts": 1776649792666676, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666688, "dur": 0, + "args": { + "External id": 86422, + "cbid": 200, "correlation": 86422 + } + }, + { + "ph": "f", "id": 86422, "pid": 435, "tid": 435, "ts": 1776649792666688, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666689, "dur": 0, + "args": { + "External id": 86423, + "cbid": 273, "correlation": 86423 + } + }, + { + "ph": "f", "id": 86423, "pid": 435, "tid": 435, "ts": 1776649792666689, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792666698, "dur": 19, + "args": { + "External id": 86424, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86424, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86424, "pid": 0, "tid": 7, "ts": 1776649792666698, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666690, "dur": 5, + "args": { + "External id": 86424, + "cbid": 211, "correlation": 86424 + } + }, + { + "ph": "s", "id": 86424, "pid": 435, "tid": 435, "ts": 1776649792666690, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792666734, "dur": 1, + "args": { + "External id": 11621, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86442, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86442, "pid": 0, "tid": 7, "ts": 1776649792666734, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666725, "dur": 6, + "args": { + "External id": 11621, + "cbid": 211, "correlation": 86442 + } + }, + { + "ph": "s", "id": 86442, "pid": 435, "tid": 435, "ts": 1776649792666725, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666737, "dur": 0, + "args": { + "External id": 86448, + "cbid": 200, "correlation": 86448 + } + }, + { + "ph": "f", "id": 86448, "pid": 435, "tid": 435, "ts": 1776649792666737, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666738, "dur": 0, + "args": { + "External id": 86449, + "cbid": 273, "correlation": 86449 + } + }, + { + "ph": "f", "id": 86449, "pid": 435, "tid": 435, "ts": 1776649792666738, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792666747, "dur": 19, + "args": { + "External id": 86450, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86450, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86450, "pid": 0, "tid": 7, "ts": 1776649792666747, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666739, "dur": 5, + "args": { + "External id": 86450, + "cbid": 211, "correlation": 86450 + } + }, + { + "ph": "s", "id": 86450, "pid": 435, "tid": 435, "ts": 1776649792666739, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792666782, "dur": 1, + "args": { + "External id": 11626, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86468, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86468, "pid": 0, "tid": 7, "ts": 1776649792666782, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666773, "dur": 6, + "args": { + "External id": 11626, + "cbid": 211, "correlation": 86468 + } + }, + { + "ph": "s", "id": 86468, "pid": 435, "tid": 435, "ts": 1776649792666773, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666785, "dur": 0, + "args": { + "External id": 86474, + "cbid": 200, "correlation": 86474 + } + }, + { + "ph": "f", "id": 86474, "pid": 435, "tid": 435, "ts": 1776649792666785, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666786, "dur": 0, + "args": { + "External id": 86475, + "cbid": 273, "correlation": 86475 + } + }, + { + "ph": "f", "id": 86475, "pid": 435, "tid": 435, "ts": 1776649792666786, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792666800, "dur": 19, + "args": { + "External id": 86476, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86476, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86476, "pid": 0, "tid": 7, "ts": 1776649792666800, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666787, "dur": 10, + "args": { + "External id": 86476, + "cbid": 211, "correlation": 86476 + } + }, + { + "ph": "s", "id": 86476, "pid": 435, "tid": 435, "ts": 1776649792666787, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792666839, "dur": 1, + "args": { + "External id": 11631, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86494, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86494, "pid": 0, "tid": 7, "ts": 1776649792666839, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666829, "dur": 7, + "args": { + "External id": 11631, + "cbid": 211, "correlation": 86494 + } + }, + { + "ph": "s", "id": 86494, "pid": 435, "tid": 435, "ts": 1776649792666829, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666842, "dur": 0, + "args": { + "External id": 86500, + "cbid": 200, "correlation": 86500 + } + }, + { + "ph": "f", "id": 86500, "pid": 435, "tid": 435, "ts": 1776649792666842, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666842, "dur": 0, + "args": { + "External id": 86501, + "cbid": 273, "correlation": 86501 + } + }, + { + "ph": "f", "id": 86501, "pid": 435, "tid": 435, "ts": 1776649792666842, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792666852, "dur": 19, + "args": { + "External id": 86502, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86502, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86502, "pid": 0, "tid": 7, "ts": 1776649792666852, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666844, "dur": 5, + "args": { + "External id": 86502, + "cbid": 211, "correlation": 86502 + } + }, + { + "ph": "s", "id": 86502, "pid": 435, "tid": 435, "ts": 1776649792666844, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792666866, "dur": 9, + "args": { + "External id": 86505, + "cbid": 165, "correlation": 86505 + } + }, + { + "ph": "s", "id": 86505, "pid": 435, "tid": 435, "ts": 1776649792666866, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649792666897, "dur": 3, + "args": { + "External id": 86514, + "cbid": 135, "correlation": 86514 + } + }, + { + "ph": "f", "id": 86514, "pid": 435, "tid": 435, "ts": 1776649792666897, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792666945, "dur": 1, + "args": { + "External id": 11636, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86534, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86534, "pid": 0, "tid": 7, "ts": 1776649792666945, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666934, "dur": 8, + "args": { + "External id": 11636, + "cbid": 211, "correlation": 86534 + } + }, + { + "ph": "s", "id": 86534, "pid": 435, "tid": 435, "ts": 1776649792666934, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666949, "dur": 0, + "args": { + "External id": 86540, + "cbid": 200, "correlation": 86540 + } + }, + { + "ph": "f", "id": 86540, "pid": 435, "tid": 435, "ts": 1776649792666949, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792666950, "dur": 0, + "args": { + "External id": 86541, + "cbid": 273, "correlation": 86541 + } + }, + { + "ph": "f", "id": 86541, "pid": 435, "tid": 435, "ts": 1776649792666950, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792666960, "dur": 19, + "args": { + "External id": 86542, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86542, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86542, "pid": 0, "tid": 7, "ts": 1776649792666960, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666951, "dur": 5, + "args": { + "External id": 86542, + "cbid": 211, "correlation": 86542 + } + }, + { + "ph": "s", "id": 86542, "pid": 435, "tid": 435, "ts": 1776649792666951, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792666997, "dur": 1, + "args": { + "External id": 11641, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86560, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86560, "pid": 0, "tid": 7, "ts": 1776649792666997, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792666988, "dur": 6, + "args": { + "External id": 11641, + "cbid": 211, "correlation": 86560 + } + }, + { + "ph": "s", "id": 86560, "pid": 435, "tid": 435, "ts": 1776649792666988, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667000, "dur": 0, + "args": { + "External id": 86566, + "cbid": 200, "correlation": 86566 + } + }, + { + "ph": "f", "id": 86566, "pid": 435, "tid": 435, "ts": 1776649792667000, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667001, "dur": 0, + "args": { + "External id": 86567, + "cbid": 273, "correlation": 86567 + } + }, + { + "ph": "f", "id": 86567, "pid": 435, "tid": 435, "ts": 1776649792667001, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792667011, "dur": 19, + "args": { + "External id": 86568, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86568, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86568, "pid": 0, "tid": 7, "ts": 1776649792667011, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667003, "dur": 4, + "args": { + "External id": 86568, + "cbid": 211, "correlation": 86568 + } + }, + { + "ph": "s", "id": 86568, "pid": 435, "tid": 435, "ts": 1776649792667003, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792667046, "dur": 1, + "args": { + "External id": 11646, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86586, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86586, "pid": 0, "tid": 7, "ts": 1776649792667046, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667037, "dur": 6, + "args": { + "External id": 11646, + "cbid": 211, "correlation": 86586 + } + }, + { + "ph": "s", "id": 86586, "pid": 435, "tid": 435, "ts": 1776649792667037, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667050, "dur": 0, + "args": { + "External id": 86592, + "cbid": 200, "correlation": 86592 + } + }, + { + "ph": "f", "id": 86592, "pid": 435, "tid": 435, "ts": 1776649792667050, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667050, "dur": 0, + "args": { + "External id": 86593, + "cbid": 273, "correlation": 86593 + } + }, + { + "ph": "f", "id": 86593, "pid": 435, "tid": 435, "ts": 1776649792667050, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792667059, "dur": 19, + "args": { + "External id": 86594, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86594, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86594, "pid": 0, "tid": 7, "ts": 1776649792667059, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667051, "dur": 4, + "args": { + "External id": 86594, + "cbid": 211, "correlation": 86594 + } + }, + { + "ph": "s", "id": 86594, "pid": 435, "tid": 435, "ts": 1776649792667051, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792667094, "dur": 1, + "args": { + "External id": 11651, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86612, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86612, "pid": 0, "tid": 7, "ts": 1776649792667094, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667085, "dur": 6, + "args": { + "External id": 11651, + "cbid": 211, "correlation": 86612 + } + }, + { + "ph": "s", "id": 86612, "pid": 435, "tid": 435, "ts": 1776649792667085, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667098, "dur": 0, + "args": { + "External id": 86618, + "cbid": 200, "correlation": 86618 + } + }, + { + "ph": "f", "id": 86618, "pid": 435, "tid": 435, "ts": 1776649792667098, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667098, "dur": 0, + "args": { + "External id": 86619, + "cbid": 273, "correlation": 86619 + } + }, + { + "ph": "f", "id": 86619, "pid": 435, "tid": 435, "ts": 1776649792667098, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792667107, "dur": 19, + "args": { + "External id": 86620, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86620, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86620, "pid": 0, "tid": 7, "ts": 1776649792667107, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667099, "dur": 4, + "args": { + "External id": 86620, + "cbid": 211, "correlation": 86620 + } + }, + { + "ph": "s", "id": 86620, "pid": 435, "tid": 435, "ts": 1776649792667099, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792667143, "dur": 1, + "args": { + "External id": 11656, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86638, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86638, "pid": 0, "tid": 7, "ts": 1776649792667143, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667133, "dur": 6, + "args": { + "External id": 11656, + "cbid": 211, "correlation": 86638 + } + }, + { + "ph": "s", "id": 86638, "pid": 435, "tid": 435, "ts": 1776649792667133, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667146, "dur": 0, + "args": { + "External id": 86644, + "cbid": 200, "correlation": 86644 + } + }, + { + "ph": "f", "id": 86644, "pid": 435, "tid": 435, "ts": 1776649792667146, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667147, "dur": 0, + "args": { + "External id": 86645, + "cbid": 273, "correlation": 86645 + } + }, + { + "ph": "f", "id": 86645, "pid": 435, "tid": 435, "ts": 1776649792667147, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792667156, "dur": 19, + "args": { + "External id": 86646, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86646, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86646, "pid": 0, "tid": 7, "ts": 1776649792667156, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667148, "dur": 4, + "args": { + "External id": 86646, + "cbid": 211, "correlation": 86646 + } + }, + { + "ph": "s", "id": 86646, "pid": 435, "tid": 435, "ts": 1776649792667148, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792667191, "dur": 1, + "args": { + "External id": 11661, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86664, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86664, "pid": 0, "tid": 7, "ts": 1776649792667191, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667182, "dur": 6, + "args": { + "External id": 11661, + "cbid": 211, "correlation": 86664 + } + }, + { + "ph": "s", "id": 86664, "pid": 435, "tid": 435, "ts": 1776649792667182, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667194, "dur": 0, + "args": { + "External id": 86670, + "cbid": 200, "correlation": 86670 + } + }, + { + "ph": "f", "id": 86670, "pid": 435, "tid": 435, "ts": 1776649792667194, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667195, "dur": 0, + "args": { + "External id": 86671, + "cbid": 273, "correlation": 86671 + } + }, + { + "ph": "f", "id": 86671, "pid": 435, "tid": 435, "ts": 1776649792667195, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792667204, "dur": 19, + "args": { + "External id": 86672, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86672, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86672, "pid": 0, "tid": 7, "ts": 1776649792667204, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667196, "dur": 4, + "args": { + "External id": 86672, + "cbid": 211, "correlation": 86672 + } + }, + { + "ph": "s", "id": 86672, "pid": 435, "tid": 435, "ts": 1776649792667196, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792667239, "dur": 1, + "args": { + "External id": 11666, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86690, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86690, "pid": 0, "tid": 7, "ts": 1776649792667239, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667230, "dur": 6, + "args": { + "External id": 11666, + "cbid": 211, "correlation": 86690 + } + }, + { + "ph": "s", "id": 86690, "pid": 435, "tid": 435, "ts": 1776649792667230, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667243, "dur": 0, + "args": { + "External id": 86696, + "cbid": 200, "correlation": 86696 + } + }, + { + "ph": "f", "id": 86696, "pid": 435, "tid": 435, "ts": 1776649792667243, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667243, "dur": 0, + "args": { + "External id": 86697, + "cbid": 273, "correlation": 86697 + } + }, + { + "ph": "f", "id": 86697, "pid": 435, "tid": 435, "ts": 1776649792667243, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792667252, "dur": 19, + "args": { + "External id": 86698, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86698, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86698, "pid": 0, "tid": 7, "ts": 1776649792667252, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667244, "dur": 4, + "args": { + "External id": 86698, + "cbid": 211, "correlation": 86698 + } + }, + { + "ph": "s", "id": 86698, "pid": 435, "tid": 435, "ts": 1776649792667244, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792667287, "dur": 1, + "args": { + "External id": 11671, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86716, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86716, "pid": 0, "tid": 7, "ts": 1776649792667287, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667278, "dur": 6, + "args": { + "External id": 11671, + "cbid": 211, "correlation": 86716 + } + }, + { + "ph": "s", "id": 86716, "pid": 435, "tid": 435, "ts": 1776649792667278, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667291, "dur": 0, + "args": { + "External id": 86722, + "cbid": 200, "correlation": 86722 + } + }, + { + "ph": "f", "id": 86722, "pid": 435, "tid": 435, "ts": 1776649792667291, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667291, "dur": 0, + "args": { + "External id": 86723, + "cbid": 273, "correlation": 86723 + } + }, + { + "ph": "f", "id": 86723, "pid": 435, "tid": 435, "ts": 1776649792667291, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792667300, "dur": 19, + "args": { + "External id": 86724, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86724, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86724, "pid": 0, "tid": 7, "ts": 1776649792667300, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667292, "dur": 4, + "args": { + "External id": 86724, + "cbid": 211, "correlation": 86724 + } + }, + { + "ph": "s", "id": 86724, "pid": 435, "tid": 435, "ts": 1776649792667292, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792667335, "dur": 1, + "args": { + "External id": 11676, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86742, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86742, "pid": 0, "tid": 7, "ts": 1776649792667335, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667326, "dur": 6, + "args": { + "External id": 11676, + "cbid": 211, "correlation": 86742 + } + }, + { + "ph": "s", "id": 86742, "pid": 435, "tid": 435, "ts": 1776649792667326, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667338, "dur": 0, + "args": { + "External id": 86748, + "cbid": 200, "correlation": 86748 + } + }, + { + "ph": "f", "id": 86748, "pid": 435, "tid": 435, "ts": 1776649792667338, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667339, "dur": 0, + "args": { + "External id": 86749, + "cbid": 273, "correlation": 86749 + } + }, + { + "ph": "f", "id": 86749, "pid": 435, "tid": 435, "ts": 1776649792667339, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792667348, "dur": 19, + "args": { + "External id": 86750, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86750, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86750, "pid": 0, "tid": 7, "ts": 1776649792667348, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667340, "dur": 4, + "args": { + "External id": 86750, + "cbid": 211, "correlation": 86750 + } + }, + { + "ph": "s", "id": 86750, "pid": 435, "tid": 435, "ts": 1776649792667340, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792667384, "dur": 1, + "args": { + "External id": 11681, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86768, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86768, "pid": 0, "tid": 7, "ts": 1776649792667384, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667375, "dur": 6, + "args": { + "External id": 11681, + "cbid": 211, "correlation": 86768 + } + }, + { + "ph": "s", "id": 86768, "pid": 435, "tid": 435, "ts": 1776649792667375, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667387, "dur": 0, + "args": { + "External id": 86774, + "cbid": 200, "correlation": 86774 + } + }, + { + "ph": "f", "id": 86774, "pid": 435, "tid": 435, "ts": 1776649792667387, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667388, "dur": 0, + "args": { + "External id": 86775, + "cbid": 273, "correlation": 86775 + } + }, + { + "ph": "f", "id": 86775, "pid": 435, "tid": 435, "ts": 1776649792667388, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792667397, "dur": 19, + "args": { + "External id": 86776, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86776, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86776, "pid": 0, "tid": 7, "ts": 1776649792667397, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667389, "dur": 4, + "args": { + "External id": 86776, + "cbid": 211, "correlation": 86776 + } + }, + { + "ph": "s", "id": 86776, "pid": 435, "tid": 435, "ts": 1776649792667389, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792667432, "dur": 1, + "args": { + "External id": 11686, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86794, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86794, "pid": 0, "tid": 7, "ts": 1776649792667432, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667423, "dur": 6, + "args": { + "External id": 11686, + "cbid": 211, "correlation": 86794 + } + }, + { + "ph": "s", "id": 86794, "pid": 435, "tid": 435, "ts": 1776649792667423, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667435, "dur": 0, + "args": { + "External id": 86800, + "cbid": 200, "correlation": 86800 + } + }, + { + "ph": "f", "id": 86800, "pid": 435, "tid": 435, "ts": 1776649792667435, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667436, "dur": 0, + "args": { + "External id": 86801, + "cbid": 273, "correlation": 86801 + } + }, + { + "ph": "f", "id": 86801, "pid": 435, "tid": 435, "ts": 1776649792667436, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792667445, "dur": 19, + "args": { + "External id": 86802, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86802, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86802, "pid": 0, "tid": 7, "ts": 1776649792667445, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667437, "dur": 5, + "args": { + "External id": 86802, + "cbid": 211, "correlation": 86802 + } + }, + { + "ph": "s", "id": 86802, "pid": 435, "tid": 435, "ts": 1776649792667437, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792667480, "dur": 1, + "args": { + "External id": 11691, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86820, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86820, "pid": 0, "tid": 7, "ts": 1776649792667480, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667471, "dur": 6, + "args": { + "External id": 11691, + "cbid": 211, "correlation": 86820 + } + }, + { + "ph": "s", "id": 86820, "pid": 435, "tid": 435, "ts": 1776649792667471, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667483, "dur": 0, + "args": { + "External id": 86826, + "cbid": 200, "correlation": 86826 + } + }, + { + "ph": "f", "id": 86826, "pid": 435, "tid": 435, "ts": 1776649792667483, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667484, "dur": 0, + "args": { + "External id": 86827, + "cbid": 273, "correlation": 86827 + } + }, + { + "ph": "f", "id": 86827, "pid": 435, "tid": 435, "ts": 1776649792667484, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792667493, "dur": 19, + "args": { + "External id": 86828, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86828, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86828, "pid": 0, "tid": 7, "ts": 1776649792667493, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667485, "dur": 4, + "args": { + "External id": 86828, + "cbid": 211, "correlation": 86828 + } + }, + { + "ph": "s", "id": 86828, "pid": 435, "tid": 435, "ts": 1776649792667485, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792667528, "dur": 1, + "args": { + "External id": 11696, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86846, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86846, "pid": 0, "tid": 7, "ts": 1776649792667528, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667519, "dur": 6, + "args": { + "External id": 11696, + "cbid": 211, "correlation": 86846 + } + }, + { + "ph": "s", "id": 86846, "pid": 435, "tid": 435, "ts": 1776649792667519, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667532, "dur": 0, + "args": { + "External id": 86852, + "cbid": 200, "correlation": 86852 + } + }, + { + "ph": "f", "id": 86852, "pid": 435, "tid": 435, "ts": 1776649792667532, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667532, "dur": 0, + "args": { + "External id": 86853, + "cbid": 273, "correlation": 86853 + } + }, + { + "ph": "f", "id": 86853, "pid": 435, "tid": 435, "ts": 1776649792667532, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792667542, "dur": 19, + "args": { + "External id": 86854, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86854, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86854, "pid": 0, "tid": 7, "ts": 1776649792667542, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667534, "dur": 4, + "args": { + "External id": 86854, + "cbid": 211, "correlation": 86854 + } + }, + { + "ph": "s", "id": 86854, "pid": 435, "tid": 435, "ts": 1776649792667534, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792667577, "dur": 1, + "args": { + "External id": 11701, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86872, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86872, "pid": 0, "tid": 7, "ts": 1776649792667577, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667568, "dur": 6, + "args": { + "External id": 11701, + "cbid": 211, "correlation": 86872 + } + }, + { + "ph": "s", "id": 86872, "pid": 435, "tid": 435, "ts": 1776649792667568, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667580, "dur": 0, + "args": { + "External id": 86878, + "cbid": 200, "correlation": 86878 + } + }, + { + "ph": "f", "id": 86878, "pid": 435, "tid": 435, "ts": 1776649792667580, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667581, "dur": 0, + "args": { + "External id": 86879, + "cbid": 273, "correlation": 86879 + } + }, + { + "ph": "f", "id": 86879, "pid": 435, "tid": 435, "ts": 1776649792667581, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792667590, "dur": 19, + "args": { + "External id": 86880, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86880, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86880, "pid": 0, "tid": 7, "ts": 1776649792667590, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667582, "dur": 4, + "args": { + "External id": 86880, + "cbid": 211, "correlation": 86880 + } + }, + { + "ph": "s", "id": 86880, "pid": 435, "tid": 435, "ts": 1776649792667582, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792667625, "dur": 1, + "args": { + "External id": 11706, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86898, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86898, "pid": 0, "tid": 7, "ts": 1776649792667625, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667616, "dur": 6, + "args": { + "External id": 11706, + "cbid": 211, "correlation": 86898 + } + }, + { + "ph": "s", "id": 86898, "pid": 435, "tid": 435, "ts": 1776649792667616, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667628, "dur": 0, + "args": { + "External id": 86904, + "cbid": 200, "correlation": 86904 + } + }, + { + "ph": "f", "id": 86904, "pid": 435, "tid": 435, "ts": 1776649792667628, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667629, "dur": 0, + "args": { + "External id": 86905, + "cbid": 273, "correlation": 86905 + } + }, + { + "ph": "f", "id": 86905, "pid": 435, "tid": 435, "ts": 1776649792667629, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792667638, "dur": 19, + "args": { + "External id": 86906, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86906, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86906, "pid": 0, "tid": 7, "ts": 1776649792667638, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667630, "dur": 4, + "args": { + "External id": 86906, + "cbid": 211, "correlation": 86906 + } + }, + { + "ph": "s", "id": 86906, "pid": 435, "tid": 435, "ts": 1776649792667630, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792667673, "dur": 1, + "args": { + "External id": 11711, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86924, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86924, "pid": 0, "tid": 7, "ts": 1776649792667673, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667664, "dur": 6, + "args": { + "External id": 11711, + "cbid": 211, "correlation": 86924 + } + }, + { + "ph": "s", "id": 86924, "pid": 435, "tid": 435, "ts": 1776649792667664, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667676, "dur": 0, + "args": { + "External id": 86930, + "cbid": 200, "correlation": 86930 + } + }, + { + "ph": "f", "id": 86930, "pid": 435, "tid": 435, "ts": 1776649792667676, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667677, "dur": 0, + "args": { + "External id": 86931, + "cbid": 273, "correlation": 86931 + } + }, + { + "ph": "f", "id": 86931, "pid": 435, "tid": 435, "ts": 1776649792667677, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792667686, "dur": 19, + "args": { + "External id": 86932, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86932, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86932, "pid": 0, "tid": 7, "ts": 1776649792667686, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667678, "dur": 4, + "args": { + "External id": 86932, + "cbid": 211, "correlation": 86932 + } + }, + { + "ph": "s", "id": 86932, "pid": 435, "tid": 435, "ts": 1776649792667678, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792667721, "dur": 1, + "args": { + "External id": 11716, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86950, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86950, "pid": 0, "tid": 7, "ts": 1776649792667721, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667711, "dur": 6, + "args": { + "External id": 11716, + "cbid": 211, "correlation": 86950 + } + }, + { + "ph": "s", "id": 86950, "pid": 435, "tid": 435, "ts": 1776649792667711, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667724, "dur": 0, + "args": { + "External id": 86956, + "cbid": 200, "correlation": 86956 + } + }, + { + "ph": "f", "id": 86956, "pid": 435, "tid": 435, "ts": 1776649792667724, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667725, "dur": 0, + "args": { + "External id": 86957, + "cbid": 273, "correlation": 86957 + } + }, + { + "ph": "f", "id": 86957, "pid": 435, "tid": 435, "ts": 1776649792667725, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792667734, "dur": 19, + "args": { + "External id": 86958, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86958, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86958, "pid": 0, "tid": 7, "ts": 1776649792667734, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667726, "dur": 4, + "args": { + "External id": 86958, + "cbid": 211, "correlation": 86958 + } + }, + { + "ph": "s", "id": 86958, "pid": 435, "tid": 435, "ts": 1776649792667726, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792667769, "dur": 1, + "args": { + "External id": 11721, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86976, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86976, "pid": 0, "tid": 7, "ts": 1776649792667769, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667759, "dur": 6, + "args": { + "External id": 11721, + "cbid": 211, "correlation": 86976 + } + }, + { + "ph": "s", "id": 86976, "pid": 435, "tid": 435, "ts": 1776649792667759, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667772, "dur": 0, + "args": { + "External id": 86982, + "cbid": 200, "correlation": 86982 + } + }, + { + "ph": "f", "id": 86982, "pid": 435, "tid": 435, "ts": 1776649792667772, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667772, "dur": 0, + "args": { + "External id": 86983, + "cbid": 273, "correlation": 86983 + } + }, + { + "ph": "f", "id": 86983, "pid": 435, "tid": 435, "ts": 1776649792667772, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792667781, "dur": 19, + "args": { + "External id": 86984, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 86984, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 86984, "pid": 0, "tid": 7, "ts": 1776649792667781, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667773, "dur": 4, + "args": { + "External id": 86984, + "cbid": 211, "correlation": 86984 + } + }, + { + "ph": "s", "id": 86984, "pid": 435, "tid": 435, "ts": 1776649792667773, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792667818, "dur": 1, + "args": { + "External id": 11726, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87002, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87002, "pid": 0, "tid": 7, "ts": 1776649792667818, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667809, "dur": 6, + "args": { + "External id": 11726, + "cbid": 211, "correlation": 87002 + } + }, + { + "ph": "s", "id": 87002, "pid": 435, "tid": 435, "ts": 1776649792667809, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667821, "dur": 0, + "args": { + "External id": 87008, + "cbid": 200, "correlation": 87008 + } + }, + { + "ph": "f", "id": 87008, "pid": 435, "tid": 435, "ts": 1776649792667821, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667822, "dur": 0, + "args": { + "External id": 87009, + "cbid": 273, "correlation": 87009 + } + }, + { + "ph": "f", "id": 87009, "pid": 435, "tid": 435, "ts": 1776649792667822, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792667831, "dur": 19, + "args": { + "External id": 87010, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87010, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87010, "pid": 0, "tid": 7, "ts": 1776649792667831, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667823, "dur": 5, + "args": { + "External id": 87010, + "cbid": 211, "correlation": 87010 + } + }, + { + "ph": "s", "id": 87010, "pid": 435, "tid": 435, "ts": 1776649792667823, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792667866, "dur": 1, + "args": { + "External id": 11731, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87028, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87028, "pid": 0, "tid": 7, "ts": 1776649792667866, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667857, "dur": 6, + "args": { + "External id": 11731, + "cbid": 211, "correlation": 87028 + } + }, + { + "ph": "s", "id": 87028, "pid": 435, "tid": 435, "ts": 1776649792667857, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667869, "dur": 0, + "args": { + "External id": 87034, + "cbid": 200, "correlation": 87034 + } + }, + { + "ph": "f", "id": 87034, "pid": 435, "tid": 435, "ts": 1776649792667869, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667870, "dur": 0, + "args": { + "External id": 87035, + "cbid": 273, "correlation": 87035 + } + }, + { + "ph": "f", "id": 87035, "pid": 435, "tid": 435, "ts": 1776649792667870, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792667879, "dur": 19, + "args": { + "External id": 87036, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87036, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87036, "pid": 0, "tid": 7, "ts": 1776649792667879, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667871, "dur": 4, + "args": { + "External id": 87036, + "cbid": 211, "correlation": 87036 + } + }, + { + "ph": "s", "id": 87036, "pid": 435, "tid": 435, "ts": 1776649792667871, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792667914, "dur": 1, + "args": { + "External id": 11736, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87054, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87054, "pid": 0, "tid": 7, "ts": 1776649792667914, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667905, "dur": 6, + "args": { + "External id": 11736, + "cbid": 211, "correlation": 87054 + } + }, + { + "ph": "s", "id": 87054, "pid": 435, "tid": 435, "ts": 1776649792667905, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667917, "dur": 0, + "args": { + "External id": 87060, + "cbid": 200, "correlation": 87060 + } + }, + { + "ph": "f", "id": 87060, "pid": 435, "tid": 435, "ts": 1776649792667917, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667918, "dur": 0, + "args": { + "External id": 87061, + "cbid": 273, "correlation": 87061 + } + }, + { + "ph": "f", "id": 87061, "pid": 435, "tid": 435, "ts": 1776649792667918, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792667927, "dur": 19, + "args": { + "External id": 87062, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87062, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87062, "pid": 0, "tid": 7, "ts": 1776649792667927, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667919, "dur": 5, + "args": { + "External id": 87062, + "cbid": 211, "correlation": 87062 + } + }, + { + "ph": "s", "id": 87062, "pid": 435, "tid": 435, "ts": 1776649792667919, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792667962, "dur": 1, + "args": { + "External id": 11741, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87080, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87080, "pid": 0, "tid": 7, "ts": 1776649792667962, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667953, "dur": 6, + "args": { + "External id": 11741, + "cbid": 211, "correlation": 87080 + } + }, + { + "ph": "s", "id": 87080, "pid": 435, "tid": 435, "ts": 1776649792667953, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667965, "dur": 0, + "args": { + "External id": 87086, + "cbid": 200, "correlation": 87086 + } + }, + { + "ph": "f", "id": 87086, "pid": 435, "tid": 435, "ts": 1776649792667965, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792667966, "dur": 0, + "args": { + "External id": 87087, + "cbid": 273, "correlation": 87087 + } + }, + { + "ph": "f", "id": 87087, "pid": 435, "tid": 435, "ts": 1776649792667966, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792667975, "dur": 19, + "args": { + "External id": 87088, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87088, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87088, "pid": 0, "tid": 7, "ts": 1776649792667975, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792667967, "dur": 4, + "args": { + "External id": 87088, + "cbid": 211, "correlation": 87088 + } + }, + { + "ph": "s", "id": 87088, "pid": 435, "tid": 435, "ts": 1776649792667967, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792668010, "dur": 1, + "args": { + "External id": 11746, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87106, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87106, "pid": 0, "tid": 7, "ts": 1776649792668010, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668001, "dur": 6, + "args": { + "External id": 11746, + "cbid": 211, "correlation": 87106 + } + }, + { + "ph": "s", "id": 87106, "pid": 435, "tid": 435, "ts": 1776649792668001, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668013, "dur": 0, + "args": { + "External id": 87112, + "cbid": 200, "correlation": 87112 + } + }, + { + "ph": "f", "id": 87112, "pid": 435, "tid": 435, "ts": 1776649792668013, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668014, "dur": 0, + "args": { + "External id": 87113, + "cbid": 273, "correlation": 87113 + } + }, + { + "ph": "f", "id": 87113, "pid": 435, "tid": 435, "ts": 1776649792668014, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792668023, "dur": 19, + "args": { + "External id": 87114, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87114, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87114, "pid": 0, "tid": 7, "ts": 1776649792668023, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668015, "dur": 4, + "args": { + "External id": 87114, + "cbid": 211, "correlation": 87114 + } + }, + { + "ph": "s", "id": 87114, "pid": 435, "tid": 435, "ts": 1776649792668015, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792668058, "dur": 1, + "args": { + "External id": 11751, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87132, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87132, "pid": 0, "tid": 7, "ts": 1776649792668058, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668048, "dur": 6, + "args": { + "External id": 11751, + "cbid": 211, "correlation": 87132 + } + }, + { + "ph": "s", "id": 87132, "pid": 435, "tid": 435, "ts": 1776649792668048, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668061, "dur": 0, + "args": { + "External id": 87138, + "cbid": 200, "correlation": 87138 + } + }, + { + "ph": "f", "id": 87138, "pid": 435, "tid": 435, "ts": 1776649792668061, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668061, "dur": 0, + "args": { + "External id": 87139, + "cbid": 273, "correlation": 87139 + } + }, + { + "ph": "f", "id": 87139, "pid": 435, "tid": 435, "ts": 1776649792668061, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792668070, "dur": 19, + "args": { + "External id": 87140, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87140, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87140, "pid": 0, "tid": 7, "ts": 1776649792668070, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668062, "dur": 4, + "args": { + "External id": 87140, + "cbid": 211, "correlation": 87140 + } + }, + { + "ph": "s", "id": 87140, "pid": 435, "tid": 435, "ts": 1776649792668062, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792668104, "dur": 1, + "args": { + "External id": 11756, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87158, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87158, "pid": 0, "tid": 7, "ts": 1776649792668104, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668095, "dur": 6, + "args": { + "External id": 11756, + "cbid": 211, "correlation": 87158 + } + }, + { + "ph": "s", "id": 87158, "pid": 435, "tid": 435, "ts": 1776649792668095, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668108, "dur": 0, + "args": { + "External id": 87164, + "cbid": 200, "correlation": 87164 + } + }, + { + "ph": "f", "id": 87164, "pid": 435, "tid": 435, "ts": 1776649792668108, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668108, "dur": 0, + "args": { + "External id": 87165, + "cbid": 273, "correlation": 87165 + } + }, + { + "ph": "f", "id": 87165, "pid": 435, "tid": 435, "ts": 1776649792668108, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792668118, "dur": 19, + "args": { + "External id": 87166, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87166, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87166, "pid": 0, "tid": 7, "ts": 1776649792668118, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668109, "dur": 4, + "args": { + "External id": 87166, + "cbid": 211, "correlation": 87166 + } + }, + { + "ph": "s", "id": 87166, "pid": 435, "tid": 435, "ts": 1776649792668109, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792668153, "dur": 1, + "args": { + "External id": 11761, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87184, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87184, "pid": 0, "tid": 7, "ts": 1776649792668153, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668143, "dur": 6, + "args": { + "External id": 11761, + "cbid": 211, "correlation": 87184 + } + }, + { + "ph": "s", "id": 87184, "pid": 435, "tid": 435, "ts": 1776649792668143, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668155, "dur": 0, + "args": { + "External id": 87190, + "cbid": 200, "correlation": 87190 + } + }, + { + "ph": "f", "id": 87190, "pid": 435, "tid": 435, "ts": 1776649792668155, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668156, "dur": 0, + "args": { + "External id": 87191, + "cbid": 273, "correlation": 87191 + } + }, + { + "ph": "f", "id": 87191, "pid": 435, "tid": 435, "ts": 1776649792668156, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792668165, "dur": 19, + "args": { + "External id": 87192, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87192, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87192, "pid": 0, "tid": 7, "ts": 1776649792668165, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668157, "dur": 4, + "args": { + "External id": 87192, + "cbid": 211, "correlation": 87192 + } + }, + { + "ph": "s", "id": 87192, "pid": 435, "tid": 435, "ts": 1776649792668157, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792668200, "dur": 1, + "args": { + "External id": 11766, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87210, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87210, "pid": 0, "tid": 7, "ts": 1776649792668200, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668191, "dur": 6, + "args": { + "External id": 11766, + "cbid": 211, "correlation": 87210 + } + }, + { + "ph": "s", "id": 87210, "pid": 435, "tid": 435, "ts": 1776649792668191, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668203, "dur": 0, + "args": { + "External id": 87216, + "cbid": 200, "correlation": 87216 + } + }, + { + "ph": "f", "id": 87216, "pid": 435, "tid": 435, "ts": 1776649792668203, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668204, "dur": 0, + "args": { + "External id": 87217, + "cbid": 273, "correlation": 87217 + } + }, + { + "ph": "f", "id": 87217, "pid": 435, "tid": 435, "ts": 1776649792668204, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792668213, "dur": 19, + "args": { + "External id": 87218, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87218, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87218, "pid": 0, "tid": 7, "ts": 1776649792668213, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668205, "dur": 4, + "args": { + "External id": 87218, + "cbid": 211, "correlation": 87218 + } + }, + { + "ph": "s", "id": 87218, "pid": 435, "tid": 435, "ts": 1776649792668205, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792668270, "dur": 1, + "args": { + "External id": 11771, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87236, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87236, "pid": 0, "tid": 7, "ts": 1776649792668270, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668260, "dur": 7, + "args": { + "External id": 11771, + "cbid": 211, "correlation": 87236 + } + }, + { + "ph": "s", "id": 87236, "pid": 435, "tid": 435, "ts": 1776649792668260, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668274, "dur": 0, + "args": { + "External id": 87242, + "cbid": 200, "correlation": 87242 + } + }, + { + "ph": "f", "id": 87242, "pid": 435, "tid": 435, "ts": 1776649792668274, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668274, "dur": 0, + "args": { + "External id": 87243, + "cbid": 273, "correlation": 87243 + } + }, + { + "ph": "f", "id": 87243, "pid": 435, "tid": 435, "ts": 1776649792668274, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792668283, "dur": 19, + "args": { + "External id": 87244, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87244, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87244, "pid": 0, "tid": 7, "ts": 1776649792668283, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668276, "dur": 4, + "args": { + "External id": 87244, + "cbid": 211, "correlation": 87244 + } + }, + { + "ph": "s", "id": 87244, "pid": 435, "tid": 435, "ts": 1776649792668276, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792668320, "dur": 1, + "args": { + "External id": 11776, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87262, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87262, "pid": 0, "tid": 7, "ts": 1776649792668320, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668310, "dur": 6, + "args": { + "External id": 11776, + "cbid": 211, "correlation": 87262 + } + }, + { + "ph": "s", "id": 87262, "pid": 435, "tid": 435, "ts": 1776649792668310, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668323, "dur": 0, + "args": { + "External id": 87268, + "cbid": 200, "correlation": 87268 + } + }, + { + "ph": "f", "id": 87268, "pid": 435, "tid": 435, "ts": 1776649792668323, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668323, "dur": 0, + "args": { + "External id": 87269, + "cbid": 273, "correlation": 87269 + } + }, + { + "ph": "f", "id": 87269, "pid": 435, "tid": 435, "ts": 1776649792668323, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792668333, "dur": 19, + "args": { + "External id": 87270, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87270, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87270, "pid": 0, "tid": 7, "ts": 1776649792668333, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668324, "dur": 4, + "args": { + "External id": 87270, + "cbid": 211, "correlation": 87270 + } + }, + { + "ph": "s", "id": 87270, "pid": 435, "tid": 435, "ts": 1776649792668324, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792668378, "dur": 1, + "args": { + "External id": 11781, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87288, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87288, "pid": 0, "tid": 7, "ts": 1776649792668378, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668369, "dur": 6, + "args": { + "External id": 11781, + "cbid": 211, "correlation": 87288 + } + }, + { + "ph": "s", "id": 87288, "pid": 435, "tid": 435, "ts": 1776649792668369, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668381, "dur": 0, + "args": { + "External id": 87294, + "cbid": 200, "correlation": 87294 + } + }, + { + "ph": "f", "id": 87294, "pid": 435, "tid": 435, "ts": 1776649792668381, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668382, "dur": 0, + "args": { + "External id": 87295, + "cbid": 273, "correlation": 87295 + } + }, + { + "ph": "f", "id": 87295, "pid": 435, "tid": 435, "ts": 1776649792668382, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792668391, "dur": 19, + "args": { + "External id": 87296, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87296, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87296, "pid": 0, "tid": 7, "ts": 1776649792668391, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668383, "dur": 5, + "args": { + "External id": 87296, + "cbid": 211, "correlation": 87296 + } + }, + { + "ph": "s", "id": 87296, "pid": 435, "tid": 435, "ts": 1776649792668383, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792668426, "dur": 1, + "args": { + "External id": 11786, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87314, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87314, "pid": 0, "tid": 7, "ts": 1776649792668426, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668417, "dur": 6, + "args": { + "External id": 11786, + "cbid": 211, "correlation": 87314 + } + }, + { + "ph": "s", "id": 87314, "pid": 435, "tid": 435, "ts": 1776649792668417, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668429, "dur": 0, + "args": { + "External id": 87320, + "cbid": 200, "correlation": 87320 + } + }, + { + "ph": "f", "id": 87320, "pid": 435, "tid": 435, "ts": 1776649792668429, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668430, "dur": 0, + "args": { + "External id": 87321, + "cbid": 273, "correlation": 87321 + } + }, + { + "ph": "f", "id": 87321, "pid": 435, "tid": 435, "ts": 1776649792668430, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792668439, "dur": 19, + "args": { + "External id": 87322, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87322, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87322, "pid": 0, "tid": 7, "ts": 1776649792668439, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668431, "dur": 5, + "args": { + "External id": 87322, + "cbid": 211, "correlation": 87322 + } + }, + { + "ph": "s", "id": 87322, "pid": 435, "tid": 435, "ts": 1776649792668431, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792668474, "dur": 1, + "args": { + "External id": 11791, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87340, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87340, "pid": 0, "tid": 7, "ts": 1776649792668474, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668465, "dur": 6, + "args": { + "External id": 11791, + "cbid": 211, "correlation": 87340 + } + }, + { + "ph": "s", "id": 87340, "pid": 435, "tid": 435, "ts": 1776649792668465, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668477, "dur": 0, + "args": { + "External id": 87346, + "cbid": 200, "correlation": 87346 + } + }, + { + "ph": "f", "id": 87346, "pid": 435, "tid": 435, "ts": 1776649792668477, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668478, "dur": 0, + "args": { + "External id": 87347, + "cbid": 273, "correlation": 87347 + } + }, + { + "ph": "f", "id": 87347, "pid": 435, "tid": 435, "ts": 1776649792668478, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792668487, "dur": 19, + "args": { + "External id": 87348, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87348, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87348, "pid": 0, "tid": 7, "ts": 1776649792668487, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668479, "dur": 4, + "args": { + "External id": 87348, + "cbid": 211, "correlation": 87348 + } + }, + { + "ph": "s", "id": 87348, "pid": 435, "tid": 435, "ts": 1776649792668479, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792668523, "dur": 1, + "args": { + "External id": 11796, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87366, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87366, "pid": 0, "tid": 7, "ts": 1776649792668523, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668513, "dur": 7, + "args": { + "External id": 11796, + "cbid": 211, "correlation": 87366 + } + }, + { + "ph": "s", "id": 87366, "pid": 435, "tid": 435, "ts": 1776649792668513, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668526, "dur": 0, + "args": { + "External id": 87372, + "cbid": 200, "correlation": 87372 + } + }, + { + "ph": "f", "id": 87372, "pid": 435, "tid": 435, "ts": 1776649792668526, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668526, "dur": 0, + "args": { + "External id": 87373, + "cbid": 273, "correlation": 87373 + } + }, + { + "ph": "f", "id": 87373, "pid": 435, "tid": 435, "ts": 1776649792668526, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792668535, "dur": 19, + "args": { + "External id": 87374, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87374, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87374, "pid": 0, "tid": 7, "ts": 1776649792668535, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668527, "dur": 4, + "args": { + "External id": 87374, + "cbid": 211, "correlation": 87374 + } + }, + { + "ph": "s", "id": 87374, "pid": 435, "tid": 435, "ts": 1776649792668527, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792668570, "dur": 1, + "args": { + "External id": 11801, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87392, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87392, "pid": 0, "tid": 7, "ts": 1776649792668570, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668561, "dur": 6, + "args": { + "External id": 11801, + "cbid": 211, "correlation": 87392 + } + }, + { + "ph": "s", "id": 87392, "pid": 435, "tid": 435, "ts": 1776649792668561, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668573, "dur": 0, + "args": { + "External id": 87398, + "cbid": 200, "correlation": 87398 + } + }, + { + "ph": "f", "id": 87398, "pid": 435, "tid": 435, "ts": 1776649792668573, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668574, "dur": 0, + "args": { + "External id": 87399, + "cbid": 273, "correlation": 87399 + } + }, + { + "ph": "f", "id": 87399, "pid": 435, "tid": 435, "ts": 1776649792668574, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792668583, "dur": 19, + "args": { + "External id": 87400, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87400, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87400, "pid": 0, "tid": 7, "ts": 1776649792668583, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668575, "dur": 5, + "args": { + "External id": 87400, + "cbid": 211, "correlation": 87400 + } + }, + { + "ph": "s", "id": 87400, "pid": 435, "tid": 435, "ts": 1776649792668575, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792668618, "dur": 1, + "args": { + "External id": 11806, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87418, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87418, "pid": 0, "tid": 7, "ts": 1776649792668618, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668609, "dur": 6, + "args": { + "External id": 11806, + "cbid": 211, "correlation": 87418 + } + }, + { + "ph": "s", "id": 87418, "pid": 435, "tid": 435, "ts": 1776649792668609, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668621, "dur": 0, + "args": { + "External id": 87424, + "cbid": 200, "correlation": 87424 + } + }, + { + "ph": "f", "id": 87424, "pid": 435, "tid": 435, "ts": 1776649792668621, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668622, "dur": 0, + "args": { + "External id": 87425, + "cbid": 273, "correlation": 87425 + } + }, + { + "ph": "f", "id": 87425, "pid": 435, "tid": 435, "ts": 1776649792668622, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792668631, "dur": 19, + "args": { + "External id": 87426, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87426, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87426, "pid": 0, "tid": 7, "ts": 1776649792668631, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668623, "dur": 4, + "args": { + "External id": 87426, + "cbid": 211, "correlation": 87426 + } + }, + { + "ph": "s", "id": 87426, "pid": 435, "tid": 435, "ts": 1776649792668623, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792668666, "dur": 1, + "args": { + "External id": 11811, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87444, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87444, "pid": 0, "tid": 7, "ts": 1776649792668666, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668657, "dur": 6, + "args": { + "External id": 11811, + "cbid": 211, "correlation": 87444 + } + }, + { + "ph": "s", "id": 87444, "pid": 435, "tid": 435, "ts": 1776649792668657, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668669, "dur": 0, + "args": { + "External id": 87450, + "cbid": 200, "correlation": 87450 + } + }, + { + "ph": "f", "id": 87450, "pid": 435, "tid": 435, "ts": 1776649792668669, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668670, "dur": 0, + "args": { + "External id": 87451, + "cbid": 273, "correlation": 87451 + } + }, + { + "ph": "f", "id": 87451, "pid": 435, "tid": 435, "ts": 1776649792668670, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792668678, "dur": 19, + "args": { + "External id": 87452, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87452, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87452, "pid": 0, "tid": 7, "ts": 1776649792668678, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668671, "dur": 5, + "args": { + "External id": 87452, + "cbid": 211, "correlation": 87452 + } + }, + { + "ph": "s", "id": 87452, "pid": 435, "tid": 435, "ts": 1776649792668671, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792668715, "dur": 1, + "args": { + "External id": 11816, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87470, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87470, "pid": 0, "tid": 7, "ts": 1776649792668715, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668704, "dur": 8, + "args": { + "External id": 11816, + "cbid": 211, "correlation": 87470 + } + }, + { + "ph": "s", "id": 87470, "pid": 435, "tid": 435, "ts": 1776649792668704, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668718, "dur": 0, + "args": { + "External id": 87476, + "cbid": 200, "correlation": 87476 + } + }, + { + "ph": "f", "id": 87476, "pid": 435, "tid": 435, "ts": 1776649792668718, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668719, "dur": 0, + "args": { + "External id": 87477, + "cbid": 273, "correlation": 87477 + } + }, + { + "ph": "f", "id": 87477, "pid": 435, "tid": 435, "ts": 1776649792668719, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792668728, "dur": 19, + "args": { + "External id": 87478, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87478, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87478, "pid": 0, "tid": 7, "ts": 1776649792668728, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668720, "dur": 4, + "args": { + "External id": 87478, + "cbid": 211, "correlation": 87478 + } + }, + { + "ph": "s", "id": 87478, "pid": 435, "tid": 435, "ts": 1776649792668720, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792668763, "dur": 1, + "args": { + "External id": 11821, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87496, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87496, "pid": 0, "tid": 7, "ts": 1776649792668763, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668754, "dur": 6, + "args": { + "External id": 11821, + "cbid": 211, "correlation": 87496 + } + }, + { + "ph": "s", "id": 87496, "pid": 435, "tid": 435, "ts": 1776649792668754, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668766, "dur": 0, + "args": { + "External id": 87502, + "cbid": 200, "correlation": 87502 + } + }, + { + "ph": "f", "id": 87502, "pid": 435, "tid": 435, "ts": 1776649792668766, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668767, "dur": 0, + "args": { + "External id": 87503, + "cbid": 273, "correlation": 87503 + } + }, + { + "ph": "f", "id": 87503, "pid": 435, "tid": 435, "ts": 1776649792668767, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792668776, "dur": 19, + "args": { + "External id": 87504, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87504, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87504, "pid": 0, "tid": 7, "ts": 1776649792668776, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668768, "dur": 4, + "args": { + "External id": 87504, + "cbid": 211, "correlation": 87504 + } + }, + { + "ph": "s", "id": 87504, "pid": 435, "tid": 435, "ts": 1776649792668768, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792668811, "dur": 1, + "args": { + "External id": 11826, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87522, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87522, "pid": 0, "tid": 7, "ts": 1776649792668811, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668802, "dur": 6, + "args": { + "External id": 11826, + "cbid": 211, "correlation": 87522 + } + }, + { + "ph": "s", "id": 87522, "pid": 435, "tid": 435, "ts": 1776649792668802, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668814, "dur": 0, + "args": { + "External id": 87528, + "cbid": 200, "correlation": 87528 + } + }, + { + "ph": "f", "id": 87528, "pid": 435, "tid": 435, "ts": 1776649792668814, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668815, "dur": 0, + "args": { + "External id": 87529, + "cbid": 273, "correlation": 87529 + } + }, + { + "ph": "f", "id": 87529, "pid": 435, "tid": 435, "ts": 1776649792668815, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792668824, "dur": 19, + "args": { + "External id": 87530, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87530, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87530, "pid": 0, "tid": 7, "ts": 1776649792668824, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668816, "dur": 5, + "args": { + "External id": 87530, + "cbid": 211, "correlation": 87530 + } + }, + { + "ph": "s", "id": 87530, "pid": 435, "tid": 435, "ts": 1776649792668816, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792668858, "dur": 1, + "args": { + "External id": 11831, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87548, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87548, "pid": 0, "tid": 7, "ts": 1776649792668858, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668850, "dur": 6, + "args": { + "External id": 11831, + "cbid": 211, "correlation": 87548 + } + }, + { + "ph": "s", "id": 87548, "pid": 435, "tid": 435, "ts": 1776649792668850, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668862, "dur": 0, + "args": { + "External id": 87554, + "cbid": 200, "correlation": 87554 + } + }, + { + "ph": "f", "id": 87554, "pid": 435, "tid": 435, "ts": 1776649792668862, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668863, "dur": 0, + "args": { + "External id": 87555, + "cbid": 273, "correlation": 87555 + } + }, + { + "ph": "f", "id": 87555, "pid": 435, "tid": 435, "ts": 1776649792668863, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792668871, "dur": 19, + "args": { + "External id": 87556, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87556, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87556, "pid": 0, "tid": 7, "ts": 1776649792668871, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668864, "dur": 4, + "args": { + "External id": 87556, + "cbid": 211, "correlation": 87556 + } + }, + { + "ph": "s", "id": 87556, "pid": 435, "tid": 435, "ts": 1776649792668864, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792668907, "dur": 1, + "args": { + "External id": 11836, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87574, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87574, "pid": 0, "tid": 7, "ts": 1776649792668907, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668898, "dur": 6, + "args": { + "External id": 11836, + "cbid": 211, "correlation": 87574 + } + }, + { + "ph": "s", "id": 87574, "pid": 435, "tid": 435, "ts": 1776649792668898, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668910, "dur": 0, + "args": { + "External id": 87580, + "cbid": 200, "correlation": 87580 + } + }, + { + "ph": "f", "id": 87580, "pid": 435, "tid": 435, "ts": 1776649792668910, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668910, "dur": 0, + "args": { + "External id": 87581, + "cbid": 273, "correlation": 87581 + } + }, + { + "ph": "f", "id": 87581, "pid": 435, "tid": 435, "ts": 1776649792668910, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792668920, "dur": 19, + "args": { + "External id": 87582, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87582, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87582, "pid": 0, "tid": 7, "ts": 1776649792668920, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668912, "dur": 4, + "args": { + "External id": 87582, + "cbid": 211, "correlation": 87582 + } + }, + { + "ph": "s", "id": 87582, "pid": 435, "tid": 435, "ts": 1776649792668912, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792668955, "dur": 1, + "args": { + "External id": 11841, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87600, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87600, "pid": 0, "tid": 7, "ts": 1776649792668955, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668945, "dur": 6, + "args": { + "External id": 11841, + "cbid": 211, "correlation": 87600 + } + }, + { + "ph": "s", "id": 87600, "pid": 435, "tid": 435, "ts": 1776649792668945, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668958, "dur": 0, + "args": { + "External id": 87606, + "cbid": 200, "correlation": 87606 + } + }, + { + "ph": "f", "id": 87606, "pid": 435, "tid": 435, "ts": 1776649792668958, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792668958, "dur": 0, + "args": { + "External id": 87607, + "cbid": 273, "correlation": 87607 + } + }, + { + "ph": "f", "id": 87607, "pid": 435, "tid": 435, "ts": 1776649792668958, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792668967, "dur": 19, + "args": { + "External id": 87608, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87608, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87608, "pid": 0, "tid": 7, "ts": 1776649792668967, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668959, "dur": 4, + "args": { + "External id": 87608, + "cbid": 211, "correlation": 87608 + } + }, + { + "ph": "s", "id": 87608, "pid": 435, "tid": 435, "ts": 1776649792668959, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792669002, "dur": 1, + "args": { + "External id": 11846, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87626, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87626, "pid": 0, "tid": 7, "ts": 1776649792669002, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792668993, "dur": 6, + "args": { + "External id": 11846, + "cbid": 211, "correlation": 87626 + } + }, + { + "ph": "s", "id": 87626, "pid": 435, "tid": 435, "ts": 1776649792668993, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792669005, "dur": 0, + "args": { + "External id": 87632, + "cbid": 200, "correlation": 87632 + } + }, + { + "ph": "f", "id": 87632, "pid": 435, "tid": 435, "ts": 1776649792669005, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792669006, "dur": 0, + "args": { + "External id": 87633, + "cbid": 273, "correlation": 87633 + } + }, + { + "ph": "f", "id": 87633, "pid": 435, "tid": 435, "ts": 1776649792669006, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792669016, "dur": 19, + "args": { + "External id": 87634, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87634, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87634, "pid": 0, "tid": 7, "ts": 1776649792669016, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792669007, "dur": 4, + "args": { + "External id": 87634, + "cbid": 211, "correlation": 87634 + } + }, + { + "ph": "s", "id": 87634, "pid": 435, "tid": 435, "ts": 1776649792669007, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792669050, "dur": 1, + "args": { + "External id": 11851, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87652, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87652, "pid": 0, "tid": 7, "ts": 1776649792669050, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792669041, "dur": 6, + "args": { + "External id": 11851, + "cbid": 211, "correlation": 87652 + } + }, + { + "ph": "s", "id": 87652, "pid": 435, "tid": 435, "ts": 1776649792669041, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792669053, "dur": 0, + "args": { + "External id": 87658, + "cbid": 200, "correlation": 87658 + } + }, + { + "ph": "f", "id": 87658, "pid": 435, "tid": 435, "ts": 1776649792669053, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792669054, "dur": 0, + "args": { + "External id": 87659, + "cbid": 273, "correlation": 87659 + } + }, + { + "ph": "f", "id": 87659, "pid": 435, "tid": 435, "ts": 1776649792669054, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792669063, "dur": 19, + "args": { + "External id": 87660, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87660, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87660, "pid": 0, "tid": 7, "ts": 1776649792669063, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792669055, "dur": 4, + "args": { + "External id": 87660, + "cbid": 211, "correlation": 87660 + } + }, + { + "ph": "s", "id": 87660, "pid": 435, "tid": 435, "ts": 1776649792669055, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792669097, "dur": 1, + "args": { + "External id": 11856, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87678, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87678, "pid": 0, "tid": 7, "ts": 1776649792669097, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792669088, "dur": 6, + "args": { + "External id": 11856, + "cbid": 211, "correlation": 87678 + } + }, + { + "ph": "s", "id": 87678, "pid": 435, "tid": 435, "ts": 1776649792669088, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792669101, "dur": 0, + "args": { + "External id": 87684, + "cbid": 200, "correlation": 87684 + } + }, + { + "ph": "f", "id": 87684, "pid": 435, "tid": 435, "ts": 1776649792669101, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792669101, "dur": 0, + "args": { + "External id": 87685, + "cbid": 273, "correlation": 87685 + } + }, + { + "ph": "f", "id": 87685, "pid": 435, "tid": 435, "ts": 1776649792669101, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792669110, "dur": 19, + "args": { + "External id": 87686, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87686, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87686, "pid": 0, "tid": 7, "ts": 1776649792669110, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792669102, "dur": 4, + "args": { + "External id": 87686, + "cbid": 211, "correlation": 87686 + } + }, + { + "ph": "s", "id": 87686, "pid": 435, "tid": 435, "ts": 1776649792669102, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792669145, "dur": 1, + "args": { + "External id": 11861, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87704, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87704, "pid": 0, "tid": 7, "ts": 1776649792669145, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792669136, "dur": 6, + "args": { + "External id": 11861, + "cbid": 211, "correlation": 87704 + } + }, + { + "ph": "s", "id": 87704, "pid": 435, "tid": 435, "ts": 1776649792669136, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792669148, "dur": 0, + "args": { + "External id": 87710, + "cbid": 200, "correlation": 87710 + } + }, + { + "ph": "f", "id": 87710, "pid": 435, "tid": 435, "ts": 1776649792669148, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792669149, "dur": 0, + "args": { + "External id": 87711, + "cbid": 273, "correlation": 87711 + } + }, + { + "ph": "f", "id": 87711, "pid": 435, "tid": 435, "ts": 1776649792669149, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792669158, "dur": 19, + "args": { + "External id": 87712, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87712, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87712, "pid": 0, "tid": 7, "ts": 1776649792669158, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792669150, "dur": 4, + "args": { + "External id": 87712, + "cbid": 211, "correlation": 87712 + } + }, + { + "ph": "s", "id": 87712, "pid": 435, "tid": 435, "ts": 1776649792669150, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792669193, "dur": 1, + "args": { + "External id": 11866, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87730, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87730, "pid": 0, "tid": 7, "ts": 1776649792669193, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792669184, "dur": 6, + "args": { + "External id": 11866, + "cbid": 211, "correlation": 87730 + } + }, + { + "ph": "s", "id": 87730, "pid": 435, "tid": 435, "ts": 1776649792669184, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792669196, "dur": 0, + "args": { + "External id": 87736, + "cbid": 200, "correlation": 87736 + } + }, + { + "ph": "f", "id": 87736, "pid": 435, "tid": 435, "ts": 1776649792669196, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792669197, "dur": 0, + "args": { + "External id": 87737, + "cbid": 273, "correlation": 87737 + } + }, + { + "ph": "f", "id": 87737, "pid": 435, "tid": 435, "ts": 1776649792669197, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792669206, "dur": 19, + "args": { + "External id": 87738, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87738, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87738, "pid": 0, "tid": 7, "ts": 1776649792669206, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792669198, "dur": 5, + "args": { + "External id": 87738, + "cbid": 211, "correlation": 87738 + } + }, + { + "ph": "s", "id": 87738, "pid": 435, "tid": 435, "ts": 1776649792669198, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792669241, "dur": 1, + "args": { + "External id": 11871, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87756, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87756, "pid": 0, "tid": 7, "ts": 1776649792669241, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792669232, "dur": 6, + "args": { + "External id": 11871, + "cbid": 211, "correlation": 87756 + } + }, + { + "ph": "s", "id": 87756, "pid": 435, "tid": 435, "ts": 1776649792669232, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792669244, "dur": 0, + "args": { + "External id": 87762, + "cbid": 200, "correlation": 87762 + } + }, + { + "ph": "f", "id": 87762, "pid": 435, "tid": 435, "ts": 1776649792669244, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792669244, "dur": 0, + "args": { + "External id": 87763, + "cbid": 273, "correlation": 87763 + } + }, + { + "ph": "f", "id": 87763, "pid": 435, "tid": 435, "ts": 1776649792669244, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792669254, "dur": 19, + "args": { + "External id": 87764, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87764, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87764, "pid": 0, "tid": 7, "ts": 1776649792669254, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792669246, "dur": 5, + "args": { + "External id": 87764, + "cbid": 211, "correlation": 87764 + } + }, + { + "ph": "s", "id": 87764, "pid": 435, "tid": 435, "ts": 1776649792669246, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792669288, "dur": 1, + "args": { + "External id": 11876, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87782, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87782, "pid": 0, "tid": 7, "ts": 1776649792669288, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792669279, "dur": 6, + "args": { + "External id": 11876, + "cbid": 211, "correlation": 87782 + } + }, + { + "ph": "s", "id": 87782, "pid": 435, "tid": 435, "ts": 1776649792669279, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792669291, "dur": 0, + "args": { + "External id": 87788, + "cbid": 200, "correlation": 87788 + } + }, + { + "ph": "f", "id": 87788, "pid": 435, "tid": 435, "ts": 1776649792669291, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792669292, "dur": 0, + "args": { + "External id": 87789, + "cbid": 273, "correlation": 87789 + } + }, + { + "ph": "f", "id": 87789, "pid": 435, "tid": 435, "ts": 1776649792669292, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792669301, "dur": 19, + "args": { + "External id": 87790, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87790, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87790, "pid": 0, "tid": 7, "ts": 1776649792669301, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792669293, "dur": 4, + "args": { + "External id": 87790, + "cbid": 211, "correlation": 87790 + } + }, + { + "ph": "s", "id": 87790, "pid": 435, "tid": 435, "ts": 1776649792669293, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792669336, "dur": 1, + "args": { + "External id": 11881, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87808, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87808, "pid": 0, "tid": 7, "ts": 1776649792669336, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792669327, "dur": 6, + "args": { + "External id": 11881, + "cbid": 211, "correlation": 87808 + } + }, + { + "ph": "s", "id": 87808, "pid": 435, "tid": 435, "ts": 1776649792669327, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792669339, "dur": 0, + "args": { + "External id": 87814, + "cbid": 200, "correlation": 87814 + } + }, + { + "ph": "f", "id": 87814, "pid": 435, "tid": 435, "ts": 1776649792669339, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649792669340, "dur": 0, + "args": { + "External id": 87815, + "cbid": 273, "correlation": 87815 + } + }, + { + "ph": "f", "id": 87815, "pid": 435, "tid": 435, "ts": 1776649792669340, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649792669349, "dur": 19, + "args": { + "External id": 87816, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87816, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 87816, "pid": 0, "tid": 7, "ts": 1776649792669349, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792669341, "dur": 4, + "args": { + "External id": 87816, + "cbid": 211, "correlation": 87816 + } + }, + { + "ph": "s", "id": 87816, "pid": 435, "tid": 435, "ts": 1776649792669341, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649792669371, "dur": 3, + "args": { + "External id": 87824, + "cbid": 135, "correlation": 87824 + } + }, + { + "ph": "f", "id": 87824, "pid": 435, "tid": 435, "ts": 1776649792669371, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792669386, "dur": 5, + "args": { + "External id": 87829, + "cbid": 165, "correlation": 87829 + } + }, + { + "ph": "s", "id": 87829, "pid": 435, "tid": 435, "ts": 1776649792669386, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649792669394, "dur": 5, + "args": { + "External id": 87832, + "cbid": 139, "correlation": 87832 + } + }, + { + "ph": "f", "id": 87832, "pid": 435, "tid": 435, "ts": 1776649792669394, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649792669523, "dur": 1, + "args": { + "External id": 11884, + "cbid": 317, "correlation": 87857 + } + }, + { + "ph": "f", "id": 87857, "pid": 435, "tid": 435, "ts": 1776649792669523, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649792669527, "dur": 0, + "args": { + "External id": 11884, + "cbid": 317, "correlation": 87861 + } + }, + { + "ph": "f", "id": 87861, "pid": 435, "tid": 435, "ts": 1776649792669527, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::distribution_elementwise_grid_stride_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::(anonymous namespace)::distribution_nullary_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2} const&, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(int, float)#1}>(int, at::PhiloxCudaState, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::(anonymous namespace)::distribution_nullary_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2} const&, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(int, float)#1})", "pid": 0, "tid": 7, + "ts": 1776649792669543, "dur": 4, + "args": { + "External id": 11884, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87863, + "registers per thread": 38, + "shared memory": 0, + "blocks per SM": 1.185185, + "warps per SM": 9.481482, + "grid": [128, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 15 + } + }, + { + "ph": "f", "id": 87863, "pid": 0, "tid": 7, "ts": 1776649792669543, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792669530, "dur": 12, + "args": { + "External id": 11884, + "cbid": 211, "correlation": 87863 + } + }, + { + "ph": "s", "id": 87863, "pid": 435, "tid": 435, "ts": 1776649792669530, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649792669555, "dur": 0, + "args": { + "External id": 87870, + "cbid": 317, "correlation": 87870 + } + }, + { + "ph": "f", "id": 87870, "pid": 435, "tid": 435, "ts": 1776649792669555, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649792669560, "dur": 0, + "args": { + "External id": 87874, + "cbid": 317, "correlation": 87874 + } + }, + { + "ph": "f", "id": 87874, "pid": 435, "tid": 435, "ts": 1776649792669560, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649792669583, "dur": 0, + "args": { + "External id": 11887, + "cbid": 317, "correlation": 87889 + } + }, + { + "ph": "f", "id": 87889, "pid": 435, "tid": 435, "ts": 1776649792669583, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649792669585, "dur": 0, + "args": { + "External id": 11887, + "cbid": 317, "correlation": 87893 + } + }, + { + "ph": "f", "id": 87893, "pid": 435, "tid": 435, "ts": 1776649792669585, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::distribution_elementwise_grid_stride_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::(anonymous namespace)::distribution_nullary_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2} const&, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(int, float)#1}>(int, at::PhiloxCudaState, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::(anonymous namespace)::distribution_nullary_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2} const&, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(int, float)#1})", "pid": 0, "tid": 7, + "ts": 1776649792669596, "dur": 165, + "args": { + "External id": 11887, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87895, + "registers per thread": 38, + "shared memory": 0, + "blocks per SM": 8.000000, + "warps per SM": 64.000000, + "grid": [864, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 75 + } + }, + { + "ph": "f", "id": 87895, "pid": 0, "tid": 7, "ts": 1776649792669596, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792669587, "dur": 7, + "args": { + "External id": 11887, + "cbid": 211, "correlation": 87895 + } + }, + { + "ph": "s", "id": 87895, "pid": 435, "tid": 435, "ts": 1776649792669587, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor >, at::detail::Array >(int, at::native::AUnaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792669763, "dur": 78, + "args": { + "External id": 11888, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87905, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 87905, "pid": 0, "tid": 7, "ts": 1776649792669763, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792669651, "dur": 9, + "args": { + "External id": 11888, + "cbid": 211, "correlation": 87905 + } + }, + { + "ph": "s", "id": 87905, "pid": 435, "tid": 435, "ts": 1776649792669651, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AbsFunctor, at::detail::Array >(int, at::native::AbsFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792669843, "dur": 77, + "args": { + "External id": 11893, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87925, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 87925, "pid": 0, "tid": 7, "ts": 1776649792669843, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792669730, "dur": 9, + "args": { + "External id": 11893, + "cbid": 211, "correlation": 87925 + } + }, + { + "ph": "s", "id": 87925, "pid": 435, "tid": 435, "ts": 1776649792669730, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<128, 4, at::native::ReduceOp >, unsigned int, c10::Half, 4> >(at::native::ReduceOp >, unsigned int, c10::Half, 4>)", "pid": 0, "tid": 7, + "ts": 1776649792669921, "dur": 72, + "args": { + "External id": 11895, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87938, + "registers per thread": 49, + "shared memory": 1040, + "blocks per SM": 18.962963, + "warps per SM": 75.851852, + "grid": [2048, 1, 1], + "block": [32, 4, 1], + "est. achieved occupancy %": 56 + } + }, + { + "ph": "f", "id": 87938, "pid": 0, "tid": 7, "ts": 1776649792669921, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792669778, "dur": 11, + "args": { + "External id": 11895, + "cbid": 211, "correlation": 87938 + } + }, + { + "ph": "s", "id": 87938, "pid": 435, "tid": 435, "ts": 1776649792669778, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BUnaryFunctor >, at::detail::Array >(int, at::native::BUnaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792669995, "dur": 3, + "args": { + "External id": 11897, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87948, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 4.740741, + "warps per SM": 18.962963, + "grid": [512, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 30 + } + }, + { + "ph": "f", "id": 87948, "pid": 0, "tid": 7, "ts": 1776649792669995, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792669819, "dur": 8, + "args": { + "External id": 11897, + "cbid": 211, "correlation": 87948 + } + }, + { + "ph": "s", "id": 87948, "pid": 435, "tid": 435, "ts": 1776649792669819, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#8}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#8}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792670000, "dur": 3, + "args": { + "External id": 11898, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87958, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 4.740741, + "warps per SM": 18.962963, + "grid": [512, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 30 + } + }, + { + "ph": "f", "id": 87958, "pid": 0, "tid": 7, "ts": 1776649792670000, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792669860, "dur": 8, + "args": { + "External id": 11898, + "cbid": 211, "correlation": 87958 + } + }, + { + "ph": "s", "id": 87958, "pid": 435, "tid": 435, "ts": 1776649792669860, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649792670004, "dur": 192, + "args": { + "External id": 11903, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87968, + "registers per thread": 20, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 87968, "pid": 0, "tid": 7, "ts": 1776649792670004, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792669907, "dur": 8, + "args": { + "External id": 11903, + "cbid": 211, "correlation": 87968 + } + }, + { + "ph": "s", "id": 87968, "pid": 435, "tid": 435, "ts": 1776649792669907, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792670198, "dur": 76, + "args": { + "External id": 11904, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87978, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 87978, "pid": 0, "tid": 7, "ts": 1776649792670198, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792669935, "dur": 7, + "args": { + "External id": 11904, + "cbid": 211, "correlation": 87978 + } + }, + { + "ph": "s", "id": 87978, "pid": 435, "tid": 435, "ts": 1776649792669935, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#8}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#8}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792670276, "dur": 80, + "args": { + "External id": 11905, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 87988, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 87988, "pid": 0, "tid": 7, "ts": 1776649792670276, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792669959, "dur": 6, + "args": { + "External id": 11905, + "cbid": 211, "correlation": 87988 + } + }, + { + "ph": "s", "id": 87988, "pid": 435, "tid": 435, "ts": 1776649792669959, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649792670358, "dur": 183, + "args": { + "External id": 11910, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88001, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 88001, "pid": 0, "tid": 7, "ts": 1776649792670358, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792669999, "dur": 9, + "args": { + "External id": 11910, + "cbid": 211, "correlation": 88001 + } + }, + { + "ph": "s", "id": 88001, "pid": 435, "tid": 435, "ts": 1776649792669999, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, at::detail::Array >(int, at::native::CUDAFunctorOnSelf_add, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792670543, "dur": 63, + "args": { + "External id": 11911, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88012, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 88012, "pid": 0, "tid": 7, "ts": 1776649792670543, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792670043, "dur": 9, + "args": { + "External id": 11911, + "cbid": 211, "correlation": 88012 + } + }, + { + "ph": "s", "id": 88012, "pid": 435, "tid": 435, "ts": 1776649792670043, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#1}::operator()() const::{lambda(unsigned char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649792670607, "dur": 111, + "args": { + "External id": 11915, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88025, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 88025, "pid": 0, "tid": 7, "ts": 1776649792670607, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792670073, "dur": 8, + "args": { + "External id": 11915, + "cbid": 211, "correlation": 88025 + } + }, + { + "ph": "s", "id": 88025, "pid": 435, "tid": 435, "ts": 1776649792670073, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast >(at::TensorIteratorBase&, at::native::BUnaryFunctor const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast >(at::TensorIteratorBase&, at::native::BUnaryFunctor const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649792670720, "dur": 77, + "args": { + "External id": 11926, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88045, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 88045, "pid": 0, "tid": 7, "ts": 1776649792670720, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792670164, "dur": 11, + "args": { + "External id": 11926, + "cbid": 211, "correlation": 88045 + } + }, + { + "ph": "s", "id": 88045, "pid": 435, "tid": 435, "ts": 1776649792670164, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649792670799, "dur": 83, + "args": { + "External id": 11929, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88055, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 88055, "pid": 0, "tid": 7, "ts": 1776649792670799, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792670204, "dur": 7, + "args": { + "External id": 11929, + "cbid": 211, "correlation": 88055 + } + }, + { + "ph": "s", "id": 88055, "pid": 435, "tid": 435, "ts": 1776649792670204, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor >, at::detail::Array >(int, at::native::AUnaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792670884, "dur": 33, + "args": { + "External id": 11933, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88065, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 88065, "pid": 0, "tid": 7, "ts": 1776649792670884, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792670257, "dur": 8, + "args": { + "External id": 11933, + "cbid": 211, "correlation": 88065 + } + }, + { + "ph": "s", "id": 88065, "pid": 435, "tid": 435, "ts": 1776649792670257, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649792670919, "dur": 57, + "args": { + "External id": 11937, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88078, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 88078, "pid": 0, "tid": 7, "ts": 1776649792670919, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792670289, "dur": 7, + "args": { + "External id": 11937, + "cbid": 211, "correlation": 88078 + } + }, + { + "ph": "s", "id": 88078, "pid": 435, "tid": 435, "ts": 1776649792670289, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BUnaryFunctor, at::detail::Array >(int, at::native::BUnaryFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792670977, "dur": 33, + "args": { + "External id": 11938, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88092, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 88092, "pid": 0, "tid": 7, "ts": 1776649792670977, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792670323, "dur": 8, + "args": { + "External id": 11938, + "cbid": 211, "correlation": 88092 + } + }, + { + "ph": "s", "id": 88092, "pid": 435, "tid": 435, "ts": 1776649792670323, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649792671012, "dur": 57, + "args": { + "External id": 11943, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88105, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 88105, "pid": 0, "tid": 7, "ts": 1776649792671012, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792670350, "dur": 6, + "args": { + "External id": 11943, + "cbid": 211, "correlation": 88105 + } + }, + { + "ph": "s", "id": 88105, "pid": 435, "tid": 435, "ts": 1776649792670350, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy_contig, unsigned int, 3, 128, 1>(at::native::(anonymous namespace)::OpaqueType<1u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata, unsigned int, 128, 1>, at::native::(anonymous namespace)::TensorSizeStride, int, unsigned int)", "pid": 0, "tid": 7, + "ts": 1776649792671071, "dur": 126, + "args": { + "External id": 11949, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88117, + "registers per thread": 21, + "shared memory": 0, + "blocks per SM": 4.000000, + "warps per SM": 64.000000, + "grid": [216, 2, 1], + "block": [512, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 88117, "pid": 0, "tid": 7, "ts": 1776649792671071, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792670412, "dur": 11, + "args": { + "External id": 11949, + "cbid": 211, "correlation": 88117 + } + }, + { + "ph": "s", "id": 88117, "pid": 435, "tid": 435, "ts": 1776649792670412, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649792671199, "dur": 136, + "args": { + "External id": 11960, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88130, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 88130, "pid": 0, "tid": 7, "ts": 1776649792671199, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792670491, "dur": 9, + "args": { + "External id": 11960, + "cbid": 211, "correlation": 88130 + } + }, + { + "ph": "s", "id": 88130, "pid": 435, "tid": 435, "ts": 1776649792670491, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, at::detail::Array >(int, at::native::CUDAFunctorOnSelf_add, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792671337, "dur": 63, + "args": { + "External id": 11963, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88141, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 88141, "pid": 0, "tid": 7, "ts": 1776649792671337, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792670533, "dur": 7, + "args": { + "External id": 11963, + "cbid": 211, "correlation": 88141 + } + }, + { + "ph": "s", "id": 88141, "pid": 435, "tid": 435, "ts": 1776649792670533, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649792671401, "dur": 179, + "args": { + "External id": 11967, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88154, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 88154, "pid": 0, "tid": 7, "ts": 1776649792671401, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792670562, "dur": 8, + "args": { + "External id": 11967, + "cbid": 211, "correlation": 88154 + } + }, + { + "ph": "s", "id": 88154, "pid": 435, "tid": 435, "ts": 1776649792670562, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor >, at::detail::Array >(int, at::native::BinaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792671582, "dur": 114, + "args": { + "External id": 11968, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88165, + "registers per thread": 22, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 88165, "pid": 0, "tid": 7, "ts": 1776649792671582, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792670589, "dur": 7, + "args": { + "External id": 11968, + "cbid": 211, "correlation": 88165 + } + }, + { + "ph": "s", "id": 88165, "pid": 435, "tid": 435, "ts": 1776649792670589, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792670773, "dur": 6, + "args": { + "External id": 11970, + "cbid": 210, "correlation": 88192 + } + }, + { + "ph": "f", "id": 88192, "pid": 435, "tid": 435, "ts": 1776649792670773, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792670791, "dur": 0, + "args": { + "External id": 11970, + "cbid": 210, "correlation": 88193 + } + }, + { + "ph": "f", "id": 88193, "pid": 435, "tid": 435, "ts": 1776649792670791, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792671698, "dur": 49, + "args": { + "External id": 11970, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88194, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88194, "pid": 0, "tid": 7, "ts": 1776649792671698, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792670797, "dur": 13, + "args": { + "External id": 11970, + "cbid": 211, "correlation": 88194 + } + }, + { + "ph": "s", "id": 88194, "pid": 435, "tid": 435, "ts": 1776649792670797, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792670842, "dur": 1, + "args": { + "External id": 11972, + "cbid": 210, "correlation": 88221 + } + }, + { + "ph": "f", "id": 88221, "pid": 435, "tid": 435, "ts": 1776649792670842, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792671749, "dur": 46, + "args": { + "External id": 11972, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88222, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88222, "pid": 0, "tid": 7, "ts": 1776649792671749, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792670844, "dur": 6, + "args": { + "External id": 11972, + "cbid": 211, "correlation": 88222 + } + }, + { + "ph": "s", "id": 88222, "pid": 435, "tid": 435, "ts": 1776649792670844, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792670875, "dur": 0, + "args": { + "External id": 11974, + "cbid": 210, "correlation": 88249 + } + }, + { + "ph": "f", "id": 88249, "pid": 435, "tid": 435, "ts": 1776649792670875, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792671797, "dur": 52, + "args": { + "External id": 11974, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88250, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88250, "pid": 0, "tid": 7, "ts": 1776649792671797, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792670876, "dur": 6, + "args": { + "External id": 11974, + "cbid": 211, "correlation": 88250 + } + }, + { + "ph": "s", "id": 88250, "pid": 435, "tid": 435, "ts": 1776649792670876, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792670906, "dur": 0, + "args": { + "External id": 11976, + "cbid": 210, "correlation": 88277 + } + }, + { + "ph": "f", "id": 88277, "pid": 435, "tid": 435, "ts": 1776649792670906, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792671851, "dur": 52, + "args": { + "External id": 11976, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88278, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88278, "pid": 0, "tid": 7, "ts": 1776649792671851, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792670907, "dur": 6, + "args": { + "External id": 11976, + "cbid": 211, "correlation": 88278 + } + }, + { + "ph": "s", "id": 88278, "pid": 435, "tid": 435, "ts": 1776649792670907, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792670936, "dur": 0, + "args": { + "External id": 11978, + "cbid": 210, "correlation": 88305 + } + }, + { + "ph": "f", "id": 88305, "pid": 435, "tid": 435, "ts": 1776649792670936, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792671905, "dur": 58, + "args": { + "External id": 11978, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88306, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88306, "pid": 0, "tid": 7, "ts": 1776649792671905, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792670938, "dur": 6, + "args": { + "External id": 11978, + "cbid": 211, "correlation": 88306 + } + }, + { + "ph": "s", "id": 88306, "pid": 435, "tid": 435, "ts": 1776649792670938, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792670966, "dur": 0, + "args": { + "External id": 11980, + "cbid": 210, "correlation": 88333 + } + }, + { + "ph": "f", "id": 88333, "pid": 435, "tid": 435, "ts": 1776649792670966, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792671965, "dur": 49, + "args": { + "External id": 11980, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88334, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88334, "pid": 0, "tid": 7, "ts": 1776649792671965, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792670968, "dur": 5, + "args": { + "External id": 11980, + "cbid": 211, "correlation": 88334 + } + }, + { + "ph": "s", "id": 88334, "pid": 435, "tid": 435, "ts": 1776649792670968, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792670996, "dur": 0, + "args": { + "External id": 11982, + "cbid": 210, "correlation": 88361 + } + }, + { + "ph": "f", "id": 88361, "pid": 435, "tid": 435, "ts": 1776649792670996, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792672016, "dur": 49, + "args": { + "External id": 11982, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88362, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88362, "pid": 0, "tid": 7, "ts": 1776649792672016, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792670998, "dur": 6, + "args": { + "External id": 11982, + "cbid": 211, "correlation": 88362 + } + }, + { + "ph": "s", "id": 88362, "pid": 435, "tid": 435, "ts": 1776649792670998, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792671026, "dur": 0, + "args": { + "External id": 11984, + "cbid": 210, "correlation": 88389 + } + }, + { + "ph": "f", "id": 88389, "pid": 435, "tid": 435, "ts": 1776649792671026, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792672066, "dur": 47, + "args": { + "External id": 11984, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88390, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88390, "pid": 0, "tid": 7, "ts": 1776649792672066, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792671028, "dur": 6, + "args": { + "External id": 11984, + "cbid": 211, "correlation": 88390 + } + }, + { + "ph": "s", "id": 88390, "pid": 435, "tid": 435, "ts": 1776649792671028, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792671056, "dur": 0, + "args": { + "External id": 11986, + "cbid": 210, "correlation": 88417 + } + }, + { + "ph": "f", "id": 88417, "pid": 435, "tid": 435, "ts": 1776649792671056, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792672115, "dur": 44, + "args": { + "External id": 11986, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88418, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88418, "pid": 0, "tid": 7, "ts": 1776649792672115, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792671057, "dur": 6, + "args": { + "External id": 11986, + "cbid": 211, "correlation": 88418 + } + }, + { + "ph": "s", "id": 88418, "pid": 435, "tid": 435, "ts": 1776649792671057, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792671086, "dur": 0, + "args": { + "External id": 11988, + "cbid": 210, "correlation": 88445 + } + }, + { + "ph": "f", "id": 88445, "pid": 435, "tid": 435, "ts": 1776649792671086, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792672160, "dur": 47, + "args": { + "External id": 11988, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88446, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88446, "pid": 0, "tid": 7, "ts": 1776649792672160, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792671087, "dur": 5, + "args": { + "External id": 11988, + "cbid": 211, "correlation": 88446 + } + }, + { + "ph": "s", "id": 88446, "pid": 435, "tid": 435, "ts": 1776649792671087, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792671116, "dur": 1096, + "args": { + "External id": 88452, + "cbid": 165, "correlation": 88452 + } + }, + { + "ph": "s", "id": 88452, "pid": 435, "tid": 435, "ts": 1776649792671116, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649792672258, "dur": 4, + "args": { + "External id": 88461, + "cbid": 135, "correlation": 88461 + } + }, + { + "ph": "f", "id": 88461, "pid": 435, "tid": 435, "ts": 1776649792672258, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792672293, "dur": 1, + "args": { + "External id": 11990, + "cbid": 210, "correlation": 88487 + } + }, + { + "ph": "f", "id": 88487, "pid": 435, "tid": 435, "ts": 1776649792672293, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792672307, "dur": 40, + "args": { + "External id": 11990, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88488, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88488, "pid": 0, "tid": 7, "ts": 1776649792672307, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792672296, "dur": 9, + "args": { + "External id": 11990, + "cbid": 211, "correlation": 88488 + } + }, + { + "ph": "s", "id": 88488, "pid": 435, "tid": 435, "ts": 1776649792672296, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792672331, "dur": 0, + "args": { + "External id": 11992, + "cbid": 210, "correlation": 88515 + } + }, + { + "ph": "f", "id": 88515, "pid": 435, "tid": 435, "ts": 1776649792672331, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792672349, "dur": 42, + "args": { + "External id": 11992, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88516, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88516, "pid": 0, "tid": 7, "ts": 1776649792672349, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792672332, "dur": 6, + "args": { + "External id": 11992, + "cbid": 211, "correlation": 88516 + } + }, + { + "ph": "s", "id": 88516, "pid": 435, "tid": 435, "ts": 1776649792672332, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792672362, "dur": 0, + "args": { + "External id": 11994, + "cbid": 210, "correlation": 88543 + } + }, + { + "ph": "f", "id": 88543, "pid": 435, "tid": 435, "ts": 1776649792672362, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792672392, "dur": 41, + "args": { + "External id": 11994, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88544, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88544, "pid": 0, "tid": 7, "ts": 1776649792672392, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792672363, "dur": 6, + "args": { + "External id": 11994, + "cbid": 211, "correlation": 88544 + } + }, + { + "ph": "s", "id": 88544, "pid": 435, "tid": 435, "ts": 1776649792672363, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792672392, "dur": 0, + "args": { + "External id": 11996, + "cbid": 210, "correlation": 88571 + } + }, + { + "ph": "f", "id": 88571, "pid": 435, "tid": 435, "ts": 1776649792672392, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792672435, "dur": 41, + "args": { + "External id": 11996, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88572, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88572, "pid": 0, "tid": 7, "ts": 1776649792672435, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792672394, "dur": 6, + "args": { + "External id": 11996, + "cbid": 211, "correlation": 88572 + } + }, + { + "ph": "s", "id": 88572, "pid": 435, "tid": 435, "ts": 1776649792672394, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792672422, "dur": 0, + "args": { + "External id": 11998, + "cbid": 210, "correlation": 88599 + } + }, + { + "ph": "f", "id": 88599, "pid": 435, "tid": 435, "ts": 1776649792672422, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792672478, "dur": 41, + "args": { + "External id": 11998, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88600, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88600, "pid": 0, "tid": 7, "ts": 1776649792672478, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792672423, "dur": 6, + "args": { + "External id": 11998, + "cbid": 211, "correlation": 88600 + } + }, + { + "ph": "s", "id": 88600, "pid": 435, "tid": 435, "ts": 1776649792672423, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792672452, "dur": 0, + "args": { + "External id": 12000, + "cbid": 210, "correlation": 88627 + } + }, + { + "ph": "f", "id": 88627, "pid": 435, "tid": 435, "ts": 1776649792672452, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792672521, "dur": 44, + "args": { + "External id": 12000, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88628, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88628, "pid": 0, "tid": 7, "ts": 1776649792672521, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792672453, "dur": 6, + "args": { + "External id": 12000, + "cbid": 211, "correlation": 88628 + } + }, + { + "ph": "s", "id": 88628, "pid": 435, "tid": 435, "ts": 1776649792672453, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792672482, "dur": 0, + "args": { + "External id": 12002, + "cbid": 210, "correlation": 88655 + } + }, + { + "ph": "f", "id": 88655, "pid": 435, "tid": 435, "ts": 1776649792672482, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792672567, "dur": 50, + "args": { + "External id": 12002, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88656, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88656, "pid": 0, "tid": 7, "ts": 1776649792672567, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792672483, "dur": 6, + "args": { + "External id": 12002, + "cbid": 211, "correlation": 88656 + } + }, + { + "ph": "s", "id": 88656, "pid": 435, "tid": 435, "ts": 1776649792672483, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792672512, "dur": 0, + "args": { + "External id": 12004, + "cbid": 210, "correlation": 88683 + } + }, + { + "ph": "f", "id": 88683, "pid": 435, "tid": 435, "ts": 1776649792672512, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792672619, "dur": 47, + "args": { + "External id": 12004, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88684, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88684, "pid": 0, "tid": 7, "ts": 1776649792672619, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792672514, "dur": 6, + "args": { + "External id": 12004, + "cbid": 211, "correlation": 88684 + } + }, + { + "ph": "s", "id": 88684, "pid": 435, "tid": 435, "ts": 1776649792672514, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792672543, "dur": 0, + "args": { + "External id": 12006, + "cbid": 210, "correlation": 88711 + } + }, + { + "ph": "f", "id": 88711, "pid": 435, "tid": 435, "ts": 1776649792672543, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792672668, "dur": 46, + "args": { + "External id": 12006, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88712, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88712, "pid": 0, "tid": 7, "ts": 1776649792672668, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792672544, "dur": 6, + "args": { + "External id": 12006, + "cbid": 211, "correlation": 88712 + } + }, + { + "ph": "s", "id": 88712, "pid": 435, "tid": 435, "ts": 1776649792672544, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792672573, "dur": 0, + "args": { + "External id": 12008, + "cbid": 210, "correlation": 88739 + } + }, + { + "ph": "f", "id": 88739, "pid": 435, "tid": 435, "ts": 1776649792672573, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792672715, "dur": 50, + "args": { + "External id": 12008, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88740, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88740, "pid": 0, "tid": 7, "ts": 1776649792672715, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792672575, "dur": 6, + "args": { + "External id": 12008, + "cbid": 211, "correlation": 88740 + } + }, + { + "ph": "s", "id": 88740, "pid": 435, "tid": 435, "ts": 1776649792672575, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792672603, "dur": 0, + "args": { + "External id": 12010, + "cbid": 210, "correlation": 88767 + } + }, + { + "ph": "f", "id": 88767, "pid": 435, "tid": 435, "ts": 1776649792672603, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792672766, "dur": 49, + "args": { + "External id": 12010, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88768, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88768, "pid": 0, "tid": 7, "ts": 1776649792672766, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792672605, "dur": 6, + "args": { + "External id": 12010, + "cbid": 211, "correlation": 88768 + } + }, + { + "ph": "s", "id": 88768, "pid": 435, "tid": 435, "ts": 1776649792672605, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792672634, "dur": 0, + "args": { + "External id": 12012, + "cbid": 210, "correlation": 88795 + } + }, + { + "ph": "f", "id": 88795, "pid": 435, "tid": 435, "ts": 1776649792672634, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792672817, "dur": 51, + "args": { + "External id": 12012, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88796, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88796, "pid": 0, "tid": 7, "ts": 1776649792672817, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792672635, "dur": 6, + "args": { + "External id": 12012, + "cbid": 211, "correlation": 88796 + } + }, + { + "ph": "s", "id": 88796, "pid": 435, "tid": 435, "ts": 1776649792672635, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792672663, "dur": 0, + "args": { + "External id": 12014, + "cbid": 210, "correlation": 88823 + } + }, + { + "ph": "f", "id": 88823, "pid": 435, "tid": 435, "ts": 1776649792672663, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792672870, "dur": 44, + "args": { + "External id": 12014, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88824, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88824, "pid": 0, "tid": 7, "ts": 1776649792672870, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792672664, "dur": 6, + "args": { + "External id": 12014, + "cbid": 211, "correlation": 88824 + } + }, + { + "ph": "s", "id": 88824, "pid": 435, "tid": 435, "ts": 1776649792672664, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792672693, "dur": 0, + "args": { + "External id": 12016, + "cbid": 210, "correlation": 88851 + } + }, + { + "ph": "f", "id": 88851, "pid": 435, "tid": 435, "ts": 1776649792672693, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792672916, "dur": 46, + "args": { + "External id": 12016, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88852, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88852, "pid": 0, "tid": 7, "ts": 1776649792672916, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792672694, "dur": 6, + "args": { + "External id": 12016, + "cbid": 211, "correlation": 88852 + } + }, + { + "ph": "s", "id": 88852, "pid": 435, "tid": 435, "ts": 1776649792672694, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792672723, "dur": 0, + "args": { + "External id": 12018, + "cbid": 210, "correlation": 88879 + } + }, + { + "ph": "f", "id": 88879, "pid": 435, "tid": 435, "ts": 1776649792672723, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792672963, "dur": 46, + "args": { + "External id": 12018, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88880, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88880, "pid": 0, "tid": 7, "ts": 1776649792672963, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792672724, "dur": 6, + "args": { + "External id": 12018, + "cbid": 211, "correlation": 88880 + } + }, + { + "ph": "s", "id": 88880, "pid": 435, "tid": 435, "ts": 1776649792672724, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792672752, "dur": 0, + "args": { + "External id": 12020, + "cbid": 210, "correlation": 88907 + } + }, + { + "ph": "f", "id": 88907, "pid": 435, "tid": 435, "ts": 1776649792672752, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792673011, "dur": 50, + "args": { + "External id": 12020, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88908, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88908, "pid": 0, "tid": 7, "ts": 1776649792673011, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792672754, "dur": 5, + "args": { + "External id": 12020, + "cbid": 211, "correlation": 88908 + } + }, + { + "ph": "s", "id": 88908, "pid": 435, "tid": 435, "ts": 1776649792672754, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792672782, "dur": 0, + "args": { + "External id": 12022, + "cbid": 210, "correlation": 88935 + } + }, + { + "ph": "f", "id": 88935, "pid": 435, "tid": 435, "ts": 1776649792672782, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792673063, "dur": 49, + "args": { + "External id": 12022, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88936, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88936, "pid": 0, "tid": 7, "ts": 1776649792673063, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792672784, "dur": 6, + "args": { + "External id": 12022, + "cbid": 211, "correlation": 88936 + } + }, + { + "ph": "s", "id": 88936, "pid": 435, "tid": 435, "ts": 1776649792672784, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792672812, "dur": 0, + "args": { + "External id": 12024, + "cbid": 210, "correlation": 88963 + } + }, + { + "ph": "f", "id": 88963, "pid": 435, "tid": 435, "ts": 1776649792672812, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792673114, "dur": 45, + "args": { + "External id": 12024, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88964, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88964, "pid": 0, "tid": 7, "ts": 1776649792673114, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792672813, "dur": 5, + "args": { + "External id": 12024, + "cbid": 211, "correlation": 88964 + } + }, + { + "ph": "s", "id": 88964, "pid": 435, "tid": 435, "ts": 1776649792672813, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792672842, "dur": 0, + "args": { + "External id": 12026, + "cbid": 210, "correlation": 88991 + } + }, + { + "ph": "f", "id": 88991, "pid": 435, "tid": 435, "ts": 1776649792672842, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792673160, "dur": 47, + "args": { + "External id": 12026, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 88992, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 88992, "pid": 0, "tid": 7, "ts": 1776649792673160, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792672843, "dur": 6, + "args": { + "External id": 12026, + "cbid": 211, "correlation": 88992 + } + }, + { + "ph": "s", "id": 88992, "pid": 435, "tid": 435, "ts": 1776649792672843, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792672872, "dur": 0, + "args": { + "External id": 12028, + "cbid": 210, "correlation": 89019 + } + }, + { + "ph": "f", "id": 89019, "pid": 435, "tid": 435, "ts": 1776649792672872, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792673209, "dur": 47, + "args": { + "External id": 12028, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89020, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89020, "pid": 0, "tid": 7, "ts": 1776649792673209, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792672873, "dur": 6, + "args": { + "External id": 12028, + "cbid": 211, "correlation": 89020 + } + }, + { + "ph": "s", "id": 89020, "pid": 435, "tid": 435, "ts": 1776649792672873, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792672902, "dur": 0, + "args": { + "External id": 12030, + "cbid": 210, "correlation": 89047 + } + }, + { + "ph": "f", "id": 89047, "pid": 435, "tid": 435, "ts": 1776649792672902, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792673258, "dur": 49, + "args": { + "External id": 12030, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89048, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89048, "pid": 0, "tid": 7, "ts": 1776649792673258, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792672903, "dur": 5, + "args": { + "External id": 12030, + "cbid": 211, "correlation": 89048 + } + }, + { + "ph": "s", "id": 89048, "pid": 435, "tid": 435, "ts": 1776649792672903, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792672931, "dur": 0, + "args": { + "External id": 12032, + "cbid": 210, "correlation": 89075 + } + }, + { + "ph": "f", "id": 89075, "pid": 435, "tid": 435, "ts": 1776649792672931, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792673309, "dur": 47, + "args": { + "External id": 12032, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89076, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89076, "pid": 0, "tid": 7, "ts": 1776649792673309, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792672933, "dur": 5, + "args": { + "External id": 12032, + "cbid": 211, "correlation": 89076 + } + }, + { + "ph": "s", "id": 89076, "pid": 435, "tid": 435, "ts": 1776649792672933, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792672961, "dur": 0, + "args": { + "External id": 12034, + "cbid": 210, "correlation": 89103 + } + }, + { + "ph": "f", "id": 89103, "pid": 435, "tid": 435, "ts": 1776649792672961, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792673357, "dur": 47, + "args": { + "External id": 12034, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89104, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89104, "pid": 0, "tid": 7, "ts": 1776649792673357, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792672962, "dur": 6, + "args": { + "External id": 12034, + "cbid": 211, "correlation": 89104 + } + }, + { + "ph": "s", "id": 89104, "pid": 435, "tid": 435, "ts": 1776649792672962, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792672991, "dur": 0, + "args": { + "External id": 12036, + "cbid": 210, "correlation": 89131 + } + }, + { + "ph": "f", "id": 89131, "pid": 435, "tid": 435, "ts": 1776649792672991, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792673407, "dur": 44, + "args": { + "External id": 12036, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89132, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89132, "pid": 0, "tid": 7, "ts": 1776649792673407, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792672992, "dur": 5, + "args": { + "External id": 12036, + "cbid": 211, "correlation": 89132 + } + }, + { + "ph": "s", "id": 89132, "pid": 435, "tid": 435, "ts": 1776649792672992, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792673020, "dur": 0, + "args": { + "External id": 12038, + "cbid": 210, "correlation": 89159 + } + }, + { + "ph": "f", "id": 89159, "pid": 435, "tid": 435, "ts": 1776649792673020, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792673453, "dur": 46, + "args": { + "External id": 12038, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89160, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89160, "pid": 0, "tid": 7, "ts": 1776649792673453, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792673022, "dur": 6, + "args": { + "External id": 12038, + "cbid": 211, "correlation": 89160 + } + }, + { + "ph": "s", "id": 89160, "pid": 435, "tid": 435, "ts": 1776649792673022, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792673050, "dur": 0, + "args": { + "External id": 12040, + "cbid": 210, "correlation": 89187 + } + }, + { + "ph": "f", "id": 89187, "pid": 435, "tid": 435, "ts": 1776649792673050, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792673500, "dur": 52, + "args": { + "External id": 12040, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89188, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89188, "pid": 0, "tid": 7, "ts": 1776649792673500, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792673051, "dur": 5, + "args": { + "External id": 12040, + "cbid": 211, "correlation": 89188 + } + }, + { + "ph": "s", "id": 89188, "pid": 435, "tid": 435, "ts": 1776649792673051, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792673080, "dur": 0, + "args": { + "External id": 12042, + "cbid": 210, "correlation": 89215 + } + }, + { + "ph": "f", "id": 89215, "pid": 435, "tid": 435, "ts": 1776649792673080, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792673554, "dur": 43, + "args": { + "External id": 12042, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89216, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89216, "pid": 0, "tid": 7, "ts": 1776649792673554, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792673081, "dur": 6, + "args": { + "External id": 12042, + "cbid": 211, "correlation": 89216 + } + }, + { + "ph": "s", "id": 89216, "pid": 435, "tid": 435, "ts": 1776649792673081, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792673109, "dur": 0, + "args": { + "External id": 12044, + "cbid": 210, "correlation": 89243 + } + }, + { + "ph": "f", "id": 89243, "pid": 435, "tid": 435, "ts": 1776649792673109, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792673599, "dur": 47, + "args": { + "External id": 12044, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89244, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89244, "pid": 0, "tid": 7, "ts": 1776649792673599, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792673110, "dur": 5, + "args": { + "External id": 12044, + "cbid": 211, "correlation": 89244 + } + }, + { + "ph": "s", "id": 89244, "pid": 435, "tid": 435, "ts": 1776649792673110, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792673139, "dur": 0, + "args": { + "External id": 12046, + "cbid": 210, "correlation": 89271 + } + }, + { + "ph": "f", "id": 89271, "pid": 435, "tid": 435, "ts": 1776649792673139, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792673647, "dur": 44, + "args": { + "External id": 12046, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89272, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89272, "pid": 0, "tid": 7, "ts": 1776649792673647, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792673140, "dur": 5, + "args": { + "External id": 12046, + "cbid": 211, "correlation": 89272 + } + }, + { + "ph": "s", "id": 89272, "pid": 435, "tid": 435, "ts": 1776649792673140, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792673168, "dur": 0, + "args": { + "External id": 12048, + "cbid": 210, "correlation": 89299 + } + }, + { + "ph": "f", "id": 89299, "pid": 435, "tid": 435, "ts": 1776649792673168, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792673693, "dur": 47, + "args": { + "External id": 12048, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89300, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89300, "pid": 0, "tid": 7, "ts": 1776649792673693, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792673170, "dur": 6, + "args": { + "External id": 12048, + "cbid": 211, "correlation": 89300 + } + }, + { + "ph": "s", "id": 89300, "pid": 435, "tid": 435, "ts": 1776649792673170, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792673198, "dur": 0, + "args": { + "External id": 12050, + "cbid": 210, "correlation": 89327 + } + }, + { + "ph": "f", "id": 89327, "pid": 435, "tid": 435, "ts": 1776649792673198, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792673742, "dur": 49, + "args": { + "External id": 12050, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89328, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89328, "pid": 0, "tid": 7, "ts": 1776649792673742, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792673199, "dur": 5, + "args": { + "External id": 12050, + "cbid": 211, "correlation": 89328 + } + }, + { + "ph": "s", "id": 89328, "pid": 435, "tid": 435, "ts": 1776649792673199, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792673228, "dur": 0, + "args": { + "External id": 12052, + "cbid": 210, "correlation": 89355 + } + }, + { + "ph": "f", "id": 89355, "pid": 435, "tid": 435, "ts": 1776649792673228, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792673793, "dur": 53, + "args": { + "External id": 12052, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89356, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89356, "pid": 0, "tid": 7, "ts": 1776649792673793, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792673229, "dur": 5, + "args": { + "External id": 12052, + "cbid": 211, "correlation": 89356 + } + }, + { + "ph": "s", "id": 89356, "pid": 435, "tid": 435, "ts": 1776649792673229, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792673257, "dur": 0, + "args": { + "External id": 12054, + "cbid": 210, "correlation": 89383 + } + }, + { + "ph": "f", "id": 89383, "pid": 435, "tid": 435, "ts": 1776649792673257, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792673847, "dur": 49, + "args": { + "External id": 12054, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89384, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89384, "pid": 0, "tid": 7, "ts": 1776649792673847, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792673259, "dur": 5, + "args": { + "External id": 12054, + "cbid": 211, "correlation": 89384 + } + }, + { + "ph": "s", "id": 89384, "pid": 435, "tid": 435, "ts": 1776649792673259, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792673287, "dur": 0, + "args": { + "External id": 12056, + "cbid": 210, "correlation": 89411 + } + }, + { + "ph": "f", "id": 89411, "pid": 435, "tid": 435, "ts": 1776649792673287, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792673898, "dur": 50, + "args": { + "External id": 12056, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89412, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89412, "pid": 0, "tid": 7, "ts": 1776649792673898, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792673289, "dur": 5, + "args": { + "External id": 12056, + "cbid": 211, "correlation": 89412 + } + }, + { + "ph": "s", "id": 89412, "pid": 435, "tid": 435, "ts": 1776649792673289, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792673317, "dur": 0, + "args": { + "External id": 12058, + "cbid": 210, "correlation": 89439 + } + }, + { + "ph": "f", "id": 89439, "pid": 435, "tid": 435, "ts": 1776649792673317, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792673949, "dur": 47, + "args": { + "External id": 12058, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89440, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89440, "pid": 0, "tid": 7, "ts": 1776649792673949, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792673318, "dur": 5, + "args": { + "External id": 12058, + "cbid": 211, "correlation": 89440 + } + }, + { + "ph": "s", "id": 89440, "pid": 435, "tid": 435, "ts": 1776649792673318, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792673346, "dur": 0, + "args": { + "External id": 12060, + "cbid": 210, "correlation": 89467 + } + }, + { + "ph": "f", "id": 89467, "pid": 435, "tid": 435, "ts": 1776649792673346, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792673998, "dur": 46, + "args": { + "External id": 12060, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89468, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89468, "pid": 0, "tid": 7, "ts": 1776649792673998, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792673348, "dur": 6, + "args": { + "External id": 12060, + "cbid": 211, "correlation": 89468 + } + }, + { + "ph": "s", "id": 89468, "pid": 435, "tid": 435, "ts": 1776649792673348, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792673376, "dur": 0, + "args": { + "External id": 12062, + "cbid": 210, "correlation": 89495 + } + }, + { + "ph": "f", "id": 89495, "pid": 435, "tid": 435, "ts": 1776649792673376, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792674046, "dur": 44, + "args": { + "External id": 12062, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89496, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89496, "pid": 0, "tid": 7, "ts": 1776649792674046, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792673377, "dur": 6, + "args": { + "External id": 12062, + "cbid": 211, "correlation": 89496 + } + }, + { + "ph": "s", "id": 89496, "pid": 435, "tid": 435, "ts": 1776649792673377, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792673406, "dur": 0, + "args": { + "External id": 12064, + "cbid": 210, "correlation": 89523 + } + }, + { + "ph": "f", "id": 89523, "pid": 435, "tid": 435, "ts": 1776649792673406, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792674092, "dur": 50, + "args": { + "External id": 12064, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89524, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89524, "pid": 0, "tid": 7, "ts": 1776649792674092, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792673407, "dur": 5, + "args": { + "External id": 12064, + "cbid": 211, "correlation": 89524 + } + }, + { + "ph": "s", "id": 89524, "pid": 435, "tid": 435, "ts": 1776649792673407, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792673436, "dur": 0, + "args": { + "External id": 12066, + "cbid": 210, "correlation": 89551 + } + }, + { + "ph": "f", "id": 89551, "pid": 435, "tid": 435, "ts": 1776649792673436, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792674143, "dur": 46, + "args": { + "External id": 12066, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89552, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89552, "pid": 0, "tid": 7, "ts": 1776649792674143, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792673437, "dur": 5, + "args": { + "External id": 12066, + "cbid": 211, "correlation": 89552 + } + }, + { + "ph": "s", "id": 89552, "pid": 435, "tid": 435, "ts": 1776649792673437, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792673465, "dur": 0, + "args": { + "External id": 12068, + "cbid": 210, "correlation": 89579 + } + }, + { + "ph": "f", "id": 89579, "pid": 435, "tid": 435, "ts": 1776649792673465, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792674191, "dur": 46, + "args": { + "External id": 12068, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89580, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89580, "pid": 0, "tid": 7, "ts": 1776649792674191, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792673466, "dur": 5, + "args": { + "External id": 12068, + "cbid": 211, "correlation": 89580 + } + }, + { + "ph": "s", "id": 89580, "pid": 435, "tid": 435, "ts": 1776649792673466, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792673495, "dur": 0, + "args": { + "External id": 12070, + "cbid": 210, "correlation": 89607 + } + }, + { + "ph": "f", "id": 89607, "pid": 435, "tid": 435, "ts": 1776649792673495, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792674239, "dur": 48, + "args": { + "External id": 12070, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89608, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89608, "pid": 0, "tid": 7, "ts": 1776649792674239, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792673496, "dur": 5, + "args": { + "External id": 12070, + "cbid": 211, "correlation": 89608 + } + }, + { + "ph": "s", "id": 89608, "pid": 435, "tid": 435, "ts": 1776649792673496, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792673525, "dur": 0, + "args": { + "External id": 12072, + "cbid": 210, "correlation": 89635 + } + }, + { + "ph": "f", "id": 89635, "pid": 435, "tid": 435, "ts": 1776649792673525, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792674288, "dur": 47, + "args": { + "External id": 12072, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89636, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89636, "pid": 0, "tid": 7, "ts": 1776649792674288, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792673526, "dur": 5, + "args": { + "External id": 12072, + "cbid": 211, "correlation": 89636 + } + }, + { + "ph": "s", "id": 89636, "pid": 435, "tid": 435, "ts": 1776649792673526, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792673554, "dur": 0, + "args": { + "External id": 12074, + "cbid": 210, "correlation": 89663 + } + }, + { + "ph": "f", "id": 89663, "pid": 435, "tid": 435, "ts": 1776649792673554, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792674336, "dur": 49, + "args": { + "External id": 12074, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89664, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89664, "pid": 0, "tid": 7, "ts": 1776649792674336, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792673556, "dur": 6, + "args": { + "External id": 12074, + "cbid": 211, "correlation": 89664 + } + }, + { + "ph": "s", "id": 89664, "pid": 435, "tid": 435, "ts": 1776649792673556, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792673584, "dur": 0, + "args": { + "External id": 12076, + "cbid": 210, "correlation": 89691 + } + }, + { + "ph": "f", "id": 89691, "pid": 435, "tid": 435, "ts": 1776649792673584, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792674387, "dur": 46, + "args": { + "External id": 12076, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89692, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89692, "pid": 0, "tid": 7, "ts": 1776649792674387, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792673585, "dur": 6, + "args": { + "External id": 12076, + "cbid": 211, "correlation": 89692 + } + }, + { + "ph": "s", "id": 89692, "pid": 435, "tid": 435, "ts": 1776649792673585, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792673614, "dur": 0, + "args": { + "External id": 12078, + "cbid": 210, "correlation": 89719 + } + }, + { + "ph": "f", "id": 89719, "pid": 435, "tid": 435, "ts": 1776649792673614, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792674435, "dur": 47, + "args": { + "External id": 12078, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89720, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89720, "pid": 0, "tid": 7, "ts": 1776649792674435, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792673615, "dur": 6, + "args": { + "External id": 12078, + "cbid": 211, "correlation": 89720 + } + }, + { + "ph": "s", "id": 89720, "pid": 435, "tid": 435, "ts": 1776649792673615, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792673643, "dur": 0, + "args": { + "External id": 12080, + "cbid": 210, "correlation": 89747 + } + }, + { + "ph": "f", "id": 89747, "pid": 435, "tid": 435, "ts": 1776649792673643, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792674484, "dur": 47, + "args": { + "External id": 12080, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89748, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89748, "pid": 0, "tid": 7, "ts": 1776649792674484, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792673645, "dur": 6, + "args": { + "External id": 12080, + "cbid": 211, "correlation": 89748 + } + }, + { + "ph": "s", "id": 89748, "pid": 435, "tid": 435, "ts": 1776649792673645, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792673673, "dur": 0, + "args": { + "External id": 12082, + "cbid": 210, "correlation": 89775 + } + }, + { + "ph": "f", "id": 89775, "pid": 435, "tid": 435, "ts": 1776649792673673, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792674532, "dur": 44, + "args": { + "External id": 12082, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89776, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89776, "pid": 0, "tid": 7, "ts": 1776649792674532, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792673675, "dur": 5, + "args": { + "External id": 12082, + "cbid": 211, "correlation": 89776 + } + }, + { + "ph": "s", "id": 89776, "pid": 435, "tid": 435, "ts": 1776649792673675, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792673703, "dur": 0, + "args": { + "External id": 12084, + "cbid": 210, "correlation": 89803 + } + }, + { + "ph": "f", "id": 89803, "pid": 435, "tid": 435, "ts": 1776649792673703, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792674578, "dur": 49, + "args": { + "External id": 12084, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89804, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89804, "pid": 0, "tid": 7, "ts": 1776649792674578, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792673704, "dur": 6, + "args": { + "External id": 12084, + "cbid": 211, "correlation": 89804 + } + }, + { + "ph": "s", "id": 89804, "pid": 435, "tid": 435, "ts": 1776649792673704, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792673733, "dur": 0, + "args": { + "External id": 12086, + "cbid": 210, "correlation": 89831 + } + }, + { + "ph": "f", "id": 89831, "pid": 435, "tid": 435, "ts": 1776649792673733, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792674629, "dur": 44, + "args": { + "External id": 12086, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89832, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89832, "pid": 0, "tid": 7, "ts": 1776649792674629, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792673734, "dur": 5, + "args": { + "External id": 12086, + "cbid": 211, "correlation": 89832 + } + }, + { + "ph": "s", "id": 89832, "pid": 435, "tid": 435, "ts": 1776649792673734, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792673763, "dur": 0, + "args": { + "External id": 12088, + "cbid": 210, "correlation": 89859 + } + }, + { + "ph": "f", "id": 89859, "pid": 435, "tid": 435, "ts": 1776649792673763, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792674674, "dur": 52, + "args": { + "External id": 12088, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89860, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89860, "pid": 0, "tid": 7, "ts": 1776649792674674, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792673764, "dur": 6, + "args": { + "External id": 12088, + "cbid": 211, "correlation": 89860 + } + }, + { + "ph": "s", "id": 89860, "pid": 435, "tid": 435, "ts": 1776649792673764, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649792673798, "dur": 4, + "args": { + "External id": 89871, + "cbid": 135, "correlation": 89871 + } + }, + { + "ph": "f", "id": 89871, "pid": 435, "tid": 435, "ts": 1776649792673798, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792673814, "dur": 922, + "args": { + "External id": 89876, + "cbid": 165, "correlation": 89876 + } + }, + { + "ph": "s", "id": 89876, "pid": 435, "tid": 435, "ts": 1776649792673814, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649792674739, "dur": 5, + "args": { + "External id": 89879, + "cbid": 139, "correlation": 89879 + } + }, + { + "ph": "f", "id": 89879, "pid": 435, "tid": 435, "ts": 1776649792674739, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792674777, "dur": 1, + "args": { + "External id": 12090, + "cbid": 210, "correlation": 89913 + } + }, + { + "ph": "f", "id": 89913, "pid": 435, "tid": 435, "ts": 1776649792674777, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792674791, "dur": 40, + "args": { + "External id": 12090, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89914, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89914, "pid": 0, "tid": 7, "ts": 1776649792674791, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792674780, "dur": 8, + "args": { + "External id": 12090, + "cbid": 211, "correlation": 89914 + } + }, + { + "ph": "s", "id": 89914, "pid": 435, "tid": 435, "ts": 1776649792674780, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792674813, "dur": 0, + "args": { + "External id": 12092, + "cbid": 210, "correlation": 89941 + } + }, + { + "ph": "f", "id": 89941, "pid": 435, "tid": 435, "ts": 1776649792674813, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792674832, "dur": 47, + "args": { + "External id": 12092, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89942, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89942, "pid": 0, "tid": 7, "ts": 1776649792674832, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792674815, "dur": 6, + "args": { + "External id": 12092, + "cbid": 211, "correlation": 89942 + } + }, + { + "ph": "s", "id": 89942, "pid": 435, "tid": 435, "ts": 1776649792674815, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792674844, "dur": 0, + "args": { + "External id": 12094, + "cbid": 210, "correlation": 89969 + } + }, + { + "ph": "f", "id": 89969, "pid": 435, "tid": 435, "ts": 1776649792674844, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792674881, "dur": 44, + "args": { + "External id": 12094, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89970, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89970, "pid": 0, "tid": 7, "ts": 1776649792674881, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792674846, "dur": 6, + "args": { + "External id": 12094, + "cbid": 211, "correlation": 89970 + } + }, + { + "ph": "s", "id": 89970, "pid": 435, "tid": 435, "ts": 1776649792674846, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792674875, "dur": 0, + "args": { + "External id": 12096, + "cbid": 210, "correlation": 89997 + } + }, + { + "ph": "f", "id": 89997, "pid": 435, "tid": 435, "ts": 1776649792674875, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792674926, "dur": 52, + "args": { + "External id": 12096, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 89998, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 89998, "pid": 0, "tid": 7, "ts": 1776649792674926, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792674876, "dur": 6, + "args": { + "External id": 12096, + "cbid": 211, "correlation": 89998 + } + }, + { + "ph": "s", "id": 89998, "pid": 435, "tid": 435, "ts": 1776649792674876, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792674905, "dur": 0, + "args": { + "External id": 12098, + "cbid": 210, "correlation": 90025 + } + }, + { + "ph": "f", "id": 90025, "pid": 435, "tid": 435, "ts": 1776649792674905, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792674980, "dur": 50, + "args": { + "External id": 12098, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90026, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90026, "pid": 0, "tid": 7, "ts": 1776649792674980, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792674906, "dur": 6, + "args": { + "External id": 12098, + "cbid": 211, "correlation": 90026 + } + }, + { + "ph": "s", "id": 90026, "pid": 435, "tid": 435, "ts": 1776649792674906, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792674934, "dur": 0, + "args": { + "External id": 12100, + "cbid": 210, "correlation": 90053 + } + }, + { + "ph": "f", "id": 90053, "pid": 435, "tid": 435, "ts": 1776649792674934, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792675032, "dur": 45, + "args": { + "External id": 12100, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90054, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90054, "pid": 0, "tid": 7, "ts": 1776649792675032, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792674936, "dur": 5, + "args": { + "External id": 12100, + "cbid": 211, "correlation": 90054 + } + }, + { + "ph": "s", "id": 90054, "pid": 435, "tid": 435, "ts": 1776649792674936, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792674964, "dur": 0, + "args": { + "External id": 12102, + "cbid": 210, "correlation": 90081 + } + }, + { + "ph": "f", "id": 90081, "pid": 435, "tid": 435, "ts": 1776649792674964, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792675079, "dur": 44, + "args": { + "External id": 12102, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90082, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90082, "pid": 0, "tid": 7, "ts": 1776649792675079, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792674966, "dur": 5, + "args": { + "External id": 12102, + "cbid": 211, "correlation": 90082 + } + }, + { + "ph": "s", "id": 90082, "pid": 435, "tid": 435, "ts": 1776649792674966, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792674994, "dur": 0, + "args": { + "External id": 12104, + "cbid": 210, "correlation": 90109 + } + }, + { + "ph": "f", "id": 90109, "pid": 435, "tid": 435, "ts": 1776649792674994, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792675125, "dur": 51, + "args": { + "External id": 12104, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90110, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90110, "pid": 0, "tid": 7, "ts": 1776649792675125, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792674995, "dur": 6, + "args": { + "External id": 12104, + "cbid": 211, "correlation": 90110 + } + }, + { + "ph": "s", "id": 90110, "pid": 435, "tid": 435, "ts": 1776649792674995, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792675024, "dur": 0, + "args": { + "External id": 12106, + "cbid": 210, "correlation": 90137 + } + }, + { + "ph": "f", "id": 90137, "pid": 435, "tid": 435, "ts": 1776649792675024, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792675178, "dur": 48, + "args": { + "External id": 12106, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90138, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90138, "pid": 0, "tid": 7, "ts": 1776649792675178, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792675025, "dur": 5, + "args": { + "External id": 12106, + "cbid": 211, "correlation": 90138 + } + }, + { + "ph": "s", "id": 90138, "pid": 435, "tid": 435, "ts": 1776649792675025, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792675053, "dur": 0, + "args": { + "External id": 12108, + "cbid": 210, "correlation": 90165 + } + }, + { + "ph": "f", "id": 90165, "pid": 435, "tid": 435, "ts": 1776649792675053, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792675227, "dur": 47, + "args": { + "External id": 12108, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90166, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90166, "pid": 0, "tid": 7, "ts": 1776649792675227, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792675054, "dur": 6, + "args": { + "External id": 12108, + "cbid": 211, "correlation": 90166 + } + }, + { + "ph": "s", "id": 90166, "pid": 435, "tid": 435, "ts": 1776649792675054, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792675079, "dur": 199, + "args": { + "External id": 90172, + "cbid": 165, "correlation": 90172 + } + }, + { + "ph": "s", "id": 90172, "pid": 435, "tid": 435, "ts": 1776649792675079, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649792675299, "dur": 3, + "args": { + "External id": 90181, + "cbid": 135, "correlation": 90181 + } + }, + { + "ph": "f", "id": 90181, "pid": 435, "tid": 435, "ts": 1776649792675299, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792675329, "dur": 1, + "args": { + "External id": 12110, + "cbid": 210, "correlation": 90207 + } + }, + { + "ph": "f", "id": 90207, "pid": 435, "tid": 435, "ts": 1776649792675329, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792675341, "dur": 41, + "args": { + "External id": 12110, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90208, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90208, "pid": 0, "tid": 7, "ts": 1776649792675341, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792675331, "dur": 8, + "args": { + "External id": 12110, + "cbid": 211, "correlation": 90208 + } + }, + { + "ph": "s", "id": 90208, "pid": 435, "tid": 435, "ts": 1776649792675331, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792675364, "dur": 0, + "args": { + "External id": 12112, + "cbid": 210, "correlation": 90235 + } + }, + { + "ph": "f", "id": 90235, "pid": 435, "tid": 435, "ts": 1776649792675364, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792675384, "dur": 40, + "args": { + "External id": 12112, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90236, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90236, "pid": 0, "tid": 7, "ts": 1776649792675384, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792675365, "dur": 6, + "args": { + "External id": 12112, + "cbid": 211, "correlation": 90236 + } + }, + { + "ph": "s", "id": 90236, "pid": 435, "tid": 435, "ts": 1776649792675365, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792675397, "dur": 0, + "args": { + "External id": 12114, + "cbid": 210, "correlation": 90263 + } + }, + { + "ph": "f", "id": 90263, "pid": 435, "tid": 435, "ts": 1776649792675397, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792675426, "dur": 49, + "args": { + "External id": 12114, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90264, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90264, "pid": 0, "tid": 7, "ts": 1776649792675426, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792675398, "dur": 6, + "args": { + "External id": 12114, + "cbid": 211, "correlation": 90264 + } + }, + { + "ph": "s", "id": 90264, "pid": 435, "tid": 435, "ts": 1776649792675398, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792675429, "dur": 0, + "args": { + "External id": 12116, + "cbid": 210, "correlation": 90291 + } + }, + { + "ph": "f", "id": 90291, "pid": 435, "tid": 435, "ts": 1776649792675429, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792675477, "dur": 44, + "args": { + "External id": 12116, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90292, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90292, "pid": 0, "tid": 7, "ts": 1776649792675477, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792675431, "dur": 6, + "args": { + "External id": 12116, + "cbid": 211, "correlation": 90292 + } + }, + { + "ph": "s", "id": 90292, "pid": 435, "tid": 435, "ts": 1776649792675431, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792675460, "dur": 0, + "args": { + "External id": 12118, + "cbid": 210, "correlation": 90319 + } + }, + { + "ph": "f", "id": 90319, "pid": 435, "tid": 435, "ts": 1776649792675460, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792675522, "dur": 49, + "args": { + "External id": 12118, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90320, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90320, "pid": 0, "tid": 7, "ts": 1776649792675522, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792675461, "dur": 6, + "args": { + "External id": 12118, + "cbid": 211, "correlation": 90320 + } + }, + { + "ph": "s", "id": 90320, "pid": 435, "tid": 435, "ts": 1776649792675461, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792675489, "dur": 0, + "args": { + "External id": 12120, + "cbid": 210, "correlation": 90347 + } + }, + { + "ph": "f", "id": 90347, "pid": 435, "tid": 435, "ts": 1776649792675489, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792675573, "dur": 50, + "args": { + "External id": 12120, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90348, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90348, "pid": 0, "tid": 7, "ts": 1776649792675573, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792675491, "dur": 6, + "args": { + "External id": 12120, + "cbid": 211, "correlation": 90348 + } + }, + { + "ph": "s", "id": 90348, "pid": 435, "tid": 435, "ts": 1776649792675491, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792675519, "dur": 0, + "args": { + "External id": 12122, + "cbid": 210, "correlation": 90375 + } + }, + { + "ph": "f", "id": 90375, "pid": 435, "tid": 435, "ts": 1776649792675519, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792675626, "dur": 44, + "args": { + "External id": 12122, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90376, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90376, "pid": 0, "tid": 7, "ts": 1776649792675626, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792675520, "dur": 6, + "args": { + "External id": 12122, + "cbid": 211, "correlation": 90376 + } + }, + { + "ph": "s", "id": 90376, "pid": 435, "tid": 435, "ts": 1776649792675520, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792675549, "dur": 0, + "args": { + "External id": 12124, + "cbid": 210, "correlation": 90403 + } + }, + { + "ph": "f", "id": 90403, "pid": 435, "tid": 435, "ts": 1776649792675549, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792675672, "dur": 50, + "args": { + "External id": 12124, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90404, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90404, "pid": 0, "tid": 7, "ts": 1776649792675672, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792675550, "dur": 6, + "args": { + "External id": 12124, + "cbid": 211, "correlation": 90404 + } + }, + { + "ph": "s", "id": 90404, "pid": 435, "tid": 435, "ts": 1776649792675550, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792675578, "dur": 0, + "args": { + "External id": 12126, + "cbid": 210, "correlation": 90431 + } + }, + { + "ph": "f", "id": 90431, "pid": 435, "tid": 435, "ts": 1776649792675578, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792675723, "dur": 46, + "args": { + "External id": 12126, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90432, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90432, "pid": 0, "tid": 7, "ts": 1776649792675723, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792675580, "dur": 6, + "args": { + "External id": 12126, + "cbid": 211, "correlation": 90432 + } + }, + { + "ph": "s", "id": 90432, "pid": 435, "tid": 435, "ts": 1776649792675580, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792675608, "dur": 0, + "args": { + "External id": 12128, + "cbid": 210, "correlation": 90459 + } + }, + { + "ph": "f", "id": 90459, "pid": 435, "tid": 435, "ts": 1776649792675608, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792675771, "dur": 46, + "args": { + "External id": 12128, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90460, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90460, "pid": 0, "tid": 7, "ts": 1776649792675771, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792675609, "dur": 5, + "args": { + "External id": 12128, + "cbid": 211, "correlation": 90460 + } + }, + { + "ph": "s", "id": 90460, "pid": 435, "tid": 435, "ts": 1776649792675609, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792675638, "dur": 0, + "args": { + "External id": 12130, + "cbid": 210, "correlation": 90487 + } + }, + { + "ph": "f", "id": 90487, "pid": 435, "tid": 435, "ts": 1776649792675638, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792675819, "dur": 43, + "args": { + "External id": 12130, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90488, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90488, "pid": 0, "tid": 7, "ts": 1776649792675819, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792675639, "dur": 5, + "args": { + "External id": 12130, + "cbid": 211, "correlation": 90488 + } + }, + { + "ph": "s", "id": 90488, "pid": 435, "tid": 435, "ts": 1776649792675639, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792675667, "dur": 0, + "args": { + "External id": 12132, + "cbid": 210, "correlation": 90515 + } + }, + { + "ph": "f", "id": 90515, "pid": 435, "tid": 435, "ts": 1776649792675667, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792675864, "dur": 52, + "args": { + "External id": 12132, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90516, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90516, "pid": 0, "tid": 7, "ts": 1776649792675864, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792675668, "dur": 5, + "args": { + "External id": 12132, + "cbid": 211, "correlation": 90516 + } + }, + { + "ph": "s", "id": 90516, "pid": 435, "tid": 435, "ts": 1776649792675668, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792675697, "dur": 0, + "args": { + "External id": 12134, + "cbid": 210, "correlation": 90543 + } + }, + { + "ph": "f", "id": 90543, "pid": 435, "tid": 435, "ts": 1776649792675697, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792675917, "dur": 44, + "args": { + "External id": 12134, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90544, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90544, "pid": 0, "tid": 7, "ts": 1776649792675917, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792675698, "dur": 5, + "args": { + "External id": 12134, + "cbid": 211, "correlation": 90544 + } + }, + { + "ph": "s", "id": 90544, "pid": 435, "tid": 435, "ts": 1776649792675698, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792675727, "dur": 0, + "args": { + "External id": 12136, + "cbid": 210, "correlation": 90571 + } + }, + { + "ph": "f", "id": 90571, "pid": 435, "tid": 435, "ts": 1776649792675727, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792675963, "dur": 50, + "args": { + "External id": 12136, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90572, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90572, "pid": 0, "tid": 7, "ts": 1776649792675963, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792675728, "dur": 6, + "args": { + "External id": 12136, + "cbid": 211, "correlation": 90572 + } + }, + { + "ph": "s", "id": 90572, "pid": 435, "tid": 435, "ts": 1776649792675728, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792675757, "dur": 0, + "args": { + "External id": 12138, + "cbid": 210, "correlation": 90599 + } + }, + { + "ph": "f", "id": 90599, "pid": 435, "tid": 435, "ts": 1776649792675757, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792676015, "dur": 51, + "args": { + "External id": 12138, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90600, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90600, "pid": 0, "tid": 7, "ts": 1776649792676015, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792675758, "dur": 5, + "args": { + "External id": 12138, + "cbid": 211, "correlation": 90600 + } + }, + { + "ph": "s", "id": 90600, "pid": 435, "tid": 435, "ts": 1776649792675758, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792675787, "dur": 0, + "args": { + "External id": 12140, + "cbid": 210, "correlation": 90627 + } + }, + { + "ph": "f", "id": 90627, "pid": 435, "tid": 435, "ts": 1776649792675787, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792676068, "dur": 49, + "args": { + "External id": 12140, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90628, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90628, "pid": 0, "tid": 7, "ts": 1776649792676068, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792675788, "dur": 6, + "args": { + "External id": 12140, + "cbid": 211, "correlation": 90628 + } + }, + { + "ph": "s", "id": 90628, "pid": 435, "tid": 435, "ts": 1776649792675788, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792675816, "dur": 0, + "args": { + "External id": 12142, + "cbid": 210, "correlation": 90655 + } + }, + { + "ph": "f", "id": 90655, "pid": 435, "tid": 435, "ts": 1776649792675816, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792676118, "dur": 51, + "args": { + "External id": 12142, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90656, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90656, "pid": 0, "tid": 7, "ts": 1776649792676118, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792675818, "dur": 5, + "args": { + "External id": 12142, + "cbid": 211, "correlation": 90656 + } + }, + { + "ph": "s", "id": 90656, "pid": 435, "tid": 435, "ts": 1776649792675818, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792675846, "dur": 0, + "args": { + "External id": 12144, + "cbid": 210, "correlation": 90683 + } + }, + { + "ph": "f", "id": 90683, "pid": 435, "tid": 435, "ts": 1776649792675846, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792676171, "dur": 47, + "args": { + "External id": 12144, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90684, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90684, "pid": 0, "tid": 7, "ts": 1776649792676171, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792675847, "dur": 5, + "args": { + "External id": 12144, + "cbid": 211, "correlation": 90684 + } + }, + { + "ph": "s", "id": 90684, "pid": 435, "tid": 435, "ts": 1776649792675847, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792675875, "dur": 0, + "args": { + "External id": 12146, + "cbid": 210, "correlation": 90711 + } + }, + { + "ph": "f", "id": 90711, "pid": 435, "tid": 435, "ts": 1776649792675875, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792676220, "dur": 46, + "args": { + "External id": 12146, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90712, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90712, "pid": 0, "tid": 7, "ts": 1776649792676220, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792675877, "dur": 6, + "args": { + "External id": 12146, + "cbid": 211, "correlation": 90712 + } + }, + { + "ph": "s", "id": 90712, "pid": 435, "tid": 435, "ts": 1776649792675877, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792675905, "dur": 0, + "args": { + "External id": 12148, + "cbid": 210, "correlation": 90739 + } + }, + { + "ph": "f", "id": 90739, "pid": 435, "tid": 435, "ts": 1776649792675905, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792676268, "dur": 47, + "args": { + "External id": 12148, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90740, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90740, "pid": 0, "tid": 7, "ts": 1776649792676268, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792675906, "dur": 5, + "args": { + "External id": 12148, + "cbid": 211, "correlation": 90740 + } + }, + { + "ph": "s", "id": 90740, "pid": 435, "tid": 435, "ts": 1776649792675906, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792675935, "dur": 0, + "args": { + "External id": 12150, + "cbid": 210, "correlation": 90767 + } + }, + { + "ph": "f", "id": 90767, "pid": 435, "tid": 435, "ts": 1776649792675935, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792676317, "dur": 49, + "args": { + "External id": 12150, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90768, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90768, "pid": 0, "tid": 7, "ts": 1776649792676317, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792675936, "dur": 5, + "args": { + "External id": 12150, + "cbid": 211, "correlation": 90768 + } + }, + { + "ph": "s", "id": 90768, "pid": 435, "tid": 435, "ts": 1776649792675936, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792675964, "dur": 0, + "args": { + "External id": 12152, + "cbid": 210, "correlation": 90795 + } + }, + { + "ph": "f", "id": 90795, "pid": 435, "tid": 435, "ts": 1776649792675964, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792676368, "dur": 45, + "args": { + "External id": 12152, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90796, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90796, "pid": 0, "tid": 7, "ts": 1776649792676368, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792675965, "dur": 5, + "args": { + "External id": 12152, + "cbid": 211, "correlation": 90796 + } + }, + { + "ph": "s", "id": 90796, "pid": 435, "tid": 435, "ts": 1776649792675965, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792675994, "dur": 0, + "args": { + "External id": 12154, + "cbid": 210, "correlation": 90823 + } + }, + { + "ph": "f", "id": 90823, "pid": 435, "tid": 435, "ts": 1776649792675994, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792676414, "dur": 44, + "args": { + "External id": 12154, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90824, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90824, "pid": 0, "tid": 7, "ts": 1776649792676414, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792675995, "dur": 5, + "args": { + "External id": 12154, + "cbid": 211, "correlation": 90824 + } + }, + { + "ph": "s", "id": 90824, "pid": 435, "tid": 435, "ts": 1776649792675995, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676023, "dur": 0, + "args": { + "External id": 12156, + "cbid": 210, "correlation": 90851 + } + }, + { + "ph": "f", "id": 90851, "pid": 435, "tid": 435, "ts": 1776649792676023, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792676460, "dur": 51, + "args": { + "External id": 12156, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90852, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90852, "pid": 0, "tid": 7, "ts": 1776649792676460, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676025, "dur": 5, + "args": { + "External id": 12156, + "cbid": 211, "correlation": 90852 + } + }, + { + "ph": "s", "id": 90852, "pid": 435, "tid": 435, "ts": 1776649792676025, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676053, "dur": 0, + "args": { + "External id": 12158, + "cbid": 210, "correlation": 90879 + } + }, + { + "ph": "f", "id": 90879, "pid": 435, "tid": 435, "ts": 1776649792676053, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792676513, "dur": 49, + "args": { + "External id": 12158, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90880, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90880, "pid": 0, "tid": 7, "ts": 1776649792676513, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676054, "dur": 5, + "args": { + "External id": 12158, + "cbid": 211, "correlation": 90880 + } + }, + { + "ph": "s", "id": 90880, "pid": 435, "tid": 435, "ts": 1776649792676054, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676083, "dur": 0, + "args": { + "External id": 12160, + "cbid": 210, "correlation": 90907 + } + }, + { + "ph": "f", "id": 90907, "pid": 435, "tid": 435, "ts": 1776649792676083, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792676564, "dur": 47, + "args": { + "External id": 12160, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90908, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90908, "pid": 0, "tid": 7, "ts": 1776649792676564, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676084, "dur": 6, + "args": { + "External id": 12160, + "cbid": 211, "correlation": 90908 + } + }, + { + "ph": "s", "id": 90908, "pid": 435, "tid": 435, "ts": 1776649792676084, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676112, "dur": 0, + "args": { + "External id": 12162, + "cbid": 210, "correlation": 90935 + } + }, + { + "ph": "f", "id": 90935, "pid": 435, "tid": 435, "ts": 1776649792676112, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792676613, "dur": 52, + "args": { + "External id": 12162, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90936, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90936, "pid": 0, "tid": 7, "ts": 1776649792676613, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676114, "dur": 5, + "args": { + "External id": 12162, + "cbid": 211, "correlation": 90936 + } + }, + { + "ph": "s", "id": 90936, "pid": 435, "tid": 435, "ts": 1776649792676114, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676142, "dur": 0, + "args": { + "External id": 12164, + "cbid": 210, "correlation": 90963 + } + }, + { + "ph": "f", "id": 90963, "pid": 435, "tid": 435, "ts": 1776649792676142, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792676667, "dur": 47, + "args": { + "External id": 12164, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90964, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90964, "pid": 0, "tid": 7, "ts": 1776649792676667, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676143, "dur": 6, + "args": { + "External id": 12164, + "cbid": 211, "correlation": 90964 + } + }, + { + "ph": "s", "id": 90964, "pid": 435, "tid": 435, "ts": 1776649792676143, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676172, "dur": 0, + "args": { + "External id": 12166, + "cbid": 210, "correlation": 90991 + } + }, + { + "ph": "f", "id": 90991, "pid": 435, "tid": 435, "ts": 1776649792676172, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792676716, "dur": 44, + "args": { + "External id": 12166, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 90992, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 90992, "pid": 0, "tid": 7, "ts": 1776649792676716, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676173, "dur": 5, + "args": { + "External id": 12166, + "cbid": 211, "correlation": 90992 + } + }, + { + "ph": "s", "id": 90992, "pid": 435, "tid": 435, "ts": 1776649792676173, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676201, "dur": 0, + "args": { + "External id": 12168, + "cbid": 210, "correlation": 91019 + } + }, + { + "ph": "f", "id": 91019, "pid": 435, "tid": 435, "ts": 1776649792676201, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792676761, "dur": 50, + "args": { + "External id": 12168, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91020, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91020, "pid": 0, "tid": 7, "ts": 1776649792676761, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676202, "dur": 5, + "args": { + "External id": 12168, + "cbid": 211, "correlation": 91020 + } + }, + { + "ph": "s", "id": 91020, "pid": 435, "tid": 435, "ts": 1776649792676202, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676290, "dur": 1, + "args": { + "External id": 12170, + "cbid": 210, "correlation": 91047 + } + }, + { + "ph": "f", "id": 91047, "pid": 435, "tid": 435, "ts": 1776649792676290, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792676813, "dur": 45, + "args": { + "External id": 12170, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91048, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91048, "pid": 0, "tid": 7, "ts": 1776649792676813, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676292, "dur": 7, + "args": { + "External id": 12170, + "cbid": 211, "correlation": 91048 + } + }, + { + "ph": "s", "id": 91048, "pid": 435, "tid": 435, "ts": 1776649792676292, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676324, "dur": 0, + "args": { + "External id": 12172, + "cbid": 210, "correlation": 91075 + } + }, + { + "ph": "f", "id": 91075, "pid": 435, "tid": 435, "ts": 1776649792676324, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792676860, "dur": 46, + "args": { + "External id": 12172, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91076, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91076, "pid": 0, "tid": 7, "ts": 1776649792676860, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676326, "dur": 6, + "args": { + "External id": 12172, + "cbid": 211, "correlation": 91076 + } + }, + { + "ph": "s", "id": 91076, "pid": 435, "tid": 435, "ts": 1776649792676326, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676355, "dur": 0, + "args": { + "External id": 12174, + "cbid": 210, "correlation": 91103 + } + }, + { + "ph": "f", "id": 91103, "pid": 435, "tid": 435, "ts": 1776649792676355, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792676908, "dur": 51, + "args": { + "External id": 12174, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91104, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91104, "pid": 0, "tid": 7, "ts": 1776649792676908, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676356, "dur": 5, + "args": { + "External id": 12174, + "cbid": 211, "correlation": 91104 + } + }, + { + "ph": "s", "id": 91104, "pid": 435, "tid": 435, "ts": 1776649792676356, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676385, "dur": 0, + "args": { + "External id": 12176, + "cbid": 210, "correlation": 91131 + } + }, + { + "ph": "f", "id": 91131, "pid": 435, "tid": 435, "ts": 1776649792676385, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792676961, "dur": 44, + "args": { + "External id": 12176, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91132, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91132, "pid": 0, "tid": 7, "ts": 1776649792676961, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676386, "dur": 6, + "args": { + "External id": 12176, + "cbid": 211, "correlation": 91132 + } + }, + { + "ph": "s", "id": 91132, "pid": 435, "tid": 435, "ts": 1776649792676386, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676414, "dur": 0, + "args": { + "External id": 12178, + "cbid": 210, "correlation": 91159 + } + }, + { + "ph": "f", "id": 91159, "pid": 435, "tid": 435, "ts": 1776649792676414, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792677007, "dur": 44, + "args": { + "External id": 12178, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91160, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91160, "pid": 0, "tid": 7, "ts": 1776649792677007, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676415, "dur": 5, + "args": { + "External id": 12178, + "cbid": 211, "correlation": 91160 + } + }, + { + "ph": "s", "id": 91160, "pid": 435, "tid": 435, "ts": 1776649792676415, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676444, "dur": 0, + "args": { + "External id": 12180, + "cbid": 210, "correlation": 91187 + } + }, + { + "ph": "f", "id": 91187, "pid": 435, "tid": 435, "ts": 1776649792676444, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792677052, "dur": 48, + "args": { + "External id": 12180, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91188, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91188, "pid": 0, "tid": 7, "ts": 1776649792677052, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676445, "dur": 6, + "args": { + "External id": 12180, + "cbid": 211, "correlation": 91188 + } + }, + { + "ph": "s", "id": 91188, "pid": 435, "tid": 435, "ts": 1776649792676445, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676474, "dur": 0, + "args": { + "External id": 12182, + "cbid": 210, "correlation": 91215 + } + }, + { + "ph": "f", "id": 91215, "pid": 435, "tid": 435, "ts": 1776649792676474, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792677102, "dur": 46, + "args": { + "External id": 12182, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91216, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91216, "pid": 0, "tid": 7, "ts": 1776649792677102, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676475, "dur": 6, + "args": { + "External id": 12182, + "cbid": 211, "correlation": 91216 + } + }, + { + "ph": "s", "id": 91216, "pid": 435, "tid": 435, "ts": 1776649792676475, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676504, "dur": 0, + "args": { + "External id": 12184, + "cbid": 210, "correlation": 91243 + } + }, + { + "ph": "f", "id": 91243, "pid": 435, "tid": 435, "ts": 1776649792676504, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792677150, "dur": 49, + "args": { + "External id": 12184, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91244, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91244, "pid": 0, "tid": 7, "ts": 1776649792677150, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676505, "dur": 6, + "args": { + "External id": 12184, + "cbid": 211, "correlation": 91244 + } + }, + { + "ph": "s", "id": 91244, "pid": 435, "tid": 435, "ts": 1776649792676505, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676534, "dur": 0, + "args": { + "External id": 12186, + "cbid": 210, "correlation": 91271 + } + }, + { + "ph": "f", "id": 91271, "pid": 435, "tid": 435, "ts": 1776649792676534, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792677201, "dur": 47, + "args": { + "External id": 12186, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91272, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91272, "pid": 0, "tid": 7, "ts": 1776649792677201, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676535, "dur": 5, + "args": { + "External id": 12186, + "cbid": 211, "correlation": 91272 + } + }, + { + "ph": "s", "id": 91272, "pid": 435, "tid": 435, "ts": 1776649792676535, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676564, "dur": 0, + "args": { + "External id": 12188, + "cbid": 210, "correlation": 91299 + } + }, + { + "ph": "f", "id": 91299, "pid": 435, "tid": 435, "ts": 1776649792676564, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792677249, "dur": 49, + "args": { + "External id": 12188, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91300, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91300, "pid": 0, "tid": 7, "ts": 1776649792677249, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676565, "dur": 5, + "args": { + "External id": 12188, + "cbid": 211, "correlation": 91300 + } + }, + { + "ph": "s", "id": 91300, "pid": 435, "tid": 435, "ts": 1776649792676565, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676594, "dur": 0, + "args": { + "External id": 12190, + "cbid": 210, "correlation": 91327 + } + }, + { + "ph": "f", "id": 91327, "pid": 435, "tid": 435, "ts": 1776649792676594, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792677300, "dur": 44, + "args": { + "External id": 12190, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91328, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91328, "pid": 0, "tid": 7, "ts": 1776649792677300, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676595, "dur": 5, + "args": { + "External id": 12190, + "cbid": 211, "correlation": 91328 + } + }, + { + "ph": "s", "id": 91328, "pid": 435, "tid": 435, "ts": 1776649792676595, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676623, "dur": 0, + "args": { + "External id": 12192, + "cbid": 210, "correlation": 91355 + } + }, + { + "ph": "f", "id": 91355, "pid": 435, "tid": 435, "ts": 1776649792676623, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792677346, "dur": 50, + "args": { + "External id": 12192, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91356, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91356, "pid": 0, "tid": 7, "ts": 1776649792677346, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676624, "dur": 5, + "args": { + "External id": 12192, + "cbid": 211, "correlation": 91356 + } + }, + { + "ph": "s", "id": 91356, "pid": 435, "tid": 435, "ts": 1776649792676624, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676653, "dur": 0, + "args": { + "External id": 12194, + "cbid": 210, "correlation": 91383 + } + }, + { + "ph": "f", "id": 91383, "pid": 435, "tid": 435, "ts": 1776649792676653, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792677398, "dur": 47, + "args": { + "External id": 12194, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91384, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91384, "pid": 0, "tid": 7, "ts": 1776649792677398, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676654, "dur": 5, + "args": { + "External id": 12194, + "cbid": 211, "correlation": 91384 + } + }, + { + "ph": "s", "id": 91384, "pid": 435, "tid": 435, "ts": 1776649792676654, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676682, "dur": 0, + "args": { + "External id": 12196, + "cbid": 210, "correlation": 91411 + } + }, + { + "ph": "f", "id": 91411, "pid": 435, "tid": 435, "ts": 1776649792676682, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792677447, "dur": 50, + "args": { + "External id": 12196, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91412, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91412, "pid": 0, "tid": 7, "ts": 1776649792677447, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676683, "dur": 5, + "args": { + "External id": 12196, + "cbid": 211, "correlation": 91412 + } + }, + { + "ph": "s", "id": 91412, "pid": 435, "tid": 435, "ts": 1776649792676683, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676712, "dur": 0, + "args": { + "External id": 12198, + "cbid": 210, "correlation": 91439 + } + }, + { + "ph": "f", "id": 91439, "pid": 435, "tid": 435, "ts": 1776649792676712, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792677499, "dur": 44, + "args": { + "External id": 12198, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91440, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91440, "pid": 0, "tid": 7, "ts": 1776649792677499, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676713, "dur": 6, + "args": { + "External id": 12198, + "cbid": 211, "correlation": 91440 + } + }, + { + "ph": "s", "id": 91440, "pid": 435, "tid": 435, "ts": 1776649792676713, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676742, "dur": 0, + "args": { + "External id": 12200, + "cbid": 210, "correlation": 91467 + } + }, + { + "ph": "f", "id": 91467, "pid": 435, "tid": 435, "ts": 1776649792676742, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792677544, "dur": 47, + "args": { + "External id": 12200, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91468, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91468, "pid": 0, "tid": 7, "ts": 1776649792677544, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676743, "dur": 5, + "args": { + "External id": 12200, + "cbid": 211, "correlation": 91468 + } + }, + { + "ph": "s", "id": 91468, "pid": 435, "tid": 435, "ts": 1776649792676743, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676771, "dur": 0, + "args": { + "External id": 12202, + "cbid": 210, "correlation": 91495 + } + }, + { + "ph": "f", "id": 91495, "pid": 435, "tid": 435, "ts": 1776649792676771, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792677593, "dur": 47, + "args": { + "External id": 12202, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91496, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91496, "pid": 0, "tid": 7, "ts": 1776649792677593, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676772, "dur": 6, + "args": { + "External id": 12202, + "cbid": 211, "correlation": 91496 + } + }, + { + "ph": "s", "id": 91496, "pid": 435, "tid": 435, "ts": 1776649792676772, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676806, "dur": 0, + "args": { + "External id": 12204, + "cbid": 210, "correlation": 91523 + } + }, + { + "ph": "f", "id": 91523, "pid": 435, "tid": 435, "ts": 1776649792676806, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792677641, "dur": 49, + "args": { + "External id": 12204, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91524, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91524, "pid": 0, "tid": 7, "ts": 1776649792677641, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676808, "dur": 6, + "args": { + "External id": 12204, + "cbid": 211, "correlation": 91524 + } + }, + { + "ph": "s", "id": 91524, "pid": 435, "tid": 435, "ts": 1776649792676808, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676838, "dur": 0, + "args": { + "External id": 12206, + "cbid": 210, "correlation": 91551 + } + }, + { + "ph": "f", "id": 91551, "pid": 435, "tid": 435, "ts": 1776649792676838, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792677692, "dur": 47, + "args": { + "External id": 12206, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91552, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91552, "pid": 0, "tid": 7, "ts": 1776649792677692, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676839, "dur": 6, + "args": { + "External id": 12206, + "cbid": 211, "correlation": 91552 + } + }, + { + "ph": "s", "id": 91552, "pid": 435, "tid": 435, "ts": 1776649792676839, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792676868, "dur": 0, + "args": { + "External id": 12208, + "cbid": 210, "correlation": 91579 + } + }, + { + "ph": "f", "id": 91579, "pid": 435, "tid": 435, "ts": 1776649792676868, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792677740, "dur": 44, + "args": { + "External id": 12208, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91580, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91580, "pid": 0, "tid": 7, "ts": 1776649792677740, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792676869, "dur": 6, + "args": { + "External id": 12208, + "cbid": 211, "correlation": 91580 + } + }, + { + "ph": "s", "id": 91580, "pid": 435, "tid": 435, "ts": 1776649792676869, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649792676904, "dur": 4, + "args": { + "External id": 91591, + "cbid": 135, "correlation": 91591 + } + }, + { + "ph": "f", "id": 91591, "pid": 435, "tid": 435, "ts": 1776649792676904, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792676919, "dur": 875, + "args": { + "External id": 91596, + "cbid": 165, "correlation": 91596 + } + }, + { + "ph": "s", "id": 91596, "pid": 435, "tid": 435, "ts": 1776649792676919, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649792677797, "dur": 4, + "args": { + "External id": 91599, + "cbid": 139, "correlation": 91599 + } + }, + { + "ph": "f", "id": 91599, "pid": 435, "tid": 435, "ts": 1776649792677797, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792677835, "dur": 1, + "args": { + "External id": 12210, + "cbid": 210, "correlation": 91633 + } + }, + { + "ph": "f", "id": 91633, "pid": 435, "tid": 435, "ts": 1776649792677835, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792677848, "dur": 41, + "args": { + "External id": 12210, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91634, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91634, "pid": 0, "tid": 7, "ts": 1776649792677848, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792677837, "dur": 8, + "args": { + "External id": 12210, + "cbid": 211, "correlation": 91634 + } + }, + { + "ph": "s", "id": 91634, "pid": 435, "tid": 435, "ts": 1776649792677837, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792677870, "dur": 0, + "args": { + "External id": 12212, + "cbid": 210, "correlation": 91661 + } + }, + { + "ph": "f", "id": 91661, "pid": 435, "tid": 435, "ts": 1776649792677870, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792677890, "dur": 43, + "args": { + "External id": 12212, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91662, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91662, "pid": 0, "tid": 7, "ts": 1776649792677890, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792677872, "dur": 6, + "args": { + "External id": 12212, + "cbid": 211, "correlation": 91662 + } + }, + { + "ph": "s", "id": 91662, "pid": 435, "tid": 435, "ts": 1776649792677872, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792677901, "dur": 0, + "args": { + "External id": 12214, + "cbid": 210, "correlation": 91689 + } + }, + { + "ph": "f", "id": 91689, "pid": 435, "tid": 435, "ts": 1776649792677901, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792677935, "dur": 48, + "args": { + "External id": 12214, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91690, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91690, "pid": 0, "tid": 7, "ts": 1776649792677935, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792677903, "dur": 6, + "args": { + "External id": 12214, + "cbid": 211, "correlation": 91690 + } + }, + { + "ph": "s", "id": 91690, "pid": 435, "tid": 435, "ts": 1776649792677903, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792677932, "dur": 0, + "args": { + "External id": 12216, + "cbid": 210, "correlation": 91717 + } + }, + { + "ph": "f", "id": 91717, "pid": 435, "tid": 435, "ts": 1776649792677932, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792677985, "dur": 48, + "args": { + "External id": 12216, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91718, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91718, "pid": 0, "tid": 7, "ts": 1776649792677985, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792677933, "dur": 5, + "args": { + "External id": 12216, + "cbid": 211, "correlation": 91718 + } + }, + { + "ph": "s", "id": 91718, "pid": 435, "tid": 435, "ts": 1776649792677933, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792677961, "dur": 0, + "args": { + "External id": 12218, + "cbid": 210, "correlation": 91745 + } + }, + { + "ph": "f", "id": 91745, "pid": 435, "tid": 435, "ts": 1776649792677961, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792678034, "dur": 47, + "args": { + "External id": 12218, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91746, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91746, "pid": 0, "tid": 7, "ts": 1776649792678034, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792677963, "dur": 5, + "args": { + "External id": 12218, + "cbid": 211, "correlation": 91746 + } + }, + { + "ph": "s", "id": 91746, "pid": 435, "tid": 435, "ts": 1776649792677963, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792677991, "dur": 0, + "args": { + "External id": 12220, + "cbid": 210, "correlation": 91773 + } + }, + { + "ph": "f", "id": 91773, "pid": 435, "tid": 435, "ts": 1776649792677991, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792678082, "dur": 49, + "args": { + "External id": 12220, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91774, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91774, "pid": 0, "tid": 7, "ts": 1776649792678082, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792677992, "dur": 5, + "args": { + "External id": 12220, + "cbid": 211, "correlation": 91774 + } + }, + { + "ph": "s", "id": 91774, "pid": 435, "tid": 435, "ts": 1776649792677992, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792678021, "dur": 0, + "args": { + "External id": 12222, + "cbid": 210, "correlation": 91801 + } + }, + { + "ph": "f", "id": 91801, "pid": 435, "tid": 435, "ts": 1776649792678021, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792678133, "dur": 44, + "args": { + "External id": 12222, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91802, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91802, "pid": 0, "tid": 7, "ts": 1776649792678133, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792678022, "dur": 5, + "args": { + "External id": 12222, + "cbid": 211, "correlation": 91802 + } + }, + { + "ph": "s", "id": 91802, "pid": 435, "tid": 435, "ts": 1776649792678022, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792678050, "dur": 0, + "args": { + "External id": 12224, + "cbid": 210, "correlation": 91829 + } + }, + { + "ph": "f", "id": 91829, "pid": 435, "tid": 435, "ts": 1776649792678050, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792678179, "dur": 50, + "args": { + "External id": 12224, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91830, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91830, "pid": 0, "tid": 7, "ts": 1776649792678179, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792678051, "dur": 5, + "args": { + "External id": 12224, + "cbid": 211, "correlation": 91830 + } + }, + { + "ph": "s", "id": 91830, "pid": 435, "tid": 435, "ts": 1776649792678051, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792678080, "dur": 0, + "args": { + "External id": 12226, + "cbid": 210, "correlation": 91857 + } + }, + { + "ph": "f", "id": 91857, "pid": 435, "tid": 435, "ts": 1776649792678080, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792678230, "dur": 51, + "args": { + "External id": 12226, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91858, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91858, "pid": 0, "tid": 7, "ts": 1776649792678230, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792678081, "dur": 6, + "args": { + "External id": 12226, + "cbid": 211, "correlation": 91858 + } + }, + { + "ph": "s", "id": 91858, "pid": 435, "tid": 435, "ts": 1776649792678081, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792678110, "dur": 0, + "args": { + "External id": 12228, + "cbid": 210, "correlation": 91885 + } + }, + { + "ph": "f", "id": 91885, "pid": 435, "tid": 435, "ts": 1776649792678110, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792678283, "dur": 47, + "args": { + "External id": 12228, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91886, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91886, "pid": 0, "tid": 7, "ts": 1776649792678283, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792678111, "dur": 5, + "args": { + "External id": 12228, + "cbid": 211, "correlation": 91886 + } + }, + { + "ph": "s", "id": 91886, "pid": 435, "tid": 435, "ts": 1776649792678111, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792678135, "dur": 198, + "args": { + "External id": 91892, + "cbid": 165, "correlation": 91892 + } + }, + { + "ph": "s", "id": 91892, "pid": 435, "tid": 435, "ts": 1776649792678135, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649792678355, "dur": 3, + "args": { + "External id": 91901, + "cbid": 135, "correlation": 91901 + } + }, + { + "ph": "f", "id": 91901, "pid": 435, "tid": 435, "ts": 1776649792678355, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792678385, "dur": 1, + "args": { + "External id": 12230, + "cbid": 210, "correlation": 91927 + } + }, + { + "ph": "f", "id": 91927, "pid": 435, "tid": 435, "ts": 1776649792678385, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792678396, "dur": 41, + "args": { + "External id": 12230, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91928, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91928, "pid": 0, "tid": 7, "ts": 1776649792678396, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792678386, "dur": 7, + "args": { + "External id": 12230, + "cbid": 211, "correlation": 91928 + } + }, + { + "ph": "s", "id": 91928, "pid": 435, "tid": 435, "ts": 1776649792678386, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792678419, "dur": 0, + "args": { + "External id": 12232, + "cbid": 210, "correlation": 91955 + } + }, + { + "ph": "f", "id": 91955, "pid": 435, "tid": 435, "ts": 1776649792678419, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792678439, "dur": 45, + "args": { + "External id": 12232, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91956, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91956, "pid": 0, "tid": 7, "ts": 1776649792678439, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792678420, "dur": 6, + "args": { + "External id": 12232, + "cbid": 211, "correlation": 91956 + } + }, + { + "ph": "s", "id": 91956, "pid": 435, "tid": 435, "ts": 1776649792678420, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792678449, "dur": 0, + "args": { + "External id": 12234, + "cbid": 210, "correlation": 91983 + } + }, + { + "ph": "f", "id": 91983, "pid": 435, "tid": 435, "ts": 1776649792678449, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792678486, "dur": 43, + "args": { + "External id": 12234, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 91984, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 91984, "pid": 0, "tid": 7, "ts": 1776649792678486, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792678450, "dur": 5, + "args": { + "External id": 12234, + "cbid": 211, "correlation": 91984 + } + }, + { + "ph": "s", "id": 91984, "pid": 435, "tid": 435, "ts": 1776649792678450, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792678479, "dur": 0, + "args": { + "External id": 12236, + "cbid": 210, "correlation": 92011 + } + }, + { + "ph": "f", "id": 92011, "pid": 435, "tid": 435, "ts": 1776649792678479, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792678532, "dur": 51, + "args": { + "External id": 12236, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92012, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92012, "pid": 0, "tid": 7, "ts": 1776649792678532, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792678480, "dur": 5, + "args": { + "External id": 12236, + "cbid": 211, "correlation": 92012 + } + }, + { + "ph": "s", "id": 92012, "pid": 435, "tid": 435, "ts": 1776649792678480, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792678509, "dur": 0, + "args": { + "External id": 12238, + "cbid": 210, "correlation": 92039 + } + }, + { + "ph": "f", "id": 92039, "pid": 435, "tid": 435, "ts": 1776649792678509, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792678584, "dur": 51, + "args": { + "External id": 12238, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92040, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92040, "pid": 0, "tid": 7, "ts": 1776649792678584, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792678510, "dur": 5, + "args": { + "External id": 12238, + "cbid": 211, "correlation": 92040 + } + }, + { + "ph": "s", "id": 92040, "pid": 435, "tid": 435, "ts": 1776649792678510, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792678538, "dur": 0, + "args": { + "External id": 12240, + "cbid": 210, "correlation": 92067 + } + }, + { + "ph": "f", "id": 92067, "pid": 435, "tid": 435, "ts": 1776649792678538, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792678637, "dur": 41, + "args": { + "External id": 12240, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92068, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92068, "pid": 0, "tid": 7, "ts": 1776649792678637, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792678540, "dur": 5, + "args": { + "External id": 12240, + "cbid": 211, "correlation": 92068 + } + }, + { + "ph": "s", "id": 92068, "pid": 435, "tid": 435, "ts": 1776649792678540, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792678568, "dur": 0, + "args": { + "External id": 12242, + "cbid": 210, "correlation": 92095 + } + }, + { + "ph": "f", "id": 92095, "pid": 435, "tid": 435, "ts": 1776649792678568, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792678680, "dur": 46, + "args": { + "External id": 12242, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92096, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92096, "pid": 0, "tid": 7, "ts": 1776649792678680, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792678569, "dur": 5, + "args": { + "External id": 12242, + "cbid": 211, "correlation": 92096 + } + }, + { + "ph": "s", "id": 92096, "pid": 435, "tid": 435, "ts": 1776649792678569, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792678597, "dur": 0, + "args": { + "External id": 12244, + "cbid": 210, "correlation": 92123 + } + }, + { + "ph": "f", "id": 92123, "pid": 435, "tid": 435, "ts": 1776649792678597, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792678728, "dur": 44, + "args": { + "External id": 12244, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92124, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92124, "pid": 0, "tid": 7, "ts": 1776649792678728, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792678599, "dur": 5, + "args": { + "External id": 12244, + "cbid": 211, "correlation": 92124 + } + }, + { + "ph": "s", "id": 92124, "pid": 435, "tid": 435, "ts": 1776649792678599, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792678627, "dur": 0, + "args": { + "External id": 12246, + "cbid": 210, "correlation": 92151 + } + }, + { + "ph": "f", "id": 92151, "pid": 435, "tid": 435, "ts": 1776649792678627, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792678774, "dur": 49, + "args": { + "External id": 12246, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92152, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92152, "pid": 0, "tid": 7, "ts": 1776649792678774, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792678628, "dur": 5, + "args": { + "External id": 12246, + "cbid": 211, "correlation": 92152 + } + }, + { + "ph": "s", "id": 92152, "pid": 435, "tid": 435, "ts": 1776649792678628, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792678656, "dur": 0, + "args": { + "External id": 12248, + "cbid": 210, "correlation": 92179 + } + }, + { + "ph": "f", "id": 92179, "pid": 435, "tid": 435, "ts": 1776649792678656, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792678825, "dur": 46, + "args": { + "External id": 12248, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92180, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92180, "pid": 0, "tid": 7, "ts": 1776649792678825, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792678658, "dur": 5, + "args": { + "External id": 12248, + "cbid": 211, "correlation": 92180 + } + }, + { + "ph": "s", "id": 92180, "pid": 435, "tid": 435, "ts": 1776649792678658, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792678686, "dur": 0, + "args": { + "External id": 12250, + "cbid": 210, "correlation": 92207 + } + }, + { + "ph": "f", "id": 92207, "pid": 435, "tid": 435, "ts": 1776649792678686, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792678873, "dur": 44, + "args": { + "External id": 12250, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92208, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92208, "pid": 0, "tid": 7, "ts": 1776649792678873, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792678687, "dur": 5, + "args": { + "External id": 12250, + "cbid": 211, "correlation": 92208 + } + }, + { + "ph": "s", "id": 92208, "pid": 435, "tid": 435, "ts": 1776649792678687, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792678716, "dur": 0, + "args": { + "External id": 12252, + "cbid": 210, "correlation": 92235 + } + }, + { + "ph": "f", "id": 92235, "pid": 435, "tid": 435, "ts": 1776649792678716, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792678919, "dur": 47, + "args": { + "External id": 12252, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92236, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92236, "pid": 0, "tid": 7, "ts": 1776649792678919, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792678717, "dur": 5, + "args": { + "External id": 12252, + "cbid": 211, "correlation": 92236 + } + }, + { + "ph": "s", "id": 92236, "pid": 435, "tid": 435, "ts": 1776649792678717, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792678745, "dur": 0, + "args": { + "External id": 12254, + "cbid": 210, "correlation": 92263 + } + }, + { + "ph": "f", "id": 92263, "pid": 435, "tid": 435, "ts": 1776649792678745, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792678968, "dur": 52, + "args": { + "External id": 12254, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92264, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92264, "pid": 0, "tid": 7, "ts": 1776649792678968, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792678746, "dur": 5, + "args": { + "External id": 12254, + "cbid": 211, "correlation": 92264 + } + }, + { + "ph": "s", "id": 92264, "pid": 435, "tid": 435, "ts": 1776649792678746, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792678775, "dur": 0, + "args": { + "External id": 12256, + "cbid": 210, "correlation": 92291 + } + }, + { + "ph": "f", "id": 92291, "pid": 435, "tid": 435, "ts": 1776649792678775, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792679021, "dur": 49, + "args": { + "External id": 12256, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92292, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92292, "pid": 0, "tid": 7, "ts": 1776649792679021, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792678776, "dur": 5, + "args": { + "External id": 12256, + "cbid": 211, "correlation": 92292 + } + }, + { + "ph": "s", "id": 92292, "pid": 435, "tid": 435, "ts": 1776649792678776, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792678804, "dur": 0, + "args": { + "External id": 12258, + "cbid": 210, "correlation": 92319 + } + }, + { + "ph": "f", "id": 92319, "pid": 435, "tid": 435, "ts": 1776649792678804, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792679072, "dur": 47, + "args": { + "External id": 12258, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92320, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92320, "pid": 0, "tid": 7, "ts": 1776649792679072, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792678805, "dur": 5, + "args": { + "External id": 12258, + "cbid": 211, "correlation": 92320 + } + }, + { + "ph": "s", "id": 92320, "pid": 435, "tid": 435, "ts": 1776649792678805, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792678834, "dur": 0, + "args": { + "External id": 12260, + "cbid": 210, "correlation": 92347 + } + }, + { + "ph": "f", "id": 92347, "pid": 435, "tid": 435, "ts": 1776649792678834, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792679121, "dur": 44, + "args": { + "External id": 12260, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92348, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92348, "pid": 0, "tid": 7, "ts": 1776649792679121, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792678835, "dur": 6, + "args": { + "External id": 12260, + "cbid": 211, "correlation": 92348 + } + }, + { + "ph": "s", "id": 92348, "pid": 435, "tid": 435, "ts": 1776649792678835, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792678864, "dur": 0, + "args": { + "External id": 12262, + "cbid": 210, "correlation": 92375 + } + }, + { + "ph": "f", "id": 92375, "pid": 435, "tid": 435, "ts": 1776649792678864, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792679167, "dur": 46, + "args": { + "External id": 12262, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92376, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92376, "pid": 0, "tid": 7, "ts": 1776649792679167, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792678865, "dur": 5, + "args": { + "External id": 12262, + "cbid": 211, "correlation": 92376 + } + }, + { + "ph": "s", "id": 92376, "pid": 435, "tid": 435, "ts": 1776649792678865, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792678895, "dur": 0, + "args": { + "External id": 12264, + "cbid": 210, "correlation": 92403 + } + }, + { + "ph": "f", "id": 92403, "pid": 435, "tid": 435, "ts": 1776649792678895, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792679215, "dur": 49, + "args": { + "External id": 12264, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92404, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92404, "pid": 0, "tid": 7, "ts": 1776649792679215, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792678896, "dur": 6, + "args": { + "External id": 12264, + "cbid": 211, "correlation": 92404 + } + }, + { + "ph": "s", "id": 92404, "pid": 435, "tid": 435, "ts": 1776649792678896, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792678925, "dur": 0, + "args": { + "External id": 12266, + "cbid": 210, "correlation": 92431 + } + }, + { + "ph": "f", "id": 92431, "pid": 435, "tid": 435, "ts": 1776649792678925, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792679266, "dur": 50, + "args": { + "External id": 12266, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92432, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92432, "pid": 0, "tid": 7, "ts": 1776649792679266, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792678926, "dur": 6, + "args": { + "External id": 12266, + "cbid": 211, "correlation": 92432 + } + }, + { + "ph": "s", "id": 92432, "pid": 435, "tid": 435, "ts": 1776649792678926, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792678955, "dur": 0, + "args": { + "External id": 12268, + "cbid": 210, "correlation": 92459 + } + }, + { + "ph": "f", "id": 92459, "pid": 435, "tid": 435, "ts": 1776649792678955, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792679318, "dur": 44, + "args": { + "External id": 12268, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92460, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92460, "pid": 0, "tid": 7, "ts": 1776649792679318, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792678956, "dur": 5, + "args": { + "External id": 12268, + "cbid": 211, "correlation": 92460 + } + }, + { + "ph": "s", "id": 92460, "pid": 435, "tid": 435, "ts": 1776649792678956, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792678984, "dur": 0, + "args": { + "External id": 12270, + "cbid": 210, "correlation": 92487 + } + }, + { + "ph": "f", "id": 92487, "pid": 435, "tid": 435, "ts": 1776649792678984, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792679364, "dur": 51, + "args": { + "External id": 12270, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92488, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92488, "pid": 0, "tid": 7, "ts": 1776649792679364, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792678985, "dur": 5, + "args": { + "External id": 12270, + "cbid": 211, "correlation": 92488 + } + }, + { + "ph": "s", "id": 92488, "pid": 435, "tid": 435, "ts": 1776649792678985, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679014, "dur": 0, + "args": { + "External id": 12272, + "cbid": 210, "correlation": 92515 + } + }, + { + "ph": "f", "id": 92515, "pid": 435, "tid": 435, "ts": 1776649792679014, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792679417, "dur": 49, + "args": { + "External id": 12272, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92516, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92516, "pid": 0, "tid": 7, "ts": 1776649792679417, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679015, "dur": 5, + "args": { + "External id": 12272, + "cbid": 211, "correlation": 92516 + } + }, + { + "ph": "s", "id": 92516, "pid": 435, "tid": 435, "ts": 1776649792679015, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679044, "dur": 0, + "args": { + "External id": 12274, + "cbid": 210, "correlation": 92543 + } + }, + { + "ph": "f", "id": 92543, "pid": 435, "tid": 435, "ts": 1776649792679044, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792679468, "dur": 43, + "args": { + "External id": 12274, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92544, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92544, "pid": 0, "tid": 7, "ts": 1776649792679468, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679045, "dur": 5, + "args": { + "External id": 12274, + "cbid": 211, "correlation": 92544 + } + }, + { + "ph": "s", "id": 92544, "pid": 435, "tid": 435, "ts": 1776649792679045, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679073, "dur": 0, + "args": { + "External id": 12276, + "cbid": 210, "correlation": 92571 + } + }, + { + "ph": "f", "id": 92571, "pid": 435, "tid": 435, "ts": 1776649792679073, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792679513, "dur": 49, + "args": { + "External id": 12276, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92572, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92572, "pid": 0, "tid": 7, "ts": 1776649792679513, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679075, "dur": 6, + "args": { + "External id": 12276, + "cbid": 211, "correlation": 92572 + } + }, + { + "ph": "s", "id": 92572, "pid": 435, "tid": 435, "ts": 1776649792679075, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679103, "dur": 0, + "args": { + "External id": 12278, + "cbid": 210, "correlation": 92599 + } + }, + { + "ph": "f", "id": 92599, "pid": 435, "tid": 435, "ts": 1776649792679103, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792679564, "dur": 47, + "args": { + "External id": 12278, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92600, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92600, "pid": 0, "tid": 7, "ts": 1776649792679564, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679105, "dur": 5, + "args": { + "External id": 12278, + "cbid": 211, "correlation": 92600 + } + }, + { + "ph": "s", "id": 92600, "pid": 435, "tid": 435, "ts": 1776649792679105, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679133, "dur": 0, + "args": { + "External id": 12280, + "cbid": 210, "correlation": 92627 + } + }, + { + "ph": "f", "id": 92627, "pid": 435, "tid": 435, "ts": 1776649792679133, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792679612, "dur": 46, + "args": { + "External id": 12280, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92628, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92628, "pid": 0, "tid": 7, "ts": 1776649792679612, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679134, "dur": 5, + "args": { + "External id": 12280, + "cbid": 211, "correlation": 92628 + } + }, + { + "ph": "s", "id": 92628, "pid": 435, "tid": 435, "ts": 1776649792679134, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679163, "dur": 0, + "args": { + "External id": 12282, + "cbid": 210, "correlation": 92655 + } + }, + { + "ph": "f", "id": 92655, "pid": 435, "tid": 435, "ts": 1776649792679163, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792679660, "dur": 47, + "args": { + "External id": 12282, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92656, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92656, "pid": 0, "tid": 7, "ts": 1776649792679660, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679164, "dur": 6, + "args": { + "External id": 12282, + "cbid": 211, "correlation": 92656 + } + }, + { + "ph": "s", "id": 92656, "pid": 435, "tid": 435, "ts": 1776649792679164, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679192, "dur": 0, + "args": { + "External id": 12284, + "cbid": 210, "correlation": 92683 + } + }, + { + "ph": "f", "id": 92683, "pid": 435, "tid": 435, "ts": 1776649792679192, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792679709, "dur": 48, + "args": { + "External id": 12284, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92684, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92684, "pid": 0, "tid": 7, "ts": 1776649792679709, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679194, "dur": 5, + "args": { + "External id": 12284, + "cbid": 211, "correlation": 92684 + } + }, + { + "ph": "s", "id": 92684, "pid": 435, "tid": 435, "ts": 1776649792679194, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679222, "dur": 0, + "args": { + "External id": 12286, + "cbid": 210, "correlation": 92711 + } + }, + { + "ph": "f", "id": 92711, "pid": 435, "tid": 435, "ts": 1776649792679222, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792679759, "dur": 47, + "args": { + "External id": 12286, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92712, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92712, "pid": 0, "tid": 7, "ts": 1776649792679759, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679223, "dur": 5, + "args": { + "External id": 12286, + "cbid": 211, "correlation": 92712 + } + }, + { + "ph": "s", "id": 92712, "pid": 435, "tid": 435, "ts": 1776649792679223, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679252, "dur": 0, + "args": { + "External id": 12288, + "cbid": 210, "correlation": 92739 + } + }, + { + "ph": "f", "id": 92739, "pid": 435, "tid": 435, "ts": 1776649792679252, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792679808, "dur": 47, + "args": { + "External id": 12288, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92740, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92740, "pid": 0, "tid": 7, "ts": 1776649792679808, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679253, "dur": 5, + "args": { + "External id": 12288, + "cbid": 211, "correlation": 92740 + } + }, + { + "ph": "s", "id": 92740, "pid": 435, "tid": 435, "ts": 1776649792679253, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679292, "dur": 0, + "args": { + "External id": 12290, + "cbid": 210, "correlation": 92767 + } + }, + { + "ph": "f", "id": 92767, "pid": 435, "tid": 435, "ts": 1776649792679292, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792679857, "dur": 47, + "args": { + "External id": 12290, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92768, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92768, "pid": 0, "tid": 7, "ts": 1776649792679857, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679293, "dur": 6, + "args": { + "External id": 12290, + "cbid": 211, "correlation": 92768 + } + }, + { + "ph": "s", "id": 92768, "pid": 435, "tid": 435, "ts": 1776649792679293, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679322, "dur": 0, + "args": { + "External id": 12292, + "cbid": 210, "correlation": 92795 + } + }, + { + "ph": "f", "id": 92795, "pid": 435, "tid": 435, "ts": 1776649792679322, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792679906, "dur": 45, + "args": { + "External id": 12292, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92796, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92796, "pid": 0, "tid": 7, "ts": 1776649792679906, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679324, "dur": 6, + "args": { + "External id": 12292, + "cbid": 211, "correlation": 92796 + } + }, + { + "ph": "s", "id": 92796, "pid": 435, "tid": 435, "ts": 1776649792679324, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679353, "dur": 0, + "args": { + "External id": 12294, + "cbid": 210, "correlation": 92823 + } + }, + { + "ph": "f", "id": 92823, "pid": 435, "tid": 435, "ts": 1776649792679353, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792679952, "dur": 51, + "args": { + "External id": 12294, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92824, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92824, "pid": 0, "tid": 7, "ts": 1776649792679952, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679354, "dur": 5, + "args": { + "External id": 12294, + "cbid": 211, "correlation": 92824 + } + }, + { + "ph": "s", "id": 92824, "pid": 435, "tid": 435, "ts": 1776649792679354, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679382, "dur": 0, + "args": { + "External id": 12296, + "cbid": 210, "correlation": 92851 + } + }, + { + "ph": "f", "id": 92851, "pid": 435, "tid": 435, "ts": 1776649792679382, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792680005, "dur": 50, + "args": { + "External id": 12296, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92852, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92852, "pid": 0, "tid": 7, "ts": 1776649792680005, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679383, "dur": 6, + "args": { + "External id": 12296, + "cbid": 211, "correlation": 92852 + } + }, + { + "ph": "s", "id": 92852, "pid": 435, "tid": 435, "ts": 1776649792679383, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679415, "dur": 0, + "args": { + "External id": 12298, + "cbid": 210, "correlation": 92879 + } + }, + { + "ph": "f", "id": 92879, "pid": 435, "tid": 435, "ts": 1776649792679415, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792680057, "dur": 44, + "args": { + "External id": 12298, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92880, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92880, "pid": 0, "tid": 7, "ts": 1776649792680057, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679417, "dur": 6, + "args": { + "External id": 12298, + "cbid": 211, "correlation": 92880 + } + }, + { + "ph": "s", "id": 92880, "pid": 435, "tid": 435, "ts": 1776649792679417, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679446, "dur": 0, + "args": { + "External id": 12300, + "cbid": 210, "correlation": 92907 + } + }, + { + "ph": "f", "id": 92907, "pid": 435, "tid": 435, "ts": 1776649792679446, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792680103, "dur": 50, + "args": { + "External id": 12300, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92908, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92908, "pid": 0, "tid": 7, "ts": 1776649792680103, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679447, "dur": 5, + "args": { + "External id": 12300, + "cbid": 211, "correlation": 92908 + } + }, + { + "ph": "s", "id": 92908, "pid": 435, "tid": 435, "ts": 1776649792679447, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679476, "dur": 0, + "args": { + "External id": 12302, + "cbid": 210, "correlation": 92935 + } + }, + { + "ph": "f", "id": 92935, "pid": 435, "tid": 435, "ts": 1776649792679476, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792680155, "dur": 47, + "args": { + "External id": 12302, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92936, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92936, "pid": 0, "tid": 7, "ts": 1776649792680155, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679477, "dur": 5, + "args": { + "External id": 12302, + "cbid": 211, "correlation": 92936 + } + }, + { + "ph": "s", "id": 92936, "pid": 435, "tid": 435, "ts": 1776649792679477, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679505, "dur": 0, + "args": { + "External id": 12304, + "cbid": 210, "correlation": 92963 + } + }, + { + "ph": "f", "id": 92963, "pid": 435, "tid": 435, "ts": 1776649792679505, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792680203, "dur": 47, + "args": { + "External id": 12304, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92964, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92964, "pid": 0, "tid": 7, "ts": 1776649792680203, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679506, "dur": 5, + "args": { + "External id": 12304, + "cbid": 211, "correlation": 92964 + } + }, + { + "ph": "s", "id": 92964, "pid": 435, "tid": 435, "ts": 1776649792679506, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679535, "dur": 0, + "args": { + "External id": 12306, + "cbid": 210, "correlation": 92991 + } + }, + { + "ph": "f", "id": 92991, "pid": 435, "tid": 435, "ts": 1776649792679535, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792680251, "dur": 46, + "args": { + "External id": 12306, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 92992, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 92992, "pid": 0, "tid": 7, "ts": 1776649792680251, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679536, "dur": 5, + "args": { + "External id": 12306, + "cbid": 211, "correlation": 92992 + } + }, + { + "ph": "s", "id": 92992, "pid": 435, "tid": 435, "ts": 1776649792679536, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679564, "dur": 0, + "args": { + "External id": 12308, + "cbid": 210, "correlation": 93019 + } + }, + { + "ph": "f", "id": 93019, "pid": 435, "tid": 435, "ts": 1776649792679564, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792680300, "dur": 46, + "args": { + "External id": 12308, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93020, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93020, "pid": 0, "tid": 7, "ts": 1776649792680300, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679566, "dur": 5, + "args": { + "External id": 12308, + "cbid": 211, "correlation": 93020 + } + }, + { + "ph": "s", "id": 93020, "pid": 435, "tid": 435, "ts": 1776649792679566, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679594, "dur": 0, + "args": { + "External id": 12310, + "cbid": 210, "correlation": 93047 + } + }, + { + "ph": "f", "id": 93047, "pid": 435, "tid": 435, "ts": 1776649792679594, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792680348, "dur": 47, + "args": { + "External id": 12310, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93048, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93048, "pid": 0, "tid": 7, "ts": 1776649792680348, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679595, "dur": 5, + "args": { + "External id": 12310, + "cbid": 211, "correlation": 93048 + } + }, + { + "ph": "s", "id": 93048, "pid": 435, "tid": 435, "ts": 1776649792679595, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679624, "dur": 0, + "args": { + "External id": 12312, + "cbid": 210, "correlation": 93075 + } + }, + { + "ph": "f", "id": 93075, "pid": 435, "tid": 435, "ts": 1776649792679624, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792680396, "dur": 50, + "args": { + "External id": 12312, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93076, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93076, "pid": 0, "tid": 7, "ts": 1776649792680396, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679625, "dur": 5, + "args": { + "External id": 12312, + "cbid": 211, "correlation": 93076 + } + }, + { + "ph": "s", "id": 93076, "pid": 435, "tid": 435, "ts": 1776649792679625, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679653, "dur": 0, + "args": { + "External id": 12314, + "cbid": 210, "correlation": 93103 + } + }, + { + "ph": "f", "id": 93103, "pid": 435, "tid": 435, "ts": 1776649792679653, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792680447, "dur": 47, + "args": { + "External id": 12314, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93104, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93104, "pid": 0, "tid": 7, "ts": 1776649792680447, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679654, "dur": 5, + "args": { + "External id": 12314, + "cbid": 211, "correlation": 93104 + } + }, + { + "ph": "s", "id": 93104, "pid": 435, "tid": 435, "ts": 1776649792679654, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679683, "dur": 0, + "args": { + "External id": 12316, + "cbid": 210, "correlation": 93131 + } + }, + { + "ph": "f", "id": 93131, "pid": 435, "tid": 435, "ts": 1776649792679683, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792680497, "dur": 44, + "args": { + "External id": 12316, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93132, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93132, "pid": 0, "tid": 7, "ts": 1776649792680497, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679684, "dur": 5, + "args": { + "External id": 12316, + "cbid": 211, "correlation": 93132 + } + }, + { + "ph": "s", "id": 93132, "pid": 435, "tid": 435, "ts": 1776649792679684, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679713, "dur": 0, + "args": { + "External id": 12318, + "cbid": 210, "correlation": 93159 + } + }, + { + "ph": "f", "id": 93159, "pid": 435, "tid": 435, "ts": 1776649792679713, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792680542, "dur": 47, + "args": { + "External id": 12318, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93160, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93160, "pid": 0, "tid": 7, "ts": 1776649792680542, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679714, "dur": 5, + "args": { + "External id": 12318, + "cbid": 211, "correlation": 93160 + } + }, + { + "ph": "s", "id": 93160, "pid": 435, "tid": 435, "ts": 1776649792679714, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679742, "dur": 0, + "args": { + "External id": 12320, + "cbid": 210, "correlation": 93187 + } + }, + { + "ph": "f", "id": 93187, "pid": 435, "tid": 435, "ts": 1776649792679742, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792680591, "dur": 46, + "args": { + "External id": 12320, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93188, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93188, "pid": 0, "tid": 7, "ts": 1776649792680591, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679743, "dur": 5, + "args": { + "External id": 12320, + "cbid": 211, "correlation": 93188 + } + }, + { + "ph": "s", "id": 93188, "pid": 435, "tid": 435, "ts": 1776649792679743, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679772, "dur": 0, + "args": { + "External id": 12322, + "cbid": 210, "correlation": 93215 + } + }, + { + "ph": "f", "id": 93215, "pid": 435, "tid": 435, "ts": 1776649792679772, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792680639, "dur": 49, + "args": { + "External id": 12322, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93216, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93216, "pid": 0, "tid": 7, "ts": 1776649792680639, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679773, "dur": 5, + "args": { + "External id": 12322, + "cbid": 211, "correlation": 93216 + } + }, + { + "ph": "s", "id": 93216, "pid": 435, "tid": 435, "ts": 1776649792679773, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679801, "dur": 0, + "args": { + "External id": 12324, + "cbid": 210, "correlation": 93243 + } + }, + { + "ph": "f", "id": 93243, "pid": 435, "tid": 435, "ts": 1776649792679801, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792680689, "dur": 49, + "args": { + "External id": 12324, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93244, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93244, "pid": 0, "tid": 7, "ts": 1776649792680689, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679803, "dur": 5, + "args": { + "External id": 12324, + "cbid": 211, "correlation": 93244 + } + }, + { + "ph": "s", "id": 93244, "pid": 435, "tid": 435, "ts": 1776649792679803, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679831, "dur": 0, + "args": { + "External id": 12326, + "cbid": 210, "correlation": 93271 + } + }, + { + "ph": "f", "id": 93271, "pid": 435, "tid": 435, "ts": 1776649792679831, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792680740, "dur": 49, + "args": { + "External id": 12326, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93272, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93272, "pid": 0, "tid": 7, "ts": 1776649792680740, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679832, "dur": 5, + "args": { + "External id": 12326, + "cbid": 211, "correlation": 93272 + } + }, + { + "ph": "s", "id": 93272, "pid": 435, "tid": 435, "ts": 1776649792679832, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792679861, "dur": 0, + "args": { + "External id": 12328, + "cbid": 210, "correlation": 93299 + } + }, + { + "ph": "f", "id": 93299, "pid": 435, "tid": 435, "ts": 1776649792679861, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792680791, "dur": 49, + "args": { + "External id": 12328, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93300, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93300, "pid": 0, "tid": 7, "ts": 1776649792680791, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792679862, "dur": 5, + "args": { + "External id": 12328, + "cbid": 211, "correlation": 93300 + } + }, + { + "ph": "s", "id": 93300, "pid": 435, "tid": 435, "ts": 1776649792679862, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649792679894, "dur": 4, + "args": { + "External id": 93311, + "cbid": 135, "correlation": 93311 + } + }, + { + "ph": "f", "id": 93311, "pid": 435, "tid": 435, "ts": 1776649792679894, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792679908, "dur": 942, + "args": { + "External id": 93316, + "cbid": 165, "correlation": 93316 + } + }, + { + "ph": "s", "id": 93316, "pid": 435, "tid": 435, "ts": 1776649792679908, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649792680855, "dur": 4, + "args": { + "External id": 93319, + "cbid": 139, "correlation": 93319 + } + }, + { + "ph": "f", "id": 93319, "pid": 435, "tid": 435, "ts": 1776649792680855, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792680894, "dur": 1, + "args": { + "External id": 12330, + "cbid": 210, "correlation": 93353 + } + }, + { + "ph": "f", "id": 93353, "pid": 435, "tid": 435, "ts": 1776649792680894, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792680908, "dur": 40, + "args": { + "External id": 12330, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93354, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93354, "pid": 0, "tid": 7, "ts": 1776649792680908, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792680896, "dur": 9, + "args": { + "External id": 12330, + "cbid": 211, "correlation": 93354 + } + }, + { + "ph": "s", "id": 93354, "pid": 435, "tid": 435, "ts": 1776649792680896, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792680932, "dur": 0, + "args": { + "External id": 12332, + "cbid": 210, "correlation": 93381 + } + }, + { + "ph": "f", "id": 93381, "pid": 435, "tid": 435, "ts": 1776649792680932, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792680950, "dur": 43, + "args": { + "External id": 12332, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93382, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93382, "pid": 0, "tid": 7, "ts": 1776649792680950, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792680933, "dur": 6, + "args": { + "External id": 12332, + "cbid": 211, "correlation": 93382 + } + }, + { + "ph": "s", "id": 93382, "pid": 435, "tid": 435, "ts": 1776649792680933, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792680965, "dur": 0, + "args": { + "External id": 12334, + "cbid": 210, "correlation": 93409 + } + }, + { + "ph": "f", "id": 93409, "pid": 435, "tid": 435, "ts": 1776649792680965, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792680995, "dur": 52, + "args": { + "External id": 12334, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93410, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93410, "pid": 0, "tid": 7, "ts": 1776649792680995, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792680967, "dur": 6, + "args": { + "External id": 12334, + "cbid": 211, "correlation": 93410 + } + }, + { + "ph": "s", "id": 93410, "pid": 435, "tid": 435, "ts": 1776649792680967, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792680996, "dur": 0, + "args": { + "External id": 12336, + "cbid": 210, "correlation": 93437 + } + }, + { + "ph": "f", "id": 93437, "pid": 435, "tid": 435, "ts": 1776649792680996, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792681049, "dur": 43, + "args": { + "External id": 12336, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93438, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93438, "pid": 0, "tid": 7, "ts": 1776649792681049, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792680998, "dur": 6, + "args": { + "External id": 12336, + "cbid": 211, "correlation": 93438 + } + }, + { + "ph": "s", "id": 93438, "pid": 435, "tid": 435, "ts": 1776649792680998, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792681026, "dur": 0, + "args": { + "External id": 12338, + "cbid": 210, "correlation": 93465 + } + }, + { + "ph": "f", "id": 93465, "pid": 435, "tid": 435, "ts": 1776649792681026, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792681095, "dur": 47, + "args": { + "External id": 12338, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93466, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93466, "pid": 0, "tid": 7, "ts": 1776649792681095, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792681028, "dur": 6, + "args": { + "External id": 12338, + "cbid": 211, "correlation": 93466 + } + }, + { + "ph": "s", "id": 93466, "pid": 435, "tid": 435, "ts": 1776649792681028, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792681056, "dur": 0, + "args": { + "External id": 12340, + "cbid": 210, "correlation": 93493 + } + }, + { + "ph": "f", "id": 93493, "pid": 435, "tid": 435, "ts": 1776649792681056, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792681143, "dur": 44, + "args": { + "External id": 12340, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93494, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93494, "pid": 0, "tid": 7, "ts": 1776649792681143, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792681058, "dur": 6, + "args": { + "External id": 12340, + "cbid": 211, "correlation": 93494 + } + }, + { + "ph": "s", "id": 93494, "pid": 435, "tid": 435, "ts": 1776649792681058, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792681086, "dur": 0, + "args": { + "External id": 12342, + "cbid": 210, "correlation": 93521 + } + }, + { + "ph": "f", "id": 93521, "pid": 435, "tid": 435, "ts": 1776649792681086, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792681189, "dur": 46, + "args": { + "External id": 12342, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93522, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93522, "pid": 0, "tid": 7, "ts": 1776649792681189, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792681087, "dur": 5, + "args": { + "External id": 12342, + "cbid": 211, "correlation": 93522 + } + }, + { + "ph": "s", "id": 93522, "pid": 435, "tid": 435, "ts": 1776649792681087, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792681116, "dur": 0, + "args": { + "External id": 12344, + "cbid": 210, "correlation": 93549 + } + }, + { + "ph": "f", "id": 93549, "pid": 435, "tid": 435, "ts": 1776649792681116, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792681237, "dur": 50, + "args": { + "External id": 12344, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93550, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93550, "pid": 0, "tid": 7, "ts": 1776649792681237, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792681117, "dur": 5, + "args": { + "External id": 12344, + "cbid": 211, "correlation": 93550 + } + }, + { + "ph": "s", "id": 93550, "pid": 435, "tid": 435, "ts": 1776649792681117, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792681145, "dur": 0, + "args": { + "External id": 12346, + "cbid": 210, "correlation": 93577 + } + }, + { + "ph": "f", "id": 93577, "pid": 435, "tid": 435, "ts": 1776649792681145, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792681288, "dur": 46, + "args": { + "External id": 12346, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93578, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93578, "pid": 0, "tid": 7, "ts": 1776649792681288, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792681147, "dur": 6, + "args": { + "External id": 12346, + "cbid": 211, "correlation": 93578 + } + }, + { + "ph": "s", "id": 93578, "pid": 435, "tid": 435, "ts": 1776649792681147, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792681175, "dur": 0, + "args": { + "External id": 12348, + "cbid": 210, "correlation": 93605 + } + }, + { + "ph": "f", "id": 93605, "pid": 435, "tid": 435, "ts": 1776649792681175, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792681336, "dur": 47, + "args": { + "External id": 12348, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93606, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93606, "pid": 0, "tid": 7, "ts": 1776649792681336, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792681176, "dur": 6, + "args": { + "External id": 12348, + "cbid": 211, "correlation": 93606 + } + }, + { + "ph": "s", "id": 93606, "pid": 435, "tid": 435, "ts": 1776649792681176, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792681202, "dur": 184, + "args": { + "External id": 93612, + "cbid": 165, "correlation": 93612 + } + }, + { + "ph": "s", "id": 93612, "pid": 435, "tid": 435, "ts": 1776649792681202, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649792681408, "dur": 3, + "args": { + "External id": 93621, + "cbid": 135, "correlation": 93621 + } + }, + { + "ph": "f", "id": 93621, "pid": 435, "tid": 435, "ts": 1776649792681408, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792681438, "dur": 1, + "args": { + "External id": 12350, + "cbid": 210, "correlation": 93647 + } + }, + { + "ph": "f", "id": 93647, "pid": 435, "tid": 435, "ts": 1776649792681438, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792681450, "dur": 40, + "args": { + "External id": 12350, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93648, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93648, "pid": 0, "tid": 7, "ts": 1776649792681450, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792681440, "dur": 8, + "args": { + "External id": 12350, + "cbid": 211, "correlation": 93648 + } + }, + { + "ph": "s", "id": 93648, "pid": 435, "tid": 435, "ts": 1776649792681440, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792681473, "dur": 0, + "args": { + "External id": 12352, + "cbid": 210, "correlation": 93675 + } + }, + { + "ph": "f", "id": 93675, "pid": 435, "tid": 435, "ts": 1776649792681473, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792681493, "dur": 43, + "args": { + "External id": 12352, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93676, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93676, "pid": 0, "tid": 7, "ts": 1776649792681493, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792681474, "dur": 6, + "args": { + "External id": 12352, + "cbid": 211, "correlation": 93676 + } + }, + { + "ph": "s", "id": 93676, "pid": 435, "tid": 435, "ts": 1776649792681474, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792681503, "dur": 0, + "args": { + "External id": 12354, + "cbid": 210, "correlation": 93703 + } + }, + { + "ph": "f", "id": 93703, "pid": 435, "tid": 435, "ts": 1776649792681503, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792681538, "dur": 47, + "args": { + "External id": 12354, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93704, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93704, "pid": 0, "tid": 7, "ts": 1776649792681538, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792681505, "dur": 6, + "args": { + "External id": 12354, + "cbid": 211, "correlation": 93704 + } + }, + { + "ph": "s", "id": 93704, "pid": 435, "tid": 435, "ts": 1776649792681505, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792681534, "dur": 0, + "args": { + "External id": 12356, + "cbid": 210, "correlation": 93731 + } + }, + { + "ph": "f", "id": 93731, "pid": 435, "tid": 435, "ts": 1776649792681534, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792681586, "dur": 51, + "args": { + "External id": 12356, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93732, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93732, "pid": 0, "tid": 7, "ts": 1776649792681586, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792681535, "dur": 6, + "args": { + "External id": 12356, + "cbid": 211, "correlation": 93732 + } + }, + { + "ph": "s", "id": 93732, "pid": 435, "tid": 435, "ts": 1776649792681535, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792681564, "dur": 0, + "args": { + "External id": 12358, + "cbid": 210, "correlation": 93759 + } + }, + { + "ph": "f", "id": 93759, "pid": 435, "tid": 435, "ts": 1776649792681564, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792681639, "dur": 46, + "args": { + "External id": 12358, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93760, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93760, "pid": 0, "tid": 7, "ts": 1776649792681639, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792681565, "dur": 6, + "args": { + "External id": 12358, + "cbid": 211, "correlation": 93760 + } + }, + { + "ph": "s", "id": 93760, "pid": 435, "tid": 435, "ts": 1776649792681565, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792681594, "dur": 0, + "args": { + "External id": 12360, + "cbid": 210, "correlation": 93787 + } + }, + { + "ph": "f", "id": 93787, "pid": 435, "tid": 435, "ts": 1776649792681594, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792681687, "dur": 44, + "args": { + "External id": 12360, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93788, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93788, "pid": 0, "tid": 7, "ts": 1776649792681687, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792681595, "dur": 6, + "args": { + "External id": 12360, + "cbid": 211, "correlation": 93788 + } + }, + { + "ph": "s", "id": 93788, "pid": 435, "tid": 435, "ts": 1776649792681595, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792681624, "dur": 0, + "args": { + "External id": 12362, + "cbid": 210, "correlation": 93815 + } + }, + { + "ph": "f", "id": 93815, "pid": 435, "tid": 435, "ts": 1776649792681624, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792681733, "dur": 51, + "args": { + "External id": 12362, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93816, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93816, "pid": 0, "tid": 7, "ts": 1776649792681733, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792681625, "dur": 5, + "args": { + "External id": 12362, + "cbid": 211, "correlation": 93816 + } + }, + { + "ph": "s", "id": 93816, "pid": 435, "tid": 435, "ts": 1776649792681625, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792681653, "dur": 0, + "args": { + "External id": 12364, + "cbid": 210, "correlation": 93843 + } + }, + { + "ph": "f", "id": 93843, "pid": 435, "tid": 435, "ts": 1776649792681653, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792681785, "dur": 49, + "args": { + "External id": 12364, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93844, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93844, "pid": 0, "tid": 7, "ts": 1776649792681785, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792681655, "dur": 6, + "args": { + "External id": 12364, + "cbid": 211, "correlation": 93844 + } + }, + { + "ph": "s", "id": 93844, "pid": 435, "tid": 435, "ts": 1776649792681655, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792681683, "dur": 0, + "args": { + "External id": 12366, + "cbid": 210, "correlation": 93871 + } + }, + { + "ph": "f", "id": 93871, "pid": 435, "tid": 435, "ts": 1776649792681683, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792681837, "dur": 43, + "args": { + "External id": 12366, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93872, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93872, "pid": 0, "tid": 7, "ts": 1776649792681837, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792681684, "dur": 6, + "args": { + "External id": 12366, + "cbid": 211, "correlation": 93872 + } + }, + { + "ph": "s", "id": 93872, "pid": 435, "tid": 435, "ts": 1776649792681684, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792681713, "dur": 0, + "args": { + "External id": 12368, + "cbid": 210, "correlation": 93899 + } + }, + { + "ph": "f", "id": 93899, "pid": 435, "tid": 435, "ts": 1776649792681713, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792681882, "dur": 46, + "args": { + "External id": 12368, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93900, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93900, "pid": 0, "tid": 7, "ts": 1776649792681882, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792681715, "dur": 5, + "args": { + "External id": 12368, + "cbid": 211, "correlation": 93900 + } + }, + { + "ph": "s", "id": 93900, "pid": 435, "tid": 435, "ts": 1776649792681715, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792681743, "dur": 0, + "args": { + "External id": 12370, + "cbid": 210, "correlation": 93927 + } + }, + { + "ph": "f", "id": 93927, "pid": 435, "tid": 435, "ts": 1776649792681743, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792681930, "dur": 47, + "args": { + "External id": 12370, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93928, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93928, "pid": 0, "tid": 7, "ts": 1776649792681930, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792681744, "dur": 5, + "args": { + "External id": 12370, + "cbid": 211, "correlation": 93928 + } + }, + { + "ph": "s", "id": 93928, "pid": 435, "tid": 435, "ts": 1776649792681744, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792681772, "dur": 0, + "args": { + "External id": 12372, + "cbid": 210, "correlation": 93955 + } + }, + { + "ph": "f", "id": 93955, "pid": 435, "tid": 435, "ts": 1776649792681772, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792681979, "dur": 47, + "args": { + "External id": 12372, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93956, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93956, "pid": 0, "tid": 7, "ts": 1776649792681979, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792681774, "dur": 5, + "args": { + "External id": 12372, + "cbid": 211, "correlation": 93956 + } + }, + { + "ph": "s", "id": 93956, "pid": 435, "tid": 435, "ts": 1776649792681774, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792681802, "dur": 0, + "args": { + "External id": 12374, + "cbid": 210, "correlation": 93983 + } + }, + { + "ph": "f", "id": 93983, "pid": 435, "tid": 435, "ts": 1776649792681802, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792682028, "dur": 44, + "args": { + "External id": 12374, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 93984, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 93984, "pid": 0, "tid": 7, "ts": 1776649792682028, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792681803, "dur": 5, + "args": { + "External id": 12374, + "cbid": 211, "correlation": 93984 + } + }, + { + "ph": "s", "id": 93984, "pid": 435, "tid": 435, "ts": 1776649792681803, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792681832, "dur": 0, + "args": { + "External id": 12376, + "cbid": 210, "correlation": 94011 + } + }, + { + "ph": "f", "id": 94011, "pid": 435, "tid": 435, "ts": 1776649792681832, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792682073, "dur": 46, + "args": { + "External id": 12376, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94012, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94012, "pid": 0, "tid": 7, "ts": 1776649792682073, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792681833, "dur": 6, + "args": { + "External id": 12376, + "cbid": 211, "correlation": 94012 + } + }, + { + "ph": "s", "id": 94012, "pid": 435, "tid": 435, "ts": 1776649792681833, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792681862, "dur": 0, + "args": { + "External id": 12378, + "cbid": 210, "correlation": 94039 + } + }, + { + "ph": "f", "id": 94039, "pid": 435, "tid": 435, "ts": 1776649792681862, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792682122, "dur": 51, + "args": { + "External id": 12378, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94040, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94040, "pid": 0, "tid": 7, "ts": 1776649792682122, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792681863, "dur": 5, + "args": { + "External id": 12378, + "cbid": 211, "correlation": 94040 + } + }, + { + "ph": "s", "id": 94040, "pid": 435, "tid": 435, "ts": 1776649792681863, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792681891, "dur": 0, + "args": { + "External id": 12380, + "cbid": 210, "correlation": 94067 + } + }, + { + "ph": "f", "id": 94067, "pid": 435, "tid": 435, "ts": 1776649792681891, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792682175, "dur": 45, + "args": { + "External id": 12380, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94068, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94068, "pid": 0, "tid": 7, "ts": 1776649792682175, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792681893, "dur": 5, + "args": { + "External id": 12380, + "cbid": 211, "correlation": 94068 + } + }, + { + "ph": "s", "id": 94068, "pid": 435, "tid": 435, "ts": 1776649792681893, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792681921, "dur": 0, + "args": { + "External id": 12382, + "cbid": 210, "correlation": 94095 + } + }, + { + "ph": "f", "id": 94095, "pid": 435, "tid": 435, "ts": 1776649792681921, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792682221, "dur": 49, + "args": { + "External id": 12382, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94096, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94096, "pid": 0, "tid": 7, "ts": 1776649792682221, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792681922, "dur": 5, + "args": { + "External id": 12382, + "cbid": 211, "correlation": 94096 + } + }, + { + "ph": "s", "id": 94096, "pid": 435, "tid": 435, "ts": 1776649792681922, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792681951, "dur": 0, + "args": { + "External id": 12384, + "cbid": 210, "correlation": 94123 + } + }, + { + "ph": "f", "id": 94123, "pid": 435, "tid": 435, "ts": 1776649792681951, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792682272, "dur": 47, + "args": { + "External id": 12384, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94124, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94124, "pid": 0, "tid": 7, "ts": 1776649792682272, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792681952, "dur": 5, + "args": { + "External id": 12384, + "cbid": 211, "correlation": 94124 + } + }, + { + "ph": "s", "id": 94124, "pid": 435, "tid": 435, "ts": 1776649792681952, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792681981, "dur": 0, + "args": { + "External id": 12386, + "cbid": 210, "correlation": 94151 + } + }, + { + "ph": "f", "id": 94151, "pid": 435, "tid": 435, "ts": 1776649792681981, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792682321, "dur": 50, + "args": { + "External id": 12386, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94152, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94152, "pid": 0, "tid": 7, "ts": 1776649792682321, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792681982, "dur": 5, + "args": { + "External id": 12386, + "cbid": 211, "correlation": 94152 + } + }, + { + "ph": "s", "id": 94152, "pid": 435, "tid": 435, "ts": 1776649792681982, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682010, "dur": 0, + "args": { + "External id": 12388, + "cbid": 210, "correlation": 94179 + } + }, + { + "ph": "f", "id": 94179, "pid": 435, "tid": 435, "ts": 1776649792682010, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792682373, "dur": 50, + "args": { + "External id": 12388, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94180, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94180, "pid": 0, "tid": 7, "ts": 1776649792682373, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682011, "dur": 5, + "args": { + "External id": 12388, + "cbid": 211, "correlation": 94180 + } + }, + { + "ph": "s", "id": 94180, "pid": 435, "tid": 435, "ts": 1776649792682011, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682040, "dur": 0, + "args": { + "External id": 12390, + "cbid": 210, "correlation": 94207 + } + }, + { + "ph": "f", "id": 94207, "pid": 435, "tid": 435, "ts": 1776649792682040, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792682424, "dur": 49, + "args": { + "External id": 12390, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94208, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94208, "pid": 0, "tid": 7, "ts": 1776649792682424, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682041, "dur": 5, + "args": { + "External id": 12390, + "cbid": 211, "correlation": 94208 + } + }, + { + "ph": "s", "id": 94208, "pid": 435, "tid": 435, "ts": 1776649792682041, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682070, "dur": 0, + "args": { + "External id": 12392, + "cbid": 210, "correlation": 94235 + } + }, + { + "ph": "f", "id": 94235, "pid": 435, "tid": 435, "ts": 1776649792682070, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792682475, "dur": 43, + "args": { + "External id": 12392, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94236, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94236, "pid": 0, "tid": 7, "ts": 1776649792682475, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682072, "dur": 5, + "args": { + "External id": 12392, + "cbid": 211, "correlation": 94236 + } + }, + { + "ph": "s", "id": 94236, "pid": 435, "tid": 435, "ts": 1776649792682072, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682100, "dur": 0, + "args": { + "External id": 12394, + "cbid": 210, "correlation": 94263 + } + }, + { + "ph": "f", "id": 94263, "pid": 435, "tid": 435, "ts": 1776649792682100, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792682520, "dur": 47, + "args": { + "External id": 12394, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94264, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94264, "pid": 0, "tid": 7, "ts": 1776649792682520, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682101, "dur": 5, + "args": { + "External id": 12394, + "cbid": 211, "correlation": 94264 + } + }, + { + "ph": "s", "id": 94264, "pid": 435, "tid": 435, "ts": 1776649792682101, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682130, "dur": 0, + "args": { + "External id": 12396, + "cbid": 210, "correlation": 94291 + } + }, + { + "ph": "f", "id": 94291, "pid": 435, "tid": 435, "ts": 1776649792682130, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792682569, "dur": 41, + "args": { + "External id": 12396, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94292, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94292, "pid": 0, "tid": 7, "ts": 1776649792682569, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682131, "dur": 5, + "args": { + "External id": 12396, + "cbid": 211, "correlation": 94292 + } + }, + { + "ph": "s", "id": 94292, "pid": 435, "tid": 435, "ts": 1776649792682131, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682160, "dur": 0, + "args": { + "External id": 12398, + "cbid": 210, "correlation": 94319 + } + }, + { + "ph": "f", "id": 94319, "pid": 435, "tid": 435, "ts": 1776649792682160, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792682612, "dur": 49, + "args": { + "External id": 12398, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94320, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94320, "pid": 0, "tid": 7, "ts": 1776649792682612, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682161, "dur": 5, + "args": { + "External id": 12398, + "cbid": 211, "correlation": 94320 + } + }, + { + "ph": "s", "id": 94320, "pid": 435, "tid": 435, "ts": 1776649792682161, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682189, "dur": 0, + "args": { + "External id": 12400, + "cbid": 210, "correlation": 94347 + } + }, + { + "ph": "f", "id": 94347, "pid": 435, "tid": 435, "ts": 1776649792682189, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792682663, "dur": 47, + "args": { + "External id": 12400, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94348, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94348, "pid": 0, "tid": 7, "ts": 1776649792682663, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682191, "dur": 6, + "args": { + "External id": 12400, + "cbid": 211, "correlation": 94348 + } + }, + { + "ph": "s", "id": 94348, "pid": 435, "tid": 435, "ts": 1776649792682191, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682219, "dur": 0, + "args": { + "External id": 12402, + "cbid": 210, "correlation": 94375 + } + }, + { + "ph": "f", "id": 94375, "pid": 435, "tid": 435, "ts": 1776649792682219, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792682712, "dur": 44, + "args": { + "External id": 12402, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94376, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94376, "pid": 0, "tid": 7, "ts": 1776649792682712, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682220, "dur": 6, + "args": { + "External id": 12402, + "cbid": 211, "correlation": 94376 + } + }, + { + "ph": "s", "id": 94376, "pid": 435, "tid": 435, "ts": 1776649792682220, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682249, "dur": 0, + "args": { + "External id": 12404, + "cbid": 210, "correlation": 94403 + } + }, + { + "ph": "f", "id": 94403, "pid": 435, "tid": 435, "ts": 1776649792682249, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792682758, "dur": 50, + "args": { + "External id": 12404, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94404, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94404, "pid": 0, "tid": 7, "ts": 1776649792682758, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682250, "dur": 5, + "args": { + "External id": 12404, + "cbid": 211, "correlation": 94404 + } + }, + { + "ph": "s", "id": 94404, "pid": 435, "tid": 435, "ts": 1776649792682250, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682279, "dur": 0, + "args": { + "External id": 12406, + "cbid": 210, "correlation": 94431 + } + }, + { + "ph": "f", "id": 94431, "pid": 435, "tid": 435, "ts": 1776649792682279, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792682809, "dur": 46, + "args": { + "External id": 12406, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94432, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94432, "pid": 0, "tid": 7, "ts": 1776649792682809, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682280, "dur": 6, + "args": { + "External id": 12406, + "cbid": 211, "correlation": 94432 + } + }, + { + "ph": "s", "id": 94432, "pid": 435, "tid": 435, "ts": 1776649792682280, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682308, "dur": 0, + "args": { + "External id": 12408, + "cbid": 210, "correlation": 94459 + } + }, + { + "ph": "f", "id": 94459, "pid": 435, "tid": 435, "ts": 1776649792682308, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792682857, "dur": 49, + "args": { + "External id": 12408, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94460, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94460, "pid": 0, "tid": 7, "ts": 1776649792682857, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682310, "dur": 6, + "args": { + "External id": 12408, + "cbid": 211, "correlation": 94460 + } + }, + { + "ph": "s", "id": 94460, "pid": 435, "tid": 435, "ts": 1776649792682310, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682338, "dur": 0, + "args": { + "External id": 12410, + "cbid": 210, "correlation": 94487 + } + }, + { + "ph": "f", "id": 94487, "pid": 435, "tid": 435, "ts": 1776649792682338, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792682908, "dur": 47, + "args": { + "External id": 12410, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94488, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94488, "pid": 0, "tid": 7, "ts": 1776649792682908, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682340, "dur": 6, + "args": { + "External id": 12410, + "cbid": 211, "correlation": 94488 + } + }, + { + "ph": "s", "id": 94488, "pid": 435, "tid": 435, "ts": 1776649792682340, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682368, "dur": 0, + "args": { + "External id": 12412, + "cbid": 210, "correlation": 94515 + } + }, + { + "ph": "f", "id": 94515, "pid": 435, "tid": 435, "ts": 1776649792682368, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792682957, "dur": 47, + "args": { + "External id": 12412, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94516, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94516, "pid": 0, "tid": 7, "ts": 1776649792682957, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682369, "dur": 5, + "args": { + "External id": 12412, + "cbid": 211, "correlation": 94516 + } + }, + { + "ph": "s", "id": 94516, "pid": 435, "tid": 435, "ts": 1776649792682369, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682398, "dur": 0, + "args": { + "External id": 12414, + "cbid": 210, "correlation": 94543 + } + }, + { + "ph": "f", "id": 94543, "pid": 435, "tid": 435, "ts": 1776649792682398, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792683006, "dur": 46, + "args": { + "External id": 12414, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94544, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94544, "pid": 0, "tid": 7, "ts": 1776649792683006, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682399, "dur": 6, + "args": { + "External id": 12414, + "cbid": 211, "correlation": 94544 + } + }, + { + "ph": "s", "id": 94544, "pid": 435, "tid": 435, "ts": 1776649792682399, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682427, "dur": 0, + "args": { + "External id": 12416, + "cbid": 210, "correlation": 94571 + } + }, + { + "ph": "f", "id": 94571, "pid": 435, "tid": 435, "ts": 1776649792682427, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792683054, "dur": 49, + "args": { + "External id": 12416, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94572, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94572, "pid": 0, "tid": 7, "ts": 1776649792683054, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682429, "dur": 6, + "args": { + "External id": 12416, + "cbid": 211, "correlation": 94572 + } + }, + { + "ph": "s", "id": 94572, "pid": 435, "tid": 435, "ts": 1776649792682429, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682458, "dur": 0, + "args": { + "External id": 12418, + "cbid": 210, "correlation": 94599 + } + }, + { + "ph": "f", "id": 94599, "pid": 435, "tid": 435, "ts": 1776649792682458, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792683105, "dur": 44, + "args": { + "External id": 12418, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94600, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94600, "pid": 0, "tid": 7, "ts": 1776649792683105, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682459, "dur": 5, + "args": { + "External id": 12418, + "cbid": 211, "correlation": 94600 + } + }, + { + "ph": "s", "id": 94600, "pid": 435, "tid": 435, "ts": 1776649792682459, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682487, "dur": 0, + "args": { + "External id": 12420, + "cbid": 210, "correlation": 94627 + } + }, + { + "ph": "f", "id": 94627, "pid": 435, "tid": 435, "ts": 1776649792682487, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792683150, "dur": 49, + "args": { + "External id": 12420, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94628, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94628, "pid": 0, "tid": 7, "ts": 1776649792683150, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682488, "dur": 5, + "args": { + "External id": 12420, + "cbid": 211, "correlation": 94628 + } + }, + { + "ph": "s", "id": 94628, "pid": 435, "tid": 435, "ts": 1776649792682488, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682517, "dur": 0, + "args": { + "External id": 12422, + "cbid": 210, "correlation": 94655 + } + }, + { + "ph": "f", "id": 94655, "pid": 435, "tid": 435, "ts": 1776649792682517, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792683201, "dur": 44, + "args": { + "External id": 12422, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94656, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94656, "pid": 0, "tid": 7, "ts": 1776649792683201, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682518, "dur": 6, + "args": { + "External id": 12422, + "cbid": 211, "correlation": 94656 + } + }, + { + "ph": "s", "id": 94656, "pid": 435, "tid": 435, "ts": 1776649792682518, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682550, "dur": 0, + "args": { + "External id": 12424, + "cbid": 210, "correlation": 94683 + } + }, + { + "ph": "f", "id": 94683, "pid": 435, "tid": 435, "ts": 1776649792682550, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792683247, "dur": 43, + "args": { + "External id": 12424, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94684, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94684, "pid": 0, "tid": 7, "ts": 1776649792683247, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682551, "dur": 6, + "args": { + "External id": 12424, + "cbid": 211, "correlation": 94684 + } + }, + { + "ph": "s", "id": 94684, "pid": 435, "tid": 435, "ts": 1776649792682551, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682581, "dur": 0, + "args": { + "External id": 12426, + "cbid": 210, "correlation": 94711 + } + }, + { + "ph": "f", "id": 94711, "pid": 435, "tid": 435, "ts": 1776649792682581, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792683293, "dur": 47, + "args": { + "External id": 12426, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94712, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94712, "pid": 0, "tid": 7, "ts": 1776649792683293, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682583, "dur": 5, + "args": { + "External id": 12426, + "cbid": 211, "correlation": 94712 + } + }, + { + "ph": "s", "id": 94712, "pid": 435, "tid": 435, "ts": 1776649792682583, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682611, "dur": 0, + "args": { + "External id": 12428, + "cbid": 210, "correlation": 94739 + } + }, + { + "ph": "f", "id": 94739, "pid": 435, "tid": 435, "ts": 1776649792682611, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792683341, "dur": 52, + "args": { + "External id": 12428, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94740, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94740, "pid": 0, "tid": 7, "ts": 1776649792683341, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682612, "dur": 5, + "args": { + "External id": 12428, + "cbid": 211, "correlation": 94740 + } + }, + { + "ph": "s", "id": 94740, "pid": 435, "tid": 435, "ts": 1776649792682612, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682641, "dur": 0, + "args": { + "External id": 12430, + "cbid": 210, "correlation": 94767 + } + }, + { + "ph": "f", "id": 94767, "pid": 435, "tid": 435, "ts": 1776649792682641, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792683394, "dur": 46, + "args": { + "External id": 12430, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94768, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94768, "pid": 0, "tid": 7, "ts": 1776649792683394, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682642, "dur": 5, + "args": { + "External id": 12430, + "cbid": 211, "correlation": 94768 + } + }, + { + "ph": "s", "id": 94768, "pid": 435, "tid": 435, "ts": 1776649792682642, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682670, "dur": 0, + "args": { + "External id": 12432, + "cbid": 210, "correlation": 94795 + } + }, + { + "ph": "f", "id": 94795, "pid": 435, "tid": 435, "ts": 1776649792682670, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792683442, "dur": 49, + "args": { + "External id": 12432, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94796, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94796, "pid": 0, "tid": 7, "ts": 1776649792683442, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682672, "dur": 5, + "args": { + "External id": 12432, + "cbid": 211, "correlation": 94796 + } + }, + { + "ph": "s", "id": 94796, "pid": 435, "tid": 435, "ts": 1776649792682672, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682700, "dur": 0, + "args": { + "External id": 12434, + "cbid": 210, "correlation": 94823 + } + }, + { + "ph": "f", "id": 94823, "pid": 435, "tid": 435, "ts": 1776649792682700, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792683493, "dur": 45, + "args": { + "External id": 12434, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94824, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94824, "pid": 0, "tid": 7, "ts": 1776649792683493, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682701, "dur": 5, + "args": { + "External id": 12434, + "cbid": 211, "correlation": 94824 + } + }, + { + "ph": "s", "id": 94824, "pid": 435, "tid": 435, "ts": 1776649792682701, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682730, "dur": 0, + "args": { + "External id": 12436, + "cbid": 210, "correlation": 94851 + } + }, + { + "ph": "f", "id": 94851, "pid": 435, "tid": 435, "ts": 1776649792682730, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792683540, "dur": 50, + "args": { + "External id": 12436, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94852, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94852, "pid": 0, "tid": 7, "ts": 1776649792683540, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682731, "dur": 5, + "args": { + "External id": 12436, + "cbid": 211, "correlation": 94852 + } + }, + { + "ph": "s", "id": 94852, "pid": 435, "tid": 435, "ts": 1776649792682731, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682759, "dur": 0, + "args": { + "External id": 12438, + "cbid": 210, "correlation": 94879 + } + }, + { + "ph": "f", "id": 94879, "pid": 435, "tid": 435, "ts": 1776649792682759, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792683591, "dur": 50, + "args": { + "External id": 12438, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94880, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94880, "pid": 0, "tid": 7, "ts": 1776649792683591, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682761, "dur": 5, + "args": { + "External id": 12438, + "cbid": 211, "correlation": 94880 + } + }, + { + "ph": "s", "id": 94880, "pid": 435, "tid": 435, "ts": 1776649792682761, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682789, "dur": 0, + "args": { + "External id": 12440, + "cbid": 210, "correlation": 94907 + } + }, + { + "ph": "f", "id": 94907, "pid": 435, "tid": 435, "ts": 1776649792682789, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792683643, "dur": 47, + "args": { + "External id": 12440, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94908, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94908, "pid": 0, "tid": 7, "ts": 1776649792683643, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682790, "dur": 5, + "args": { + "External id": 12440, + "cbid": 211, "correlation": 94908 + } + }, + { + "ph": "s", "id": 94908, "pid": 435, "tid": 435, "ts": 1776649792682790, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682819, "dur": 0, + "args": { + "External id": 12442, + "cbid": 210, "correlation": 94935 + } + }, + { + "ph": "f", "id": 94935, "pid": 435, "tid": 435, "ts": 1776649792682819, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792683692, "dur": 48, + "args": { + "External id": 12442, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94936, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94936, "pid": 0, "tid": 7, "ts": 1776649792683692, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682820, "dur": 5, + "args": { + "External id": 12442, + "cbid": 211, "correlation": 94936 + } + }, + { + "ph": "s", "id": 94936, "pid": 435, "tid": 435, "ts": 1776649792682820, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682848, "dur": 0, + "args": { + "External id": 12444, + "cbid": 210, "correlation": 94963 + } + }, + { + "ph": "f", "id": 94963, "pid": 435, "tid": 435, "ts": 1776649792682848, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792683741, "dur": 46, + "args": { + "External id": 12444, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94964, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94964, "pid": 0, "tid": 7, "ts": 1776649792683741, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682850, "dur": 6, + "args": { + "External id": 12444, + "cbid": 211, "correlation": 94964 + } + }, + { + "ph": "s", "id": 94964, "pid": 435, "tid": 435, "ts": 1776649792682850, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682879, "dur": 0, + "args": { + "External id": 12446, + "cbid": 210, "correlation": 94991 + } + }, + { + "ph": "f", "id": 94991, "pid": 435, "tid": 435, "ts": 1776649792682879, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792683789, "dur": 47, + "args": { + "External id": 12446, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 94992, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 94992, "pid": 0, "tid": 7, "ts": 1776649792683789, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682880, "dur": 5, + "args": { + "External id": 12446, + "cbid": 211, "correlation": 94992 + } + }, + { + "ph": "s", "id": 94992, "pid": 435, "tid": 435, "ts": 1776649792682880, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792682908, "dur": 0, + "args": { + "External id": 12448, + "cbid": 210, "correlation": 95019 + } + }, + { + "ph": "f", "id": 95019, "pid": 435, "tid": 435, "ts": 1776649792682908, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792683838, "dur": 47, + "args": { + "External id": 12448, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95020, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95020, "pid": 0, "tid": 7, "ts": 1776649792683838, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792682909, "dur": 5, + "args": { + "External id": 12448, + "cbid": 211, "correlation": 95020 + } + }, + { + "ph": "s", "id": 95020, "pid": 435, "tid": 435, "ts": 1776649792682909, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649792682942, "dur": 4, + "args": { + "External id": 95031, + "cbid": 135, "correlation": 95031 + } + }, + { + "ph": "f", "id": 95031, "pid": 435, "tid": 435, "ts": 1776649792682942, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792682956, "dur": 937, + "args": { + "External id": 95036, + "cbid": 165, "correlation": 95036 + } + }, + { + "ph": "s", "id": 95036, "pid": 435, "tid": 435, "ts": 1776649792682956, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649792683897, "dur": 4, + "args": { + "External id": 95039, + "cbid": 139, "correlation": 95039 + } + }, + { + "ph": "f", "id": 95039, "pid": 435, "tid": 435, "ts": 1776649792683897, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792683934, "dur": 1, + "args": { + "External id": 12450, + "cbid": 210, "correlation": 95073 + } + }, + { + "ph": "f", "id": 95073, "pid": 435, "tid": 435, "ts": 1776649792683934, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792683946, "dur": 41, + "args": { + "External id": 12450, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95074, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95074, "pid": 0, "tid": 7, "ts": 1776649792683946, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792683936, "dur": 8, + "args": { + "External id": 12450, + "cbid": 211, "correlation": 95074 + } + }, + { + "ph": "s", "id": 95074, "pid": 435, "tid": 435, "ts": 1776649792683936, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792683969, "dur": 0, + "args": { + "External id": 12452, + "cbid": 210, "correlation": 95101 + } + }, + { + "ph": "f", "id": 95101, "pid": 435, "tid": 435, "ts": 1776649792683969, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792683989, "dur": 44, + "args": { + "External id": 12452, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95102, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95102, "pid": 0, "tid": 7, "ts": 1776649792683989, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792683970, "dur": 6, + "args": { + "External id": 12452, + "cbid": 211, "correlation": 95102 + } + }, + { + "ph": "s", "id": 95102, "pid": 435, "tid": 435, "ts": 1776649792683970, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792684000, "dur": 0, + "args": { + "External id": 12454, + "cbid": 210, "correlation": 95129 + } + }, + { + "ph": "f", "id": 95129, "pid": 435, "tid": 435, "ts": 1776649792684000, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792684034, "dur": 44, + "args": { + "External id": 12454, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95130, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95130, "pid": 0, "tid": 7, "ts": 1776649792684034, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792684001, "dur": 6, + "args": { + "External id": 12454, + "cbid": 211, "correlation": 95130 + } + }, + { + "ph": "s", "id": 95130, "pid": 435, "tid": 435, "ts": 1776649792684001, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792684030, "dur": 0, + "args": { + "External id": 12456, + "cbid": 210, "correlation": 95157 + } + }, + { + "ph": "f", "id": 95157, "pid": 435, "tid": 435, "ts": 1776649792684030, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792684080, "dur": 49, + "args": { + "External id": 12456, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95158, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95158, "pid": 0, "tid": 7, "ts": 1776649792684080, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792684032, "dur": 6, + "args": { + "External id": 12456, + "cbid": 211, "correlation": 95158 + } + }, + { + "ph": "s", "id": 95158, "pid": 435, "tid": 435, "ts": 1776649792684032, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792684061, "dur": 0, + "args": { + "External id": 12458, + "cbid": 210, "correlation": 95185 + } + }, + { + "ph": "f", "id": 95185, "pid": 435, "tid": 435, "ts": 1776649792684061, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792684130, "dur": 50, + "args": { + "External id": 12458, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95186, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95186, "pid": 0, "tid": 7, "ts": 1776649792684130, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792684062, "dur": 6, + "args": { + "External id": 12458, + "cbid": 211, "correlation": 95186 + } + }, + { + "ph": "s", "id": 95186, "pid": 435, "tid": 435, "ts": 1776649792684062, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792684091, "dur": 0, + "args": { + "External id": 12460, + "cbid": 210, "correlation": 95213 + } + }, + { + "ph": "f", "id": 95213, "pid": 435, "tid": 435, "ts": 1776649792684091, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792684182, "dur": 43, + "args": { + "External id": 12460, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95214, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95214, "pid": 0, "tid": 7, "ts": 1776649792684182, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792684092, "dur": 6, + "args": { + "External id": 12460, + "cbid": 211, "correlation": 95214 + } + }, + { + "ph": "s", "id": 95214, "pid": 435, "tid": 435, "ts": 1776649792684092, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792684121, "dur": 0, + "args": { + "External id": 12462, + "cbid": 210, "correlation": 95241 + } + }, + { + "ph": "f", "id": 95241, "pid": 435, "tid": 435, "ts": 1776649792684121, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792684228, "dur": 50, + "args": { + "External id": 12462, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95242, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95242, "pid": 0, "tid": 7, "ts": 1776649792684228, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792684122, "dur": 5, + "args": { + "External id": 12462, + "cbid": 211, "correlation": 95242 + } + }, + { + "ph": "s", "id": 95242, "pid": 435, "tid": 435, "ts": 1776649792684122, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792684151, "dur": 0, + "args": { + "External id": 12464, + "cbid": 210, "correlation": 95269 + } + }, + { + "ph": "f", "id": 95269, "pid": 435, "tid": 435, "ts": 1776649792684151, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792684279, "dur": 50, + "args": { + "External id": 12464, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95270, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95270, "pid": 0, "tid": 7, "ts": 1776649792684279, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792684152, "dur": 6, + "args": { + "External id": 12464, + "cbid": 211, "correlation": 95270 + } + }, + { + "ph": "s", "id": 95270, "pid": 435, "tid": 435, "ts": 1776649792684152, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792684181, "dur": 0, + "args": { + "External id": 12466, + "cbid": 210, "correlation": 95297 + } + }, + { + "ph": "f", "id": 95297, "pid": 435, "tid": 435, "ts": 1776649792684181, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792684331, "dur": 48, + "args": { + "External id": 12466, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95298, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95298, "pid": 0, "tid": 7, "ts": 1776649792684331, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792684182, "dur": 6, + "args": { + "External id": 12466, + "cbid": 211, "correlation": 95298 + } + }, + { + "ph": "s", "id": 95298, "pid": 435, "tid": 435, "ts": 1776649792684182, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792684211, "dur": 0, + "args": { + "External id": 12468, + "cbid": 210, "correlation": 95325 + } + }, + { + "ph": "f", "id": 95325, "pid": 435, "tid": 435, "ts": 1776649792684211, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792684381, "dur": 47, + "args": { + "External id": 12468, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95326, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95326, "pid": 0, "tid": 7, "ts": 1776649792684381, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792684212, "dur": 25, + "args": { + "External id": 12468, + "cbid": 211, "correlation": 95326 + } + }, + { + "ph": "s", "id": 95326, "pid": 435, "tid": 435, "ts": 1776649792684212, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792684259, "dur": 173, + "args": { + "External id": 95332, + "cbid": 165, "correlation": 95332 + } + }, + { + "ph": "s", "id": 95332, "pid": 435, "tid": 435, "ts": 1776649792684259, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649792684453, "dur": 3, + "args": { + "External id": 95341, + "cbid": 135, "correlation": 95341 + } + }, + { + "ph": "f", "id": 95341, "pid": 435, "tid": 435, "ts": 1776649792684453, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792684485, "dur": 1, + "args": { + "External id": 12470, + "cbid": 210, "correlation": 95367 + } + }, + { + "ph": "f", "id": 95367, "pid": 435, "tid": 435, "ts": 1776649792684485, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792684497, "dur": 41, + "args": { + "External id": 12470, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95368, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95368, "pid": 0, "tid": 7, "ts": 1776649792684497, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792684487, "dur": 8, + "args": { + "External id": 12470, + "cbid": 211, "correlation": 95368 + } + }, + { + "ph": "s", "id": 95368, "pid": 435, "tid": 435, "ts": 1776649792684487, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792684520, "dur": 0, + "args": { + "External id": 12472, + "cbid": 210, "correlation": 95395 + } + }, + { + "ph": "f", "id": 95395, "pid": 435, "tid": 435, "ts": 1776649792684520, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792684540, "dur": 44, + "args": { + "External id": 12472, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95396, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95396, "pid": 0, "tid": 7, "ts": 1776649792684540, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792684522, "dur": 6, + "args": { + "External id": 12472, + "cbid": 211, "correlation": 95396 + } + }, + { + "ph": "s", "id": 95396, "pid": 435, "tid": 435, "ts": 1776649792684522, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792684551, "dur": 0, + "args": { + "External id": 12474, + "cbid": 210, "correlation": 95423 + } + }, + { + "ph": "f", "id": 95423, "pid": 435, "tid": 435, "ts": 1776649792684551, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792684586, "dur": 49, + "args": { + "External id": 12474, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95424, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95424, "pid": 0, "tid": 7, "ts": 1776649792684586, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792684553, "dur": 6, + "args": { + "External id": 12474, + "cbid": 211, "correlation": 95424 + } + }, + { + "ph": "s", "id": 95424, "pid": 435, "tid": 435, "ts": 1776649792684553, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792684582, "dur": 0, + "args": { + "External id": 12476, + "cbid": 210, "correlation": 95451 + } + }, + { + "ph": "f", "id": 95451, "pid": 435, "tid": 435, "ts": 1776649792684582, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792684637, "dur": 49, + "args": { + "External id": 12476, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95452, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95452, "pid": 0, "tid": 7, "ts": 1776649792684637, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792684583, "dur": 5, + "args": { + "External id": 12476, + "cbid": 211, "correlation": 95452 + } + }, + { + "ph": "s", "id": 95452, "pid": 435, "tid": 435, "ts": 1776649792684583, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792684612, "dur": 0, + "args": { + "External id": 12478, + "cbid": 210, "correlation": 95479 + } + }, + { + "ph": "f", "id": 95479, "pid": 435, "tid": 435, "ts": 1776649792684612, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792684688, "dur": 47, + "args": { + "External id": 12478, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95480, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95480, "pid": 0, "tid": 7, "ts": 1776649792684688, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792684613, "dur": 5, + "args": { + "External id": 12478, + "cbid": 211, "correlation": 95480 + } + }, + { + "ph": "s", "id": 95480, "pid": 435, "tid": 435, "ts": 1776649792684613, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792684642, "dur": 0, + "args": { + "External id": 12480, + "cbid": 210, "correlation": 95507 + } + }, + { + "ph": "f", "id": 95507, "pid": 435, "tid": 435, "ts": 1776649792684642, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792684737, "dur": 45, + "args": { + "External id": 12480, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95508, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95508, "pid": 0, "tid": 7, "ts": 1776649792684737, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792684643, "dur": 5, + "args": { + "External id": 12480, + "cbid": 211, "correlation": 95508 + } + }, + { + "ph": "s", "id": 95508, "pid": 435, "tid": 435, "ts": 1776649792684643, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792684672, "dur": 0, + "args": { + "External id": 12482, + "cbid": 210, "correlation": 95535 + } + }, + { + "ph": "f", "id": 95535, "pid": 435, "tid": 435, "ts": 1776649792684672, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792684784, "dur": 49, + "args": { + "External id": 12482, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95536, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95536, "pid": 0, "tid": 7, "ts": 1776649792684784, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792684673, "dur": 6, + "args": { + "External id": 12482, + "cbid": 211, "correlation": 95536 + } + }, + { + "ph": "s", "id": 95536, "pid": 435, "tid": 435, "ts": 1776649792684673, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792684702, "dur": 0, + "args": { + "External id": 12484, + "cbid": 210, "correlation": 95563 + } + }, + { + "ph": "f", "id": 95563, "pid": 435, "tid": 435, "ts": 1776649792684702, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792684835, "dur": 47, + "args": { + "External id": 12484, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95564, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95564, "pid": 0, "tid": 7, "ts": 1776649792684835, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792684703, "dur": 6, + "args": { + "External id": 12484, + "cbid": 211, "correlation": 95564 + } + }, + { + "ph": "s", "id": 95564, "pid": 435, "tid": 435, "ts": 1776649792684703, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792684732, "dur": 0, + "args": { + "External id": 12486, + "cbid": 210, "correlation": 95591 + } + }, + { + "ph": "f", "id": 95591, "pid": 435, "tid": 435, "ts": 1776649792684732, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792684884, "dur": 52, + "args": { + "External id": 12486, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95592, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95592, "pid": 0, "tid": 7, "ts": 1776649792684884, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792684733, "dur": 6, + "args": { + "External id": 12486, + "cbid": 211, "correlation": 95592 + } + }, + { + "ph": "s", "id": 95592, "pid": 435, "tid": 435, "ts": 1776649792684733, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792684762, "dur": 0, + "args": { + "External id": 12488, + "cbid": 210, "correlation": 95619 + } + }, + { + "ph": "f", "id": 95619, "pid": 435, "tid": 435, "ts": 1776649792684762, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792684938, "dur": 44, + "args": { + "External id": 12488, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95620, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95620, "pid": 0, "tid": 7, "ts": 1776649792684938, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792684763, "dur": 6, + "args": { + "External id": 12488, + "cbid": 211, "correlation": 95620 + } + }, + { + "ph": "s", "id": 95620, "pid": 435, "tid": 435, "ts": 1776649792684763, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792684792, "dur": 0, + "args": { + "External id": 12490, + "cbid": 210, "correlation": 95647 + } + }, + { + "ph": "f", "id": 95647, "pid": 435, "tid": 435, "ts": 1776649792684792, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792684983, "dur": 44, + "args": { + "External id": 12490, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95648, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95648, "pid": 0, "tid": 7, "ts": 1776649792684983, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792684793, "dur": 6, + "args": { + "External id": 12490, + "cbid": 211, "correlation": 95648 + } + }, + { + "ph": "s", "id": 95648, "pid": 435, "tid": 435, "ts": 1776649792684793, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792684822, "dur": 0, + "args": { + "External id": 12492, + "cbid": 210, "correlation": 95675 + } + }, + { + "ph": "f", "id": 95675, "pid": 435, "tid": 435, "ts": 1776649792684822, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792685029, "dur": 47, + "args": { + "External id": 12492, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95676, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95676, "pid": 0, "tid": 7, "ts": 1776649792685029, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792684824, "dur": 5, + "args": { + "External id": 12492, + "cbid": 211, "correlation": 95676 + } + }, + { + "ph": "s", "id": 95676, "pid": 435, "tid": 435, "ts": 1776649792684824, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792684853, "dur": 0, + "args": { + "External id": 12494, + "cbid": 210, "correlation": 95703 + } + }, + { + "ph": "f", "id": 95703, "pid": 435, "tid": 435, "ts": 1776649792684853, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792685077, "dur": 49, + "args": { + "External id": 12494, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95704, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95704, "pid": 0, "tid": 7, "ts": 1776649792685077, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792684854, "dur": 6, + "args": { + "External id": 12494, + "cbid": 211, "correlation": 95704 + } + }, + { + "ph": "s", "id": 95704, "pid": 435, "tid": 435, "ts": 1776649792684854, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792684883, "dur": 0, + "args": { + "External id": 12496, + "cbid": 210, "correlation": 95731 + } + }, + { + "ph": "f", "id": 95731, "pid": 435, "tid": 435, "ts": 1776649792684883, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792685128, "dur": 44, + "args": { + "External id": 12496, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95732, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95732, "pid": 0, "tid": 7, "ts": 1776649792685128, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792684884, "dur": 5, + "args": { + "External id": 12496, + "cbid": 211, "correlation": 95732 + } + }, + { + "ph": "s", "id": 95732, "pid": 435, "tid": 435, "ts": 1776649792684884, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792684913, "dur": 0, + "args": { + "External id": 12498, + "cbid": 210, "correlation": 95759 + } + }, + { + "ph": "f", "id": 95759, "pid": 435, "tid": 435, "ts": 1776649792684913, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792685174, "dur": 49, + "args": { + "External id": 12498, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95760, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95760, "pid": 0, "tid": 7, "ts": 1776649792685174, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792684914, "dur": 6, + "args": { + "External id": 12498, + "cbid": 211, "correlation": 95760 + } + }, + { + "ph": "s", "id": 95760, "pid": 435, "tid": 435, "ts": 1776649792684914, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792684943, "dur": 0, + "args": { + "External id": 12500, + "cbid": 210, "correlation": 95787 + } + }, + { + "ph": "f", "id": 95787, "pid": 435, "tid": 435, "ts": 1776649792684943, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792685224, "dur": 47, + "args": { + "External id": 12500, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95788, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95788, "pid": 0, "tid": 7, "ts": 1776649792685224, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792684944, "dur": 6, + "args": { + "External id": 12500, + "cbid": 211, "correlation": 95788 + } + }, + { + "ph": "s", "id": 95788, "pid": 435, "tid": 435, "ts": 1776649792684944, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792684973, "dur": 0, + "args": { + "External id": 12502, + "cbid": 210, "correlation": 95815 + } + }, + { + "ph": "f", "id": 95815, "pid": 435, "tid": 435, "ts": 1776649792684973, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792685273, "dur": 49, + "args": { + "External id": 12502, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95816, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95816, "pid": 0, "tid": 7, "ts": 1776649792685273, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792684974, "dur": 5, + "args": { + "External id": 12502, + "cbid": 211, "correlation": 95816 + } + }, + { + "ph": "s", "id": 95816, "pid": 435, "tid": 435, "ts": 1776649792684974, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685002, "dur": 0, + "args": { + "External id": 12504, + "cbid": 210, "correlation": 95843 + } + }, + { + "ph": "f", "id": 95843, "pid": 435, "tid": 435, "ts": 1776649792685002, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792685324, "dur": 46, + "args": { + "External id": 12504, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95844, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95844, "pid": 0, "tid": 7, "ts": 1776649792685324, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685004, "dur": 6, + "args": { + "External id": 12504, + "cbid": 211, "correlation": 95844 + } + }, + { + "ph": "s", "id": 95844, "pid": 435, "tid": 435, "ts": 1776649792685004, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685032, "dur": 0, + "args": { + "External id": 12506, + "cbid": 210, "correlation": 95871 + } + }, + { + "ph": "f", "id": 95871, "pid": 435, "tid": 435, "ts": 1776649792685032, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792685372, "dur": 46, + "args": { + "External id": 12506, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95872, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95872, "pid": 0, "tid": 7, "ts": 1776649792685372, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685034, "dur": 6, + "args": { + "External id": 12506, + "cbid": 211, "correlation": 95872 + } + }, + { + "ph": "s", "id": 95872, "pid": 435, "tid": 435, "ts": 1776649792685034, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685062, "dur": 0, + "args": { + "External id": 12508, + "cbid": 210, "correlation": 95899 + } + }, + { + "ph": "f", "id": 95899, "pid": 435, "tid": 435, "ts": 1776649792685062, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792685420, "dur": 50, + "args": { + "External id": 12508, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95900, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95900, "pid": 0, "tid": 7, "ts": 1776649792685420, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685064, "dur": 5, + "args": { + "External id": 12508, + "cbid": 211, "correlation": 95900 + } + }, + { + "ph": "s", "id": 95900, "pid": 435, "tid": 435, "ts": 1776649792685064, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685093, "dur": 0, + "args": { + "External id": 12510, + "cbid": 210, "correlation": 95927 + } + }, + { + "ph": "f", "id": 95927, "pid": 435, "tid": 435, "ts": 1776649792685093, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792685473, "dur": 46, + "args": { + "External id": 12510, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95928, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95928, "pid": 0, "tid": 7, "ts": 1776649792685473, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685094, "dur": 6, + "args": { + "External id": 12510, + "cbid": 211, "correlation": 95928 + } + }, + { + "ph": "s", "id": 95928, "pid": 435, "tid": 435, "ts": 1776649792685094, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685123, "dur": 0, + "args": { + "External id": 12512, + "cbid": 210, "correlation": 95955 + } + }, + { + "ph": "f", "id": 95955, "pid": 435, "tid": 435, "ts": 1776649792685123, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792685520, "dur": 47, + "args": { + "External id": 12512, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95956, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95956, "pid": 0, "tid": 7, "ts": 1776649792685520, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685125, "dur": 6, + "args": { + "External id": 12512, + "cbid": 211, "correlation": 95956 + } + }, + { + "ph": "s", "id": 95956, "pid": 435, "tid": 435, "ts": 1776649792685125, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685153, "dur": 0, + "args": { + "External id": 12514, + "cbid": 210, "correlation": 95983 + } + }, + { + "ph": "f", "id": 95983, "pid": 435, "tid": 435, "ts": 1776649792685153, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792685569, "dur": 46, + "args": { + "External id": 12514, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 95984, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 95984, "pid": 0, "tid": 7, "ts": 1776649792685569, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685155, "dur": 6, + "args": { + "External id": 12514, + "cbid": 211, "correlation": 95984 + } + }, + { + "ph": "s", "id": 95984, "pid": 435, "tid": 435, "ts": 1776649792685155, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685183, "dur": 0, + "args": { + "External id": 12516, + "cbid": 210, "correlation": 96011 + } + }, + { + "ph": "f", "id": 96011, "pid": 435, "tid": 435, "ts": 1776649792685183, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792685617, "dur": 46, + "args": { + "External id": 12516, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96012, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96012, "pid": 0, "tid": 7, "ts": 1776649792685617, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685185, "dur": 6, + "args": { + "External id": 12516, + "cbid": 211, "correlation": 96012 + } + }, + { + "ph": "s", "id": 96012, "pid": 435, "tid": 435, "ts": 1776649792685185, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685213, "dur": 0, + "args": { + "External id": 12518, + "cbid": 210, "correlation": 96039 + } + }, + { + "ph": "f", "id": 96039, "pid": 435, "tid": 435, "ts": 1776649792685213, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792685665, "dur": 50, + "args": { + "External id": 12518, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96040, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96040, "pid": 0, "tid": 7, "ts": 1776649792685665, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685215, "dur": 5, + "args": { + "External id": 12518, + "cbid": 211, "correlation": 96040 + } + }, + { + "ph": "s", "id": 96040, "pid": 435, "tid": 435, "ts": 1776649792685215, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685244, "dur": 0, + "args": { + "External id": 12520, + "cbid": 210, "correlation": 96067 + } + }, + { + "ph": "f", "id": 96067, "pid": 435, "tid": 435, "ts": 1776649792685244, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792685717, "dur": 46, + "args": { + "External id": 12520, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96068, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96068, "pid": 0, "tid": 7, "ts": 1776649792685717, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685245, "dur": 6, + "args": { + "External id": 12520, + "cbid": 211, "correlation": 96068 + } + }, + { + "ph": "s", "id": 96068, "pid": 435, "tid": 435, "ts": 1776649792685245, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685274, "dur": 0, + "args": { + "External id": 12522, + "cbid": 210, "correlation": 96095 + } + }, + { + "ph": "f", "id": 96095, "pid": 435, "tid": 435, "ts": 1776649792685274, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792685764, "dur": 46, + "args": { + "External id": 12522, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96096, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96096, "pid": 0, "tid": 7, "ts": 1776649792685764, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685275, "dur": 5, + "args": { + "External id": 12522, + "cbid": 211, "correlation": 96096 + } + }, + { + "ph": "s", "id": 96096, "pid": 435, "tid": 435, "ts": 1776649792685275, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685304, "dur": 0, + "args": { + "External id": 12524, + "cbid": 210, "correlation": 96123 + } + }, + { + "ph": "f", "id": 96123, "pid": 435, "tid": 435, "ts": 1776649792685304, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792685812, "dur": 49, + "args": { + "External id": 12524, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96124, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96124, "pid": 0, "tid": 7, "ts": 1776649792685812, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685305, "dur": 5, + "args": { + "External id": 12524, + "cbid": 211, "correlation": 96124 + } + }, + { + "ph": "s", "id": 96124, "pid": 435, "tid": 435, "ts": 1776649792685305, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685334, "dur": 0, + "args": { + "External id": 12526, + "cbid": 210, "correlation": 96151 + } + }, + { + "ph": "f", "id": 96151, "pid": 435, "tid": 435, "ts": 1776649792685334, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792685863, "dur": 43, + "args": { + "External id": 12526, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96152, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96152, "pid": 0, "tid": 7, "ts": 1776649792685863, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685335, "dur": 5, + "args": { + "External id": 12526, + "cbid": 211, "correlation": 96152 + } + }, + { + "ph": "s", "id": 96152, "pid": 435, "tid": 435, "ts": 1776649792685335, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685364, "dur": 0, + "args": { + "External id": 12528, + "cbid": 210, "correlation": 96179 + } + }, + { + "ph": "f", "id": 96179, "pid": 435, "tid": 435, "ts": 1776649792685364, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792685908, "dur": 47, + "args": { + "External id": 12528, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96180, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96180, "pid": 0, "tid": 7, "ts": 1776649792685908, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685365, "dur": 6, + "args": { + "External id": 12528, + "cbid": 211, "correlation": 96180 + } + }, + { + "ph": "s", "id": 96180, "pid": 435, "tid": 435, "ts": 1776649792685365, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685394, "dur": 0, + "args": { + "External id": 12530, + "cbid": 210, "correlation": 96207 + } + }, + { + "ph": "f", "id": 96207, "pid": 435, "tid": 435, "ts": 1776649792685394, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792685957, "dur": 46, + "args": { + "External id": 12530, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96208, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96208, "pid": 0, "tid": 7, "ts": 1776649792685957, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685395, "dur": 5, + "args": { + "External id": 12530, + "cbid": 211, "correlation": 96208 + } + }, + { + "ph": "s", "id": 96208, "pid": 435, "tid": 435, "ts": 1776649792685395, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685424, "dur": 0, + "args": { + "External id": 12532, + "cbid": 210, "correlation": 96235 + } + }, + { + "ph": "f", "id": 96235, "pid": 435, "tid": 435, "ts": 1776649792685424, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792686005, "dur": 47, + "args": { + "External id": 12532, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96236, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96236, "pid": 0, "tid": 7, "ts": 1776649792686005, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685425, "dur": 6, + "args": { + "External id": 12532, + "cbid": 211, "correlation": 96236 + } + }, + { + "ph": "s", "id": 96236, "pid": 435, "tid": 435, "ts": 1776649792685425, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685454, "dur": 0, + "args": { + "External id": 12534, + "cbid": 210, "correlation": 96263 + } + }, + { + "ph": "f", "id": 96263, "pid": 435, "tid": 435, "ts": 1776649792685454, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792686054, "dur": 49, + "args": { + "External id": 12534, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96264, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96264, "pid": 0, "tid": 7, "ts": 1776649792686054, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685455, "dur": 5, + "args": { + "External id": 12534, + "cbid": 211, "correlation": 96264 + } + }, + { + "ph": "s", "id": 96264, "pid": 435, "tid": 435, "ts": 1776649792685455, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685483, "dur": 0, + "args": { + "External id": 12536, + "cbid": 210, "correlation": 96291 + } + }, + { + "ph": "f", "id": 96291, "pid": 435, "tid": 435, "ts": 1776649792685483, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792686105, "dur": 48, + "args": { + "External id": 12536, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96292, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96292, "pid": 0, "tid": 7, "ts": 1776649792686105, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685485, "dur": 6, + "args": { + "External id": 12536, + "cbid": 211, "correlation": 96292 + } + }, + { + "ph": "s", "id": 96292, "pid": 435, "tid": 435, "ts": 1776649792685485, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685513, "dur": 0, + "args": { + "External id": 12538, + "cbid": 210, "correlation": 96319 + } + }, + { + "ph": "f", "id": 96319, "pid": 435, "tid": 435, "ts": 1776649792685513, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792686156, "dur": 47, + "args": { + "External id": 12538, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96320, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96320, "pid": 0, "tid": 7, "ts": 1776649792686156, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685515, "dur": 5, + "args": { + "External id": 12538, + "cbid": 211, "correlation": 96320 + } + }, + { + "ph": "s", "id": 96320, "pid": 435, "tid": 435, "ts": 1776649792685515, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685543, "dur": 0, + "args": { + "External id": 12540, + "cbid": 210, "correlation": 96347 + } + }, + { + "ph": "f", "id": 96347, "pid": 435, "tid": 435, "ts": 1776649792685543, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792686204, "dur": 47, + "args": { + "External id": 12540, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96348, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96348, "pid": 0, "tid": 7, "ts": 1776649792686204, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685545, "dur": 5, + "args": { + "External id": 12540, + "cbid": 211, "correlation": 96348 + } + }, + { + "ph": "s", "id": 96348, "pid": 435, "tid": 435, "ts": 1776649792685545, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685573, "dur": 0, + "args": { + "External id": 12542, + "cbid": 210, "correlation": 96375 + } + }, + { + "ph": "f", "id": 96375, "pid": 435, "tid": 435, "ts": 1776649792685573, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792686253, "dur": 46, + "args": { + "External id": 12542, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96376, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96376, "pid": 0, "tid": 7, "ts": 1776649792686253, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685575, "dur": 5, + "args": { + "External id": 12542, + "cbid": 211, "correlation": 96376 + } + }, + { + "ph": "s", "id": 96376, "pid": 435, "tid": 435, "ts": 1776649792685575, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685603, "dur": 0, + "args": { + "External id": 12544, + "cbid": 210, "correlation": 96403 + } + }, + { + "ph": "f", "id": 96403, "pid": 435, "tid": 435, "ts": 1776649792685603, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792686301, "dur": 44, + "args": { + "External id": 12544, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96404, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96404, "pid": 0, "tid": 7, "ts": 1776649792686301, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685605, "dur": 5, + "args": { + "External id": 12544, + "cbid": 211, "correlation": 96404 + } + }, + { + "ph": "s", "id": 96404, "pid": 435, "tid": 435, "ts": 1776649792685605, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685633, "dur": 0, + "args": { + "External id": 12546, + "cbid": 210, "correlation": 96431 + } + }, + { + "ph": "f", "id": 96431, "pid": 435, "tid": 435, "ts": 1776649792685633, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792686346, "dur": 49, + "args": { + "External id": 12546, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96432, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96432, "pid": 0, "tid": 7, "ts": 1776649792686346, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685635, "dur": 6, + "args": { + "External id": 12546, + "cbid": 211, "correlation": 96432 + } + }, + { + "ph": "s", "id": 96432, "pid": 435, "tid": 435, "ts": 1776649792685635, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685663, "dur": 0, + "args": { + "External id": 12548, + "cbid": 210, "correlation": 96459 + } + }, + { + "ph": "f", "id": 96459, "pid": 435, "tid": 435, "ts": 1776649792685663, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792686398, "dur": 49, + "args": { + "External id": 12548, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96460, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96460, "pid": 0, "tid": 7, "ts": 1776649792686398, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685665, "dur": 5, + "args": { + "External id": 12548, + "cbid": 211, "correlation": 96460 + } + }, + { + "ph": "s", "id": 96460, "pid": 435, "tid": 435, "ts": 1776649792685665, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685694, "dur": 0, + "args": { + "External id": 12550, + "cbid": 210, "correlation": 96487 + } + }, + { + "ph": "f", "id": 96487, "pid": 435, "tid": 435, "ts": 1776649792685694, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792686448, "dur": 50, + "args": { + "External id": 12550, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96488, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96488, "pid": 0, "tid": 7, "ts": 1776649792686448, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685695, "dur": 5, + "args": { + "External id": 12550, + "cbid": 211, "correlation": 96488 + } + }, + { + "ph": "s", "id": 96488, "pid": 435, "tid": 435, "ts": 1776649792685695, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685723, "dur": 0, + "args": { + "External id": 12552, + "cbid": 210, "correlation": 96515 + } + }, + { + "ph": "f", "id": 96515, "pid": 435, "tid": 435, "ts": 1776649792685723, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792686500, "dur": 44, + "args": { + "External id": 12552, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96516, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96516, "pid": 0, "tid": 7, "ts": 1776649792686500, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685725, "dur": 5, + "args": { + "External id": 12552, + "cbid": 211, "correlation": 96516 + } + }, + { + "ph": "s", "id": 96516, "pid": 435, "tid": 435, "ts": 1776649792685725, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685753, "dur": 0, + "args": { + "External id": 12554, + "cbid": 210, "correlation": 96543 + } + }, + { + "ph": "f", "id": 96543, "pid": 435, "tid": 435, "ts": 1776649792685753, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792686546, "dur": 45, + "args": { + "External id": 12554, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96544, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96544, "pid": 0, "tid": 7, "ts": 1776649792686546, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685755, "dur": 5, + "args": { + "External id": 12554, + "cbid": 211, "correlation": 96544 + } + }, + { + "ph": "s", "id": 96544, "pid": 435, "tid": 435, "ts": 1776649792685755, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685783, "dur": 0, + "args": { + "External id": 12556, + "cbid": 210, "correlation": 96571 + } + }, + { + "ph": "f", "id": 96571, "pid": 435, "tid": 435, "ts": 1776649792685783, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792686592, "dur": 46, + "args": { + "External id": 12556, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96572, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96572, "pid": 0, "tid": 7, "ts": 1776649792686592, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685785, "dur": 5, + "args": { + "External id": 12556, + "cbid": 211, "correlation": 96572 + } + }, + { + "ph": "s", "id": 96572, "pid": 435, "tid": 435, "ts": 1776649792685785, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685813, "dur": 0, + "args": { + "External id": 12558, + "cbid": 210, "correlation": 96599 + } + }, + { + "ph": "f", "id": 96599, "pid": 435, "tid": 435, "ts": 1776649792685813, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792686640, "dur": 47, + "args": { + "External id": 12558, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96600, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96600, "pid": 0, "tid": 7, "ts": 1776649792686640, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685814, "dur": 5, + "args": { + "External id": 12558, + "cbid": 211, "correlation": 96600 + } + }, + { + "ph": "s", "id": 96600, "pid": 435, "tid": 435, "ts": 1776649792685814, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685842, "dur": 0, + "args": { + "External id": 12560, + "cbid": 210, "correlation": 96627 + } + }, + { + "ph": "f", "id": 96627, "pid": 435, "tid": 435, "ts": 1776649792685842, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792686689, "dur": 50, + "args": { + "External id": 12560, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96628, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96628, "pid": 0, "tid": 7, "ts": 1776649792686689, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685844, "dur": 5, + "args": { + "External id": 12560, + "cbid": 211, "correlation": 96628 + } + }, + { + "ph": "s", "id": 96628, "pid": 435, "tid": 435, "ts": 1776649792685844, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685872, "dur": 0, + "args": { + "External id": 12562, + "cbid": 210, "correlation": 96655 + } + }, + { + "ph": "f", "id": 96655, "pid": 435, "tid": 435, "ts": 1776649792685872, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792686740, "dur": 46, + "args": { + "External id": 12562, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96656, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96656, "pid": 0, "tid": 7, "ts": 1776649792686740, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685874, "dur": 5, + "args": { + "External id": 12562, + "cbid": 211, "correlation": 96656 + } + }, + { + "ph": "s", "id": 96656, "pid": 435, "tid": 435, "ts": 1776649792685874, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685902, "dur": 0, + "args": { + "External id": 12564, + "cbid": 210, "correlation": 96683 + } + }, + { + "ph": "f", "id": 96683, "pid": 435, "tid": 435, "ts": 1776649792685902, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792686788, "dur": 51, + "args": { + "External id": 12564, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96684, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96684, "pid": 0, "tid": 7, "ts": 1776649792686788, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685904, "dur": 5, + "args": { + "External id": 12564, + "cbid": 211, "correlation": 96684 + } + }, + { + "ph": "s", "id": 96684, "pid": 435, "tid": 435, "ts": 1776649792685904, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685932, "dur": 0, + "args": { + "External id": 12566, + "cbid": 210, "correlation": 96711 + } + }, + { + "ph": "f", "id": 96711, "pid": 435, "tid": 435, "ts": 1776649792685932, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792686842, "dur": 46, + "args": { + "External id": 12566, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96712, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96712, "pid": 0, "tid": 7, "ts": 1776649792686842, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685933, "dur": 6, + "args": { + "External id": 12566, + "cbid": 211, "correlation": 96712 + } + }, + { + "ph": "s", "id": 96712, "pid": 435, "tid": 435, "ts": 1776649792685933, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649792685962, "dur": 0, + "args": { + "External id": 12568, + "cbid": 210, "correlation": 96739 + } + }, + { + "ph": "f", "id": 96739, "pid": 435, "tid": 435, "ts": 1776649792685962, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649792686889, "dur": 48, + "args": { + "External id": 12568, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96740, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 96740, "pid": 0, "tid": 7, "ts": 1776649792686889, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792685963, "dur": 5, + "args": { + "External id": 12568, + "cbid": 211, "correlation": 96740 + } + }, + { + "ph": "s", "id": 96740, "pid": 435, "tid": 435, "ts": 1776649792685963, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649792685996, "dur": 4, + "args": { + "External id": 96751, + "cbid": 135, "correlation": 96751 + } + }, + { + "ph": "f", "id": 96751, "pid": 435, "tid": 435, "ts": 1776649792685996, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649792686010, "dur": 936, + "args": { + "External id": 96756, + "cbid": 165, "correlation": 96756 + } + }, + { + "ph": "s", "id": 96756, "pid": 435, "tid": 435, "ts": 1776649792686010, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649792686950, "dur": 4, + "args": { + "External id": 96759, + "cbid": 139, "correlation": 96759 + } + }, + { + "ph": "f", "id": 96759, "pid": 435, "tid": 435, "ts": 1776649792686950, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor >, at::detail::Array >(int, at::native::AUnaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792687029, "dur": 33, + "args": { + "External id": 12571, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96776, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 96776, "pid": 0, "tid": 7, "ts": 1776649792687029, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792687017, "dur": 10, + "args": { + "External id": 12571, + "cbid": 211, "correlation": 96776 + } + }, + { + "ph": "s", "id": 96776, "pid": 435, "tid": 435, "ts": 1776649792687017, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649792687068, "dur": 57, + "args": { + "External id": 12575, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96789, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 96789, "pid": 0, "tid": 7, "ts": 1776649792687068, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792687056, "dur": 7, + "args": { + "External id": 12575, + "cbid": 211, "correlation": 96789 + } + }, + { + "ph": "s", "id": 96789, "pid": 435, "tid": 435, "ts": 1776649792687056, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BUnaryFunctor, at::detail::Array >(int, at::native::BUnaryFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792687127, "dur": 33, + "args": { + "External id": 12576, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96803, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 96803, "pid": 0, "tid": 7, "ts": 1776649792687127, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792687089, "dur": 7, + "args": { + "External id": 12576, + "cbid": 211, "correlation": 96803 + } + }, + { + "ph": "s", "id": 96803, "pid": 435, "tid": 435, "ts": 1776649792687089, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649792687161, "dur": 57, + "args": { + "External id": 12581, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96816, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 96816, "pid": 0, "tid": 7, "ts": 1776649792687161, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792687115, "dur": 6, + "args": { + "External id": 12581, + "cbid": 211, "correlation": 96816 + } + }, + { + "ph": "s", "id": 96816, "pid": 435, "tid": 435, "ts": 1776649792687115, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy_contig, unsigned int, 3, 128, 1>(at::native::(anonymous namespace)::OpaqueType<1u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata, unsigned int, 128, 1>, at::native::(anonymous namespace)::TensorSizeStride, int, unsigned int)", "pid": 0, "tid": 7, + "ts": 1776649792687220, "dur": 126, + "args": { + "External id": 12587, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96828, + "registers per thread": 21, + "shared memory": 0, + "blocks per SM": 4.000000, + "warps per SM": 64.000000, + "grid": [216, 2, 1], + "block": [512, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 96828, "pid": 0, "tid": 7, "ts": 1776649792687220, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792687156, "dur": 8, + "args": { + "External id": 12587, + "cbid": 211, "correlation": 96828 + } + }, + { + "ph": "s", "id": 96828, "pid": 435, "tid": 435, "ts": 1776649792687156, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649792687347, "dur": 136, + "args": { + "External id": 12598, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96841, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 96841, "pid": 0, "tid": 7, "ts": 1776649792687347, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792687213, "dur": 8, + "args": { + "External id": 12598, + "cbid": 211, "correlation": 96841 + } + }, + { + "ph": "s", "id": 96841, "pid": 435, "tid": 435, "ts": 1776649792687213, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, at::detail::Array >(int, at::native::CUDAFunctorOnSelf_add, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792687484, "dur": 63, + "args": { + "External id": 12601, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96852, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 96852, "pid": 0, "tid": 7, "ts": 1776649792687484, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792687248, "dur": 7, + "args": { + "External id": 12601, + "cbid": 211, "correlation": 96852 + } + }, + { + "ph": "s", "id": 96852, "pid": 435, "tid": 435, "ts": 1776649792687248, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649792687549, "dur": 178, + "args": { + "External id": 12605, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96865, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 96865, "pid": 0, "tid": 7, "ts": 1776649792687549, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792687276, "dur": 6, + "args": { + "External id": 12605, + "cbid": 211, "correlation": 96865 + } + }, + { + "ph": "s", "id": 96865, "pid": 435, "tid": 435, "ts": 1776649792687276, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor >, at::detail::Array >(int, at::native::BinaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649792687729, "dur": 114, + "args": { + "External id": 12606, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96876, + "registers per thread": 22, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 96876, "pid": 0, "tid": 7, "ts": 1776649792687729, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792687300, "dur": 6, + "args": { + "External id": 12606, + "cbid": 211, "correlation": 96876 + } + }, + { + "ph": "s", "id": 96876, "pid": 435, "tid": 435, "ts": 1776649792687300, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649792687845, "dur": 515, + "args": { + "External id": 12614, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96890, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 96890, "pid": 0, "tid": 7, "ts": 1776649792687845, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649792687353, "dur": 7, + "args": { + "External id": 12614, + "cbid": 211, "correlation": 96890 + } + }, + { + "ph": "s", "id": 96890, "pid": 435, "tid": 435, "ts": 1776649792687353, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649792976473, "dur": 103159, + "args": { + "External id": 12630, + "device": 0, "context": 1, + "stream": 7, "correlation": 96906, + "bytes": 67108864, "memory bandwidth (GB/s)": 0.6505369041457051 + } + }, + { + "ph": "f", "id": 96906, "pid": 0, "tid": 7, "ts": 1776649792976473, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649792976335, "dur": 103270, + "args": { + "External id": 12630, + "cbid": 41, "correlation": 96906 + } + }, + { + "ph": "s", "id": 96906, "pid": 435, "tid": 435, "ts": 1776649792976335, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649793079615, "dur": 31, + "args": { + "External id": 12630, + "cbid": 131, "correlation": 96907 + } + }, + { + "ph": "s", "id": 96907, "pid": 435, "tid": 435, "ts": 1776649793079615, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7, + "ts": 1776649793106412, "dur": 76, + "args": { + "External id": 12633, + "device": 0, "context": 1, + "stream": 7, "correlation": 96912, + "bytes": 67108864, "memory bandwidth (GB/s)": 880.7977845152315 + } + }, + { + "ph": "f", "id": 96912, "pid": 0, "tid": 7, "ts": 1776649793106412, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649793106351, "dur": 59, + "args": { + "External id": 12633, + "cbid": 41, "correlation": 96912 + } + }, + { + "ph": "s", "id": 96912, "pid": 435, "tid": 435, "ts": 1776649793106351, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649793106741, "dur": 1226, + "args": { + "External id": 12642, + "device": 0, "context": 1, + "stream": 7, "correlation": 96927, + "bytes": 16777216, "memory bandwidth (GB/s)": 13.673685849135799 + } + }, + { + "ph": "f", "id": 96927, "pid": 0, "tid": 7, "ts": 1776649793106741, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649793106643, "dur": 1272, + "args": { + "External id": 12642, + "cbid": 41, "correlation": 96927 + } + }, + { + "ph": "s", "id": 96927, "pid": 435, "tid": 435, "ts": 1776649793106643, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649793107920, "dur": 60, + "args": { + "External id": 12642, + "cbid": 131, "correlation": 96928 + } + }, + { + "ph": "s", "id": 96928, "pid": 435, "tid": 435, "ts": 1776649793107920, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649793108133, "dur": 31, + "args": { + "External id": 12646, + "device": 0, "context": 1, + "stream": 7, "correlation": 96940, + "bytes": 524288, "memory bandwidth (GB/s)": 16.769703172978506 + } + }, + { + "ph": "f", "id": 96940, "pid": 0, "tid": 7, "ts": 1776649793108133, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649793108078, "dur": 52, + "args": { + "External id": 12646, + "cbid": 41, "correlation": 96940 + } + }, + { + "ph": "s", "id": 96940, "pid": 435, "tid": 435, "ts": 1776649793108078, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649793108131, "dur": 40, + "args": { + "External id": 12646, + "cbid": 131, "correlation": 96941 + } + }, + { + "ph": "s", "id": 96941, "pid": 435, "tid": 435, "ts": 1776649793108131, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649793108209, "dur": 1, + "args": { + "External id": 12650, + "device": 0, "context": 1, + "stream": 7, "correlation": 96953, + "bytes": 4096, "memory bandwidth (GB/s)": 2.6122448979591835 + } + }, + { + "ph": "f", "id": 96953, "pid": 0, "tid": 7, "ts": 1776649793108209, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649793108196, "dur": 8, + "args": { + "External id": 12650, + "cbid": 41, "correlation": 96953 + } + }, + { + "ph": "s", "id": 96953, "pid": 435, "tid": 435, "ts": 1776649793108196, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649793108205, "dur": 10, + "args": { + "External id": 12650, + "cbid": 131, "correlation": 96954 + } + }, + { + "ph": "s", "id": 96954, "pid": 435, "tid": 435, "ts": 1776649793108205, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649793108490, "dur": 1041, + "args": { + "External id": 12665, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96966, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 96966, "pid": 0, "tid": 7, "ts": 1776649793108490, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793108449, "dur": 43, + "args": { + "External id": 12665, + "cbid": 211, "correlation": 96966 + } + }, + { + "ph": "s", "id": 96966, "pid": 435, "tid": 435, "ts": 1776649793108449, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649793109533, "dur": 7, + "args": { + "External id": 12671, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96980, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 4.740741, + "warps per SM": 18.962963, + "grid": [512, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 30 + } + }, + { + "ph": "f", "id": 96980, "pid": 0, "tid": 7, "ts": 1776649793109533, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793108532, "dur": 9, + "args": { + "External id": 12671, + "cbid": 211, "correlation": 96980 + } + }, + { + "ph": "s", "id": 96980, "pid": 435, "tid": 435, "ts": 1776649793108532, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649793109541, "dur": 247, + "args": { + "External id": 12673, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96991, + "registers per thread": 20, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 96991, "pid": 0, "tid": 7, "ts": 1776649793109541, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793108582, "dur": 8, + "args": { + "External id": 12673, + "cbid": 211, "correlation": 96991 + } + }, + { + "ph": "s", "id": 96991, "pid": 435, "tid": 435, "ts": 1776649793108582, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793109790, "dur": 76, + "args": { + "External id": 12674, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97001, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 97001, "pid": 0, "tid": 7, "ts": 1776649793109790, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793108615, "dur": 7, + "args": { + "External id": 12674, + "cbid": 211, "correlation": 97001 + } + }, + { + "ph": "s", "id": 97001, "pid": 435, "tid": 435, "ts": 1776649793108615, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649793109868, "dur": 192, + "args": { + "External id": 12678, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97014, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 97014, "pid": 0, "tid": 7, "ts": 1776649793109868, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793108652, "dur": 9, + "args": { + "External id": 12678, + "cbid": 211, "correlation": 97014 + } + }, + { + "ph": "s", "id": 97014, "pid": 435, "tid": 435, "ts": 1776649793108652, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, at::detail::Array >(int, at::native::CUDAFunctorOnSelf_add, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793110062, "dur": 151, + "args": { + "External id": 12679, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97021, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 97021, "pid": 0, "tid": 7, "ts": 1776649793110062, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793108704, "dur": 9, + "args": { + "External id": 12679, + "cbid": 211, "correlation": 97021 + } + }, + { + "ph": "s", "id": 97021, "pid": 435, "tid": 435, "ts": 1776649793108704, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1}, at::detail::Array >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793110215, "dur": 158, + "args": { + "External id": 12680, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97031, + "registers per thread": 22, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 97031, "pid": 0, "tid": 7, "ts": 1776649793110215, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793108748, "dur": 8, + "args": { + "External id": 12680, + "cbid": 211, "correlation": 97031 + } + }, + { + "ph": "s", "id": 97031, "pid": 435, "tid": 435, "ts": 1776649793108748, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649793110374, "dur": 198, + "args": { + "External id": 12690, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97044, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 1213.629639, + "warps per SM": 4854.518555, + "grid": [131072, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 97044, "pid": 0, "tid": 7, "ts": 1776649793110374, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793108807, "dur": 9, + "args": { + "External id": 12690, + "cbid": 211, "correlation": 97044 + } + }, + { + "ph": "s", "id": 97044, "pid": 435, "tid": 435, "ts": 1776649793108807, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649793110575, "dur": 1, + "args": { + "External id": 12700, + "device": 0, "context": 1, + "stream": 7, "correlation": 97059, + "bytes": 512, "memory bandwidth (GB/s)": 0.38095238095238093 + } + }, + { + "ph": "f", "id": 97059, "pid": 0, "tid": 7, "ts": 1776649793110575, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649793108963, "dur": 12, + "args": { + "External id": 12700, + "cbid": 41, "correlation": 97059 + } + }, + { + "ph": "s", "id": 97059, "pid": 435, "tid": 435, "ts": 1776649793108963, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649793108976, "dur": 1605, + "args": { + "External id": 12700, + "cbid": 131, "correlation": 97060 + } + }, + { + "ph": "s", "id": 97060, "pid": 435, "tid": 435, "ts": 1776649793108976, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649793110646, "dur": 9, + "args": { + "External id": 12702, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97069, + "registers per thread": 40, + "shared memory": 0, + "blocks per SM": 4.740741, + "warps per SM": 18.962963, + "grid": [512, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 30 + } + }, + { + "ph": "f", "id": 97069, "pid": 0, "tid": 7, "ts": 1776649793110646, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793110632, "dur": 12, + "args": { + "External id": 12702, + "cbid": 211, "correlation": 97069 + } + }, + { + "ph": "s", "id": 97069, "pid": 435, "tid": 435, "ts": 1776649793110632, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649793110716, "dur": 217, + "args": { + "External id": 12716, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97084, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 1213.629639, + "warps per SM": 4854.518555, + "grid": [131072, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 97084, "pid": 0, "tid": 7, "ts": 1776649793110716, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793110705, "dur": 8, + "args": { + "External id": 12716, + "cbid": 211, "correlation": 97084 + } + }, + { + "ph": "s", "id": 97084, "pid": 435, "tid": 435, "ts": 1776649793110705, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649793110936, "dur": 1, + "args": { + "External id": 12726, + "device": 0, "context": 1, + "stream": 7, "correlation": 97099, + "bytes": 8192, "memory bandwidth (GB/s)": 4.654545454545454 + } + }, + { + "ph": "f", "id": 97099, "pid": 0, "tid": 7, "ts": 1776649793110936, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649793110765, "dur": 9, + "args": { + "External id": 12726, + "cbid": 41, "correlation": 97099 + } + }, + { + "ph": "s", "id": 97099, "pid": 435, "tid": 435, "ts": 1776649793110765, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649793110775, "dur": 167, + "args": { + "External id": 12726, + "cbid": 131, "correlation": 97100 + } + }, + { + "ph": "s", "id": 97100, "pid": 435, "tid": 435, "ts": 1776649793110775, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649793110974, "dur": 245, + "args": { + "External id": 12722, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97109, + "registers per thread": 40, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 75 + } + }, + { + "ph": "f", "id": 97109, "pid": 0, "tid": 7, "ts": 1776649793110974, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793110962, "dur": 11, + "args": { + "External id": 12722, + "cbid": 211, "correlation": 97109 + } + }, + { + "ph": "s", "id": 97109, "pid": 435, "tid": 435, "ts": 1776649793110962, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoH (Device -> Pageable)", "pid": 0, "tid": 7, + "ts": 1776649793112081, "dur": 96507, + "args": { + "External id": 12735, + "device": 0, "context": 1, + "stream": 7, "correlation": 97117, + "bytes": 134217728, "memory bandwidth (GB/s)": 1.3907554643797322 + } + }, + { + "ph": "f", "id": 97117, "pid": 0, "tid": 7, "ts": 1776649793112081, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649793112054, "dur": 97976, + "args": { + "External id": 12735, + "cbid": 41, "correlation": 97117 + } + }, + { + "ph": "s", "id": 97117, "pid": 435, "tid": 435, "ts": 1776649793112054, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649793210033, "dur": 11, + "args": { + "External id": 12735, + "cbid": 131, "correlation": 97118 + } + }, + { + "ph": "s", "id": 97118, "pid": 435, "tid": 435, "ts": 1776649793210033, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649793326960, "dur": 1197, + "args": { + "External id": 12745, + "device": 0, "context": 1, + "stream": 7, "correlation": 97130, + "bytes": 16777216, "memory bandwidth (GB/s)": 14.00648178604763 + } + }, + { + "ph": "f", "id": 97130, "pid": 0, "tid": 7, "ts": 1776649793326960, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649793326798, "dur": 1308, + "args": { + "External id": 12745, + "cbid": 41, "correlation": 97130 + } + }, + { + "ph": "s", "id": 97130, "pid": 435, "tid": 435, "ts": 1776649793326798, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649793328111, "dur": 57, + "args": { + "External id": 12745, + "cbid": 131, "correlation": 97131 + } + }, + { + "ph": "s", "id": 97131, "pid": 435, "tid": 435, "ts": 1776649793328111, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7, + "ts": 1776649793328445, "dur": 15, + "args": { + "External id": 12751, + "device": 0, "context": 1, + "stream": 7, "correlation": 97136, + "bytes": 16777216, "memory bandwidth (GB/s)": 1070.0437527903566 + } + }, + { + "ph": "f", "id": 97136, "pid": 0, "tid": 7, "ts": 1776649793328445, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649793328392, "dur": 51, + "args": { + "External id": 12751, + "cbid": 41, "correlation": 97136 + } + }, + { + "ph": "s", "id": 97136, "pid": 435, "tid": 435, "ts": 1776649793328392, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7, + "ts": 1776649793328501, "dur": 3, + "args": { + "External id": 12757, + "device": 0, "context": 1, + "stream": 7, "correlation": 97144, + "bytes": 524288, "memory bandwidth (GB/s)": 147.6036036036036 + } + }, + { + "ph": "f", "id": 97144, "pid": 0, "tid": 7, "ts": 1776649793328501, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649793328482, "dur": 16, + "args": { + "External id": 12757, + "cbid": 41, "correlation": 97144 + } + }, + { + "ph": "s", "id": 97144, "pid": 435, "tid": 435, "ts": 1776649793328482, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793329727, "dur": 1, + "args": { + "External id": 12762, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97167, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97167, "pid": 0, "tid": 7, "ts": 1776649793329727, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793329705, "dur": 23, + "args": { + "External id": 12762, + "cbid": 211, "correlation": 97167 + } + }, + { + "ph": "s", "id": 97167, "pid": 435, "tid": 435, "ts": 1776649793329705, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793329746, "dur": 2, + "args": { + "External id": 97173, + "cbid": 200, "correlation": 97173 + } + }, + { + "ph": "f", "id": 97173, "pid": 435, "tid": 435, "ts": 1776649793329746, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793329749, "dur": 4, + "args": { + "External id": 97174, + "cbid": 273, "correlation": 97174 + } + }, + { + "ph": "f", "id": 97174, "pid": 435, "tid": 435, "ts": 1776649793329749, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793329764, "dur": 19, + "args": { + "External id": 97175, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97175, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97175, "pid": 0, "tid": 7, "ts": 1776649793329764, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793329755, "dur": 6, + "args": { + "External id": 97175, + "cbid": 211, "correlation": 97175 + } + }, + { + "ph": "s", "id": 97175, "pid": 435, "tid": 435, "ts": 1776649793329755, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793329807, "dur": 1, + "args": { + "External id": 12767, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97193, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97193, "pid": 0, "tid": 7, "ts": 1776649793329807, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793329798, "dur": 7, + "args": { + "External id": 12767, + "cbid": 211, "correlation": 97193 + } + }, + { + "ph": "s", "id": 97193, "pid": 435, "tid": 435, "ts": 1776649793329798, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793329812, "dur": 0, + "args": { + "External id": 97199, + "cbid": 200, "correlation": 97199 + } + }, + { + "ph": "f", "id": 97199, "pid": 435, "tid": 435, "ts": 1776649793329812, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793329812, "dur": 0, + "args": { + "External id": 97200, + "cbid": 273, "correlation": 97200 + } + }, + { + "ph": "f", "id": 97200, "pid": 435, "tid": 435, "ts": 1776649793329812, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793329821, "dur": 19, + "args": { + "External id": 97201, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97201, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97201, "pid": 0, "tid": 7, "ts": 1776649793329821, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793329813, "dur": 4, + "args": { + "External id": 97201, + "cbid": 211, "correlation": 97201 + } + }, + { + "ph": "s", "id": 97201, "pid": 435, "tid": 435, "ts": 1776649793329813, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793329859, "dur": 1, + "args": { + "External id": 12772, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97219, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97219, "pid": 0, "tid": 7, "ts": 1776649793329859, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793329850, "dur": 6, + "args": { + "External id": 12772, + "cbid": 211, "correlation": 97219 + } + }, + { + "ph": "s", "id": 97219, "pid": 435, "tid": 435, "ts": 1776649793329850, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793329863, "dur": 0, + "args": { + "External id": 97225, + "cbid": 200, "correlation": 97225 + } + }, + { + "ph": "f", "id": 97225, "pid": 435, "tid": 435, "ts": 1776649793329863, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793329863, "dur": 0, + "args": { + "External id": 97226, + "cbid": 273, "correlation": 97226 + } + }, + { + "ph": "f", "id": 97226, "pid": 435, "tid": 435, "ts": 1776649793329863, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793329872, "dur": 19, + "args": { + "External id": 97227, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97227, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97227, "pid": 0, "tid": 7, "ts": 1776649793329872, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793329865, "dur": 4, + "args": { + "External id": 97227, + "cbid": 211, "correlation": 97227 + } + }, + { + "ph": "s", "id": 97227, "pid": 435, "tid": 435, "ts": 1776649793329865, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793329908, "dur": 1, + "args": { + "External id": 12777, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97245, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97245, "pid": 0, "tid": 7, "ts": 1776649793329908, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793329900, "dur": 6, + "args": { + "External id": 12777, + "cbid": 211, "correlation": 97245 + } + }, + { + "ph": "s", "id": 97245, "pid": 435, "tid": 435, "ts": 1776649793329900, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793329912, "dur": 0, + "args": { + "External id": 97251, + "cbid": 200, "correlation": 97251 + } + }, + { + "ph": "f", "id": 97251, "pid": 435, "tid": 435, "ts": 1776649793329912, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793329913, "dur": 0, + "args": { + "External id": 97252, + "cbid": 273, "correlation": 97252 + } + }, + { + "ph": "f", "id": 97252, "pid": 435, "tid": 435, "ts": 1776649793329913, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793329921, "dur": 19, + "args": { + "External id": 97253, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97253, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97253, "pid": 0, "tid": 7, "ts": 1776649793329921, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793329914, "dur": 4, + "args": { + "External id": 97253, + "cbid": 211, "correlation": 97253 + } + }, + { + "ph": "s", "id": 97253, "pid": 435, "tid": 435, "ts": 1776649793329914, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793329957, "dur": 1, + "args": { + "External id": 12782, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97271, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97271, "pid": 0, "tid": 7, "ts": 1776649793329957, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793329948, "dur": 6, + "args": { + "External id": 12782, + "cbid": 211, "correlation": 97271 + } + }, + { + "ph": "s", "id": 97271, "pid": 435, "tid": 435, "ts": 1776649793329948, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793329961, "dur": 0, + "args": { + "External id": 97277, + "cbid": 200, "correlation": 97277 + } + }, + { + "ph": "f", "id": 97277, "pid": 435, "tid": 435, "ts": 1776649793329961, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793329961, "dur": 0, + "args": { + "External id": 97278, + "cbid": 273, "correlation": 97278 + } + }, + { + "ph": "f", "id": 97278, "pid": 435, "tid": 435, "ts": 1776649793329961, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793329970, "dur": 19, + "args": { + "External id": 97279, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97279, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97279, "pid": 0, "tid": 7, "ts": 1776649793329970, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793329962, "dur": 4, + "args": { + "External id": 97279, + "cbid": 211, "correlation": 97279 + } + }, + { + "ph": "s", "id": 97279, "pid": 435, "tid": 435, "ts": 1776649793329962, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793330005, "dur": 1, + "args": { + "External id": 12787, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97297, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97297, "pid": 0, "tid": 7, "ts": 1776649793330005, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793329997, "dur": 6, + "args": { + "External id": 12787, + "cbid": 211, "correlation": 97297 + } + }, + { + "ph": "s", "id": 97297, "pid": 435, "tid": 435, "ts": 1776649793329997, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330009, "dur": 0, + "args": { + "External id": 97303, + "cbid": 200, "correlation": 97303 + } + }, + { + "ph": "f", "id": 97303, "pid": 435, "tid": 435, "ts": 1776649793330009, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330009, "dur": 0, + "args": { + "External id": 97304, + "cbid": 273, "correlation": 97304 + } + }, + { + "ph": "f", "id": 97304, "pid": 435, "tid": 435, "ts": 1776649793330009, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793330018, "dur": 19, + "args": { + "External id": 97305, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97305, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97305, "pid": 0, "tid": 7, "ts": 1776649793330018, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330010, "dur": 4, + "args": { + "External id": 97305, + "cbid": 211, "correlation": 97305 + } + }, + { + "ph": "s", "id": 97305, "pid": 435, "tid": 435, "ts": 1776649793330010, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793330054, "dur": 1, + "args": { + "External id": 12792, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97323, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97323, "pid": 0, "tid": 7, "ts": 1776649793330054, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330045, "dur": 6, + "args": { + "External id": 12792, + "cbid": 211, "correlation": 97323 + } + }, + { + "ph": "s", "id": 97323, "pid": 435, "tid": 435, "ts": 1776649793330045, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330058, "dur": 0, + "args": { + "External id": 97329, + "cbid": 200, "correlation": 97329 + } + }, + { + "ph": "f", "id": 97329, "pid": 435, "tid": 435, "ts": 1776649793330058, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330059, "dur": 0, + "args": { + "External id": 97330, + "cbid": 273, "correlation": 97330 + } + }, + { + "ph": "f", "id": 97330, "pid": 435, "tid": 435, "ts": 1776649793330059, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793330067, "dur": 19, + "args": { + "External id": 97331, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97331, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97331, "pid": 0, "tid": 7, "ts": 1776649793330067, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330060, "dur": 4, + "args": { + "External id": 97331, + "cbid": 211, "correlation": 97331 + } + }, + { + "ph": "s", "id": 97331, "pid": 435, "tid": 435, "ts": 1776649793330060, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793330102, "dur": 1, + "args": { + "External id": 12797, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97349, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97349, "pid": 0, "tid": 7, "ts": 1776649793330102, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330094, "dur": 6, + "args": { + "External id": 12797, + "cbid": 211, "correlation": 97349 + } + }, + { + "ph": "s", "id": 97349, "pid": 435, "tid": 435, "ts": 1776649793330094, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330106, "dur": 0, + "args": { + "External id": 97355, + "cbid": 200, "correlation": 97355 + } + }, + { + "ph": "f", "id": 97355, "pid": 435, "tid": 435, "ts": 1776649793330106, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330107, "dur": 0, + "args": { + "External id": 97356, + "cbid": 273, "correlation": 97356 + } + }, + { + "ph": "f", "id": 97356, "pid": 435, "tid": 435, "ts": 1776649793330107, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793330115, "dur": 19, + "args": { + "External id": 97357, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97357, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97357, "pid": 0, "tid": 7, "ts": 1776649793330115, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330108, "dur": 4, + "args": { + "External id": 97357, + "cbid": 211, "correlation": 97357 + } + }, + { + "ph": "s", "id": 97357, "pid": 435, "tid": 435, "ts": 1776649793330108, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793330162, "dur": 1, + "args": { + "External id": 12802, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97375, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97375, "pid": 0, "tid": 7, "ts": 1776649793330162, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330153, "dur": 6, + "args": { + "External id": 12802, + "cbid": 211, "correlation": 97375 + } + }, + { + "ph": "s", "id": 97375, "pid": 435, "tid": 435, "ts": 1776649793330153, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330166, "dur": 0, + "args": { + "External id": 97381, + "cbid": 200, "correlation": 97381 + } + }, + { + "ph": "f", "id": 97381, "pid": 435, "tid": 435, "ts": 1776649793330166, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330167, "dur": 0, + "args": { + "External id": 97382, + "cbid": 273, "correlation": 97382 + } + }, + { + "ph": "f", "id": 97382, "pid": 435, "tid": 435, "ts": 1776649793330167, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793330176, "dur": 19, + "args": { + "External id": 97383, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97383, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97383, "pid": 0, "tid": 7, "ts": 1776649793330176, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330168, "dur": 4, + "args": { + "External id": 97383, + "cbid": 211, "correlation": 97383 + } + }, + { + "ph": "s", "id": 97383, "pid": 435, "tid": 435, "ts": 1776649793330168, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793330212, "dur": 1, + "args": { + "External id": 12807, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97401, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97401, "pid": 0, "tid": 7, "ts": 1776649793330212, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330203, "dur": 6, + "args": { + "External id": 12807, + "cbid": 211, "correlation": 97401 + } + }, + { + "ph": "s", "id": 97401, "pid": 435, "tid": 435, "ts": 1776649793330203, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330215, "dur": 0, + "args": { + "External id": 97407, + "cbid": 200, "correlation": 97407 + } + }, + { + "ph": "f", "id": 97407, "pid": 435, "tid": 435, "ts": 1776649793330215, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330215, "dur": 0, + "args": { + "External id": 97408, + "cbid": 273, "correlation": 97408 + } + }, + { + "ph": "f", "id": 97408, "pid": 435, "tid": 435, "ts": 1776649793330215, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793330224, "dur": 19, + "args": { + "External id": 97409, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97409, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97409, "pid": 0, "tid": 7, "ts": 1776649793330224, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330216, "dur": 4, + "args": { + "External id": 97409, + "cbid": 211, "correlation": 97409 + } + }, + { + "ph": "s", "id": 97409, "pid": 435, "tid": 435, "ts": 1776649793330216, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649793330259, "dur": 12, + "args": { + "External id": 97412, + "cbid": 165, "correlation": 97412 + } + }, + { + "ph": "s", "id": 97412, "pid": 435, "tid": 435, "ts": 1776649793330259, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649793330315, "dur": 8, + "args": { + "External id": 97421, + "cbid": 135, "correlation": 97421 + } + }, + { + "ph": "f", "id": 97421, "pid": 435, "tid": 435, "ts": 1776649793330315, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793330370, "dur": 1, + "args": { + "External id": 12812, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97441, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97441, "pid": 0, "tid": 7, "ts": 1776649793330370, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330360, "dur": 8, + "args": { + "External id": 12812, + "cbid": 211, "correlation": 97441 + } + }, + { + "ph": "s", "id": 97441, "pid": 435, "tid": 435, "ts": 1776649793330360, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330375, "dur": 0, + "args": { + "External id": 97447, + "cbid": 200, "correlation": 97447 + } + }, + { + "ph": "f", "id": 97447, "pid": 435, "tid": 435, "ts": 1776649793330375, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330376, "dur": 0, + "args": { + "External id": 97448, + "cbid": 273, "correlation": 97448 + } + }, + { + "ph": "f", "id": 97448, "pid": 435, "tid": 435, "ts": 1776649793330376, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793330385, "dur": 19, + "args": { + "External id": 97449, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97449, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97449, "pid": 0, "tid": 7, "ts": 1776649793330385, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330377, "dur": 4, + "args": { + "External id": 97449, + "cbid": 211, "correlation": 97449 + } + }, + { + "ph": "s", "id": 97449, "pid": 435, "tid": 435, "ts": 1776649793330377, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793330423, "dur": 1, + "args": { + "External id": 12817, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97467, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97467, "pid": 0, "tid": 7, "ts": 1776649793330423, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330413, "dur": 6, + "args": { + "External id": 12817, + "cbid": 211, "correlation": 97467 + } + }, + { + "ph": "s", "id": 97467, "pid": 435, "tid": 435, "ts": 1776649793330413, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330426, "dur": 0, + "args": { + "External id": 97473, + "cbid": 200, "correlation": 97473 + } + }, + { + "ph": "f", "id": 97473, "pid": 435, "tid": 435, "ts": 1776649793330426, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330426, "dur": 0, + "args": { + "External id": 97474, + "cbid": 273, "correlation": 97474 + } + }, + { + "ph": "f", "id": 97474, "pid": 435, "tid": 435, "ts": 1776649793330426, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793330435, "dur": 19, + "args": { + "External id": 97475, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97475, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97475, "pid": 0, "tid": 7, "ts": 1776649793330435, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330427, "dur": 4, + "args": { + "External id": 97475, + "cbid": 211, "correlation": 97475 + } + }, + { + "ph": "s", "id": 97475, "pid": 435, "tid": 435, "ts": 1776649793330427, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793330471, "dur": 1, + "args": { + "External id": 12822, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97493, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97493, "pid": 0, "tid": 7, "ts": 1776649793330471, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330462, "dur": 6, + "args": { + "External id": 12822, + "cbid": 211, "correlation": 97493 + } + }, + { + "ph": "s", "id": 97493, "pid": 435, "tid": 435, "ts": 1776649793330462, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330474, "dur": 0, + "args": { + "External id": 97499, + "cbid": 200, "correlation": 97499 + } + }, + { + "ph": "f", "id": 97499, "pid": 435, "tid": 435, "ts": 1776649793330474, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330475, "dur": 0, + "args": { + "External id": 97500, + "cbid": 273, "correlation": 97500 + } + }, + { + "ph": "f", "id": 97500, "pid": 435, "tid": 435, "ts": 1776649793330475, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793330483, "dur": 19, + "args": { + "External id": 97501, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97501, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97501, "pid": 0, "tid": 7, "ts": 1776649793330483, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330476, "dur": 4, + "args": { + "External id": 97501, + "cbid": 211, "correlation": 97501 + } + }, + { + "ph": "s", "id": 97501, "pid": 435, "tid": 435, "ts": 1776649793330476, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793330519, "dur": 1, + "args": { + "External id": 12827, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97519, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97519, "pid": 0, "tid": 7, "ts": 1776649793330519, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330510, "dur": 6, + "args": { + "External id": 12827, + "cbid": 211, "correlation": 97519 + } + }, + { + "ph": "s", "id": 97519, "pid": 435, "tid": 435, "ts": 1776649793330510, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330522, "dur": 0, + "args": { + "External id": 97525, + "cbid": 200, "correlation": 97525 + } + }, + { + "ph": "f", "id": 97525, "pid": 435, "tid": 435, "ts": 1776649793330522, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330522, "dur": 0, + "args": { + "External id": 97526, + "cbid": 273, "correlation": 97526 + } + }, + { + "ph": "f", "id": 97526, "pid": 435, "tid": 435, "ts": 1776649793330522, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793330531, "dur": 19, + "args": { + "External id": 97527, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97527, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97527, "pid": 0, "tid": 7, "ts": 1776649793330531, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330524, "dur": 4, + "args": { + "External id": 97527, + "cbid": 211, "correlation": 97527 + } + }, + { + "ph": "s", "id": 97527, "pid": 435, "tid": 435, "ts": 1776649793330524, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793330566, "dur": 1, + "args": { + "External id": 12832, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97545, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97545, "pid": 0, "tid": 7, "ts": 1776649793330566, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330558, "dur": 6, + "args": { + "External id": 12832, + "cbid": 211, "correlation": 97545 + } + }, + { + "ph": "s", "id": 97545, "pid": 435, "tid": 435, "ts": 1776649793330558, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330570, "dur": 0, + "args": { + "External id": 97551, + "cbid": 200, "correlation": 97551 + } + }, + { + "ph": "f", "id": 97551, "pid": 435, "tid": 435, "ts": 1776649793330570, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330570, "dur": 0, + "args": { + "External id": 97552, + "cbid": 273, "correlation": 97552 + } + }, + { + "ph": "f", "id": 97552, "pid": 435, "tid": 435, "ts": 1776649793330570, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793330579, "dur": 19, + "args": { + "External id": 97553, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97553, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97553, "pid": 0, "tid": 7, "ts": 1776649793330579, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330571, "dur": 4, + "args": { + "External id": 97553, + "cbid": 211, "correlation": 97553 + } + }, + { + "ph": "s", "id": 97553, "pid": 435, "tid": 435, "ts": 1776649793330571, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793330614, "dur": 1, + "args": { + "External id": 12837, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97571, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97571, "pid": 0, "tid": 7, "ts": 1776649793330614, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330605, "dur": 6, + "args": { + "External id": 12837, + "cbid": 211, "correlation": 97571 + } + }, + { + "ph": "s", "id": 97571, "pid": 435, "tid": 435, "ts": 1776649793330605, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330619, "dur": 0, + "args": { + "External id": 97577, + "cbid": 200, "correlation": 97577 + } + }, + { + "ph": "f", "id": 97577, "pid": 435, "tid": 435, "ts": 1776649793330619, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330620, "dur": 0, + "args": { + "External id": 97578, + "cbid": 273, "correlation": 97578 + } + }, + { + "ph": "f", "id": 97578, "pid": 435, "tid": 435, "ts": 1776649793330620, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793330629, "dur": 19, + "args": { + "External id": 97579, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97579, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97579, "pid": 0, "tid": 7, "ts": 1776649793330629, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330621, "dur": 4, + "args": { + "External id": 97579, + "cbid": 211, "correlation": 97579 + } + }, + { + "ph": "s", "id": 97579, "pid": 435, "tid": 435, "ts": 1776649793330621, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793330668, "dur": 1, + "args": { + "External id": 12842, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97597, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97597, "pid": 0, "tid": 7, "ts": 1776649793330668, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330659, "dur": 6, + "args": { + "External id": 12842, + "cbid": 211, "correlation": 97597 + } + }, + { + "ph": "s", "id": 97597, "pid": 435, "tid": 435, "ts": 1776649793330659, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330671, "dur": 0, + "args": { + "External id": 97603, + "cbid": 200, "correlation": 97603 + } + }, + { + "ph": "f", "id": 97603, "pid": 435, "tid": 435, "ts": 1776649793330671, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330672, "dur": 0, + "args": { + "External id": 97604, + "cbid": 273, "correlation": 97604 + } + }, + { + "ph": "f", "id": 97604, "pid": 435, "tid": 435, "ts": 1776649793330672, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793330680, "dur": 19, + "args": { + "External id": 97605, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97605, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97605, "pid": 0, "tid": 7, "ts": 1776649793330680, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330673, "dur": 4, + "args": { + "External id": 97605, + "cbid": 211, "correlation": 97605 + } + }, + { + "ph": "s", "id": 97605, "pid": 435, "tid": 435, "ts": 1776649793330673, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793330715, "dur": 1, + "args": { + "External id": 12847, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97623, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97623, "pid": 0, "tid": 7, "ts": 1776649793330715, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330706, "dur": 6, + "args": { + "External id": 12847, + "cbid": 211, "correlation": 97623 + } + }, + { + "ph": "s", "id": 97623, "pid": 435, "tid": 435, "ts": 1776649793330706, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330718, "dur": 0, + "args": { + "External id": 97629, + "cbid": 200, "correlation": 97629 + } + }, + { + "ph": "f", "id": 97629, "pid": 435, "tid": 435, "ts": 1776649793330718, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330719, "dur": 0, + "args": { + "External id": 97630, + "cbid": 273, "correlation": 97630 + } + }, + { + "ph": "f", "id": 97630, "pid": 435, "tid": 435, "ts": 1776649793330719, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793330727, "dur": 19, + "args": { + "External id": 97631, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97631, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97631, "pid": 0, "tid": 7, "ts": 1776649793330727, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330720, "dur": 4, + "args": { + "External id": 97631, + "cbid": 211, "correlation": 97631 + } + }, + { + "ph": "s", "id": 97631, "pid": 435, "tid": 435, "ts": 1776649793330720, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793330762, "dur": 1, + "args": { + "External id": 12852, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97649, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97649, "pid": 0, "tid": 7, "ts": 1776649793330762, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330753, "dur": 6, + "args": { + "External id": 12852, + "cbid": 211, "correlation": 97649 + } + }, + { + "ph": "s", "id": 97649, "pid": 435, "tid": 435, "ts": 1776649793330753, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330765, "dur": 0, + "args": { + "External id": 97655, + "cbid": 200, "correlation": 97655 + } + }, + { + "ph": "f", "id": 97655, "pid": 435, "tid": 435, "ts": 1776649793330765, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330766, "dur": 0, + "args": { + "External id": 97656, + "cbid": 273, "correlation": 97656 + } + }, + { + "ph": "f", "id": 97656, "pid": 435, "tid": 435, "ts": 1776649793330766, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793330775, "dur": 19, + "args": { + "External id": 97657, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97657, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97657, "pid": 0, "tid": 7, "ts": 1776649793330775, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330767, "dur": 4, + "args": { + "External id": 97657, + "cbid": 211, "correlation": 97657 + } + }, + { + "ph": "s", "id": 97657, "pid": 435, "tid": 435, "ts": 1776649793330767, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793330810, "dur": 1, + "args": { + "External id": 12857, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97675, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97675, "pid": 0, "tid": 7, "ts": 1776649793330810, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330801, "dur": 6, + "args": { + "External id": 12857, + "cbid": 211, "correlation": 97675 + } + }, + { + "ph": "s", "id": 97675, "pid": 435, "tid": 435, "ts": 1776649793330801, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330813, "dur": 0, + "args": { + "External id": 97681, + "cbid": 200, "correlation": 97681 + } + }, + { + "ph": "f", "id": 97681, "pid": 435, "tid": 435, "ts": 1776649793330813, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330813, "dur": 0, + "args": { + "External id": 97682, + "cbid": 273, "correlation": 97682 + } + }, + { + "ph": "f", "id": 97682, "pid": 435, "tid": 435, "ts": 1776649793330813, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793330822, "dur": 19, + "args": { + "External id": 97683, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97683, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97683, "pid": 0, "tid": 7, "ts": 1776649793330822, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330815, "dur": 4, + "args": { + "External id": 97683, + "cbid": 211, "correlation": 97683 + } + }, + { + "ph": "s", "id": 97683, "pid": 435, "tid": 435, "ts": 1776649793330815, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793330857, "dur": 1, + "args": { + "External id": 12862, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97701, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97701, "pid": 0, "tid": 7, "ts": 1776649793330857, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330848, "dur": 6, + "args": { + "External id": 12862, + "cbid": 211, "correlation": 97701 + } + }, + { + "ph": "s", "id": 97701, "pid": 435, "tid": 435, "ts": 1776649793330848, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330860, "dur": 0, + "args": { + "External id": 97707, + "cbid": 200, "correlation": 97707 + } + }, + { + "ph": "f", "id": 97707, "pid": 435, "tid": 435, "ts": 1776649793330860, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330861, "dur": 0, + "args": { + "External id": 97708, + "cbid": 273, "correlation": 97708 + } + }, + { + "ph": "f", "id": 97708, "pid": 435, "tid": 435, "ts": 1776649793330861, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793330870, "dur": 19, + "args": { + "External id": 97709, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97709, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97709, "pid": 0, "tid": 7, "ts": 1776649793330870, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330862, "dur": 5, + "args": { + "External id": 97709, + "cbid": 211, "correlation": 97709 + } + }, + { + "ph": "s", "id": 97709, "pid": 435, "tid": 435, "ts": 1776649793330862, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793330906, "dur": 1, + "args": { + "External id": 12867, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97727, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97727, "pid": 0, "tid": 7, "ts": 1776649793330906, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330897, "dur": 6, + "args": { + "External id": 12867, + "cbid": 211, "correlation": 97727 + } + }, + { + "ph": "s", "id": 97727, "pid": 435, "tid": 435, "ts": 1776649793330897, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330909, "dur": 0, + "args": { + "External id": 97733, + "cbid": 200, "correlation": 97733 + } + }, + { + "ph": "f", "id": 97733, "pid": 435, "tid": 435, "ts": 1776649793330909, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330909, "dur": 0, + "args": { + "External id": 97734, + "cbid": 273, "correlation": 97734 + } + }, + { + "ph": "f", "id": 97734, "pid": 435, "tid": 435, "ts": 1776649793330909, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793330918, "dur": 19, + "args": { + "External id": 97735, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97735, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97735, "pid": 0, "tid": 7, "ts": 1776649793330918, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330910, "dur": 4, + "args": { + "External id": 97735, + "cbid": 211, "correlation": 97735 + } + }, + { + "ph": "s", "id": 97735, "pid": 435, "tid": 435, "ts": 1776649793330910, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793330953, "dur": 1, + "args": { + "External id": 12872, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97753, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97753, "pid": 0, "tid": 7, "ts": 1776649793330953, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330944, "dur": 6, + "args": { + "External id": 12872, + "cbid": 211, "correlation": 97753 + } + }, + { + "ph": "s", "id": 97753, "pid": 435, "tid": 435, "ts": 1776649793330944, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330956, "dur": 0, + "args": { + "External id": 97759, + "cbid": 200, "correlation": 97759 + } + }, + { + "ph": "f", "id": 97759, "pid": 435, "tid": 435, "ts": 1776649793330956, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793330957, "dur": 0, + "args": { + "External id": 97760, + "cbid": 273, "correlation": 97760 + } + }, + { + "ph": "f", "id": 97760, "pid": 435, "tid": 435, "ts": 1776649793330957, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793330966, "dur": 19, + "args": { + "External id": 97761, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97761, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97761, "pid": 0, "tid": 7, "ts": 1776649793330966, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330958, "dur": 4, + "args": { + "External id": 97761, + "cbid": 211, "correlation": 97761 + } + }, + { + "ph": "s", "id": 97761, "pid": 435, "tid": 435, "ts": 1776649793330958, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793331001, "dur": 1, + "args": { + "External id": 12877, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97779, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97779, "pid": 0, "tid": 7, "ts": 1776649793331001, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793330992, "dur": 6, + "args": { + "External id": 12877, + "cbid": 211, "correlation": 97779 + } + }, + { + "ph": "s", "id": 97779, "pid": 435, "tid": 435, "ts": 1776649793330992, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331004, "dur": 0, + "args": { + "External id": 97785, + "cbid": 200, "correlation": 97785 + } + }, + { + "ph": "f", "id": 97785, "pid": 435, "tid": 435, "ts": 1776649793331004, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331005, "dur": 0, + "args": { + "External id": 97786, + "cbid": 273, "correlation": 97786 + } + }, + { + "ph": "f", "id": 97786, "pid": 435, "tid": 435, "ts": 1776649793331005, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793331013, "dur": 19, + "args": { + "External id": 97787, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97787, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97787, "pid": 0, "tid": 7, "ts": 1776649793331013, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331006, "dur": 5, + "args": { + "External id": 97787, + "cbid": 211, "correlation": 97787 + } + }, + { + "ph": "s", "id": 97787, "pid": 435, "tid": 435, "ts": 1776649793331006, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793331048, "dur": 1, + "args": { + "External id": 12882, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97805, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97805, "pid": 0, "tid": 7, "ts": 1776649793331048, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331040, "dur": 6, + "args": { + "External id": 12882, + "cbid": 211, "correlation": 97805 + } + }, + { + "ph": "s", "id": 97805, "pid": 435, "tid": 435, "ts": 1776649793331040, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331052, "dur": 0, + "args": { + "External id": 97811, + "cbid": 200, "correlation": 97811 + } + }, + { + "ph": "f", "id": 97811, "pid": 435, "tid": 435, "ts": 1776649793331052, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331052, "dur": 0, + "args": { + "External id": 97812, + "cbid": 273, "correlation": 97812 + } + }, + { + "ph": "f", "id": 97812, "pid": 435, "tid": 435, "ts": 1776649793331052, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793331061, "dur": 19, + "args": { + "External id": 97813, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97813, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97813, "pid": 0, "tid": 7, "ts": 1776649793331061, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331054, "dur": 4, + "args": { + "External id": 97813, + "cbid": 211, "correlation": 97813 + } + }, + { + "ph": "s", "id": 97813, "pid": 435, "tid": 435, "ts": 1776649793331054, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793331096, "dur": 1, + "args": { + "External id": 12887, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97831, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97831, "pid": 0, "tid": 7, "ts": 1776649793331096, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331087, "dur": 6, + "args": { + "External id": 12887, + "cbid": 211, "correlation": 97831 + } + }, + { + "ph": "s", "id": 97831, "pid": 435, "tid": 435, "ts": 1776649793331087, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331100, "dur": 0, + "args": { + "External id": 97837, + "cbid": 200, "correlation": 97837 + } + }, + { + "ph": "f", "id": 97837, "pid": 435, "tid": 435, "ts": 1776649793331100, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331100, "dur": 0, + "args": { + "External id": 97838, + "cbid": 273, "correlation": 97838 + } + }, + { + "ph": "f", "id": 97838, "pid": 435, "tid": 435, "ts": 1776649793331100, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793331109, "dur": 19, + "args": { + "External id": 97839, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97839, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97839, "pid": 0, "tid": 7, "ts": 1776649793331109, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331101, "dur": 4, + "args": { + "External id": 97839, + "cbid": 211, "correlation": 97839 + } + }, + { + "ph": "s", "id": 97839, "pid": 435, "tid": 435, "ts": 1776649793331101, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793331144, "dur": 1, + "args": { + "External id": 12892, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97857, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97857, "pid": 0, "tid": 7, "ts": 1776649793331144, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331135, "dur": 6, + "args": { + "External id": 12892, + "cbid": 211, "correlation": 97857 + } + }, + { + "ph": "s", "id": 97857, "pid": 435, "tid": 435, "ts": 1776649793331135, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331147, "dur": 0, + "args": { + "External id": 97863, + "cbid": 200, "correlation": 97863 + } + }, + { + "ph": "f", "id": 97863, "pid": 435, "tid": 435, "ts": 1776649793331147, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331148, "dur": 0, + "args": { + "External id": 97864, + "cbid": 273, "correlation": 97864 + } + }, + { + "ph": "f", "id": 97864, "pid": 435, "tid": 435, "ts": 1776649793331148, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793331156, "dur": 19, + "args": { + "External id": 97865, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97865, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97865, "pid": 0, "tid": 7, "ts": 1776649793331156, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331149, "dur": 4, + "args": { + "External id": 97865, + "cbid": 211, "correlation": 97865 + } + }, + { + "ph": "s", "id": 97865, "pid": 435, "tid": 435, "ts": 1776649793331149, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793331191, "dur": 1, + "args": { + "External id": 12897, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97883, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97883, "pid": 0, "tid": 7, "ts": 1776649793331191, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331182, "dur": 6, + "args": { + "External id": 12897, + "cbid": 211, "correlation": 97883 + } + }, + { + "ph": "s", "id": 97883, "pid": 435, "tid": 435, "ts": 1776649793331182, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331194, "dur": 0, + "args": { + "External id": 97889, + "cbid": 200, "correlation": 97889 + } + }, + { + "ph": "f", "id": 97889, "pid": 435, "tid": 435, "ts": 1776649793331194, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331195, "dur": 0, + "args": { + "External id": 97890, + "cbid": 273, "correlation": 97890 + } + }, + { + "ph": "f", "id": 97890, "pid": 435, "tid": 435, "ts": 1776649793331195, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793331203, "dur": 19, + "args": { + "External id": 97891, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97891, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97891, "pid": 0, "tid": 7, "ts": 1776649793331203, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331196, "dur": 4, + "args": { + "External id": 97891, + "cbid": 211, "correlation": 97891 + } + }, + { + "ph": "s", "id": 97891, "pid": 435, "tid": 435, "ts": 1776649793331196, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793331238, "dur": 1, + "args": { + "External id": 12902, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97909, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97909, "pid": 0, "tid": 7, "ts": 1776649793331238, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331229, "dur": 6, + "args": { + "External id": 12902, + "cbid": 211, "correlation": 97909 + } + }, + { + "ph": "s", "id": 97909, "pid": 435, "tid": 435, "ts": 1776649793331229, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331241, "dur": 0, + "args": { + "External id": 97915, + "cbid": 200, "correlation": 97915 + } + }, + { + "ph": "f", "id": 97915, "pid": 435, "tid": 435, "ts": 1776649793331241, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331242, "dur": 0, + "args": { + "External id": 97916, + "cbid": 273, "correlation": 97916 + } + }, + { + "ph": "f", "id": 97916, "pid": 435, "tid": 435, "ts": 1776649793331242, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793331250, "dur": 19, + "args": { + "External id": 97917, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97917, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97917, "pid": 0, "tid": 7, "ts": 1776649793331250, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331243, "dur": 4, + "args": { + "External id": 97917, + "cbid": 211, "correlation": 97917 + } + }, + { + "ph": "s", "id": 97917, "pid": 435, "tid": 435, "ts": 1776649793331243, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793331285, "dur": 1, + "args": { + "External id": 12907, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97935, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97935, "pid": 0, "tid": 7, "ts": 1776649793331285, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331276, "dur": 6, + "args": { + "External id": 12907, + "cbid": 211, "correlation": 97935 + } + }, + { + "ph": "s", "id": 97935, "pid": 435, "tid": 435, "ts": 1776649793331276, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331288, "dur": 0, + "args": { + "External id": 97941, + "cbid": 200, "correlation": 97941 + } + }, + { + "ph": "f", "id": 97941, "pid": 435, "tid": 435, "ts": 1776649793331288, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331289, "dur": 0, + "args": { + "External id": 97942, + "cbid": 273, "correlation": 97942 + } + }, + { + "ph": "f", "id": 97942, "pid": 435, "tid": 435, "ts": 1776649793331289, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793331298, "dur": 19, + "args": { + "External id": 97943, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97943, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97943, "pid": 0, "tid": 7, "ts": 1776649793331298, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331290, "dur": 4, + "args": { + "External id": 97943, + "cbid": 211, "correlation": 97943 + } + }, + { + "ph": "s", "id": 97943, "pid": 435, "tid": 435, "ts": 1776649793331290, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793331333, "dur": 1, + "args": { + "External id": 12912, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97961, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97961, "pid": 0, "tid": 7, "ts": 1776649793331333, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331324, "dur": 6, + "args": { + "External id": 12912, + "cbid": 211, "correlation": 97961 + } + }, + { + "ph": "s", "id": 97961, "pid": 435, "tid": 435, "ts": 1776649793331324, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331336, "dur": 0, + "args": { + "External id": 97967, + "cbid": 200, "correlation": 97967 + } + }, + { + "ph": "f", "id": 97967, "pid": 435, "tid": 435, "ts": 1776649793331336, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331336, "dur": 0, + "args": { + "External id": 97968, + "cbid": 273, "correlation": 97968 + } + }, + { + "ph": "f", "id": 97968, "pid": 435, "tid": 435, "ts": 1776649793331336, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793331345, "dur": 19, + "args": { + "External id": 97969, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97969, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97969, "pid": 0, "tid": 7, "ts": 1776649793331345, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331338, "dur": 4, + "args": { + "External id": 97969, + "cbid": 211, "correlation": 97969 + } + }, + { + "ph": "s", "id": 97969, "pid": 435, "tid": 435, "ts": 1776649793331338, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793331380, "dur": 1, + "args": { + "External id": 12917, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97987, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97987, "pid": 0, "tid": 7, "ts": 1776649793331380, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331371, "dur": 6, + "args": { + "External id": 12917, + "cbid": 211, "correlation": 97987 + } + }, + { + "ph": "s", "id": 97987, "pid": 435, "tid": 435, "ts": 1776649793331371, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331383, "dur": 0, + "args": { + "External id": 97993, + "cbid": 200, "correlation": 97993 + } + }, + { + "ph": "f", "id": 97993, "pid": 435, "tid": 435, "ts": 1776649793331383, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331384, "dur": 0, + "args": { + "External id": 97994, + "cbid": 273, "correlation": 97994 + } + }, + { + "ph": "f", "id": 97994, "pid": 435, "tid": 435, "ts": 1776649793331384, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793331392, "dur": 19, + "args": { + "External id": 97995, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 97995, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 97995, "pid": 0, "tid": 7, "ts": 1776649793331392, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331385, "dur": 4, + "args": { + "External id": 97995, + "cbid": 211, "correlation": 97995 + } + }, + { + "ph": "s", "id": 97995, "pid": 435, "tid": 435, "ts": 1776649793331385, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793331427, "dur": 1, + "args": { + "External id": 12922, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98013, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98013, "pid": 0, "tid": 7, "ts": 1776649793331427, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331419, "dur": 6, + "args": { + "External id": 12922, + "cbid": 211, "correlation": 98013 + } + }, + { + "ph": "s", "id": 98013, "pid": 435, "tid": 435, "ts": 1776649793331419, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331431, "dur": 0, + "args": { + "External id": 98019, + "cbid": 200, "correlation": 98019 + } + }, + { + "ph": "f", "id": 98019, "pid": 435, "tid": 435, "ts": 1776649793331431, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331431, "dur": 0, + "args": { + "External id": 98020, + "cbid": 273, "correlation": 98020 + } + }, + { + "ph": "f", "id": 98020, "pid": 435, "tid": 435, "ts": 1776649793331431, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793331440, "dur": 19, + "args": { + "External id": 98021, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98021, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98021, "pid": 0, "tid": 7, "ts": 1776649793331440, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331432, "dur": 4, + "args": { + "External id": 98021, + "cbid": 211, "correlation": 98021 + } + }, + { + "ph": "s", "id": 98021, "pid": 435, "tid": 435, "ts": 1776649793331432, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793331475, "dur": 1, + "args": { + "External id": 12927, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98039, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98039, "pid": 0, "tid": 7, "ts": 1776649793331475, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331466, "dur": 6, + "args": { + "External id": 12927, + "cbid": 211, "correlation": 98039 + } + }, + { + "ph": "s", "id": 98039, "pid": 435, "tid": 435, "ts": 1776649793331466, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331478, "dur": 0, + "args": { + "External id": 98045, + "cbid": 200, "correlation": 98045 + } + }, + { + "ph": "f", "id": 98045, "pid": 435, "tid": 435, "ts": 1776649793331478, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331479, "dur": 0, + "args": { + "External id": 98046, + "cbid": 273, "correlation": 98046 + } + }, + { + "ph": "f", "id": 98046, "pid": 435, "tid": 435, "ts": 1776649793331479, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793331487, "dur": 19, + "args": { + "External id": 98047, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98047, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98047, "pid": 0, "tid": 7, "ts": 1776649793331487, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331480, "dur": 4, + "args": { + "External id": 98047, + "cbid": 211, "correlation": 98047 + } + }, + { + "ph": "s", "id": 98047, "pid": 435, "tid": 435, "ts": 1776649793331480, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793331522, "dur": 1, + "args": { + "External id": 12932, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98065, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98065, "pid": 0, "tid": 7, "ts": 1776649793331522, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331514, "dur": 6, + "args": { + "External id": 12932, + "cbid": 211, "correlation": 98065 + } + }, + { + "ph": "s", "id": 98065, "pid": 435, "tid": 435, "ts": 1776649793331514, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331526, "dur": 0, + "args": { + "External id": 98071, + "cbid": 200, "correlation": 98071 + } + }, + { + "ph": "f", "id": 98071, "pid": 435, "tid": 435, "ts": 1776649793331526, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331526, "dur": 0, + "args": { + "External id": 98072, + "cbid": 273, "correlation": 98072 + } + }, + { + "ph": "f", "id": 98072, "pid": 435, "tid": 435, "ts": 1776649793331526, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793331535, "dur": 19, + "args": { + "External id": 98073, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98073, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98073, "pid": 0, "tid": 7, "ts": 1776649793331535, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331527, "dur": 4, + "args": { + "External id": 98073, + "cbid": 211, "correlation": 98073 + } + }, + { + "ph": "s", "id": 98073, "pid": 435, "tid": 435, "ts": 1776649793331527, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793331570, "dur": 1, + "args": { + "External id": 12937, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98091, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98091, "pid": 0, "tid": 7, "ts": 1776649793331570, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331561, "dur": 6, + "args": { + "External id": 12937, + "cbid": 211, "correlation": 98091 + } + }, + { + "ph": "s", "id": 98091, "pid": 435, "tid": 435, "ts": 1776649793331561, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331573, "dur": 0, + "args": { + "External id": 98097, + "cbid": 200, "correlation": 98097 + } + }, + { + "ph": "f", "id": 98097, "pid": 435, "tid": 435, "ts": 1776649793331573, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331574, "dur": 0, + "args": { + "External id": 98098, + "cbid": 273, "correlation": 98098 + } + }, + { + "ph": "f", "id": 98098, "pid": 435, "tid": 435, "ts": 1776649793331574, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793331582, "dur": 19, + "args": { + "External id": 98099, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98099, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98099, "pid": 0, "tid": 7, "ts": 1776649793331582, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331575, "dur": 4, + "args": { + "External id": 98099, + "cbid": 211, "correlation": 98099 + } + }, + { + "ph": "s", "id": 98099, "pid": 435, "tid": 435, "ts": 1776649793331575, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793331617, "dur": 1, + "args": { + "External id": 12942, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98117, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98117, "pid": 0, "tid": 7, "ts": 1776649793331617, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331608, "dur": 6, + "args": { + "External id": 12942, + "cbid": 211, "correlation": 98117 + } + }, + { + "ph": "s", "id": 98117, "pid": 435, "tid": 435, "ts": 1776649793331608, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331620, "dur": 0, + "args": { + "External id": 98123, + "cbid": 200, "correlation": 98123 + } + }, + { + "ph": "f", "id": 98123, "pid": 435, "tid": 435, "ts": 1776649793331620, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331621, "dur": 0, + "args": { + "External id": 98124, + "cbid": 273, "correlation": 98124 + } + }, + { + "ph": "f", "id": 98124, "pid": 435, "tid": 435, "ts": 1776649793331621, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793331630, "dur": 19, + "args": { + "External id": 98125, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98125, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98125, "pid": 0, "tid": 7, "ts": 1776649793331630, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331622, "dur": 5, + "args": { + "External id": 98125, + "cbid": 211, "correlation": 98125 + } + }, + { + "ph": "s", "id": 98125, "pid": 435, "tid": 435, "ts": 1776649793331622, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793331670, "dur": 1, + "args": { + "External id": 12947, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98143, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98143, "pid": 0, "tid": 7, "ts": 1776649793331670, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331661, "dur": 6, + "args": { + "External id": 12947, + "cbid": 211, "correlation": 98143 + } + }, + { + "ph": "s", "id": 98143, "pid": 435, "tid": 435, "ts": 1776649793331661, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331673, "dur": 0, + "args": { + "External id": 98149, + "cbid": 200, "correlation": 98149 + } + }, + { + "ph": "f", "id": 98149, "pid": 435, "tid": 435, "ts": 1776649793331673, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331674, "dur": 0, + "args": { + "External id": 98150, + "cbid": 273, "correlation": 98150 + } + }, + { + "ph": "f", "id": 98150, "pid": 435, "tid": 435, "ts": 1776649793331674, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793331682, "dur": 19, + "args": { + "External id": 98151, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98151, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98151, "pid": 0, "tid": 7, "ts": 1776649793331682, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331675, "dur": 4, + "args": { + "External id": 98151, + "cbid": 211, "correlation": 98151 + } + }, + { + "ph": "s", "id": 98151, "pid": 435, "tid": 435, "ts": 1776649793331675, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793331717, "dur": 1, + "args": { + "External id": 12952, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98169, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98169, "pid": 0, "tid": 7, "ts": 1776649793331717, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331709, "dur": 6, + "args": { + "External id": 12952, + "cbid": 211, "correlation": 98169 + } + }, + { + "ph": "s", "id": 98169, "pid": 435, "tid": 435, "ts": 1776649793331709, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331720, "dur": 0, + "args": { + "External id": 98175, + "cbid": 200, "correlation": 98175 + } + }, + { + "ph": "f", "id": 98175, "pid": 435, "tid": 435, "ts": 1776649793331720, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331721, "dur": 0, + "args": { + "External id": 98176, + "cbid": 273, "correlation": 98176 + } + }, + { + "ph": "f", "id": 98176, "pid": 435, "tid": 435, "ts": 1776649793331721, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793331730, "dur": 19, + "args": { + "External id": 98177, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98177, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98177, "pid": 0, "tid": 7, "ts": 1776649793331730, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331722, "dur": 4, + "args": { + "External id": 98177, + "cbid": 211, "correlation": 98177 + } + }, + { + "ph": "s", "id": 98177, "pid": 435, "tid": 435, "ts": 1776649793331722, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793331765, "dur": 1, + "args": { + "External id": 12957, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98195, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98195, "pid": 0, "tid": 7, "ts": 1776649793331765, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331756, "dur": 6, + "args": { + "External id": 12957, + "cbid": 211, "correlation": 98195 + } + }, + { + "ph": "s", "id": 98195, "pid": 435, "tid": 435, "ts": 1776649793331756, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331768, "dur": 0, + "args": { + "External id": 98201, + "cbid": 200, "correlation": 98201 + } + }, + { + "ph": "f", "id": 98201, "pid": 435, "tid": 435, "ts": 1776649793331768, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331769, "dur": 0, + "args": { + "External id": 98202, + "cbid": 273, "correlation": 98202 + } + }, + { + "ph": "f", "id": 98202, "pid": 435, "tid": 435, "ts": 1776649793331769, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793331777, "dur": 19, + "args": { + "External id": 98203, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98203, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98203, "pid": 0, "tid": 7, "ts": 1776649793331777, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331770, "dur": 4, + "args": { + "External id": 98203, + "cbid": 211, "correlation": 98203 + } + }, + { + "ph": "s", "id": 98203, "pid": 435, "tid": 435, "ts": 1776649793331770, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793331812, "dur": 1, + "args": { + "External id": 12962, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98221, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98221, "pid": 0, "tid": 7, "ts": 1776649793331812, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331803, "dur": 6, + "args": { + "External id": 12962, + "cbid": 211, "correlation": 98221 + } + }, + { + "ph": "s", "id": 98221, "pid": 435, "tid": 435, "ts": 1776649793331803, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331815, "dur": 0, + "args": { + "External id": 98227, + "cbid": 200, "correlation": 98227 + } + }, + { + "ph": "f", "id": 98227, "pid": 435, "tid": 435, "ts": 1776649793331815, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331816, "dur": 0, + "args": { + "External id": 98228, + "cbid": 273, "correlation": 98228 + } + }, + { + "ph": "f", "id": 98228, "pid": 435, "tid": 435, "ts": 1776649793331816, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793331824, "dur": 19, + "args": { + "External id": 98229, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98229, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98229, "pid": 0, "tid": 7, "ts": 1776649793331824, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331817, "dur": 4, + "args": { + "External id": 98229, + "cbid": 211, "correlation": 98229 + } + }, + { + "ph": "s", "id": 98229, "pid": 435, "tid": 435, "ts": 1776649793331817, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793331859, "dur": 1, + "args": { + "External id": 12967, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98247, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98247, "pid": 0, "tid": 7, "ts": 1776649793331859, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331850, "dur": 6, + "args": { + "External id": 12967, + "cbid": 211, "correlation": 98247 + } + }, + { + "ph": "s", "id": 98247, "pid": 435, "tid": 435, "ts": 1776649793331850, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331862, "dur": 0, + "args": { + "External id": 98253, + "cbid": 200, "correlation": 98253 + } + }, + { + "ph": "f", "id": 98253, "pid": 435, "tid": 435, "ts": 1776649793331862, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331863, "dur": 0, + "args": { + "External id": 98254, + "cbid": 273, "correlation": 98254 + } + }, + { + "ph": "f", "id": 98254, "pid": 435, "tid": 435, "ts": 1776649793331863, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793331871, "dur": 19, + "args": { + "External id": 98255, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98255, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98255, "pid": 0, "tid": 7, "ts": 1776649793331871, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331864, "dur": 4, + "args": { + "External id": 98255, + "cbid": 211, "correlation": 98255 + } + }, + { + "ph": "s", "id": 98255, "pid": 435, "tid": 435, "ts": 1776649793331864, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793331906, "dur": 1, + "args": { + "External id": 12972, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98273, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98273, "pid": 0, "tid": 7, "ts": 1776649793331906, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331897, "dur": 6, + "args": { + "External id": 12972, + "cbid": 211, "correlation": 98273 + } + }, + { + "ph": "s", "id": 98273, "pid": 435, "tid": 435, "ts": 1776649793331897, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331909, "dur": 0, + "args": { + "External id": 98279, + "cbid": 200, "correlation": 98279 + } + }, + { + "ph": "f", "id": 98279, "pid": 435, "tid": 435, "ts": 1776649793331909, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331910, "dur": 0, + "args": { + "External id": 98280, + "cbid": 273, "correlation": 98280 + } + }, + { + "ph": "f", "id": 98280, "pid": 435, "tid": 435, "ts": 1776649793331910, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793331919, "dur": 19, + "args": { + "External id": 98281, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98281, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98281, "pid": 0, "tid": 7, "ts": 1776649793331919, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331911, "dur": 4, + "args": { + "External id": 98281, + "cbid": 211, "correlation": 98281 + } + }, + { + "ph": "s", "id": 98281, "pid": 435, "tid": 435, "ts": 1776649793331911, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793331954, "dur": 1, + "args": { + "External id": 12977, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98299, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98299, "pid": 0, "tid": 7, "ts": 1776649793331954, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331945, "dur": 6, + "args": { + "External id": 12977, + "cbid": 211, "correlation": 98299 + } + }, + { + "ph": "s", "id": 98299, "pid": 435, "tid": 435, "ts": 1776649793331945, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331957, "dur": 0, + "args": { + "External id": 98305, + "cbid": 200, "correlation": 98305 + } + }, + { + "ph": "f", "id": 98305, "pid": 435, "tid": 435, "ts": 1776649793331957, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793331958, "dur": 0, + "args": { + "External id": 98306, + "cbid": 273, "correlation": 98306 + } + }, + { + "ph": "f", "id": 98306, "pid": 435, "tid": 435, "ts": 1776649793331958, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793331966, "dur": 19, + "args": { + "External id": 98307, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98307, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98307, "pid": 0, "tid": 7, "ts": 1776649793331966, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331959, "dur": 4, + "args": { + "External id": 98307, + "cbid": 211, "correlation": 98307 + } + }, + { + "ph": "s", "id": 98307, "pid": 435, "tid": 435, "ts": 1776649793331959, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793332001, "dur": 1, + "args": { + "External id": 12982, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98325, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98325, "pid": 0, "tid": 7, "ts": 1776649793332001, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793331993, "dur": 6, + "args": { + "External id": 12982, + "cbid": 211, "correlation": 98325 + } + }, + { + "ph": "s", "id": 98325, "pid": 435, "tid": 435, "ts": 1776649793331993, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332005, "dur": 0, + "args": { + "External id": 98331, + "cbid": 200, "correlation": 98331 + } + }, + { + "ph": "f", "id": 98331, "pid": 435, "tid": 435, "ts": 1776649793332005, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332005, "dur": 0, + "args": { + "External id": 98332, + "cbid": 273, "correlation": 98332 + } + }, + { + "ph": "f", "id": 98332, "pid": 435, "tid": 435, "ts": 1776649793332005, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793332014, "dur": 19, + "args": { + "External id": 98333, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98333, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98333, "pid": 0, "tid": 7, "ts": 1776649793332014, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332006, "dur": 4, + "args": { + "External id": 98333, + "cbid": 211, "correlation": 98333 + } + }, + { + "ph": "s", "id": 98333, "pid": 435, "tid": 435, "ts": 1776649793332006, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793332049, "dur": 1, + "args": { + "External id": 12987, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98351, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98351, "pid": 0, "tid": 7, "ts": 1776649793332049, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332040, "dur": 6, + "args": { + "External id": 12987, + "cbid": 211, "correlation": 98351 + } + }, + { + "ph": "s", "id": 98351, "pid": 435, "tid": 435, "ts": 1776649793332040, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332052, "dur": 0, + "args": { + "External id": 98357, + "cbid": 200, "correlation": 98357 + } + }, + { + "ph": "f", "id": 98357, "pid": 435, "tid": 435, "ts": 1776649793332052, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332052, "dur": 0, + "args": { + "External id": 98358, + "cbid": 273, "correlation": 98358 + } + }, + { + "ph": "f", "id": 98358, "pid": 435, "tid": 435, "ts": 1776649793332052, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793332061, "dur": 19, + "args": { + "External id": 98359, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98359, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98359, "pid": 0, "tid": 7, "ts": 1776649793332061, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332054, "dur": 4, + "args": { + "External id": 98359, + "cbid": 211, "correlation": 98359 + } + }, + { + "ph": "s", "id": 98359, "pid": 435, "tid": 435, "ts": 1776649793332054, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793332096, "dur": 1, + "args": { + "External id": 12992, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98377, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98377, "pid": 0, "tid": 7, "ts": 1776649793332096, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332087, "dur": 6, + "args": { + "External id": 12992, + "cbid": 211, "correlation": 98377 + } + }, + { + "ph": "s", "id": 98377, "pid": 435, "tid": 435, "ts": 1776649793332087, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332100, "dur": 0, + "args": { + "External id": 98383, + "cbid": 200, "correlation": 98383 + } + }, + { + "ph": "f", "id": 98383, "pid": 435, "tid": 435, "ts": 1776649793332100, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332100, "dur": 0, + "args": { + "External id": 98384, + "cbid": 273, "correlation": 98384 + } + }, + { + "ph": "f", "id": 98384, "pid": 435, "tid": 435, "ts": 1776649793332100, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793332109, "dur": 19, + "args": { + "External id": 98385, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98385, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98385, "pid": 0, "tid": 7, "ts": 1776649793332109, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332101, "dur": 4, + "args": { + "External id": 98385, + "cbid": 211, "correlation": 98385 + } + }, + { + "ph": "s", "id": 98385, "pid": 435, "tid": 435, "ts": 1776649793332101, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793332143, "dur": 1, + "args": { + "External id": 12997, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98403, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98403, "pid": 0, "tid": 7, "ts": 1776649793332143, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332135, "dur": 6, + "args": { + "External id": 12997, + "cbid": 211, "correlation": 98403 + } + }, + { + "ph": "s", "id": 98403, "pid": 435, "tid": 435, "ts": 1776649793332135, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332147, "dur": 0, + "args": { + "External id": 98409, + "cbid": 200, "correlation": 98409 + } + }, + { + "ph": "f", "id": 98409, "pid": 435, "tid": 435, "ts": 1776649793332147, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332147, "dur": 0, + "args": { + "External id": 98410, + "cbid": 273, "correlation": 98410 + } + }, + { + "ph": "f", "id": 98410, "pid": 435, "tid": 435, "ts": 1776649793332147, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793332156, "dur": 19, + "args": { + "External id": 98411, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98411, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98411, "pid": 0, "tid": 7, "ts": 1776649793332156, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332148, "dur": 4, + "args": { + "External id": 98411, + "cbid": 211, "correlation": 98411 + } + }, + { + "ph": "s", "id": 98411, "pid": 435, "tid": 435, "ts": 1776649793332148, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793332191, "dur": 1, + "args": { + "External id": 13002, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98429, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98429, "pid": 0, "tid": 7, "ts": 1776649793332191, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332182, "dur": 6, + "args": { + "External id": 13002, + "cbid": 211, "correlation": 98429 + } + }, + { + "ph": "s", "id": 98429, "pid": 435, "tid": 435, "ts": 1776649793332182, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332194, "dur": 0, + "args": { + "External id": 98435, + "cbid": 200, "correlation": 98435 + } + }, + { + "ph": "f", "id": 98435, "pid": 435, "tid": 435, "ts": 1776649793332194, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332195, "dur": 0, + "args": { + "External id": 98436, + "cbid": 273, "correlation": 98436 + } + }, + { + "ph": "f", "id": 98436, "pid": 435, "tid": 435, "ts": 1776649793332195, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793332203, "dur": 19, + "args": { + "External id": 98437, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98437, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98437, "pid": 0, "tid": 7, "ts": 1776649793332203, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332196, "dur": 4, + "args": { + "External id": 98437, + "cbid": 211, "correlation": 98437 + } + }, + { + "ph": "s", "id": 98437, "pid": 435, "tid": 435, "ts": 1776649793332196, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793332262, "dur": 1, + "args": { + "External id": 13007, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98455, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98455, "pid": 0, "tid": 7, "ts": 1776649793332262, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332253, "dur": 7, + "args": { + "External id": 13007, + "cbid": 211, "correlation": 98455 + } + }, + { + "ph": "s", "id": 98455, "pid": 435, "tid": 435, "ts": 1776649793332253, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332267, "dur": 0, + "args": { + "External id": 98461, + "cbid": 200, "correlation": 98461 + } + }, + { + "ph": "f", "id": 98461, "pid": 435, "tid": 435, "ts": 1776649793332267, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332267, "dur": 0, + "args": { + "External id": 98462, + "cbid": 273, "correlation": 98462 + } + }, + { + "ph": "f", "id": 98462, "pid": 435, "tid": 435, "ts": 1776649793332267, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793332276, "dur": 19, + "args": { + "External id": 98463, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98463, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98463, "pid": 0, "tid": 7, "ts": 1776649793332276, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332269, "dur": 4, + "args": { + "External id": 98463, + "cbid": 211, "correlation": 98463 + } + }, + { + "ph": "s", "id": 98463, "pid": 435, "tid": 435, "ts": 1776649793332269, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793332313, "dur": 1, + "args": { + "External id": 13012, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98481, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98481, "pid": 0, "tid": 7, "ts": 1776649793332313, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332304, "dur": 6, + "args": { + "External id": 13012, + "cbid": 211, "correlation": 98481 + } + }, + { + "ph": "s", "id": 98481, "pid": 435, "tid": 435, "ts": 1776649793332304, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332316, "dur": 0, + "args": { + "External id": 98487, + "cbid": 200, "correlation": 98487 + } + }, + { + "ph": "f", "id": 98487, "pid": 435, "tid": 435, "ts": 1776649793332316, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332317, "dur": 0, + "args": { + "External id": 98488, + "cbid": 273, "correlation": 98488 + } + }, + { + "ph": "f", "id": 98488, "pid": 435, "tid": 435, "ts": 1776649793332317, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793332326, "dur": 19, + "args": { + "External id": 98489, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98489, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98489, "pid": 0, "tid": 7, "ts": 1776649793332326, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332318, "dur": 4, + "args": { + "External id": 98489, + "cbid": 211, "correlation": 98489 + } + }, + { + "ph": "s", "id": 98489, "pid": 435, "tid": 435, "ts": 1776649793332318, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793332361, "dur": 1, + "args": { + "External id": 13017, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98507, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98507, "pid": 0, "tid": 7, "ts": 1776649793332361, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332352, "dur": 6, + "args": { + "External id": 13017, + "cbid": 211, "correlation": 98507 + } + }, + { + "ph": "s", "id": 98507, "pid": 435, "tid": 435, "ts": 1776649793332352, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332364, "dur": 0, + "args": { + "External id": 98513, + "cbid": 200, "correlation": 98513 + } + }, + { + "ph": "f", "id": 98513, "pid": 435, "tid": 435, "ts": 1776649793332364, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332365, "dur": 0, + "args": { + "External id": 98514, + "cbid": 273, "correlation": 98514 + } + }, + { + "ph": "f", "id": 98514, "pid": 435, "tid": 435, "ts": 1776649793332365, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793332373, "dur": 19, + "args": { + "External id": 98515, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98515, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98515, "pid": 0, "tid": 7, "ts": 1776649793332373, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332366, "dur": 4, + "args": { + "External id": 98515, + "cbid": 211, "correlation": 98515 + } + }, + { + "ph": "s", "id": 98515, "pid": 435, "tid": 435, "ts": 1776649793332366, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793332408, "dur": 1, + "args": { + "External id": 13022, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98533, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98533, "pid": 0, "tid": 7, "ts": 1776649793332408, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332400, "dur": 6, + "args": { + "External id": 13022, + "cbid": 211, "correlation": 98533 + } + }, + { + "ph": "s", "id": 98533, "pid": 435, "tid": 435, "ts": 1776649793332400, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332412, "dur": 0, + "args": { + "External id": 98539, + "cbid": 200, "correlation": 98539 + } + }, + { + "ph": "f", "id": 98539, "pid": 435, "tid": 435, "ts": 1776649793332412, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332412, "dur": 0, + "args": { + "External id": 98540, + "cbid": 273, "correlation": 98540 + } + }, + { + "ph": "f", "id": 98540, "pid": 435, "tid": 435, "ts": 1776649793332412, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793332421, "dur": 19, + "args": { + "External id": 98541, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98541, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98541, "pid": 0, "tid": 7, "ts": 1776649793332421, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332414, "dur": 4, + "args": { + "External id": 98541, + "cbid": 211, "correlation": 98541 + } + }, + { + "ph": "s", "id": 98541, "pid": 435, "tid": 435, "ts": 1776649793332414, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793332456, "dur": 1, + "args": { + "External id": 13027, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98559, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98559, "pid": 0, "tid": 7, "ts": 1776649793332456, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332447, "dur": 6, + "args": { + "External id": 13027, + "cbid": 211, "correlation": 98559 + } + }, + { + "ph": "s", "id": 98559, "pid": 435, "tid": 435, "ts": 1776649793332447, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332459, "dur": 0, + "args": { + "External id": 98565, + "cbid": 200, "correlation": 98565 + } + }, + { + "ph": "f", "id": 98565, "pid": 435, "tid": 435, "ts": 1776649793332459, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332460, "dur": 0, + "args": { + "External id": 98566, + "cbid": 273, "correlation": 98566 + } + }, + { + "ph": "f", "id": 98566, "pid": 435, "tid": 435, "ts": 1776649793332460, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793332468, "dur": 19, + "args": { + "External id": 98567, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98567, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98567, "pid": 0, "tid": 7, "ts": 1776649793332468, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332461, "dur": 4, + "args": { + "External id": 98567, + "cbid": 211, "correlation": 98567 + } + }, + { + "ph": "s", "id": 98567, "pid": 435, "tid": 435, "ts": 1776649793332461, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793332504, "dur": 1, + "args": { + "External id": 13032, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98585, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98585, "pid": 0, "tid": 7, "ts": 1776649793332504, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332495, "dur": 6, + "args": { + "External id": 13032, + "cbid": 211, "correlation": 98585 + } + }, + { + "ph": "s", "id": 98585, "pid": 435, "tid": 435, "ts": 1776649793332495, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332513, "dur": 0, + "args": { + "External id": 98591, + "cbid": 200, "correlation": 98591 + } + }, + { + "ph": "f", "id": 98591, "pid": 435, "tid": 435, "ts": 1776649793332513, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332514, "dur": 0, + "args": { + "External id": 98592, + "cbid": 273, "correlation": 98592 + } + }, + { + "ph": "f", "id": 98592, "pid": 435, "tid": 435, "ts": 1776649793332514, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793332523, "dur": 19, + "args": { + "External id": 98593, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98593, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98593, "pid": 0, "tid": 7, "ts": 1776649793332523, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332515, "dur": 5, + "args": { + "External id": 98593, + "cbid": 211, "correlation": 98593 + } + }, + { + "ph": "s", "id": 98593, "pid": 435, "tid": 435, "ts": 1776649793332515, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793332560, "dur": 1, + "args": { + "External id": 13037, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98611, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98611, "pid": 0, "tid": 7, "ts": 1776649793332560, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332551, "dur": 6, + "args": { + "External id": 13037, + "cbid": 211, "correlation": 98611 + } + }, + { + "ph": "s", "id": 98611, "pid": 435, "tid": 435, "ts": 1776649793332551, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332563, "dur": 0, + "args": { + "External id": 98617, + "cbid": 200, "correlation": 98617 + } + }, + { + "ph": "f", "id": 98617, "pid": 435, "tid": 435, "ts": 1776649793332563, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332563, "dur": 0, + "args": { + "External id": 98618, + "cbid": 273, "correlation": 98618 + } + }, + { + "ph": "f", "id": 98618, "pid": 435, "tid": 435, "ts": 1776649793332563, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793332572, "dur": 19, + "args": { + "External id": 98619, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98619, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98619, "pid": 0, "tid": 7, "ts": 1776649793332572, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332565, "dur": 4, + "args": { + "External id": 98619, + "cbid": 211, "correlation": 98619 + } + }, + { + "ph": "s", "id": 98619, "pid": 435, "tid": 435, "ts": 1776649793332565, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793332607, "dur": 1, + "args": { + "External id": 13042, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98637, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98637, "pid": 0, "tid": 7, "ts": 1776649793332607, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332598, "dur": 6, + "args": { + "External id": 13042, + "cbid": 211, "correlation": 98637 + } + }, + { + "ph": "s", "id": 98637, "pid": 435, "tid": 435, "ts": 1776649793332598, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332611, "dur": 0, + "args": { + "External id": 98643, + "cbid": 200, "correlation": 98643 + } + }, + { + "ph": "f", "id": 98643, "pid": 435, "tid": 435, "ts": 1776649793332611, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332612, "dur": 0, + "args": { + "External id": 98644, + "cbid": 273, "correlation": 98644 + } + }, + { + "ph": "f", "id": 98644, "pid": 435, "tid": 435, "ts": 1776649793332612, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793332620, "dur": 19, + "args": { + "External id": 98645, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98645, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98645, "pid": 0, "tid": 7, "ts": 1776649793332620, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332613, "dur": 4, + "args": { + "External id": 98645, + "cbid": 211, "correlation": 98645 + } + }, + { + "ph": "s", "id": 98645, "pid": 435, "tid": 435, "ts": 1776649793332613, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793332656, "dur": 1, + "args": { + "External id": 13047, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98663, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98663, "pid": 0, "tid": 7, "ts": 1776649793332656, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332647, "dur": 6, + "args": { + "External id": 13047, + "cbid": 211, "correlation": 98663 + } + }, + { + "ph": "s", "id": 98663, "pid": 435, "tid": 435, "ts": 1776649793332647, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332659, "dur": 0, + "args": { + "External id": 98669, + "cbid": 200, "correlation": 98669 + } + }, + { + "ph": "f", "id": 98669, "pid": 435, "tid": 435, "ts": 1776649793332659, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332659, "dur": 0, + "args": { + "External id": 98670, + "cbid": 273, "correlation": 98670 + } + }, + { + "ph": "f", "id": 98670, "pid": 435, "tid": 435, "ts": 1776649793332659, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793332668, "dur": 19, + "args": { + "External id": 98671, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98671, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98671, "pid": 0, "tid": 7, "ts": 1776649793332668, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332660, "dur": 4, + "args": { + "External id": 98671, + "cbid": 211, "correlation": 98671 + } + }, + { + "ph": "s", "id": 98671, "pid": 435, "tid": 435, "ts": 1776649793332660, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793332703, "dur": 1, + "args": { + "External id": 13052, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98689, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98689, "pid": 0, "tid": 7, "ts": 1776649793332703, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332694, "dur": 6, + "args": { + "External id": 13052, + "cbid": 211, "correlation": 98689 + } + }, + { + "ph": "s", "id": 98689, "pid": 435, "tid": 435, "ts": 1776649793332694, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332707, "dur": 0, + "args": { + "External id": 98695, + "cbid": 200, "correlation": 98695 + } + }, + { + "ph": "f", "id": 98695, "pid": 435, "tid": 435, "ts": 1776649793332707, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332707, "dur": 0, + "args": { + "External id": 98696, + "cbid": 273, "correlation": 98696 + } + }, + { + "ph": "f", "id": 98696, "pid": 435, "tid": 435, "ts": 1776649793332707, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793332716, "dur": 19, + "args": { + "External id": 98697, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98697, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98697, "pid": 0, "tid": 7, "ts": 1776649793332716, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332708, "dur": 4, + "args": { + "External id": 98697, + "cbid": 211, "correlation": 98697 + } + }, + { + "ph": "s", "id": 98697, "pid": 435, "tid": 435, "ts": 1776649793332708, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793332751, "dur": 1, + "args": { + "External id": 13057, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98715, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98715, "pid": 0, "tid": 7, "ts": 1776649793332751, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332742, "dur": 6, + "args": { + "External id": 13057, + "cbid": 211, "correlation": 98715 + } + }, + { + "ph": "s", "id": 98715, "pid": 435, "tid": 435, "ts": 1776649793332742, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332754, "dur": 0, + "args": { + "External id": 98721, + "cbid": 200, "correlation": 98721 + } + }, + { + "ph": "f", "id": 98721, "pid": 435, "tid": 435, "ts": 1776649793332754, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332754, "dur": 0, + "args": { + "External id": 98722, + "cbid": 273, "correlation": 98722 + } + }, + { + "ph": "f", "id": 98722, "pid": 435, "tid": 435, "ts": 1776649793332754, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793332763, "dur": 19, + "args": { + "External id": 98723, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98723, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98723, "pid": 0, "tid": 7, "ts": 1776649793332763, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332755, "dur": 4, + "args": { + "External id": 98723, + "cbid": 211, "correlation": 98723 + } + }, + { + "ph": "s", "id": 98723, "pid": 435, "tid": 435, "ts": 1776649793332755, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649793332786, "dur": 3, + "args": { + "External id": 98731, + "cbid": 135, "correlation": 98731 + } + }, + { + "ph": "f", "id": 98731, "pid": 435, "tid": 435, "ts": 1776649793332786, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649793332801, "dur": 5, + "args": { + "External id": 98736, + "cbid": 165, "correlation": 98736 + } + }, + { + "ph": "s", "id": 98736, "pid": 435, "tid": 435, "ts": 1776649793332801, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649793332810, "dur": 6, + "args": { + "External id": 98739, + "cbid": 139, "correlation": 98739 + } + }, + { + "ph": "f", "id": 98739, "pid": 435, "tid": 435, "ts": 1776649793332810, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793332871, "dur": 1, + "args": { + "External id": 13062, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98767, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98767, "pid": 0, "tid": 7, "ts": 1776649793332871, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332860, "dur": 8, + "args": { + "External id": 13062, + "cbid": 211, "correlation": 98767 + } + }, + { + "ph": "s", "id": 98767, "pid": 435, "tid": 435, "ts": 1776649793332860, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332876, "dur": 0, + "args": { + "External id": 98773, + "cbid": 200, "correlation": 98773 + } + }, + { + "ph": "f", "id": 98773, "pid": 435, "tid": 435, "ts": 1776649793332876, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332877, "dur": 1, + "args": { + "External id": 98774, + "cbid": 273, "correlation": 98774 + } + }, + { + "ph": "f", "id": 98774, "pid": 435, "tid": 435, "ts": 1776649793332877, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793332886, "dur": 19, + "args": { + "External id": 98775, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98775, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98775, "pid": 0, "tid": 7, "ts": 1776649793332886, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332878, "dur": 5, + "args": { + "External id": 98775, + "cbid": 211, "correlation": 98775 + } + }, + { + "ph": "s", "id": 98775, "pid": 435, "tid": 435, "ts": 1776649793332878, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793332923, "dur": 1, + "args": { + "External id": 13067, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98793, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98793, "pid": 0, "tid": 7, "ts": 1776649793332923, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332914, "dur": 6, + "args": { + "External id": 13067, + "cbid": 211, "correlation": 98793 + } + }, + { + "ph": "s", "id": 98793, "pid": 435, "tid": 435, "ts": 1776649793332914, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332927, "dur": 0, + "args": { + "External id": 98799, + "cbid": 200, "correlation": 98799 + } + }, + { + "ph": "f", "id": 98799, "pid": 435, "tid": 435, "ts": 1776649793332927, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332927, "dur": 0, + "args": { + "External id": 98800, + "cbid": 273, "correlation": 98800 + } + }, + { + "ph": "f", "id": 98800, "pid": 435, "tid": 435, "ts": 1776649793332927, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793332936, "dur": 19, + "args": { + "External id": 98801, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98801, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98801, "pid": 0, "tid": 7, "ts": 1776649793332936, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332928, "dur": 4, + "args": { + "External id": 98801, + "cbid": 211, "correlation": 98801 + } + }, + { + "ph": "s", "id": 98801, "pid": 435, "tid": 435, "ts": 1776649793332928, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793332972, "dur": 1, + "args": { + "External id": 13072, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98819, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98819, "pid": 0, "tid": 7, "ts": 1776649793332972, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332963, "dur": 6, + "args": { + "External id": 13072, + "cbid": 211, "correlation": 98819 + } + }, + { + "ph": "s", "id": 98819, "pid": 435, "tid": 435, "ts": 1776649793332963, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332975, "dur": 0, + "args": { + "External id": 98825, + "cbid": 200, "correlation": 98825 + } + }, + { + "ph": "f", "id": 98825, "pid": 435, "tid": 435, "ts": 1776649793332975, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793332976, "dur": 0, + "args": { + "External id": 98826, + "cbid": 273, "correlation": 98826 + } + }, + { + "ph": "f", "id": 98826, "pid": 435, "tid": 435, "ts": 1776649793332976, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793332985, "dur": 19, + "args": { + "External id": 98827, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98827, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98827, "pid": 0, "tid": 7, "ts": 1776649793332985, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793332977, "dur": 4, + "args": { + "External id": 98827, + "cbid": 211, "correlation": 98827 + } + }, + { + "ph": "s", "id": 98827, "pid": 435, "tid": 435, "ts": 1776649793332977, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793333020, "dur": 1, + "args": { + "External id": 13077, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98845, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98845, "pid": 0, "tid": 7, "ts": 1776649793333020, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333011, "dur": 6, + "args": { + "External id": 13077, + "cbid": 211, "correlation": 98845 + } + }, + { + "ph": "s", "id": 98845, "pid": 435, "tid": 435, "ts": 1776649793333011, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333023, "dur": 0, + "args": { + "External id": 98851, + "cbid": 200, "correlation": 98851 + } + }, + { + "ph": "f", "id": 98851, "pid": 435, "tid": 435, "ts": 1776649793333023, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333024, "dur": 0, + "args": { + "External id": 98852, + "cbid": 273, "correlation": 98852 + } + }, + { + "ph": "f", "id": 98852, "pid": 435, "tid": 435, "ts": 1776649793333024, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793333032, "dur": 19, + "args": { + "External id": 98853, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98853, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98853, "pid": 0, "tid": 7, "ts": 1776649793333032, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333025, "dur": 4, + "args": { + "External id": 98853, + "cbid": 211, "correlation": 98853 + } + }, + { + "ph": "s", "id": 98853, "pid": 435, "tid": 435, "ts": 1776649793333025, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793333067, "dur": 1, + "args": { + "External id": 13082, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98871, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98871, "pid": 0, "tid": 7, "ts": 1776649793333067, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333059, "dur": 6, + "args": { + "External id": 13082, + "cbid": 211, "correlation": 98871 + } + }, + { + "ph": "s", "id": 98871, "pid": 435, "tid": 435, "ts": 1776649793333059, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333071, "dur": 0, + "args": { + "External id": 98877, + "cbid": 200, "correlation": 98877 + } + }, + { + "ph": "f", "id": 98877, "pid": 435, "tid": 435, "ts": 1776649793333071, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333071, "dur": 0, + "args": { + "External id": 98878, + "cbid": 273, "correlation": 98878 + } + }, + { + "ph": "f", "id": 98878, "pid": 435, "tid": 435, "ts": 1776649793333071, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793333080, "dur": 19, + "args": { + "External id": 98879, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98879, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98879, "pid": 0, "tid": 7, "ts": 1776649793333080, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333072, "dur": 4, + "args": { + "External id": 98879, + "cbid": 211, "correlation": 98879 + } + }, + { + "ph": "s", "id": 98879, "pid": 435, "tid": 435, "ts": 1776649793333072, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793333115, "dur": 1, + "args": { + "External id": 13087, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98897, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98897, "pid": 0, "tid": 7, "ts": 1776649793333115, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333106, "dur": 6, + "args": { + "External id": 13087, + "cbid": 211, "correlation": 98897 + } + }, + { + "ph": "s", "id": 98897, "pid": 435, "tid": 435, "ts": 1776649793333106, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333118, "dur": 0, + "args": { + "External id": 98903, + "cbid": 200, "correlation": 98903 + } + }, + { + "ph": "f", "id": 98903, "pid": 435, "tid": 435, "ts": 1776649793333118, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333119, "dur": 0, + "args": { + "External id": 98904, + "cbid": 273, "correlation": 98904 + } + }, + { + "ph": "f", "id": 98904, "pid": 435, "tid": 435, "ts": 1776649793333119, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793333127, "dur": 19, + "args": { + "External id": 98905, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98905, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98905, "pid": 0, "tid": 7, "ts": 1776649793333127, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333120, "dur": 4, + "args": { + "External id": 98905, + "cbid": 211, "correlation": 98905 + } + }, + { + "ph": "s", "id": 98905, "pid": 435, "tid": 435, "ts": 1776649793333120, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793333162, "dur": 1, + "args": { + "External id": 13092, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98923, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98923, "pid": 0, "tid": 7, "ts": 1776649793333162, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333154, "dur": 6, + "args": { + "External id": 13092, + "cbid": 211, "correlation": 98923 + } + }, + { + "ph": "s", "id": 98923, "pid": 435, "tid": 435, "ts": 1776649793333154, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333165, "dur": 0, + "args": { + "External id": 98929, + "cbid": 200, "correlation": 98929 + } + }, + { + "ph": "f", "id": 98929, "pid": 435, "tid": 435, "ts": 1776649793333165, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333166, "dur": 0, + "args": { + "External id": 98930, + "cbid": 273, "correlation": 98930 + } + }, + { + "ph": "f", "id": 98930, "pid": 435, "tid": 435, "ts": 1776649793333166, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793333175, "dur": 19, + "args": { + "External id": 98931, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98931, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98931, "pid": 0, "tid": 7, "ts": 1776649793333175, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333167, "dur": 4, + "args": { + "External id": 98931, + "cbid": 211, "correlation": 98931 + } + }, + { + "ph": "s", "id": 98931, "pid": 435, "tid": 435, "ts": 1776649793333167, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793333209, "dur": 1, + "args": { + "External id": 13097, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98949, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98949, "pid": 0, "tid": 7, "ts": 1776649793333209, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333201, "dur": 6, + "args": { + "External id": 13097, + "cbid": 211, "correlation": 98949 + } + }, + { + "ph": "s", "id": 98949, "pid": 435, "tid": 435, "ts": 1776649793333201, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333213, "dur": 0, + "args": { + "External id": 98955, + "cbid": 200, "correlation": 98955 + } + }, + { + "ph": "f", "id": 98955, "pid": 435, "tid": 435, "ts": 1776649793333213, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333213, "dur": 0, + "args": { + "External id": 98956, + "cbid": 273, "correlation": 98956 + } + }, + { + "ph": "f", "id": 98956, "pid": 435, "tid": 435, "ts": 1776649793333213, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793333222, "dur": 19, + "args": { + "External id": 98957, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98957, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98957, "pid": 0, "tid": 7, "ts": 1776649793333222, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333214, "dur": 4, + "args": { + "External id": 98957, + "cbid": 211, "correlation": 98957 + } + }, + { + "ph": "s", "id": 98957, "pid": 435, "tid": 435, "ts": 1776649793333214, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793333257, "dur": 1, + "args": { + "External id": 13102, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98975, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98975, "pid": 0, "tid": 7, "ts": 1776649793333257, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333248, "dur": 6, + "args": { + "External id": 13102, + "cbid": 211, "correlation": 98975 + } + }, + { + "ph": "s", "id": 98975, "pid": 435, "tid": 435, "ts": 1776649793333248, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333260, "dur": 0, + "args": { + "External id": 98981, + "cbid": 200, "correlation": 98981 + } + }, + { + "ph": "f", "id": 98981, "pid": 435, "tid": 435, "ts": 1776649793333260, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333260, "dur": 0, + "args": { + "External id": 98982, + "cbid": 273, "correlation": 98982 + } + }, + { + "ph": "f", "id": 98982, "pid": 435, "tid": 435, "ts": 1776649793333260, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793333269, "dur": 19, + "args": { + "External id": 98983, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 98983, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 98983, "pid": 0, "tid": 7, "ts": 1776649793333269, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333262, "dur": 4, + "args": { + "External id": 98983, + "cbid": 211, "correlation": 98983 + } + }, + { + "ph": "s", "id": 98983, "pid": 435, "tid": 435, "ts": 1776649793333262, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793333304, "dur": 1, + "args": { + "External id": 13107, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99001, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99001, "pid": 0, "tid": 7, "ts": 1776649793333304, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333295, "dur": 6, + "args": { + "External id": 13107, + "cbid": 211, "correlation": 99001 + } + }, + { + "ph": "s", "id": 99001, "pid": 435, "tid": 435, "ts": 1776649793333295, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333307, "dur": 0, + "args": { + "External id": 99007, + "cbid": 200, "correlation": 99007 + } + }, + { + "ph": "f", "id": 99007, "pid": 435, "tid": 435, "ts": 1776649793333307, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333308, "dur": 0, + "args": { + "External id": 99008, + "cbid": 273, "correlation": 99008 + } + }, + { + "ph": "f", "id": 99008, "pid": 435, "tid": 435, "ts": 1776649793333308, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793333317, "dur": 19, + "args": { + "External id": 99009, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99009, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99009, "pid": 0, "tid": 7, "ts": 1776649793333317, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333309, "dur": 4, + "args": { + "External id": 99009, + "cbid": 211, "correlation": 99009 + } + }, + { + "ph": "s", "id": 99009, "pid": 435, "tid": 435, "ts": 1776649793333309, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649793333331, "dur": 9, + "args": { + "External id": 99012, + "cbid": 165, "correlation": 99012 + } + }, + { + "ph": "s", "id": 99012, "pid": 435, "tid": 435, "ts": 1776649793333331, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649793333361, "dur": 3, + "args": { + "External id": 99021, + "cbid": 135, "correlation": 99021 + } + }, + { + "ph": "f", "id": 99021, "pid": 435, "tid": 435, "ts": 1776649793333361, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793333409, "dur": 1, + "args": { + "External id": 13112, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99041, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99041, "pid": 0, "tid": 7, "ts": 1776649793333409, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333399, "dur": 8, + "args": { + "External id": 13112, + "cbid": 211, "correlation": 99041 + } + }, + { + "ph": "s", "id": 99041, "pid": 435, "tid": 435, "ts": 1776649793333399, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333414, "dur": 0, + "args": { + "External id": 99047, + "cbid": 200, "correlation": 99047 + } + }, + { + "ph": "f", "id": 99047, "pid": 435, "tid": 435, "ts": 1776649793333414, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333414, "dur": 0, + "args": { + "External id": 99048, + "cbid": 273, "correlation": 99048 + } + }, + { + "ph": "f", "id": 99048, "pid": 435, "tid": 435, "ts": 1776649793333414, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793333423, "dur": 19, + "args": { + "External id": 99049, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99049, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99049, "pid": 0, "tid": 7, "ts": 1776649793333423, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333416, "dur": 4, + "args": { + "External id": 99049, + "cbid": 211, "correlation": 99049 + } + }, + { + "ph": "s", "id": 99049, "pid": 435, "tid": 435, "ts": 1776649793333416, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793333461, "dur": 1, + "args": { + "External id": 13117, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99067, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99067, "pid": 0, "tid": 7, "ts": 1776649793333461, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333452, "dur": 6, + "args": { + "External id": 13117, + "cbid": 211, "correlation": 99067 + } + }, + { + "ph": "s", "id": 99067, "pid": 435, "tid": 435, "ts": 1776649793333452, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333464, "dur": 0, + "args": { + "External id": 99073, + "cbid": 200, "correlation": 99073 + } + }, + { + "ph": "f", "id": 99073, "pid": 435, "tid": 435, "ts": 1776649793333464, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333465, "dur": 0, + "args": { + "External id": 99074, + "cbid": 273, "correlation": 99074 + } + }, + { + "ph": "f", "id": 99074, "pid": 435, "tid": 435, "ts": 1776649793333465, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793333474, "dur": 19, + "args": { + "External id": 99075, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99075, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99075, "pid": 0, "tid": 7, "ts": 1776649793333474, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333466, "dur": 4, + "args": { + "External id": 99075, + "cbid": 211, "correlation": 99075 + } + }, + { + "ph": "s", "id": 99075, "pid": 435, "tid": 435, "ts": 1776649793333466, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793333509, "dur": 1, + "args": { + "External id": 13122, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99093, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99093, "pid": 0, "tid": 7, "ts": 1776649793333509, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333500, "dur": 6, + "args": { + "External id": 13122, + "cbid": 211, "correlation": 99093 + } + }, + { + "ph": "s", "id": 99093, "pid": 435, "tid": 435, "ts": 1776649793333500, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333512, "dur": 0, + "args": { + "External id": 99099, + "cbid": 200, "correlation": 99099 + } + }, + { + "ph": "f", "id": 99099, "pid": 435, "tid": 435, "ts": 1776649793333512, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333513, "dur": 0, + "args": { + "External id": 99100, + "cbid": 273, "correlation": 99100 + } + }, + { + "ph": "f", "id": 99100, "pid": 435, "tid": 435, "ts": 1776649793333513, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793333521, "dur": 19, + "args": { + "External id": 99101, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99101, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99101, "pid": 0, "tid": 7, "ts": 1776649793333521, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333514, "dur": 4, + "args": { + "External id": 99101, + "cbid": 211, "correlation": 99101 + } + }, + { + "ph": "s", "id": 99101, "pid": 435, "tid": 435, "ts": 1776649793333514, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793333556, "dur": 1, + "args": { + "External id": 13127, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99119, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99119, "pid": 0, "tid": 7, "ts": 1776649793333556, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333547, "dur": 6, + "args": { + "External id": 13127, + "cbid": 211, "correlation": 99119 + } + }, + { + "ph": "s", "id": 99119, "pid": 435, "tid": 435, "ts": 1776649793333547, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333559, "dur": 0, + "args": { + "External id": 99125, + "cbid": 200, "correlation": 99125 + } + }, + { + "ph": "f", "id": 99125, "pid": 435, "tid": 435, "ts": 1776649793333559, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333560, "dur": 0, + "args": { + "External id": 99126, + "cbid": 273, "correlation": 99126 + } + }, + { + "ph": "f", "id": 99126, "pid": 435, "tid": 435, "ts": 1776649793333560, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793333568, "dur": 19, + "args": { + "External id": 99127, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99127, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99127, "pid": 0, "tid": 7, "ts": 1776649793333568, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333561, "dur": 4, + "args": { + "External id": 99127, + "cbid": 211, "correlation": 99127 + } + }, + { + "ph": "s", "id": 99127, "pid": 435, "tid": 435, "ts": 1776649793333561, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793333603, "dur": 1, + "args": { + "External id": 13132, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99145, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99145, "pid": 0, "tid": 7, "ts": 1776649793333603, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333595, "dur": 6, + "args": { + "External id": 13132, + "cbid": 211, "correlation": 99145 + } + }, + { + "ph": "s", "id": 99145, "pid": 435, "tid": 435, "ts": 1776649793333595, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333606, "dur": 0, + "args": { + "External id": 99151, + "cbid": 200, "correlation": 99151 + } + }, + { + "ph": "f", "id": 99151, "pid": 435, "tid": 435, "ts": 1776649793333606, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333607, "dur": 0, + "args": { + "External id": 99152, + "cbid": 273, "correlation": 99152 + } + }, + { + "ph": "f", "id": 99152, "pid": 435, "tid": 435, "ts": 1776649793333607, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793333616, "dur": 19, + "args": { + "External id": 99153, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99153, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99153, "pid": 0, "tid": 7, "ts": 1776649793333616, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333608, "dur": 4, + "args": { + "External id": 99153, + "cbid": 211, "correlation": 99153 + } + }, + { + "ph": "s", "id": 99153, "pid": 435, "tid": 435, "ts": 1776649793333608, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793333651, "dur": 1, + "args": { + "External id": 13137, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99171, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99171, "pid": 0, "tid": 7, "ts": 1776649793333651, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333642, "dur": 6, + "args": { + "External id": 13137, + "cbid": 211, "correlation": 99171 + } + }, + { + "ph": "s", "id": 99171, "pid": 435, "tid": 435, "ts": 1776649793333642, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333654, "dur": 0, + "args": { + "External id": 99177, + "cbid": 200, "correlation": 99177 + } + }, + { + "ph": "f", "id": 99177, "pid": 435, "tid": 435, "ts": 1776649793333654, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333655, "dur": 0, + "args": { + "External id": 99178, + "cbid": 273, "correlation": 99178 + } + }, + { + "ph": "f", "id": 99178, "pid": 435, "tid": 435, "ts": 1776649793333655, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793333663, "dur": 19, + "args": { + "External id": 99179, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99179, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99179, "pid": 0, "tid": 7, "ts": 1776649793333663, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333656, "dur": 4, + "args": { + "External id": 99179, + "cbid": 211, "correlation": 99179 + } + }, + { + "ph": "s", "id": 99179, "pid": 435, "tid": 435, "ts": 1776649793333656, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793333699, "dur": 1, + "args": { + "External id": 13142, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99197, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99197, "pid": 0, "tid": 7, "ts": 1776649793333699, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333690, "dur": 6, + "args": { + "External id": 13142, + "cbid": 211, "correlation": 99197 + } + }, + { + "ph": "s", "id": 99197, "pid": 435, "tid": 435, "ts": 1776649793333690, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333702, "dur": 0, + "args": { + "External id": 99203, + "cbid": 200, "correlation": 99203 + } + }, + { + "ph": "f", "id": 99203, "pid": 435, "tid": 435, "ts": 1776649793333702, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333702, "dur": 0, + "args": { + "External id": 99204, + "cbid": 273, "correlation": 99204 + } + }, + { + "ph": "f", "id": 99204, "pid": 435, "tid": 435, "ts": 1776649793333702, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793333711, "dur": 19, + "args": { + "External id": 99205, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99205, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99205, "pid": 0, "tid": 7, "ts": 1776649793333711, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333704, "dur": 4, + "args": { + "External id": 99205, + "cbid": 211, "correlation": 99205 + } + }, + { + "ph": "s", "id": 99205, "pid": 435, "tid": 435, "ts": 1776649793333704, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793333746, "dur": 1, + "args": { + "External id": 13147, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99223, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99223, "pid": 0, "tid": 7, "ts": 1776649793333746, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333737, "dur": 6, + "args": { + "External id": 13147, + "cbid": 211, "correlation": 99223 + } + }, + { + "ph": "s", "id": 99223, "pid": 435, "tid": 435, "ts": 1776649793333737, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333749, "dur": 0, + "args": { + "External id": 99229, + "cbid": 200, "correlation": 99229 + } + }, + { + "ph": "f", "id": 99229, "pid": 435, "tid": 435, "ts": 1776649793333749, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333750, "dur": 0, + "args": { + "External id": 99230, + "cbid": 273, "correlation": 99230 + } + }, + { + "ph": "f", "id": 99230, "pid": 435, "tid": 435, "ts": 1776649793333750, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793333758, "dur": 19, + "args": { + "External id": 99231, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99231, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99231, "pid": 0, "tid": 7, "ts": 1776649793333758, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333751, "dur": 4, + "args": { + "External id": 99231, + "cbid": 211, "correlation": 99231 + } + }, + { + "ph": "s", "id": 99231, "pid": 435, "tid": 435, "ts": 1776649793333751, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793333794, "dur": 1, + "args": { + "External id": 13152, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99249, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99249, "pid": 0, "tid": 7, "ts": 1776649793333794, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333785, "dur": 6, + "args": { + "External id": 13152, + "cbid": 211, "correlation": 99249 + } + }, + { + "ph": "s", "id": 99249, "pid": 435, "tid": 435, "ts": 1776649793333785, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333797, "dur": 0, + "args": { + "External id": 99255, + "cbid": 200, "correlation": 99255 + } + }, + { + "ph": "f", "id": 99255, "pid": 435, "tid": 435, "ts": 1776649793333797, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333798, "dur": 0, + "args": { + "External id": 99256, + "cbid": 273, "correlation": 99256 + } + }, + { + "ph": "f", "id": 99256, "pid": 435, "tid": 435, "ts": 1776649793333798, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793333806, "dur": 19, + "args": { + "External id": 99257, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99257, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99257, "pid": 0, "tid": 7, "ts": 1776649793333806, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333799, "dur": 4, + "args": { + "External id": 99257, + "cbid": 211, "correlation": 99257 + } + }, + { + "ph": "s", "id": 99257, "pid": 435, "tid": 435, "ts": 1776649793333799, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793333841, "dur": 1, + "args": { + "External id": 13157, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99275, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99275, "pid": 0, "tid": 7, "ts": 1776649793333841, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333833, "dur": 6, + "args": { + "External id": 13157, + "cbid": 211, "correlation": 99275 + } + }, + { + "ph": "s", "id": 99275, "pid": 435, "tid": 435, "ts": 1776649793333833, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333844, "dur": 0, + "args": { + "External id": 99281, + "cbid": 200, "correlation": 99281 + } + }, + { + "ph": "f", "id": 99281, "pid": 435, "tid": 435, "ts": 1776649793333844, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333845, "dur": 0, + "args": { + "External id": 99282, + "cbid": 273, "correlation": 99282 + } + }, + { + "ph": "f", "id": 99282, "pid": 435, "tid": 435, "ts": 1776649793333845, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793333854, "dur": 19, + "args": { + "External id": 99283, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99283, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99283, "pid": 0, "tid": 7, "ts": 1776649793333854, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333846, "dur": 4, + "args": { + "External id": 99283, + "cbid": 211, "correlation": 99283 + } + }, + { + "ph": "s", "id": 99283, "pid": 435, "tid": 435, "ts": 1776649793333846, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793333889, "dur": 1, + "args": { + "External id": 13162, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99301, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99301, "pid": 0, "tid": 7, "ts": 1776649793333889, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333880, "dur": 6, + "args": { + "External id": 13162, + "cbid": 211, "correlation": 99301 + } + }, + { + "ph": "s", "id": 99301, "pid": 435, "tid": 435, "ts": 1776649793333880, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333892, "dur": 0, + "args": { + "External id": 99307, + "cbid": 200, "correlation": 99307 + } + }, + { + "ph": "f", "id": 99307, "pid": 435, "tid": 435, "ts": 1776649793333892, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333893, "dur": 0, + "args": { + "External id": 99308, + "cbid": 273, "correlation": 99308 + } + }, + { + "ph": "f", "id": 99308, "pid": 435, "tid": 435, "ts": 1776649793333893, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793333902, "dur": 19, + "args": { + "External id": 99309, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99309, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99309, "pid": 0, "tid": 7, "ts": 1776649793333902, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333894, "dur": 4, + "args": { + "External id": 99309, + "cbid": 211, "correlation": 99309 + } + }, + { + "ph": "s", "id": 99309, "pid": 435, "tid": 435, "ts": 1776649793333894, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793333936, "dur": 1, + "args": { + "External id": 13167, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99327, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99327, "pid": 0, "tid": 7, "ts": 1776649793333936, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333927, "dur": 6, + "args": { + "External id": 13167, + "cbid": 211, "correlation": 99327 + } + }, + { + "ph": "s", "id": 99327, "pid": 435, "tid": 435, "ts": 1776649793333927, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333939, "dur": 0, + "args": { + "External id": 99333, + "cbid": 200, "correlation": 99333 + } + }, + { + "ph": "f", "id": 99333, "pid": 435, "tid": 435, "ts": 1776649793333939, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333940, "dur": 0, + "args": { + "External id": 99334, + "cbid": 273, "correlation": 99334 + } + }, + { + "ph": "f", "id": 99334, "pid": 435, "tid": 435, "ts": 1776649793333940, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793333949, "dur": 19, + "args": { + "External id": 99335, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99335, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99335, "pid": 0, "tid": 7, "ts": 1776649793333949, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333941, "dur": 4, + "args": { + "External id": 99335, + "cbid": 211, "correlation": 99335 + } + }, + { + "ph": "s", "id": 99335, "pid": 435, "tid": 435, "ts": 1776649793333941, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793333984, "dur": 1, + "args": { + "External id": 13172, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99353, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99353, "pid": 0, "tid": 7, "ts": 1776649793333984, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333975, "dur": 6, + "args": { + "External id": 13172, + "cbid": 211, "correlation": 99353 + } + }, + { + "ph": "s", "id": 99353, "pid": 435, "tid": 435, "ts": 1776649793333975, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333987, "dur": 0, + "args": { + "External id": 99359, + "cbid": 200, "correlation": 99359 + } + }, + { + "ph": "f", "id": 99359, "pid": 435, "tid": 435, "ts": 1776649793333987, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793333988, "dur": 0, + "args": { + "External id": 99360, + "cbid": 273, "correlation": 99360 + } + }, + { + "ph": "f", "id": 99360, "pid": 435, "tid": 435, "ts": 1776649793333988, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793333996, "dur": 19, + "args": { + "External id": 99361, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99361, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99361, "pid": 0, "tid": 7, "ts": 1776649793333996, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793333989, "dur": 4, + "args": { + "External id": 99361, + "cbid": 211, "correlation": 99361 + } + }, + { + "ph": "s", "id": 99361, "pid": 435, "tid": 435, "ts": 1776649793333989, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793334031, "dur": 1, + "args": { + "External id": 13177, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99379, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99379, "pid": 0, "tid": 7, "ts": 1776649793334031, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334022, "dur": 6, + "args": { + "External id": 13177, + "cbid": 211, "correlation": 99379 + } + }, + { + "ph": "s", "id": 99379, "pid": 435, "tid": 435, "ts": 1776649793334022, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334034, "dur": 0, + "args": { + "External id": 99385, + "cbid": 200, "correlation": 99385 + } + }, + { + "ph": "f", "id": 99385, "pid": 435, "tid": 435, "ts": 1776649793334034, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334035, "dur": 0, + "args": { + "External id": 99386, + "cbid": 273, "correlation": 99386 + } + }, + { + "ph": "f", "id": 99386, "pid": 435, "tid": 435, "ts": 1776649793334035, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793334043, "dur": 19, + "args": { + "External id": 99387, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99387, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99387, "pid": 0, "tid": 7, "ts": 1776649793334043, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334036, "dur": 4, + "args": { + "External id": 99387, + "cbid": 211, "correlation": 99387 + } + }, + { + "ph": "s", "id": 99387, "pid": 435, "tid": 435, "ts": 1776649793334036, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793334079, "dur": 1, + "args": { + "External id": 13182, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99405, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99405, "pid": 0, "tid": 7, "ts": 1776649793334079, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334070, "dur": 6, + "args": { + "External id": 13182, + "cbid": 211, "correlation": 99405 + } + }, + { + "ph": "s", "id": 99405, "pid": 435, "tid": 435, "ts": 1776649793334070, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334082, "dur": 0, + "args": { + "External id": 99411, + "cbid": 200, "correlation": 99411 + } + }, + { + "ph": "f", "id": 99411, "pid": 435, "tid": 435, "ts": 1776649793334082, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334082, "dur": 0, + "args": { + "External id": 99412, + "cbid": 273, "correlation": 99412 + } + }, + { + "ph": "f", "id": 99412, "pid": 435, "tid": 435, "ts": 1776649793334082, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793334091, "dur": 19, + "args": { + "External id": 99413, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99413, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99413, "pid": 0, "tid": 7, "ts": 1776649793334091, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334083, "dur": 4, + "args": { + "External id": 99413, + "cbid": 211, "correlation": 99413 + } + }, + { + "ph": "s", "id": 99413, "pid": 435, "tid": 435, "ts": 1776649793334083, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793334126, "dur": 1, + "args": { + "External id": 13187, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99431, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99431, "pid": 0, "tid": 7, "ts": 1776649793334126, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334118, "dur": 6, + "args": { + "External id": 13187, + "cbid": 211, "correlation": 99431 + } + }, + { + "ph": "s", "id": 99431, "pid": 435, "tid": 435, "ts": 1776649793334118, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334130, "dur": 0, + "args": { + "External id": 99437, + "cbid": 200, "correlation": 99437 + } + }, + { + "ph": "f", "id": 99437, "pid": 435, "tid": 435, "ts": 1776649793334130, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334130, "dur": 0, + "args": { + "External id": 99438, + "cbid": 273, "correlation": 99438 + } + }, + { + "ph": "f", "id": 99438, "pid": 435, "tid": 435, "ts": 1776649793334130, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793334139, "dur": 19, + "args": { + "External id": 99439, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99439, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99439, "pid": 0, "tid": 7, "ts": 1776649793334139, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334132, "dur": 4, + "args": { + "External id": 99439, + "cbid": 211, "correlation": 99439 + } + }, + { + "ph": "s", "id": 99439, "pid": 435, "tid": 435, "ts": 1776649793334132, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793334174, "dur": 1, + "args": { + "External id": 13192, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99457, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99457, "pid": 0, "tid": 7, "ts": 1776649793334174, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334165, "dur": 6, + "args": { + "External id": 13192, + "cbid": 211, "correlation": 99457 + } + }, + { + "ph": "s", "id": 99457, "pid": 435, "tid": 435, "ts": 1776649793334165, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334177, "dur": 0, + "args": { + "External id": 99463, + "cbid": 200, "correlation": 99463 + } + }, + { + "ph": "f", "id": 99463, "pid": 435, "tid": 435, "ts": 1776649793334177, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334178, "dur": 0, + "args": { + "External id": 99464, + "cbid": 273, "correlation": 99464 + } + }, + { + "ph": "f", "id": 99464, "pid": 435, "tid": 435, "ts": 1776649793334178, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793334186, "dur": 19, + "args": { + "External id": 99465, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99465, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99465, "pid": 0, "tid": 7, "ts": 1776649793334186, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334179, "dur": 4, + "args": { + "External id": 99465, + "cbid": 211, "correlation": 99465 + } + }, + { + "ph": "s", "id": 99465, "pid": 435, "tid": 435, "ts": 1776649793334179, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793334221, "dur": 1, + "args": { + "External id": 13197, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99483, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99483, "pid": 0, "tid": 7, "ts": 1776649793334221, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334212, "dur": 6, + "args": { + "External id": 13197, + "cbid": 211, "correlation": 99483 + } + }, + { + "ph": "s", "id": 99483, "pid": 435, "tid": 435, "ts": 1776649793334212, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334224, "dur": 0, + "args": { + "External id": 99489, + "cbid": 200, "correlation": 99489 + } + }, + { + "ph": "f", "id": 99489, "pid": 435, "tid": 435, "ts": 1776649793334224, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334225, "dur": 0, + "args": { + "External id": 99490, + "cbid": 273, "correlation": 99490 + } + }, + { + "ph": "f", "id": 99490, "pid": 435, "tid": 435, "ts": 1776649793334225, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793334233, "dur": 19, + "args": { + "External id": 99491, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99491, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99491, "pid": 0, "tid": 7, "ts": 1776649793334233, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334226, "dur": 4, + "args": { + "External id": 99491, + "cbid": 211, "correlation": 99491 + } + }, + { + "ph": "s", "id": 99491, "pid": 435, "tid": 435, "ts": 1776649793334226, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793334268, "dur": 1, + "args": { + "External id": 13202, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99509, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99509, "pid": 0, "tid": 7, "ts": 1776649793334268, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334259, "dur": 6, + "args": { + "External id": 13202, + "cbid": 211, "correlation": 99509 + } + }, + { + "ph": "s", "id": 99509, "pid": 435, "tid": 435, "ts": 1776649793334259, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334271, "dur": 0, + "args": { + "External id": 99515, + "cbid": 200, "correlation": 99515 + } + }, + { + "ph": "f", "id": 99515, "pid": 435, "tid": 435, "ts": 1776649793334271, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334272, "dur": 0, + "args": { + "External id": 99516, + "cbid": 273, "correlation": 99516 + } + }, + { + "ph": "f", "id": 99516, "pid": 435, "tid": 435, "ts": 1776649793334272, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793334280, "dur": 19, + "args": { + "External id": 99517, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99517, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99517, "pid": 0, "tid": 7, "ts": 1776649793334280, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334273, "dur": 4, + "args": { + "External id": 99517, + "cbid": 211, "correlation": 99517 + } + }, + { + "ph": "s", "id": 99517, "pid": 435, "tid": 435, "ts": 1776649793334273, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793334315, "dur": 1, + "args": { + "External id": 13207, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99535, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99535, "pid": 0, "tid": 7, "ts": 1776649793334315, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334306, "dur": 6, + "args": { + "External id": 13207, + "cbid": 211, "correlation": 99535 + } + }, + { + "ph": "s", "id": 99535, "pid": 435, "tid": 435, "ts": 1776649793334306, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334318, "dur": 0, + "args": { + "External id": 99541, + "cbid": 200, "correlation": 99541 + } + }, + { + "ph": "f", "id": 99541, "pid": 435, "tid": 435, "ts": 1776649793334318, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334319, "dur": 0, + "args": { + "External id": 99542, + "cbid": 273, "correlation": 99542 + } + }, + { + "ph": "f", "id": 99542, "pid": 435, "tid": 435, "ts": 1776649793334319, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793334328, "dur": 19, + "args": { + "External id": 99543, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99543, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99543, "pid": 0, "tid": 7, "ts": 1776649793334328, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334320, "dur": 4, + "args": { + "External id": 99543, + "cbid": 211, "correlation": 99543 + } + }, + { + "ph": "s", "id": 99543, "pid": 435, "tid": 435, "ts": 1776649793334320, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793334362, "dur": 1, + "args": { + "External id": 13212, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99561, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99561, "pid": 0, "tid": 7, "ts": 1776649793334362, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334353, "dur": 6, + "args": { + "External id": 13212, + "cbid": 211, "correlation": 99561 + } + }, + { + "ph": "s", "id": 99561, "pid": 435, "tid": 435, "ts": 1776649793334353, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334365, "dur": 0, + "args": { + "External id": 99567, + "cbid": 200, "correlation": 99567 + } + }, + { + "ph": "f", "id": 99567, "pid": 435, "tid": 435, "ts": 1776649793334365, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334366, "dur": 0, + "args": { + "External id": 99568, + "cbid": 273, "correlation": 99568 + } + }, + { + "ph": "f", "id": 99568, "pid": 435, "tid": 435, "ts": 1776649793334366, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793334374, "dur": 19, + "args": { + "External id": 99569, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99569, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99569, "pid": 0, "tid": 7, "ts": 1776649793334374, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334367, "dur": 4, + "args": { + "External id": 99569, + "cbid": 211, "correlation": 99569 + } + }, + { + "ph": "s", "id": 99569, "pid": 435, "tid": 435, "ts": 1776649793334367, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793334410, "dur": 1, + "args": { + "External id": 13217, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99587, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99587, "pid": 0, "tid": 7, "ts": 1776649793334410, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334401, "dur": 6, + "args": { + "External id": 13217, + "cbid": 211, "correlation": 99587 + } + }, + { + "ph": "s", "id": 99587, "pid": 435, "tid": 435, "ts": 1776649793334401, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334413, "dur": 0, + "args": { + "External id": 99593, + "cbid": 200, "correlation": 99593 + } + }, + { + "ph": "f", "id": 99593, "pid": 435, "tid": 435, "ts": 1776649793334413, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334413, "dur": 0, + "args": { + "External id": 99594, + "cbid": 273, "correlation": 99594 + } + }, + { + "ph": "f", "id": 99594, "pid": 435, "tid": 435, "ts": 1776649793334413, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793334422, "dur": 19, + "args": { + "External id": 99595, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99595, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99595, "pid": 0, "tid": 7, "ts": 1776649793334422, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334414, "dur": 4, + "args": { + "External id": 99595, + "cbid": 211, "correlation": 99595 + } + }, + { + "ph": "s", "id": 99595, "pid": 435, "tid": 435, "ts": 1776649793334414, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793334457, "dur": 1, + "args": { + "External id": 13222, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99613, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99613, "pid": 0, "tid": 7, "ts": 1776649793334457, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334448, "dur": 6, + "args": { + "External id": 13222, + "cbid": 211, "correlation": 99613 + } + }, + { + "ph": "s", "id": 99613, "pid": 435, "tid": 435, "ts": 1776649793334448, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334460, "dur": 0, + "args": { + "External id": 99619, + "cbid": 200, "correlation": 99619 + } + }, + { + "ph": "f", "id": 99619, "pid": 435, "tid": 435, "ts": 1776649793334460, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334461, "dur": 0, + "args": { + "External id": 99620, + "cbid": 273, "correlation": 99620 + } + }, + { + "ph": "f", "id": 99620, "pid": 435, "tid": 435, "ts": 1776649793334461, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793334469, "dur": 19, + "args": { + "External id": 99621, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99621, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99621, "pid": 0, "tid": 7, "ts": 1776649793334469, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334462, "dur": 4, + "args": { + "External id": 99621, + "cbid": 211, "correlation": 99621 + } + }, + { + "ph": "s", "id": 99621, "pid": 435, "tid": 435, "ts": 1776649793334462, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793334505, "dur": 1, + "args": { + "External id": 13227, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99639, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99639, "pid": 0, "tid": 7, "ts": 1776649793334505, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334495, "dur": 6, + "args": { + "External id": 13227, + "cbid": 211, "correlation": 99639 + } + }, + { + "ph": "s", "id": 99639, "pid": 435, "tid": 435, "ts": 1776649793334495, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334508, "dur": 0, + "args": { + "External id": 99645, + "cbid": 200, "correlation": 99645 + } + }, + { + "ph": "f", "id": 99645, "pid": 435, "tid": 435, "ts": 1776649793334508, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334508, "dur": 0, + "args": { + "External id": 99646, + "cbid": 273, "correlation": 99646 + } + }, + { + "ph": "f", "id": 99646, "pid": 435, "tid": 435, "ts": 1776649793334508, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793334517, "dur": 19, + "args": { + "External id": 99647, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99647, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99647, "pid": 0, "tid": 7, "ts": 1776649793334517, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334509, "dur": 4, + "args": { + "External id": 99647, + "cbid": 211, "correlation": 99647 + } + }, + { + "ph": "s", "id": 99647, "pid": 435, "tid": 435, "ts": 1776649793334509, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793334552, "dur": 1, + "args": { + "External id": 13232, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99665, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99665, "pid": 0, "tid": 7, "ts": 1776649793334552, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334543, "dur": 6, + "args": { + "External id": 13232, + "cbid": 211, "correlation": 99665 + } + }, + { + "ph": "s", "id": 99665, "pid": 435, "tid": 435, "ts": 1776649793334543, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334555, "dur": 0, + "args": { + "External id": 99671, + "cbid": 200, "correlation": 99671 + } + }, + { + "ph": "f", "id": 99671, "pid": 435, "tid": 435, "ts": 1776649793334555, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334556, "dur": 0, + "args": { + "External id": 99672, + "cbid": 273, "correlation": 99672 + } + }, + { + "ph": "f", "id": 99672, "pid": 435, "tid": 435, "ts": 1776649793334556, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793334564, "dur": 19, + "args": { + "External id": 99673, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99673, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99673, "pid": 0, "tid": 7, "ts": 1776649793334564, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334557, "dur": 4, + "args": { + "External id": 99673, + "cbid": 211, "correlation": 99673 + } + }, + { + "ph": "s", "id": 99673, "pid": 435, "tid": 435, "ts": 1776649793334557, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793334599, "dur": 1, + "args": { + "External id": 13237, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99691, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99691, "pid": 0, "tid": 7, "ts": 1776649793334599, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334590, "dur": 6, + "args": { + "External id": 13237, + "cbid": 211, "correlation": 99691 + } + }, + { + "ph": "s", "id": 99691, "pid": 435, "tid": 435, "ts": 1776649793334590, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334602, "dur": 0, + "args": { + "External id": 99697, + "cbid": 200, "correlation": 99697 + } + }, + { + "ph": "f", "id": 99697, "pid": 435, "tid": 435, "ts": 1776649793334602, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334603, "dur": 0, + "args": { + "External id": 99698, + "cbid": 273, "correlation": 99698 + } + }, + { + "ph": "f", "id": 99698, "pid": 435, "tid": 435, "ts": 1776649793334603, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793334611, "dur": 19, + "args": { + "External id": 99699, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99699, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99699, "pid": 0, "tid": 7, "ts": 1776649793334611, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334604, "dur": 4, + "args": { + "External id": 99699, + "cbid": 211, "correlation": 99699 + } + }, + { + "ph": "s", "id": 99699, "pid": 435, "tid": 435, "ts": 1776649793334604, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793334649, "dur": 1, + "args": { + "External id": 13242, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99717, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99717, "pid": 0, "tid": 7, "ts": 1776649793334649, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334640, "dur": 6, + "args": { + "External id": 13242, + "cbid": 211, "correlation": 99717 + } + }, + { + "ph": "s", "id": 99717, "pid": 435, "tid": 435, "ts": 1776649793334640, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334653, "dur": 0, + "args": { + "External id": 99723, + "cbid": 200, "correlation": 99723 + } + }, + { + "ph": "f", "id": 99723, "pid": 435, "tid": 435, "ts": 1776649793334653, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334654, "dur": 0, + "args": { + "External id": 99724, + "cbid": 273, "correlation": 99724 + } + }, + { + "ph": "f", "id": 99724, "pid": 435, "tid": 435, "ts": 1776649793334654, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793334663, "dur": 19, + "args": { + "External id": 99725, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99725, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99725, "pid": 0, "tid": 7, "ts": 1776649793334663, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334655, "dur": 4, + "args": { + "External id": 99725, + "cbid": 211, "correlation": 99725 + } + }, + { + "ph": "s", "id": 99725, "pid": 435, "tid": 435, "ts": 1776649793334655, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793334698, "dur": 1, + "args": { + "External id": 13247, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99743, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99743, "pid": 0, "tid": 7, "ts": 1776649793334698, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334689, "dur": 6, + "args": { + "External id": 13247, + "cbid": 211, "correlation": 99743 + } + }, + { + "ph": "s", "id": 99743, "pid": 435, "tid": 435, "ts": 1776649793334689, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334701, "dur": 0, + "args": { + "External id": 99749, + "cbid": 200, "correlation": 99749 + } + }, + { + "ph": "f", "id": 99749, "pid": 435, "tid": 435, "ts": 1776649793334701, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334702, "dur": 0, + "args": { + "External id": 99750, + "cbid": 273, "correlation": 99750 + } + }, + { + "ph": "f", "id": 99750, "pid": 435, "tid": 435, "ts": 1776649793334702, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793334710, "dur": 19, + "args": { + "External id": 99751, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99751, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99751, "pid": 0, "tid": 7, "ts": 1776649793334710, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334703, "dur": 4, + "args": { + "External id": 99751, + "cbid": 211, "correlation": 99751 + } + }, + { + "ph": "s", "id": 99751, "pid": 435, "tid": 435, "ts": 1776649793334703, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793334746, "dur": 1, + "args": { + "External id": 13252, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99769, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99769, "pid": 0, "tid": 7, "ts": 1776649793334746, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334736, "dur": 6, + "args": { + "External id": 13252, + "cbid": 211, "correlation": 99769 + } + }, + { + "ph": "s", "id": 99769, "pid": 435, "tid": 435, "ts": 1776649793334736, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334749, "dur": 0, + "args": { + "External id": 99775, + "cbid": 200, "correlation": 99775 + } + }, + { + "ph": "f", "id": 99775, "pid": 435, "tid": 435, "ts": 1776649793334749, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334750, "dur": 0, + "args": { + "External id": 99776, + "cbid": 273, "correlation": 99776 + } + }, + { + "ph": "f", "id": 99776, "pid": 435, "tid": 435, "ts": 1776649793334750, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793334759, "dur": 19, + "args": { + "External id": 99777, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99777, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99777, "pid": 0, "tid": 7, "ts": 1776649793334759, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334751, "dur": 4, + "args": { + "External id": 99777, + "cbid": 211, "correlation": 99777 + } + }, + { + "ph": "s", "id": 99777, "pid": 435, "tid": 435, "ts": 1776649793334751, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793334794, "dur": 1, + "args": { + "External id": 13257, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99795, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99795, "pid": 0, "tid": 7, "ts": 1776649793334794, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334784, "dur": 6, + "args": { + "External id": 13257, + "cbid": 211, "correlation": 99795 + } + }, + { + "ph": "s", "id": 99795, "pid": 435, "tid": 435, "ts": 1776649793334784, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334797, "dur": 0, + "args": { + "External id": 99801, + "cbid": 200, "correlation": 99801 + } + }, + { + "ph": "f", "id": 99801, "pid": 435, "tid": 435, "ts": 1776649793334797, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334797, "dur": 0, + "args": { + "External id": 99802, + "cbid": 273, "correlation": 99802 + } + }, + { + "ph": "f", "id": 99802, "pid": 435, "tid": 435, "ts": 1776649793334797, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793334806, "dur": 19, + "args": { + "External id": 99803, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99803, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99803, "pid": 0, "tid": 7, "ts": 1776649793334806, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334798, "dur": 4, + "args": { + "External id": 99803, + "cbid": 211, "correlation": 99803 + } + }, + { + "ph": "s", "id": 99803, "pid": 435, "tid": 435, "ts": 1776649793334798, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793334842, "dur": 1, + "args": { + "External id": 13262, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99821, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99821, "pid": 0, "tid": 7, "ts": 1776649793334842, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334832, "dur": 6, + "args": { + "External id": 13262, + "cbid": 211, "correlation": 99821 + } + }, + { + "ph": "s", "id": 99821, "pid": 435, "tid": 435, "ts": 1776649793334832, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334844, "dur": 0, + "args": { + "External id": 99827, + "cbid": 200, "correlation": 99827 + } + }, + { + "ph": "f", "id": 99827, "pid": 435, "tid": 435, "ts": 1776649793334844, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334845, "dur": 0, + "args": { + "External id": 99828, + "cbid": 273, "correlation": 99828 + } + }, + { + "ph": "f", "id": 99828, "pid": 435, "tid": 435, "ts": 1776649793334845, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793334854, "dur": 19, + "args": { + "External id": 99829, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99829, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99829, "pid": 0, "tid": 7, "ts": 1776649793334854, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334846, "dur": 5, + "args": { + "External id": 99829, + "cbid": 211, "correlation": 99829 + } + }, + { + "ph": "s", "id": 99829, "pid": 435, "tid": 435, "ts": 1776649793334846, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793334890, "dur": 1, + "args": { + "External id": 13267, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99847, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99847, "pid": 0, "tid": 7, "ts": 1776649793334890, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334880, "dur": 6, + "args": { + "External id": 13267, + "cbid": 211, "correlation": 99847 + } + }, + { + "ph": "s", "id": 99847, "pid": 435, "tid": 435, "ts": 1776649793334880, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334892, "dur": 0, + "args": { + "External id": 99853, + "cbid": 200, "correlation": 99853 + } + }, + { + "ph": "f", "id": 99853, "pid": 435, "tid": 435, "ts": 1776649793334892, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334893, "dur": 0, + "args": { + "External id": 99854, + "cbid": 273, "correlation": 99854 + } + }, + { + "ph": "f", "id": 99854, "pid": 435, "tid": 435, "ts": 1776649793334893, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793334902, "dur": 19, + "args": { + "External id": 99855, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99855, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99855, "pid": 0, "tid": 7, "ts": 1776649793334902, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334894, "dur": 5, + "args": { + "External id": 99855, + "cbid": 211, "correlation": 99855 + } + }, + { + "ph": "s", "id": 99855, "pid": 435, "tid": 435, "ts": 1776649793334894, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793334938, "dur": 1, + "args": { + "External id": 13272, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99873, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99873, "pid": 0, "tid": 7, "ts": 1776649793334938, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334928, "dur": 6, + "args": { + "External id": 13272, + "cbid": 211, "correlation": 99873 + } + }, + { + "ph": "s", "id": 99873, "pid": 435, "tid": 435, "ts": 1776649793334928, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334941, "dur": 0, + "args": { + "External id": 99879, + "cbid": 200, "correlation": 99879 + } + }, + { + "ph": "f", "id": 99879, "pid": 435, "tid": 435, "ts": 1776649793334941, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334941, "dur": 0, + "args": { + "External id": 99880, + "cbid": 273, "correlation": 99880 + } + }, + { + "ph": "f", "id": 99880, "pid": 435, "tid": 435, "ts": 1776649793334941, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793334950, "dur": 19, + "args": { + "External id": 99881, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99881, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99881, "pid": 0, "tid": 7, "ts": 1776649793334950, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334943, "dur": 5, + "args": { + "External id": 99881, + "cbid": 211, "correlation": 99881 + } + }, + { + "ph": "s", "id": 99881, "pid": 435, "tid": 435, "ts": 1776649793334943, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793334986, "dur": 1, + "args": { + "External id": 13277, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99899, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99899, "pid": 0, "tid": 7, "ts": 1776649793334986, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334977, "dur": 6, + "args": { + "External id": 13277, + "cbid": 211, "correlation": 99899 + } + }, + { + "ph": "s", "id": 99899, "pid": 435, "tid": 435, "ts": 1776649793334977, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334989, "dur": 0, + "args": { + "External id": 99905, + "cbid": 200, "correlation": 99905 + } + }, + { + "ph": "f", "id": 99905, "pid": 435, "tid": 435, "ts": 1776649793334989, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793334990, "dur": 0, + "args": { + "External id": 99906, + "cbid": 273, "correlation": 99906 + } + }, + { + "ph": "f", "id": 99906, "pid": 435, "tid": 435, "ts": 1776649793334990, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793334999, "dur": 19, + "args": { + "External id": 99907, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99907, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99907, "pid": 0, "tid": 7, "ts": 1776649793334999, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793334991, "dur": 4, + "args": { + "External id": 99907, + "cbid": 211, "correlation": 99907 + } + }, + { + "ph": "s", "id": 99907, "pid": 435, "tid": 435, "ts": 1776649793334991, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793335034, "dur": 1, + "args": { + "External id": 13282, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99925, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99925, "pid": 0, "tid": 7, "ts": 1776649793335034, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335025, "dur": 6, + "args": { + "External id": 13282, + "cbid": 211, "correlation": 99925 + } + }, + { + "ph": "s", "id": 99925, "pid": 435, "tid": 435, "ts": 1776649793335025, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335037, "dur": 0, + "args": { + "External id": 99931, + "cbid": 200, "correlation": 99931 + } + }, + { + "ph": "f", "id": 99931, "pid": 435, "tid": 435, "ts": 1776649793335037, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335038, "dur": 0, + "args": { + "External id": 99932, + "cbid": 273, "correlation": 99932 + } + }, + { + "ph": "f", "id": 99932, "pid": 435, "tid": 435, "ts": 1776649793335038, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793335046, "dur": 19, + "args": { + "External id": 99933, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99933, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99933, "pid": 0, "tid": 7, "ts": 1776649793335046, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335039, "dur": 4, + "args": { + "External id": 99933, + "cbid": 211, "correlation": 99933 + } + }, + { + "ph": "s", "id": 99933, "pid": 435, "tid": 435, "ts": 1776649793335039, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793335082, "dur": 1, + "args": { + "External id": 13287, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99951, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99951, "pid": 0, "tid": 7, "ts": 1776649793335082, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335073, "dur": 6, + "args": { + "External id": 13287, + "cbid": 211, "correlation": 99951 + } + }, + { + "ph": "s", "id": 99951, "pid": 435, "tid": 435, "ts": 1776649793335073, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335085, "dur": 0, + "args": { + "External id": 99957, + "cbid": 200, "correlation": 99957 + } + }, + { + "ph": "f", "id": 99957, "pid": 435, "tid": 435, "ts": 1776649793335085, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335085, "dur": 0, + "args": { + "External id": 99958, + "cbid": 273, "correlation": 99958 + } + }, + { + "ph": "f", "id": 99958, "pid": 435, "tid": 435, "ts": 1776649793335085, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793335094, "dur": 19, + "args": { + "External id": 99959, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99959, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99959, "pid": 0, "tid": 7, "ts": 1776649793335094, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335086, "dur": 4, + "args": { + "External id": 99959, + "cbid": 211, "correlation": 99959 + } + }, + { + "ph": "s", "id": 99959, "pid": 435, "tid": 435, "ts": 1776649793335086, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793335130, "dur": 1, + "args": { + "External id": 13292, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99977, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99977, "pid": 0, "tid": 7, "ts": 1776649793335130, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335120, "dur": 6, + "args": { + "External id": 13292, + "cbid": 211, "correlation": 99977 + } + }, + { + "ph": "s", "id": 99977, "pid": 435, "tid": 435, "ts": 1776649793335120, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335133, "dur": 0, + "args": { + "External id": 99983, + "cbid": 200, "correlation": 99983 + } + }, + { + "ph": "f", "id": 99983, "pid": 435, "tid": 435, "ts": 1776649793335133, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335133, "dur": 0, + "args": { + "External id": 99984, + "cbid": 273, "correlation": 99984 + } + }, + { + "ph": "f", "id": 99984, "pid": 435, "tid": 435, "ts": 1776649793335133, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793335142, "dur": 19, + "args": { + "External id": 99985, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 99985, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 99985, "pid": 0, "tid": 7, "ts": 1776649793335142, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335134, "dur": 4, + "args": { + "External id": 99985, + "cbid": 211, "correlation": 99985 + } + }, + { + "ph": "s", "id": 99985, "pid": 435, "tid": 435, "ts": 1776649793335134, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793335178, "dur": 1, + "args": { + "External id": 13297, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100003, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100003, "pid": 0, "tid": 7, "ts": 1776649793335178, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335168, "dur": 6, + "args": { + "External id": 13297, + "cbid": 211, "correlation": 100003 + } + }, + { + "ph": "s", "id": 100003, "pid": 435, "tid": 435, "ts": 1776649793335168, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335181, "dur": 0, + "args": { + "External id": 100009, + "cbid": 200, "correlation": 100009 + } + }, + { + "ph": "f", "id": 100009, "pid": 435, "tid": 435, "ts": 1776649793335181, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335181, "dur": 0, + "args": { + "External id": 100010, + "cbid": 273, "correlation": 100010 + } + }, + { + "ph": "f", "id": 100010, "pid": 435, "tid": 435, "ts": 1776649793335181, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793335190, "dur": 19, + "args": { + "External id": 100011, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100011, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100011, "pid": 0, "tid": 7, "ts": 1776649793335190, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335183, "dur": 5, + "args": { + "External id": 100011, + "cbid": 211, "correlation": 100011 + } + }, + { + "ph": "s", "id": 100011, "pid": 435, "tid": 435, "ts": 1776649793335183, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793335226, "dur": 1, + "args": { + "External id": 13302, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100029, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100029, "pid": 0, "tid": 7, "ts": 1776649793335226, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335216, "dur": 6, + "args": { + "External id": 13302, + "cbid": 211, "correlation": 100029 + } + }, + { + "ph": "s", "id": 100029, "pid": 435, "tid": 435, "ts": 1776649793335216, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335229, "dur": 0, + "args": { + "External id": 100035, + "cbid": 200, "correlation": 100035 + } + }, + { + "ph": "f", "id": 100035, "pid": 435, "tid": 435, "ts": 1776649793335229, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335229, "dur": 0, + "args": { + "External id": 100036, + "cbid": 273, "correlation": 100036 + } + }, + { + "ph": "f", "id": 100036, "pid": 435, "tid": 435, "ts": 1776649793335229, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793335238, "dur": 19, + "args": { + "External id": 100037, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100037, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100037, "pid": 0, "tid": 7, "ts": 1776649793335238, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335230, "dur": 4, + "args": { + "External id": 100037, + "cbid": 211, "correlation": 100037 + } + }, + { + "ph": "s", "id": 100037, "pid": 435, "tid": 435, "ts": 1776649793335230, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793335273, "dur": 1, + "args": { + "External id": 13307, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100055, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100055, "pid": 0, "tid": 7, "ts": 1776649793335273, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335264, "dur": 6, + "args": { + "External id": 13307, + "cbid": 211, "correlation": 100055 + } + }, + { + "ph": "s", "id": 100055, "pid": 435, "tid": 435, "ts": 1776649793335264, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335276, "dur": 0, + "args": { + "External id": 100061, + "cbid": 200, "correlation": 100061 + } + }, + { + "ph": "f", "id": 100061, "pid": 435, "tid": 435, "ts": 1776649793335276, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335277, "dur": 0, + "args": { + "External id": 100062, + "cbid": 273, "correlation": 100062 + } + }, + { + "ph": "f", "id": 100062, "pid": 435, "tid": 435, "ts": 1776649793335277, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793335286, "dur": 19, + "args": { + "External id": 100063, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100063, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100063, "pid": 0, "tid": 7, "ts": 1776649793335286, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335278, "dur": 4, + "args": { + "External id": 100063, + "cbid": 211, "correlation": 100063 + } + }, + { + "ph": "s", "id": 100063, "pid": 435, "tid": 435, "ts": 1776649793335278, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793335321, "dur": 1, + "args": { + "External id": 13312, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100081, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100081, "pid": 0, "tid": 7, "ts": 1776649793335321, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335311, "dur": 6, + "args": { + "External id": 13312, + "cbid": 211, "correlation": 100081 + } + }, + { + "ph": "s", "id": 100081, "pid": 435, "tid": 435, "ts": 1776649793335311, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335323, "dur": 0, + "args": { + "External id": 100087, + "cbid": 200, "correlation": 100087 + } + }, + { + "ph": "f", "id": 100087, "pid": 435, "tid": 435, "ts": 1776649793335323, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335324, "dur": 0, + "args": { + "External id": 100088, + "cbid": 273, "correlation": 100088 + } + }, + { + "ph": "f", "id": 100088, "pid": 435, "tid": 435, "ts": 1776649793335324, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793335333, "dur": 19, + "args": { + "External id": 100089, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100089, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100089, "pid": 0, "tid": 7, "ts": 1776649793335333, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335325, "dur": 4, + "args": { + "External id": 100089, + "cbid": 211, "correlation": 100089 + } + }, + { + "ph": "s", "id": 100089, "pid": 435, "tid": 435, "ts": 1776649793335325, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793335378, "dur": 1, + "args": { + "External id": 13317, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100107, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100107, "pid": 0, "tid": 7, "ts": 1776649793335378, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335369, "dur": 6, + "args": { + "External id": 13317, + "cbid": 211, "correlation": 100107 + } + }, + { + "ph": "s", "id": 100107, "pid": 435, "tid": 435, "ts": 1776649793335369, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335381, "dur": 0, + "args": { + "External id": 100113, + "cbid": 200, "correlation": 100113 + } + }, + { + "ph": "f", "id": 100113, "pid": 435, "tid": 435, "ts": 1776649793335381, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335382, "dur": 0, + "args": { + "External id": 100114, + "cbid": 273, "correlation": 100114 + } + }, + { + "ph": "f", "id": 100114, "pid": 435, "tid": 435, "ts": 1776649793335382, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793335391, "dur": 19, + "args": { + "External id": 100115, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100115, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100115, "pid": 0, "tid": 7, "ts": 1776649793335391, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335383, "dur": 4, + "args": { + "External id": 100115, + "cbid": 211, "correlation": 100115 + } + }, + { + "ph": "s", "id": 100115, "pid": 435, "tid": 435, "ts": 1776649793335383, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793335427, "dur": 1, + "args": { + "External id": 13322, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100133, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100133, "pid": 0, "tid": 7, "ts": 1776649793335427, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335417, "dur": 6, + "args": { + "External id": 13322, + "cbid": 211, "correlation": 100133 + } + }, + { + "ph": "s", "id": 100133, "pid": 435, "tid": 435, "ts": 1776649793335417, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335429, "dur": 0, + "args": { + "External id": 100139, + "cbid": 200, "correlation": 100139 + } + }, + { + "ph": "f", "id": 100139, "pid": 435, "tid": 435, "ts": 1776649793335429, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335430, "dur": 0, + "args": { + "External id": 100140, + "cbid": 273, "correlation": 100140 + } + }, + { + "ph": "f", "id": 100140, "pid": 435, "tid": 435, "ts": 1776649793335430, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793335439, "dur": 19, + "args": { + "External id": 100141, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100141, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100141, "pid": 0, "tid": 7, "ts": 1776649793335439, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335431, "dur": 5, + "args": { + "External id": 100141, + "cbid": 211, "correlation": 100141 + } + }, + { + "ph": "s", "id": 100141, "pid": 435, "tid": 435, "ts": 1776649793335431, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793335475, "dur": 1, + "args": { + "External id": 13327, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100159, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100159, "pid": 0, "tid": 7, "ts": 1776649793335475, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335465, "dur": 6, + "args": { + "External id": 13327, + "cbid": 211, "correlation": 100159 + } + }, + { + "ph": "s", "id": 100159, "pid": 435, "tid": 435, "ts": 1776649793335465, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335478, "dur": 0, + "args": { + "External id": 100165, + "cbid": 200, "correlation": 100165 + } + }, + { + "ph": "f", "id": 100165, "pid": 435, "tid": 435, "ts": 1776649793335478, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335478, "dur": 0, + "args": { + "External id": 100166, + "cbid": 273, "correlation": 100166 + } + }, + { + "ph": "f", "id": 100166, "pid": 435, "tid": 435, "ts": 1776649793335478, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793335487, "dur": 19, + "args": { + "External id": 100167, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100167, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100167, "pid": 0, "tid": 7, "ts": 1776649793335487, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335480, "dur": 4, + "args": { + "External id": 100167, + "cbid": 211, "correlation": 100167 + } + }, + { + "ph": "s", "id": 100167, "pid": 435, "tid": 435, "ts": 1776649793335480, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793335523, "dur": 1, + "args": { + "External id": 13332, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100185, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100185, "pid": 0, "tid": 7, "ts": 1776649793335523, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335513, "dur": 6, + "args": { + "External id": 13332, + "cbid": 211, "correlation": 100185 + } + }, + { + "ph": "s", "id": 100185, "pid": 435, "tid": 435, "ts": 1776649793335513, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335526, "dur": 0, + "args": { + "External id": 100191, + "cbid": 200, "correlation": 100191 + } + }, + { + "ph": "f", "id": 100191, "pid": 435, "tid": 435, "ts": 1776649793335526, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335526, "dur": 0, + "args": { + "External id": 100192, + "cbid": 273, "correlation": 100192 + } + }, + { + "ph": "f", "id": 100192, "pid": 435, "tid": 435, "ts": 1776649793335526, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793335535, "dur": 19, + "args": { + "External id": 100193, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100193, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100193, "pid": 0, "tid": 7, "ts": 1776649793335535, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335528, "dur": 4, + "args": { + "External id": 100193, + "cbid": 211, "correlation": 100193 + } + }, + { + "ph": "s", "id": 100193, "pid": 435, "tid": 435, "ts": 1776649793335528, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793335570, "dur": 1, + "args": { + "External id": 13337, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100211, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100211, "pid": 0, "tid": 7, "ts": 1776649793335570, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335561, "dur": 6, + "args": { + "External id": 13337, + "cbid": 211, "correlation": 100211 + } + }, + { + "ph": "s", "id": 100211, "pid": 435, "tid": 435, "ts": 1776649793335561, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335573, "dur": 0, + "args": { + "External id": 100217, + "cbid": 200, "correlation": 100217 + } + }, + { + "ph": "f", "id": 100217, "pid": 435, "tid": 435, "ts": 1776649793335573, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335574, "dur": 0, + "args": { + "External id": 100218, + "cbid": 273, "correlation": 100218 + } + }, + { + "ph": "f", "id": 100218, "pid": 435, "tid": 435, "ts": 1776649793335574, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793335583, "dur": 19, + "args": { + "External id": 100219, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100219, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100219, "pid": 0, "tid": 7, "ts": 1776649793335583, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335575, "dur": 4, + "args": { + "External id": 100219, + "cbid": 211, "correlation": 100219 + } + }, + { + "ph": "s", "id": 100219, "pid": 435, "tid": 435, "ts": 1776649793335575, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793335618, "dur": 1, + "args": { + "External id": 13342, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100237, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100237, "pid": 0, "tid": 7, "ts": 1776649793335618, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335609, "dur": 6, + "args": { + "External id": 13342, + "cbid": 211, "correlation": 100237 + } + }, + { + "ph": "s", "id": 100237, "pid": 435, "tid": 435, "ts": 1776649793335609, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335621, "dur": 0, + "args": { + "External id": 100243, + "cbid": 200, "correlation": 100243 + } + }, + { + "ph": "f", "id": 100243, "pid": 435, "tid": 435, "ts": 1776649793335621, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335622, "dur": 0, + "args": { + "External id": 100244, + "cbid": 273, "correlation": 100244 + } + }, + { + "ph": "f", "id": 100244, "pid": 435, "tid": 435, "ts": 1776649793335622, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793335631, "dur": 19, + "args": { + "External id": 100245, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100245, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100245, "pid": 0, "tid": 7, "ts": 1776649793335631, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335623, "dur": 4, + "args": { + "External id": 100245, + "cbid": 211, "correlation": 100245 + } + }, + { + "ph": "s", "id": 100245, "pid": 435, "tid": 435, "ts": 1776649793335623, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793335671, "dur": 1, + "args": { + "External id": 13347, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100263, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100263, "pid": 0, "tid": 7, "ts": 1776649793335671, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335662, "dur": 6, + "args": { + "External id": 13347, + "cbid": 211, "correlation": 100263 + } + }, + { + "ph": "s", "id": 100263, "pid": 435, "tid": 435, "ts": 1776649793335662, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335674, "dur": 0, + "args": { + "External id": 100269, + "cbid": 200, "correlation": 100269 + } + }, + { + "ph": "f", "id": 100269, "pid": 435, "tid": 435, "ts": 1776649793335674, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335674, "dur": 0, + "args": { + "External id": 100270, + "cbid": 273, "correlation": 100270 + } + }, + { + "ph": "f", "id": 100270, "pid": 435, "tid": 435, "ts": 1776649793335674, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793335683, "dur": 19, + "args": { + "External id": 100271, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100271, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100271, "pid": 0, "tid": 7, "ts": 1776649793335683, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335675, "dur": 4, + "args": { + "External id": 100271, + "cbid": 211, "correlation": 100271 + } + }, + { + "ph": "s", "id": 100271, "pid": 435, "tid": 435, "ts": 1776649793335675, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793335718, "dur": 1, + "args": { + "External id": 13352, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100289, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100289, "pid": 0, "tid": 7, "ts": 1776649793335718, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335709, "dur": 6, + "args": { + "External id": 13352, + "cbid": 211, "correlation": 100289 + } + }, + { + "ph": "s", "id": 100289, "pid": 435, "tid": 435, "ts": 1776649793335709, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335721, "dur": 0, + "args": { + "External id": 100295, + "cbid": 200, "correlation": 100295 + } + }, + { + "ph": "f", "id": 100295, "pid": 435, "tid": 435, "ts": 1776649793335721, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335722, "dur": 0, + "args": { + "External id": 100296, + "cbid": 273, "correlation": 100296 + } + }, + { + "ph": "f", "id": 100296, "pid": 435, "tid": 435, "ts": 1776649793335722, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793335731, "dur": 19, + "args": { + "External id": 100297, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100297, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100297, "pid": 0, "tid": 7, "ts": 1776649793335731, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335723, "dur": 5, + "args": { + "External id": 100297, + "cbid": 211, "correlation": 100297 + } + }, + { + "ph": "s", "id": 100297, "pid": 435, "tid": 435, "ts": 1776649793335723, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793335768, "dur": 1, + "args": { + "External id": 13357, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100315, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100315, "pid": 0, "tid": 7, "ts": 1776649793335768, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335759, "dur": 6, + "args": { + "External id": 13357, + "cbid": 211, "correlation": 100315 + } + }, + { + "ph": "s", "id": 100315, "pid": 435, "tid": 435, "ts": 1776649793335759, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335771, "dur": 0, + "args": { + "External id": 100321, + "cbid": 200, "correlation": 100321 + } + }, + { + "ph": "f", "id": 100321, "pid": 435, "tid": 435, "ts": 1776649793335771, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335772, "dur": 0, + "args": { + "External id": 100322, + "cbid": 273, "correlation": 100322 + } + }, + { + "ph": "f", "id": 100322, "pid": 435, "tid": 435, "ts": 1776649793335772, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793335781, "dur": 19, + "args": { + "External id": 100323, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100323, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100323, "pid": 0, "tid": 7, "ts": 1776649793335781, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335773, "dur": 4, + "args": { + "External id": 100323, + "cbid": 211, "correlation": 100323 + } + }, + { + "ph": "s", "id": 100323, "pid": 435, "tid": 435, "ts": 1776649793335773, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649793335802, "dur": 3, + "args": { + "External id": 100331, + "cbid": 135, "correlation": 100331 + } + }, + { + "ph": "f", "id": 100331, "pid": 435, "tid": 435, "ts": 1776649793335802, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649793335817, "dur": 5, + "args": { + "External id": 100336, + "cbid": 165, "correlation": 100336 + } + }, + { + "ph": "s", "id": 100336, "pid": 435, "tid": 435, "ts": 1776649793335817, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649793335826, "dur": 4, + "args": { + "External id": 100339, + "cbid": 139, "correlation": 100339 + } + }, + { + "ph": "f", "id": 100339, "pid": 435, "tid": 435, "ts": 1776649793335826, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793335881, "dur": 1, + "args": { + "External id": 13362, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100367, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100367, "pid": 0, "tid": 7, "ts": 1776649793335881, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335870, "dur": 8, + "args": { + "External id": 13362, + "cbid": 211, "correlation": 100367 + } + }, + { + "ph": "s", "id": 100367, "pid": 435, "tid": 435, "ts": 1776649793335870, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335886, "dur": 0, + "args": { + "External id": 100373, + "cbid": 200, "correlation": 100373 + } + }, + { + "ph": "f", "id": 100373, "pid": 435, "tid": 435, "ts": 1776649793335886, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335886, "dur": 1, + "args": { + "External id": 100374, + "cbid": 273, "correlation": 100374 + } + }, + { + "ph": "f", "id": 100374, "pid": 435, "tid": 435, "ts": 1776649793335886, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793335896, "dur": 19, + "args": { + "External id": 100375, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100375, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100375, "pid": 0, "tid": 7, "ts": 1776649793335896, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335888, "dur": 5, + "args": { + "External id": 100375, + "cbid": 211, "correlation": 100375 + } + }, + { + "ph": "s", "id": 100375, "pid": 435, "tid": 435, "ts": 1776649793335888, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793335934, "dur": 1, + "args": { + "External id": 13367, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100393, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100393, "pid": 0, "tid": 7, "ts": 1776649793335934, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335924, "dur": 7, + "args": { + "External id": 13367, + "cbid": 211, "correlation": 100393 + } + }, + { + "ph": "s", "id": 100393, "pid": 435, "tid": 435, "ts": 1776649793335924, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335937, "dur": 0, + "args": { + "External id": 100399, + "cbid": 200, "correlation": 100399 + } + }, + { + "ph": "f", "id": 100399, "pid": 435, "tid": 435, "ts": 1776649793335937, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335938, "dur": 0, + "args": { + "External id": 100400, + "cbid": 273, "correlation": 100400 + } + }, + { + "ph": "f", "id": 100400, "pid": 435, "tid": 435, "ts": 1776649793335938, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793335947, "dur": 19, + "args": { + "External id": 100401, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100401, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100401, "pid": 0, "tid": 7, "ts": 1776649793335947, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335939, "dur": 5, + "args": { + "External id": 100401, + "cbid": 211, "correlation": 100401 + } + }, + { + "ph": "s", "id": 100401, "pid": 435, "tid": 435, "ts": 1776649793335939, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793335982, "dur": 1, + "args": { + "External id": 13372, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100419, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100419, "pid": 0, "tid": 7, "ts": 1776649793335982, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335973, "dur": 6, + "args": { + "External id": 13372, + "cbid": 211, "correlation": 100419 + } + }, + { + "ph": "s", "id": 100419, "pid": 435, "tid": 435, "ts": 1776649793335973, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335986, "dur": 0, + "args": { + "External id": 100425, + "cbid": 200, "correlation": 100425 + } + }, + { + "ph": "f", "id": 100425, "pid": 435, "tid": 435, "ts": 1776649793335986, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793335986, "dur": 0, + "args": { + "External id": 100426, + "cbid": 273, "correlation": 100426 + } + }, + { + "ph": "f", "id": 100426, "pid": 435, "tid": 435, "ts": 1776649793335986, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793335996, "dur": 19, + "args": { + "External id": 100427, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100427, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100427, "pid": 0, "tid": 7, "ts": 1776649793335996, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793335988, "dur": 5, + "args": { + "External id": 100427, + "cbid": 211, "correlation": 100427 + } + }, + { + "ph": "s", "id": 100427, "pid": 435, "tid": 435, "ts": 1776649793335988, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793336031, "dur": 1, + "args": { + "External id": 13377, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100445, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100445, "pid": 0, "tid": 7, "ts": 1776649793336031, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336022, "dur": 6, + "args": { + "External id": 13377, + "cbid": 211, "correlation": 100445 + } + }, + { + "ph": "s", "id": 100445, "pid": 435, "tid": 435, "ts": 1776649793336022, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336034, "dur": 0, + "args": { + "External id": 100451, + "cbid": 200, "correlation": 100451 + } + }, + { + "ph": "f", "id": 100451, "pid": 435, "tid": 435, "ts": 1776649793336034, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336035, "dur": 0, + "args": { + "External id": 100452, + "cbid": 273, "correlation": 100452 + } + }, + { + "ph": "f", "id": 100452, "pid": 435, "tid": 435, "ts": 1776649793336035, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793336044, "dur": 19, + "args": { + "External id": 100453, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100453, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100453, "pid": 0, "tid": 7, "ts": 1776649793336044, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336036, "dur": 5, + "args": { + "External id": 100453, + "cbid": 211, "correlation": 100453 + } + }, + { + "ph": "s", "id": 100453, "pid": 435, "tid": 435, "ts": 1776649793336036, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793336080, "dur": 1, + "args": { + "External id": 13382, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100471, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100471, "pid": 0, "tid": 7, "ts": 1776649793336080, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336070, "dur": 6, + "args": { + "External id": 13382, + "cbid": 211, "correlation": 100471 + } + }, + { + "ph": "s", "id": 100471, "pid": 435, "tid": 435, "ts": 1776649793336070, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336083, "dur": 0, + "args": { + "External id": 100477, + "cbid": 200, "correlation": 100477 + } + }, + { + "ph": "f", "id": 100477, "pid": 435, "tid": 435, "ts": 1776649793336083, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336084, "dur": 0, + "args": { + "External id": 100478, + "cbid": 273, "correlation": 100478 + } + }, + { + "ph": "f", "id": 100478, "pid": 435, "tid": 435, "ts": 1776649793336084, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793336093, "dur": 19, + "args": { + "External id": 100479, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100479, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100479, "pid": 0, "tid": 7, "ts": 1776649793336093, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336085, "dur": 5, + "args": { + "External id": 100479, + "cbid": 211, "correlation": 100479 + } + }, + { + "ph": "s", "id": 100479, "pid": 435, "tid": 435, "ts": 1776649793336085, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793336128, "dur": 1, + "args": { + "External id": 13387, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100497, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100497, "pid": 0, "tid": 7, "ts": 1776649793336128, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336119, "dur": 6, + "args": { + "External id": 13387, + "cbid": 211, "correlation": 100497 + } + }, + { + "ph": "s", "id": 100497, "pid": 435, "tid": 435, "ts": 1776649793336119, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336131, "dur": 0, + "args": { + "External id": 100503, + "cbid": 200, "correlation": 100503 + } + }, + { + "ph": "f", "id": 100503, "pid": 435, "tid": 435, "ts": 1776649793336131, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336132, "dur": 0, + "args": { + "External id": 100504, + "cbid": 273, "correlation": 100504 + } + }, + { + "ph": "f", "id": 100504, "pid": 435, "tid": 435, "ts": 1776649793336132, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793336141, "dur": 19, + "args": { + "External id": 100505, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100505, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100505, "pid": 0, "tid": 7, "ts": 1776649793336141, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336133, "dur": 5, + "args": { + "External id": 100505, + "cbid": 211, "correlation": 100505 + } + }, + { + "ph": "s", "id": 100505, "pid": 435, "tid": 435, "ts": 1776649793336133, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793336176, "dur": 1, + "args": { + "External id": 13392, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100523, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100523, "pid": 0, "tid": 7, "ts": 1776649793336176, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336167, "dur": 6, + "args": { + "External id": 13392, + "cbid": 211, "correlation": 100523 + } + }, + { + "ph": "s", "id": 100523, "pid": 435, "tid": 435, "ts": 1776649793336167, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336179, "dur": 0, + "args": { + "External id": 100529, + "cbid": 200, "correlation": 100529 + } + }, + { + "ph": "f", "id": 100529, "pid": 435, "tid": 435, "ts": 1776649793336179, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336180, "dur": 0, + "args": { + "External id": 100530, + "cbid": 273, "correlation": 100530 + } + }, + { + "ph": "f", "id": 100530, "pid": 435, "tid": 435, "ts": 1776649793336180, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793336189, "dur": 19, + "args": { + "External id": 100531, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100531, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100531, "pid": 0, "tid": 7, "ts": 1776649793336189, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336181, "dur": 4, + "args": { + "External id": 100531, + "cbid": 211, "correlation": 100531 + } + }, + { + "ph": "s", "id": 100531, "pid": 435, "tid": 435, "ts": 1776649793336181, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793336242, "dur": 1, + "args": { + "External id": 13397, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100549, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100549, "pid": 0, "tid": 7, "ts": 1776649793336242, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336215, "dur": 25, + "args": { + "External id": 13397, + "cbid": 211, "correlation": 100549 + } + }, + { + "ph": "s", "id": 100549, "pid": 435, "tid": 435, "ts": 1776649793336215, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336247, "dur": 0, + "args": { + "External id": 100555, + "cbid": 200, "correlation": 100555 + } + }, + { + "ph": "f", "id": 100555, "pid": 435, "tid": 435, "ts": 1776649793336247, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336248, "dur": 0, + "args": { + "External id": 100556, + "cbid": 273, "correlation": 100556 + } + }, + { + "ph": "f", "id": 100556, "pid": 435, "tid": 435, "ts": 1776649793336248, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793336257, "dur": 19, + "args": { + "External id": 100557, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100557, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100557, "pid": 0, "tid": 7, "ts": 1776649793336257, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336249, "dur": 5, + "args": { + "External id": 100557, + "cbid": 211, "correlation": 100557 + } + }, + { + "ph": "s", "id": 100557, "pid": 435, "tid": 435, "ts": 1776649793336249, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793336295, "dur": 1, + "args": { + "External id": 13402, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100575, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100575, "pid": 0, "tid": 7, "ts": 1776649793336295, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336286, "dur": 6, + "args": { + "External id": 13402, + "cbid": 211, "correlation": 100575 + } + }, + { + "ph": "s", "id": 100575, "pid": 435, "tid": 435, "ts": 1776649793336286, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336298, "dur": 0, + "args": { + "External id": 100581, + "cbid": 200, "correlation": 100581 + } + }, + { + "ph": "f", "id": 100581, "pid": 435, "tid": 435, "ts": 1776649793336298, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336299, "dur": 0, + "args": { + "External id": 100582, + "cbid": 273, "correlation": 100582 + } + }, + { + "ph": "f", "id": 100582, "pid": 435, "tid": 435, "ts": 1776649793336299, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793336308, "dur": 19, + "args": { + "External id": 100583, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100583, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100583, "pid": 0, "tid": 7, "ts": 1776649793336308, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336300, "dur": 4, + "args": { + "External id": 100583, + "cbid": 211, "correlation": 100583 + } + }, + { + "ph": "s", "id": 100583, "pid": 435, "tid": 435, "ts": 1776649793336300, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793336344, "dur": 1, + "args": { + "External id": 13407, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100601, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100601, "pid": 0, "tid": 7, "ts": 1776649793336344, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336335, "dur": 6, + "args": { + "External id": 13407, + "cbid": 211, "correlation": 100601 + } + }, + { + "ph": "s", "id": 100601, "pid": 435, "tid": 435, "ts": 1776649793336335, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336347, "dur": 0, + "args": { + "External id": 100607, + "cbid": 200, "correlation": 100607 + } + }, + { + "ph": "f", "id": 100607, "pid": 435, "tid": 435, "ts": 1776649793336347, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336348, "dur": 0, + "args": { + "External id": 100608, + "cbid": 273, "correlation": 100608 + } + }, + { + "ph": "f", "id": 100608, "pid": 435, "tid": 435, "ts": 1776649793336348, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793336357, "dur": 19, + "args": { + "External id": 100609, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100609, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100609, "pid": 0, "tid": 7, "ts": 1776649793336357, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336349, "dur": 4, + "args": { + "External id": 100609, + "cbid": 211, "correlation": 100609 + } + }, + { + "ph": "s", "id": 100609, "pid": 435, "tid": 435, "ts": 1776649793336349, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649793336371, "dur": 9, + "args": { + "External id": 100612, + "cbid": 165, "correlation": 100612 + } + }, + { + "ph": "s", "id": 100612, "pid": 435, "tid": 435, "ts": 1776649793336371, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649793336402, "dur": 3, + "args": { + "External id": 100621, + "cbid": 135, "correlation": 100621 + } + }, + { + "ph": "f", "id": 100621, "pid": 435, "tid": 435, "ts": 1776649793336402, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793336450, "dur": 1, + "args": { + "External id": 13412, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100641, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100641, "pid": 0, "tid": 7, "ts": 1776649793336450, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336440, "dur": 8, + "args": { + "External id": 13412, + "cbid": 211, "correlation": 100641 + } + }, + { + "ph": "s", "id": 100641, "pid": 435, "tid": 435, "ts": 1776649793336440, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336455, "dur": 0, + "args": { + "External id": 100647, + "cbid": 200, "correlation": 100647 + } + }, + { + "ph": "f", "id": 100647, "pid": 435, "tid": 435, "ts": 1776649793336455, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336456, "dur": 0, + "args": { + "External id": 100648, + "cbid": 273, "correlation": 100648 + } + }, + { + "ph": "f", "id": 100648, "pid": 435, "tid": 435, "ts": 1776649793336456, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793336465, "dur": 19, + "args": { + "External id": 100649, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100649, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100649, "pid": 0, "tid": 7, "ts": 1776649793336465, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336457, "dur": 5, + "args": { + "External id": 100649, + "cbid": 211, "correlation": 100649 + } + }, + { + "ph": "s", "id": 100649, "pid": 435, "tid": 435, "ts": 1776649793336457, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793336503, "dur": 1, + "args": { + "External id": 13417, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100667, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100667, "pid": 0, "tid": 7, "ts": 1776649793336503, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336493, "dur": 7, + "args": { + "External id": 13417, + "cbid": 211, "correlation": 100667 + } + }, + { + "ph": "s", "id": 100667, "pid": 435, "tid": 435, "ts": 1776649793336493, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336506, "dur": 0, + "args": { + "External id": 100673, + "cbid": 200, "correlation": 100673 + } + }, + { + "ph": "f", "id": 100673, "pid": 435, "tid": 435, "ts": 1776649793336506, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336507, "dur": 0, + "args": { + "External id": 100674, + "cbid": 273, "correlation": 100674 + } + }, + { + "ph": "f", "id": 100674, "pid": 435, "tid": 435, "ts": 1776649793336507, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793336516, "dur": 19, + "args": { + "External id": 100675, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100675, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100675, "pid": 0, "tid": 7, "ts": 1776649793336516, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336508, "dur": 5, + "args": { + "External id": 100675, + "cbid": 211, "correlation": 100675 + } + }, + { + "ph": "s", "id": 100675, "pid": 435, "tid": 435, "ts": 1776649793336508, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793336552, "dur": 1, + "args": { + "External id": 13422, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100693, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100693, "pid": 0, "tid": 7, "ts": 1776649793336552, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336542, "dur": 6, + "args": { + "External id": 13422, + "cbid": 211, "correlation": 100693 + } + }, + { + "ph": "s", "id": 100693, "pid": 435, "tid": 435, "ts": 1776649793336542, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336555, "dur": 0, + "args": { + "External id": 100699, + "cbid": 200, "correlation": 100699 + } + }, + { + "ph": "f", "id": 100699, "pid": 435, "tid": 435, "ts": 1776649793336555, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336555, "dur": 0, + "args": { + "External id": 100700, + "cbid": 273, "correlation": 100700 + } + }, + { + "ph": "f", "id": 100700, "pid": 435, "tid": 435, "ts": 1776649793336555, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793336564, "dur": 19, + "args": { + "External id": 100701, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100701, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100701, "pid": 0, "tid": 7, "ts": 1776649793336564, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336556, "dur": 4, + "args": { + "External id": 100701, + "cbid": 211, "correlation": 100701 + } + }, + { + "ph": "s", "id": 100701, "pid": 435, "tid": 435, "ts": 1776649793336556, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793336600, "dur": 1, + "args": { + "External id": 13427, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100719, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100719, "pid": 0, "tid": 7, "ts": 1776649793336600, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336591, "dur": 6, + "args": { + "External id": 13427, + "cbid": 211, "correlation": 100719 + } + }, + { + "ph": "s", "id": 100719, "pid": 435, "tid": 435, "ts": 1776649793336591, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336603, "dur": 0, + "args": { + "External id": 100725, + "cbid": 200, "correlation": 100725 + } + }, + { + "ph": "f", "id": 100725, "pid": 435, "tid": 435, "ts": 1776649793336603, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336604, "dur": 0, + "args": { + "External id": 100726, + "cbid": 273, "correlation": 100726 + } + }, + { + "ph": "f", "id": 100726, "pid": 435, "tid": 435, "ts": 1776649793336604, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793336613, "dur": 19, + "args": { + "External id": 100727, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100727, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100727, "pid": 0, "tid": 7, "ts": 1776649793336613, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336605, "dur": 5, + "args": { + "External id": 100727, + "cbid": 211, "correlation": 100727 + } + }, + { + "ph": "s", "id": 100727, "pid": 435, "tid": 435, "ts": 1776649793336605, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793336649, "dur": 1, + "args": { + "External id": 13432, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100745, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100745, "pid": 0, "tid": 7, "ts": 1776649793336649, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336640, "dur": 6, + "args": { + "External id": 13432, + "cbid": 211, "correlation": 100745 + } + }, + { + "ph": "s", "id": 100745, "pid": 435, "tid": 435, "ts": 1776649793336640, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336652, "dur": 0, + "args": { + "External id": 100751, + "cbid": 200, "correlation": 100751 + } + }, + { + "ph": "f", "id": 100751, "pid": 435, "tid": 435, "ts": 1776649793336652, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336652, "dur": 0, + "args": { + "External id": 100752, + "cbid": 273, "correlation": 100752 + } + }, + { + "ph": "f", "id": 100752, "pid": 435, "tid": 435, "ts": 1776649793336652, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793336661, "dur": 19, + "args": { + "External id": 100753, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100753, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100753, "pid": 0, "tid": 7, "ts": 1776649793336661, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336654, "dur": 4, + "args": { + "External id": 100753, + "cbid": 211, "correlation": 100753 + } + }, + { + "ph": "s", "id": 100753, "pid": 435, "tid": 435, "ts": 1776649793336654, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793336697, "dur": 1, + "args": { + "External id": 13437, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100771, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100771, "pid": 0, "tid": 7, "ts": 1776649793336697, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336688, "dur": 6, + "args": { + "External id": 13437, + "cbid": 211, "correlation": 100771 + } + }, + { + "ph": "s", "id": 100771, "pid": 435, "tid": 435, "ts": 1776649793336688, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336700, "dur": 0, + "args": { + "External id": 100777, + "cbid": 200, "correlation": 100777 + } + }, + { + "ph": "f", "id": 100777, "pid": 435, "tid": 435, "ts": 1776649793336700, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336701, "dur": 0, + "args": { + "External id": 100778, + "cbid": 273, "correlation": 100778 + } + }, + { + "ph": "f", "id": 100778, "pid": 435, "tid": 435, "ts": 1776649793336701, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793336709, "dur": 19, + "args": { + "External id": 100779, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100779, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100779, "pid": 0, "tid": 7, "ts": 1776649793336709, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336702, "dur": 4, + "args": { + "External id": 100779, + "cbid": 211, "correlation": 100779 + } + }, + { + "ph": "s", "id": 100779, "pid": 435, "tid": 435, "ts": 1776649793336702, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793336745, "dur": 1, + "args": { + "External id": 13442, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100797, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100797, "pid": 0, "tid": 7, "ts": 1776649793336745, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336736, "dur": 6, + "args": { + "External id": 13442, + "cbid": 211, "correlation": 100797 + } + }, + { + "ph": "s", "id": 100797, "pid": 435, "tid": 435, "ts": 1776649793336736, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336748, "dur": 0, + "args": { + "External id": 100803, + "cbid": 200, "correlation": 100803 + } + }, + { + "ph": "f", "id": 100803, "pid": 435, "tid": 435, "ts": 1776649793336748, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336749, "dur": 0, + "args": { + "External id": 100804, + "cbid": 273, "correlation": 100804 + } + }, + { + "ph": "f", "id": 100804, "pid": 435, "tid": 435, "ts": 1776649793336749, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793336758, "dur": 19, + "args": { + "External id": 100805, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100805, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100805, "pid": 0, "tid": 7, "ts": 1776649793336758, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336750, "dur": 4, + "args": { + "External id": 100805, + "cbid": 211, "correlation": 100805 + } + }, + { + "ph": "s", "id": 100805, "pid": 435, "tid": 435, "ts": 1776649793336750, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793336793, "dur": 1, + "args": { + "External id": 13447, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100823, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100823, "pid": 0, "tid": 7, "ts": 1776649793336793, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336784, "dur": 6, + "args": { + "External id": 13447, + "cbid": 211, "correlation": 100823 + } + }, + { + "ph": "s", "id": 100823, "pid": 435, "tid": 435, "ts": 1776649793336784, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336796, "dur": 0, + "args": { + "External id": 100829, + "cbid": 200, "correlation": 100829 + } + }, + { + "ph": "f", "id": 100829, "pid": 435, "tid": 435, "ts": 1776649793336796, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336797, "dur": 0, + "args": { + "External id": 100830, + "cbid": 273, "correlation": 100830 + } + }, + { + "ph": "f", "id": 100830, "pid": 435, "tid": 435, "ts": 1776649793336797, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793336806, "dur": 19, + "args": { + "External id": 100831, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100831, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100831, "pid": 0, "tid": 7, "ts": 1776649793336806, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336798, "dur": 5, + "args": { + "External id": 100831, + "cbid": 211, "correlation": 100831 + } + }, + { + "ph": "s", "id": 100831, "pid": 435, "tid": 435, "ts": 1776649793336798, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793336841, "dur": 1, + "args": { + "External id": 13452, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100849, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100849, "pid": 0, "tid": 7, "ts": 1776649793336841, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336832, "dur": 6, + "args": { + "External id": 13452, + "cbid": 211, "correlation": 100849 + } + }, + { + "ph": "s", "id": 100849, "pid": 435, "tid": 435, "ts": 1776649793336832, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336844, "dur": 0, + "args": { + "External id": 100855, + "cbid": 200, "correlation": 100855 + } + }, + { + "ph": "f", "id": 100855, "pid": 435, "tid": 435, "ts": 1776649793336844, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336845, "dur": 0, + "args": { + "External id": 100856, + "cbid": 273, "correlation": 100856 + } + }, + { + "ph": "f", "id": 100856, "pid": 435, "tid": 435, "ts": 1776649793336845, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793336854, "dur": 19, + "args": { + "External id": 100857, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100857, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100857, "pid": 0, "tid": 7, "ts": 1776649793336854, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336846, "dur": 4, + "args": { + "External id": 100857, + "cbid": 211, "correlation": 100857 + } + }, + { + "ph": "s", "id": 100857, "pid": 435, "tid": 435, "ts": 1776649793336846, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793336889, "dur": 1, + "args": { + "External id": 13457, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100875, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100875, "pid": 0, "tid": 7, "ts": 1776649793336889, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336880, "dur": 6, + "args": { + "External id": 13457, + "cbid": 211, "correlation": 100875 + } + }, + { + "ph": "s", "id": 100875, "pid": 435, "tid": 435, "ts": 1776649793336880, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336892, "dur": 0, + "args": { + "External id": 100881, + "cbid": 200, "correlation": 100881 + } + }, + { + "ph": "f", "id": 100881, "pid": 435, "tid": 435, "ts": 1776649793336892, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336893, "dur": 0, + "args": { + "External id": 100882, + "cbid": 273, "correlation": 100882 + } + }, + { + "ph": "f", "id": 100882, "pid": 435, "tid": 435, "ts": 1776649793336893, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793336902, "dur": 19, + "args": { + "External id": 100883, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100883, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100883, "pid": 0, "tid": 7, "ts": 1776649793336902, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336894, "dur": 5, + "args": { + "External id": 100883, + "cbid": 211, "correlation": 100883 + } + }, + { + "ph": "s", "id": 100883, "pid": 435, "tid": 435, "ts": 1776649793336894, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793336937, "dur": 1, + "args": { + "External id": 13462, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100901, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100901, "pid": 0, "tid": 7, "ts": 1776649793336937, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336928, "dur": 6, + "args": { + "External id": 13462, + "cbid": 211, "correlation": 100901 + } + }, + { + "ph": "s", "id": 100901, "pid": 435, "tid": 435, "ts": 1776649793336928, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336940, "dur": 0, + "args": { + "External id": 100907, + "cbid": 200, "correlation": 100907 + } + }, + { + "ph": "f", "id": 100907, "pid": 435, "tid": 435, "ts": 1776649793336940, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336941, "dur": 0, + "args": { + "External id": 100908, + "cbid": 273, "correlation": 100908 + } + }, + { + "ph": "f", "id": 100908, "pid": 435, "tid": 435, "ts": 1776649793336941, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793336950, "dur": 19, + "args": { + "External id": 100909, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100909, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100909, "pid": 0, "tid": 7, "ts": 1776649793336950, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336942, "dur": 4, + "args": { + "External id": 100909, + "cbid": 211, "correlation": 100909 + } + }, + { + "ph": "s", "id": 100909, "pid": 435, "tid": 435, "ts": 1776649793336942, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793336985, "dur": 1, + "args": { + "External id": 13467, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100927, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100927, "pid": 0, "tid": 7, "ts": 1776649793336985, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336976, "dur": 6, + "args": { + "External id": 13467, + "cbid": 211, "correlation": 100927 + } + }, + { + "ph": "s", "id": 100927, "pid": 435, "tid": 435, "ts": 1776649793336976, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336988, "dur": 0, + "args": { + "External id": 100933, + "cbid": 200, "correlation": 100933 + } + }, + { + "ph": "f", "id": 100933, "pid": 435, "tid": 435, "ts": 1776649793336988, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793336989, "dur": 0, + "args": { + "External id": 100934, + "cbid": 273, "correlation": 100934 + } + }, + { + "ph": "f", "id": 100934, "pid": 435, "tid": 435, "ts": 1776649793336989, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793336998, "dur": 19, + "args": { + "External id": 100935, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100935, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100935, "pid": 0, "tid": 7, "ts": 1776649793336998, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793336990, "dur": 4, + "args": { + "External id": 100935, + "cbid": 211, "correlation": 100935 + } + }, + { + "ph": "s", "id": 100935, "pid": 435, "tid": 435, "ts": 1776649793336990, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793337033, "dur": 1, + "args": { + "External id": 13472, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100953, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100953, "pid": 0, "tid": 7, "ts": 1776649793337033, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337024, "dur": 7, + "args": { + "External id": 13472, + "cbid": 211, "correlation": 100953 + } + }, + { + "ph": "s", "id": 100953, "pid": 435, "tid": 435, "ts": 1776649793337024, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337037, "dur": 0, + "args": { + "External id": 100959, + "cbid": 200, "correlation": 100959 + } + }, + { + "ph": "f", "id": 100959, "pid": 435, "tid": 435, "ts": 1776649793337037, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337037, "dur": 0, + "args": { + "External id": 100960, + "cbid": 273, "correlation": 100960 + } + }, + { + "ph": "f", "id": 100960, "pid": 435, "tid": 435, "ts": 1776649793337037, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793337046, "dur": 19, + "args": { + "External id": 100961, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100961, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100961, "pid": 0, "tid": 7, "ts": 1776649793337046, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337039, "dur": 5, + "args": { + "External id": 100961, + "cbid": 211, "correlation": 100961 + } + }, + { + "ph": "s", "id": 100961, "pid": 435, "tid": 435, "ts": 1776649793337039, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793337082, "dur": 1, + "args": { + "External id": 13477, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100979, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100979, "pid": 0, "tid": 7, "ts": 1776649793337082, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337073, "dur": 6, + "args": { + "External id": 13477, + "cbid": 211, "correlation": 100979 + } + }, + { + "ph": "s", "id": 100979, "pid": 435, "tid": 435, "ts": 1776649793337073, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337085, "dur": 0, + "args": { + "External id": 100985, + "cbid": 200, "correlation": 100985 + } + }, + { + "ph": "f", "id": 100985, "pid": 435, "tid": 435, "ts": 1776649793337085, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337086, "dur": 0, + "args": { + "External id": 100986, + "cbid": 273, "correlation": 100986 + } + }, + { + "ph": "f", "id": 100986, "pid": 435, "tid": 435, "ts": 1776649793337086, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793337095, "dur": 19, + "args": { + "External id": 100987, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 100987, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 100987, "pid": 0, "tid": 7, "ts": 1776649793337095, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337087, "dur": 4, + "args": { + "External id": 100987, + "cbid": 211, "correlation": 100987 + } + }, + { + "ph": "s", "id": 100987, "pid": 435, "tid": 435, "ts": 1776649793337087, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793337130, "dur": 1, + "args": { + "External id": 13482, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101005, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101005, "pid": 0, "tid": 7, "ts": 1776649793337130, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337121, "dur": 6, + "args": { + "External id": 13482, + "cbid": 211, "correlation": 101005 + } + }, + { + "ph": "s", "id": 101005, "pid": 435, "tid": 435, "ts": 1776649793337121, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337133, "dur": 0, + "args": { + "External id": 101011, + "cbid": 200, "correlation": 101011 + } + }, + { + "ph": "f", "id": 101011, "pid": 435, "tid": 435, "ts": 1776649793337133, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337134, "dur": 0, + "args": { + "External id": 101012, + "cbid": 273, "correlation": 101012 + } + }, + { + "ph": "f", "id": 101012, "pid": 435, "tid": 435, "ts": 1776649793337134, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793337143, "dur": 19, + "args": { + "External id": 101013, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101013, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101013, "pid": 0, "tid": 7, "ts": 1776649793337143, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337135, "dur": 4, + "args": { + "External id": 101013, + "cbid": 211, "correlation": 101013 + } + }, + { + "ph": "s", "id": 101013, "pid": 435, "tid": 435, "ts": 1776649793337135, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793337178, "dur": 1, + "args": { + "External id": 13487, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101031, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101031, "pid": 0, "tid": 7, "ts": 1776649793337178, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337169, "dur": 6, + "args": { + "External id": 13487, + "cbid": 211, "correlation": 101031 + } + }, + { + "ph": "s", "id": 101031, "pid": 435, "tid": 435, "ts": 1776649793337169, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337181, "dur": 0, + "args": { + "External id": 101037, + "cbid": 200, "correlation": 101037 + } + }, + { + "ph": "f", "id": 101037, "pid": 435, "tid": 435, "ts": 1776649793337181, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337182, "dur": 0, + "args": { + "External id": 101038, + "cbid": 273, "correlation": 101038 + } + }, + { + "ph": "f", "id": 101038, "pid": 435, "tid": 435, "ts": 1776649793337182, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793337191, "dur": 19, + "args": { + "External id": 101039, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101039, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101039, "pid": 0, "tid": 7, "ts": 1776649793337191, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337183, "dur": 5, + "args": { + "External id": 101039, + "cbid": 211, "correlation": 101039 + } + }, + { + "ph": "s", "id": 101039, "pid": 435, "tid": 435, "ts": 1776649793337183, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793337227, "dur": 1, + "args": { + "External id": 13492, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101057, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101057, "pid": 0, "tid": 7, "ts": 1776649793337227, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337218, "dur": 6, + "args": { + "External id": 13492, + "cbid": 211, "correlation": 101057 + } + }, + { + "ph": "s", "id": 101057, "pid": 435, "tid": 435, "ts": 1776649793337218, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337230, "dur": 0, + "args": { + "External id": 101063, + "cbid": 200, "correlation": 101063 + } + }, + { + "ph": "f", "id": 101063, "pid": 435, "tid": 435, "ts": 1776649793337230, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337230, "dur": 0, + "args": { + "External id": 101064, + "cbid": 273, "correlation": 101064 + } + }, + { + "ph": "f", "id": 101064, "pid": 435, "tid": 435, "ts": 1776649793337230, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793337239, "dur": 19, + "args": { + "External id": 101065, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101065, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101065, "pid": 0, "tid": 7, "ts": 1776649793337239, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337232, "dur": 4, + "args": { + "External id": 101065, + "cbid": 211, "correlation": 101065 + } + }, + { + "ph": "s", "id": 101065, "pid": 435, "tid": 435, "ts": 1776649793337232, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793337275, "dur": 1, + "args": { + "External id": 13497, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101083, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101083, "pid": 0, "tid": 7, "ts": 1776649793337275, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337266, "dur": 6, + "args": { + "External id": 13497, + "cbid": 211, "correlation": 101083 + } + }, + { + "ph": "s", "id": 101083, "pid": 435, "tid": 435, "ts": 1776649793337266, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337278, "dur": 0, + "args": { + "External id": 101089, + "cbid": 200, "correlation": 101089 + } + }, + { + "ph": "f", "id": 101089, "pid": 435, "tid": 435, "ts": 1776649793337278, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337279, "dur": 0, + "args": { + "External id": 101090, + "cbid": 273, "correlation": 101090 + } + }, + { + "ph": "f", "id": 101090, "pid": 435, "tid": 435, "ts": 1776649793337279, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793337288, "dur": 19, + "args": { + "External id": 101091, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101091, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101091, "pid": 0, "tid": 7, "ts": 1776649793337288, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337280, "dur": 5, + "args": { + "External id": 101091, + "cbid": 211, "correlation": 101091 + } + }, + { + "ph": "s", "id": 101091, "pid": 435, "tid": 435, "ts": 1776649793337280, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793337323, "dur": 1, + "args": { + "External id": 13502, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101109, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101109, "pid": 0, "tid": 7, "ts": 1776649793337323, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337314, "dur": 6, + "args": { + "External id": 13502, + "cbid": 211, "correlation": 101109 + } + }, + { + "ph": "s", "id": 101109, "pid": 435, "tid": 435, "ts": 1776649793337314, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337327, "dur": 0, + "args": { + "External id": 101115, + "cbid": 200, "correlation": 101115 + } + }, + { + "ph": "f", "id": 101115, "pid": 435, "tid": 435, "ts": 1776649793337327, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337327, "dur": 0, + "args": { + "External id": 101116, + "cbid": 273, "correlation": 101116 + } + }, + { + "ph": "f", "id": 101116, "pid": 435, "tid": 435, "ts": 1776649793337327, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793337336, "dur": 19, + "args": { + "External id": 101117, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101117, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101117, "pid": 0, "tid": 7, "ts": 1776649793337336, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337328, "dur": 4, + "args": { + "External id": 101117, + "cbid": 211, "correlation": 101117 + } + }, + { + "ph": "s", "id": 101117, "pid": 435, "tid": 435, "ts": 1776649793337328, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793337372, "dur": 1, + "args": { + "External id": 13507, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101135, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101135, "pid": 0, "tid": 7, "ts": 1776649793337372, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337362, "dur": 6, + "args": { + "External id": 13507, + "cbid": 211, "correlation": 101135 + } + }, + { + "ph": "s", "id": 101135, "pid": 435, "tid": 435, "ts": 1776649793337362, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337374, "dur": 0, + "args": { + "External id": 101141, + "cbid": 200, "correlation": 101141 + } + }, + { + "ph": "f", "id": 101141, "pid": 435, "tid": 435, "ts": 1776649793337374, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337375, "dur": 0, + "args": { + "External id": 101142, + "cbid": 273, "correlation": 101142 + } + }, + { + "ph": "f", "id": 101142, "pid": 435, "tid": 435, "ts": 1776649793337375, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793337384, "dur": 19, + "args": { + "External id": 101143, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101143, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101143, "pid": 0, "tid": 7, "ts": 1776649793337384, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337376, "dur": 5, + "args": { + "External id": 101143, + "cbid": 211, "correlation": 101143 + } + }, + { + "ph": "s", "id": 101143, "pid": 435, "tid": 435, "ts": 1776649793337376, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793337419, "dur": 1, + "args": { + "External id": 13512, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101161, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101161, "pid": 0, "tid": 7, "ts": 1776649793337419, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337410, "dur": 6, + "args": { + "External id": 13512, + "cbid": 211, "correlation": 101161 + } + }, + { + "ph": "s", "id": 101161, "pid": 435, "tid": 435, "ts": 1776649793337410, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337422, "dur": 0, + "args": { + "External id": 101167, + "cbid": 200, "correlation": 101167 + } + }, + { + "ph": "f", "id": 101167, "pid": 435, "tid": 435, "ts": 1776649793337422, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337423, "dur": 0, + "args": { + "External id": 101168, + "cbid": 273, "correlation": 101168 + } + }, + { + "ph": "f", "id": 101168, "pid": 435, "tid": 435, "ts": 1776649793337423, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793337432, "dur": 19, + "args": { + "External id": 101169, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101169, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101169, "pid": 0, "tid": 7, "ts": 1776649793337432, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337424, "dur": 4, + "args": { + "External id": 101169, + "cbid": 211, "correlation": 101169 + } + }, + { + "ph": "s", "id": 101169, "pid": 435, "tid": 435, "ts": 1776649793337424, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793337467, "dur": 1, + "args": { + "External id": 13517, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101187, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101187, "pid": 0, "tid": 7, "ts": 1776649793337467, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337458, "dur": 6, + "args": { + "External id": 13517, + "cbid": 211, "correlation": 101187 + } + }, + { + "ph": "s", "id": 101187, "pid": 435, "tid": 435, "ts": 1776649793337458, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337470, "dur": 0, + "args": { + "External id": 101193, + "cbid": 200, "correlation": 101193 + } + }, + { + "ph": "f", "id": 101193, "pid": 435, "tid": 435, "ts": 1776649793337470, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337471, "dur": 0, + "args": { + "External id": 101194, + "cbid": 273, "correlation": 101194 + } + }, + { + "ph": "f", "id": 101194, "pid": 435, "tid": 435, "ts": 1776649793337471, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793337480, "dur": 19, + "args": { + "External id": 101195, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101195, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101195, "pid": 0, "tid": 7, "ts": 1776649793337480, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337472, "dur": 5, + "args": { + "External id": 101195, + "cbid": 211, "correlation": 101195 + } + }, + { + "ph": "s", "id": 101195, "pid": 435, "tid": 435, "ts": 1776649793337472, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793337516, "dur": 1, + "args": { + "External id": 13522, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101213, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101213, "pid": 0, "tid": 7, "ts": 1776649793337516, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337506, "dur": 6, + "args": { + "External id": 13522, + "cbid": 211, "correlation": 101213 + } + }, + { + "ph": "s", "id": 101213, "pid": 435, "tid": 435, "ts": 1776649793337506, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337519, "dur": 0, + "args": { + "External id": 101219, + "cbid": 200, "correlation": 101219 + } + }, + { + "ph": "f", "id": 101219, "pid": 435, "tid": 435, "ts": 1776649793337519, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337519, "dur": 0, + "args": { + "External id": 101220, + "cbid": 273, "correlation": 101220 + } + }, + { + "ph": "f", "id": 101220, "pid": 435, "tid": 435, "ts": 1776649793337519, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793337528, "dur": 19, + "args": { + "External id": 101221, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101221, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101221, "pid": 0, "tid": 7, "ts": 1776649793337528, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337520, "dur": 5, + "args": { + "External id": 101221, + "cbid": 211, "correlation": 101221 + } + }, + { + "ph": "s", "id": 101221, "pid": 435, "tid": 435, "ts": 1776649793337520, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793337564, "dur": 1, + "args": { + "External id": 13527, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101239, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101239, "pid": 0, "tid": 7, "ts": 1776649793337564, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337555, "dur": 6, + "args": { + "External id": 13527, + "cbid": 211, "correlation": 101239 + } + }, + { + "ph": "s", "id": 101239, "pid": 435, "tid": 435, "ts": 1776649793337555, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337567, "dur": 0, + "args": { + "External id": 101245, + "cbid": 200, "correlation": 101245 + } + }, + { + "ph": "f", "id": 101245, "pid": 435, "tid": 435, "ts": 1776649793337567, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337568, "dur": 0, + "args": { + "External id": 101246, + "cbid": 273, "correlation": 101246 + } + }, + { + "ph": "f", "id": 101246, "pid": 435, "tid": 435, "ts": 1776649793337568, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793337577, "dur": 19, + "args": { + "External id": 101247, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101247, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101247, "pid": 0, "tid": 7, "ts": 1776649793337577, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337569, "dur": 4, + "args": { + "External id": 101247, + "cbid": 211, "correlation": 101247 + } + }, + { + "ph": "s", "id": 101247, "pid": 435, "tid": 435, "ts": 1776649793337569, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793337612, "dur": 1, + "args": { + "External id": 13532, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101265, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101265, "pid": 0, "tid": 7, "ts": 1776649793337612, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337603, "dur": 6, + "args": { + "External id": 13532, + "cbid": 211, "correlation": 101265 + } + }, + { + "ph": "s", "id": 101265, "pid": 435, "tid": 435, "ts": 1776649793337603, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337615, "dur": 0, + "args": { + "External id": 101271, + "cbid": 200, "correlation": 101271 + } + }, + { + "ph": "f", "id": 101271, "pid": 435, "tid": 435, "ts": 1776649793337615, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337616, "dur": 0, + "args": { + "External id": 101272, + "cbid": 273, "correlation": 101272 + } + }, + { + "ph": "f", "id": 101272, "pid": 435, "tid": 435, "ts": 1776649793337616, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793337625, "dur": 19, + "args": { + "External id": 101273, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101273, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101273, "pid": 0, "tid": 7, "ts": 1776649793337625, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337617, "dur": 4, + "args": { + "External id": 101273, + "cbid": 211, "correlation": 101273 + } + }, + { + "ph": "s", "id": 101273, "pid": 435, "tid": 435, "ts": 1776649793337617, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793337665, "dur": 1, + "args": { + "External id": 13537, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101291, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101291, "pid": 0, "tid": 7, "ts": 1776649793337665, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337654, "dur": 7, + "args": { + "External id": 13537, + "cbid": 211, "correlation": 101291 + } + }, + { + "ph": "s", "id": 101291, "pid": 435, "tid": 435, "ts": 1776649793337654, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337668, "dur": 0, + "args": { + "External id": 101297, + "cbid": 200, "correlation": 101297 + } + }, + { + "ph": "f", "id": 101297, "pid": 435, "tid": 435, "ts": 1776649793337668, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337668, "dur": 0, + "args": { + "External id": 101298, + "cbid": 273, "correlation": 101298 + } + }, + { + "ph": "f", "id": 101298, "pid": 435, "tid": 435, "ts": 1776649793337668, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793337677, "dur": 19, + "args": { + "External id": 101299, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101299, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101299, "pid": 0, "tid": 7, "ts": 1776649793337677, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337669, "dur": 5, + "args": { + "External id": 101299, + "cbid": 211, "correlation": 101299 + } + }, + { + "ph": "s", "id": 101299, "pid": 435, "tid": 435, "ts": 1776649793337669, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793337712, "dur": 1, + "args": { + "External id": 13542, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101317, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101317, "pid": 0, "tid": 7, "ts": 1776649793337712, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337703, "dur": 6, + "args": { + "External id": 13542, + "cbid": 211, "correlation": 101317 + } + }, + { + "ph": "s", "id": 101317, "pid": 435, "tid": 435, "ts": 1776649793337703, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337715, "dur": 0, + "args": { + "External id": 101323, + "cbid": 200, "correlation": 101323 + } + }, + { + "ph": "f", "id": 101323, "pid": 435, "tid": 435, "ts": 1776649793337715, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337716, "dur": 0, + "args": { + "External id": 101324, + "cbid": 273, "correlation": 101324 + } + }, + { + "ph": "f", "id": 101324, "pid": 435, "tid": 435, "ts": 1776649793337716, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793337725, "dur": 19, + "args": { + "External id": 101325, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101325, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101325, "pid": 0, "tid": 7, "ts": 1776649793337725, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337717, "dur": 4, + "args": { + "External id": 101325, + "cbid": 211, "correlation": 101325 + } + }, + { + "ph": "s", "id": 101325, "pid": 435, "tid": 435, "ts": 1776649793337717, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793337760, "dur": 1, + "args": { + "External id": 13547, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101343, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101343, "pid": 0, "tid": 7, "ts": 1776649793337760, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337751, "dur": 6, + "args": { + "External id": 13547, + "cbid": 211, "correlation": 101343 + } + }, + { + "ph": "s", "id": 101343, "pid": 435, "tid": 435, "ts": 1776649793337751, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337763, "dur": 0, + "args": { + "External id": 101349, + "cbid": 200, "correlation": 101349 + } + }, + { + "ph": "f", "id": 101349, "pid": 435, "tid": 435, "ts": 1776649793337763, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337764, "dur": 0, + "args": { + "External id": 101350, + "cbid": 273, "correlation": 101350 + } + }, + { + "ph": "f", "id": 101350, "pid": 435, "tid": 435, "ts": 1776649793337764, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793337773, "dur": 19, + "args": { + "External id": 101351, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101351, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101351, "pid": 0, "tid": 7, "ts": 1776649793337773, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337765, "dur": 5, + "args": { + "External id": 101351, + "cbid": 211, "correlation": 101351 + } + }, + { + "ph": "s", "id": 101351, "pid": 435, "tid": 435, "ts": 1776649793337765, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793337808, "dur": 1, + "args": { + "External id": 13552, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101369, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101369, "pid": 0, "tid": 7, "ts": 1776649793337808, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337799, "dur": 6, + "args": { + "External id": 13552, + "cbid": 211, "correlation": 101369 + } + }, + { + "ph": "s", "id": 101369, "pid": 435, "tid": 435, "ts": 1776649793337799, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337811, "dur": 0, + "args": { + "External id": 101375, + "cbid": 200, "correlation": 101375 + } + }, + { + "ph": "f", "id": 101375, "pid": 435, "tid": 435, "ts": 1776649793337811, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337812, "dur": 0, + "args": { + "External id": 101376, + "cbid": 273, "correlation": 101376 + } + }, + { + "ph": "f", "id": 101376, "pid": 435, "tid": 435, "ts": 1776649793337812, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793337820, "dur": 19, + "args": { + "External id": 101377, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101377, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101377, "pid": 0, "tid": 7, "ts": 1776649793337820, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337813, "dur": 5, + "args": { + "External id": 101377, + "cbid": 211, "correlation": 101377 + } + }, + { + "ph": "s", "id": 101377, "pid": 435, "tid": 435, "ts": 1776649793337813, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793337856, "dur": 1, + "args": { + "External id": 13557, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101395, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101395, "pid": 0, "tid": 7, "ts": 1776649793337856, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337847, "dur": 6, + "args": { + "External id": 13557, + "cbid": 211, "correlation": 101395 + } + }, + { + "ph": "s", "id": 101395, "pid": 435, "tid": 435, "ts": 1776649793337847, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337859, "dur": 0, + "args": { + "External id": 101401, + "cbid": 200, "correlation": 101401 + } + }, + { + "ph": "f", "id": 101401, "pid": 435, "tid": 435, "ts": 1776649793337859, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337860, "dur": 0, + "args": { + "External id": 101402, + "cbid": 273, "correlation": 101402 + } + }, + { + "ph": "f", "id": 101402, "pid": 435, "tid": 435, "ts": 1776649793337860, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793337869, "dur": 19, + "args": { + "External id": 101403, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101403, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101403, "pid": 0, "tid": 7, "ts": 1776649793337869, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337861, "dur": 5, + "args": { + "External id": 101403, + "cbid": 211, "correlation": 101403 + } + }, + { + "ph": "s", "id": 101403, "pid": 435, "tid": 435, "ts": 1776649793337861, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793337905, "dur": 1, + "args": { + "External id": 13562, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101421, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101421, "pid": 0, "tid": 7, "ts": 1776649793337905, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337895, "dur": 6, + "args": { + "External id": 13562, + "cbid": 211, "correlation": 101421 + } + }, + { + "ph": "s", "id": 101421, "pid": 435, "tid": 435, "ts": 1776649793337895, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337908, "dur": 0, + "args": { + "External id": 101427, + "cbid": 200, "correlation": 101427 + } + }, + { + "ph": "f", "id": 101427, "pid": 435, "tid": 435, "ts": 1776649793337908, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337909, "dur": 0, + "args": { + "External id": 101428, + "cbid": 273, "correlation": 101428 + } + }, + { + "ph": "f", "id": 101428, "pid": 435, "tid": 435, "ts": 1776649793337909, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793337918, "dur": 19, + "args": { + "External id": 101429, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101429, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101429, "pid": 0, "tid": 7, "ts": 1776649793337918, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337910, "dur": 5, + "args": { + "External id": 101429, + "cbid": 211, "correlation": 101429 + } + }, + { + "ph": "s", "id": 101429, "pid": 435, "tid": 435, "ts": 1776649793337910, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793337953, "dur": 1, + "args": { + "External id": 13567, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101447, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101447, "pid": 0, "tid": 7, "ts": 1776649793337953, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337944, "dur": 6, + "args": { + "External id": 13567, + "cbid": 211, "correlation": 101447 + } + }, + { + "ph": "s", "id": 101447, "pid": 435, "tid": 435, "ts": 1776649793337944, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337956, "dur": 0, + "args": { + "External id": 101453, + "cbid": 200, "correlation": 101453 + } + }, + { + "ph": "f", "id": 101453, "pid": 435, "tid": 435, "ts": 1776649793337956, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793337957, "dur": 0, + "args": { + "External id": 101454, + "cbid": 273, "correlation": 101454 + } + }, + { + "ph": "f", "id": 101454, "pid": 435, "tid": 435, "ts": 1776649793337957, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793337966, "dur": 19, + "args": { + "External id": 101455, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101455, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101455, "pid": 0, "tid": 7, "ts": 1776649793337966, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337958, "dur": 4, + "args": { + "External id": 101455, + "cbid": 211, "correlation": 101455 + } + }, + { + "ph": "s", "id": 101455, "pid": 435, "tid": 435, "ts": 1776649793337958, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793338001, "dur": 1, + "args": { + "External id": 13572, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101473, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101473, "pid": 0, "tid": 7, "ts": 1776649793338001, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793337992, "dur": 6, + "args": { + "External id": 13572, + "cbid": 211, "correlation": 101473 + } + }, + { + "ph": "s", "id": 101473, "pid": 435, "tid": 435, "ts": 1776649793337992, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338004, "dur": 0, + "args": { + "External id": 101479, + "cbid": 200, "correlation": 101479 + } + }, + { + "ph": "f", "id": 101479, "pid": 435, "tid": 435, "ts": 1776649793338004, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338005, "dur": 0, + "args": { + "External id": 101480, + "cbid": 273, "correlation": 101480 + } + }, + { + "ph": "f", "id": 101480, "pid": 435, "tid": 435, "ts": 1776649793338005, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793338014, "dur": 19, + "args": { + "External id": 101481, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101481, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101481, "pid": 0, "tid": 7, "ts": 1776649793338014, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338006, "dur": 4, + "args": { + "External id": 101481, + "cbid": 211, "correlation": 101481 + } + }, + { + "ph": "s", "id": 101481, "pid": 435, "tid": 435, "ts": 1776649793338006, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793338049, "dur": 1, + "args": { + "External id": 13577, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101499, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101499, "pid": 0, "tid": 7, "ts": 1776649793338049, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338040, "dur": 6, + "args": { + "External id": 13577, + "cbid": 211, "correlation": 101499 + } + }, + { + "ph": "s", "id": 101499, "pid": 435, "tid": 435, "ts": 1776649793338040, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338052, "dur": 0, + "args": { + "External id": 101505, + "cbid": 200, "correlation": 101505 + } + }, + { + "ph": "f", "id": 101505, "pid": 435, "tid": 435, "ts": 1776649793338052, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338053, "dur": 0, + "args": { + "External id": 101506, + "cbid": 273, "correlation": 101506 + } + }, + { + "ph": "f", "id": 101506, "pid": 435, "tid": 435, "ts": 1776649793338053, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793338062, "dur": 19, + "args": { + "External id": 101507, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101507, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101507, "pid": 0, "tid": 7, "ts": 1776649793338062, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338054, "dur": 4, + "args": { + "External id": 101507, + "cbid": 211, "correlation": 101507 + } + }, + { + "ph": "s", "id": 101507, "pid": 435, "tid": 435, "ts": 1776649793338054, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793338097, "dur": 1, + "args": { + "External id": 13582, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101525, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101525, "pid": 0, "tid": 7, "ts": 1776649793338097, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338088, "dur": 6, + "args": { + "External id": 13582, + "cbid": 211, "correlation": 101525 + } + }, + { + "ph": "s", "id": 101525, "pid": 435, "tid": 435, "ts": 1776649793338088, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338100, "dur": 0, + "args": { + "External id": 101531, + "cbid": 200, "correlation": 101531 + } + }, + { + "ph": "f", "id": 101531, "pid": 435, "tid": 435, "ts": 1776649793338100, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338100, "dur": 0, + "args": { + "External id": 101532, + "cbid": 273, "correlation": 101532 + } + }, + { + "ph": "f", "id": 101532, "pid": 435, "tid": 435, "ts": 1776649793338100, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793338109, "dur": 19, + "args": { + "External id": 101533, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101533, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101533, "pid": 0, "tid": 7, "ts": 1776649793338109, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338102, "dur": 4, + "args": { + "External id": 101533, + "cbid": 211, "correlation": 101533 + } + }, + { + "ph": "s", "id": 101533, "pid": 435, "tid": 435, "ts": 1776649793338102, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793338145, "dur": 1, + "args": { + "External id": 13587, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101551, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101551, "pid": 0, "tid": 7, "ts": 1776649793338145, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338136, "dur": 6, + "args": { + "External id": 13587, + "cbid": 211, "correlation": 101551 + } + }, + { + "ph": "s", "id": 101551, "pid": 435, "tid": 435, "ts": 1776649793338136, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338148, "dur": 0, + "args": { + "External id": 101557, + "cbid": 200, "correlation": 101557 + } + }, + { + "ph": "f", "id": 101557, "pid": 435, "tid": 435, "ts": 1776649793338148, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338149, "dur": 0, + "args": { + "External id": 101558, + "cbid": 273, "correlation": 101558 + } + }, + { + "ph": "f", "id": 101558, "pid": 435, "tid": 435, "ts": 1776649793338149, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793338158, "dur": 19, + "args": { + "External id": 101559, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101559, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101559, "pid": 0, "tid": 7, "ts": 1776649793338158, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338150, "dur": 5, + "args": { + "External id": 101559, + "cbid": 211, "correlation": 101559 + } + }, + { + "ph": "s", "id": 101559, "pid": 435, "tid": 435, "ts": 1776649793338150, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793338193, "dur": 1, + "args": { + "External id": 13592, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101577, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101577, "pid": 0, "tid": 7, "ts": 1776649793338193, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338184, "dur": 6, + "args": { + "External id": 13592, + "cbid": 211, "correlation": 101577 + } + }, + { + "ph": "s", "id": 101577, "pid": 435, "tid": 435, "ts": 1776649793338184, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338196, "dur": 0, + "args": { + "External id": 101583, + "cbid": 200, "correlation": 101583 + } + }, + { + "ph": "f", "id": 101583, "pid": 435, "tid": 435, "ts": 1776649793338196, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338197, "dur": 0, + "args": { + "External id": 101584, + "cbid": 273, "correlation": 101584 + } + }, + { + "ph": "f", "id": 101584, "pid": 435, "tid": 435, "ts": 1776649793338197, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793338205, "dur": 19, + "args": { + "External id": 101585, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101585, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101585, "pid": 0, "tid": 7, "ts": 1776649793338205, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338198, "dur": 4, + "args": { + "External id": 101585, + "cbid": 211, "correlation": 101585 + } + }, + { + "ph": "s", "id": 101585, "pid": 435, "tid": 435, "ts": 1776649793338198, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793338241, "dur": 1, + "args": { + "External id": 13597, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101603, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101603, "pid": 0, "tid": 7, "ts": 1776649793338241, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338232, "dur": 6, + "args": { + "External id": 13597, + "cbid": 211, "correlation": 101603 + } + }, + { + "ph": "s", "id": 101603, "pid": 435, "tid": 435, "ts": 1776649793338232, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338244, "dur": 0, + "args": { + "External id": 101609, + "cbid": 200, "correlation": 101609 + } + }, + { + "ph": "f", "id": 101609, "pid": 435, "tid": 435, "ts": 1776649793338244, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338245, "dur": 0, + "args": { + "External id": 101610, + "cbid": 273, "correlation": 101610 + } + }, + { + "ph": "f", "id": 101610, "pid": 435, "tid": 435, "ts": 1776649793338245, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793338254, "dur": 19, + "args": { + "External id": 101611, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101611, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101611, "pid": 0, "tid": 7, "ts": 1776649793338254, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338246, "dur": 5, + "args": { + "External id": 101611, + "cbid": 211, "correlation": 101611 + } + }, + { + "ph": "s", "id": 101611, "pid": 435, "tid": 435, "ts": 1776649793338246, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793338289, "dur": 1, + "args": { + "External id": 13602, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101629, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101629, "pid": 0, "tid": 7, "ts": 1776649793338289, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338280, "dur": 6, + "args": { + "External id": 13602, + "cbid": 211, "correlation": 101629 + } + }, + { + "ph": "s", "id": 101629, "pid": 435, "tid": 435, "ts": 1776649793338280, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338292, "dur": 0, + "args": { + "External id": 101635, + "cbid": 200, "correlation": 101635 + } + }, + { + "ph": "f", "id": 101635, "pid": 435, "tid": 435, "ts": 1776649793338292, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338293, "dur": 0, + "args": { + "External id": 101636, + "cbid": 273, "correlation": 101636 + } + }, + { + "ph": "f", "id": 101636, "pid": 435, "tid": 435, "ts": 1776649793338293, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793338302, "dur": 19, + "args": { + "External id": 101637, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101637, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101637, "pid": 0, "tid": 7, "ts": 1776649793338302, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338294, "dur": 4, + "args": { + "External id": 101637, + "cbid": 211, "correlation": 101637 + } + }, + { + "ph": "s", "id": 101637, "pid": 435, "tid": 435, "ts": 1776649793338294, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793338337, "dur": 1, + "args": { + "External id": 13607, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101655, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101655, "pid": 0, "tid": 7, "ts": 1776649793338337, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338328, "dur": 6, + "args": { + "External id": 13607, + "cbid": 211, "correlation": 101655 + } + }, + { + "ph": "s", "id": 101655, "pid": 435, "tid": 435, "ts": 1776649793338328, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338340, "dur": 0, + "args": { + "External id": 101661, + "cbid": 200, "correlation": 101661 + } + }, + { + "ph": "f", "id": 101661, "pid": 435, "tid": 435, "ts": 1776649793338340, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338341, "dur": 0, + "args": { + "External id": 101662, + "cbid": 273, "correlation": 101662 + } + }, + { + "ph": "f", "id": 101662, "pid": 435, "tid": 435, "ts": 1776649793338341, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793338350, "dur": 19, + "args": { + "External id": 101663, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101663, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101663, "pid": 0, "tid": 7, "ts": 1776649793338350, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338342, "dur": 5, + "args": { + "External id": 101663, + "cbid": 211, "correlation": 101663 + } + }, + { + "ph": "s", "id": 101663, "pid": 435, "tid": 435, "ts": 1776649793338342, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793338386, "dur": 1, + "args": { + "External id": 13612, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101681, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101681, "pid": 0, "tid": 7, "ts": 1776649793338386, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338376, "dur": 6, + "args": { + "External id": 13612, + "cbid": 211, "correlation": 101681 + } + }, + { + "ph": "s", "id": 101681, "pid": 435, "tid": 435, "ts": 1776649793338376, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338388, "dur": 0, + "args": { + "External id": 101687, + "cbid": 200, "correlation": 101687 + } + }, + { + "ph": "f", "id": 101687, "pid": 435, "tid": 435, "ts": 1776649793338388, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338389, "dur": 0, + "args": { + "External id": 101688, + "cbid": 273, "correlation": 101688 + } + }, + { + "ph": "f", "id": 101688, "pid": 435, "tid": 435, "ts": 1776649793338389, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793338398, "dur": 19, + "args": { + "External id": 101689, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101689, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101689, "pid": 0, "tid": 7, "ts": 1776649793338398, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338390, "dur": 4, + "args": { + "External id": 101689, + "cbid": 211, "correlation": 101689 + } + }, + { + "ph": "s", "id": 101689, "pid": 435, "tid": 435, "ts": 1776649793338390, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793338434, "dur": 1, + "args": { + "External id": 13617, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101707, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101707, "pid": 0, "tid": 7, "ts": 1776649793338434, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338424, "dur": 7, + "args": { + "External id": 13617, + "cbid": 211, "correlation": 101707 + } + }, + { + "ph": "s", "id": 101707, "pid": 435, "tid": 435, "ts": 1776649793338424, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338437, "dur": 0, + "args": { + "External id": 101713, + "cbid": 200, "correlation": 101713 + } + }, + { + "ph": "f", "id": 101713, "pid": 435, "tid": 435, "ts": 1776649793338437, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338437, "dur": 0, + "args": { + "External id": 101714, + "cbid": 273, "correlation": 101714 + } + }, + { + "ph": "f", "id": 101714, "pid": 435, "tid": 435, "ts": 1776649793338437, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793338447, "dur": 19, + "args": { + "External id": 101715, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101715, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101715, "pid": 0, "tid": 7, "ts": 1776649793338447, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338439, "dur": 4, + "args": { + "External id": 101715, + "cbid": 211, "correlation": 101715 + } + }, + { + "ph": "s", "id": 101715, "pid": 435, "tid": 435, "ts": 1776649793338439, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793338482, "dur": 1, + "args": { + "External id": 13622, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101733, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101733, "pid": 0, "tid": 7, "ts": 1776649793338482, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338473, "dur": 6, + "args": { + "External id": 13622, + "cbid": 211, "correlation": 101733 + } + }, + { + "ph": "s", "id": 101733, "pid": 435, "tid": 435, "ts": 1776649793338473, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338485, "dur": 0, + "args": { + "External id": 101739, + "cbid": 200, "correlation": 101739 + } + }, + { + "ph": "f", "id": 101739, "pid": 435, "tid": 435, "ts": 1776649793338485, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338486, "dur": 0, + "args": { + "External id": 101740, + "cbid": 273, "correlation": 101740 + } + }, + { + "ph": "f", "id": 101740, "pid": 435, "tid": 435, "ts": 1776649793338486, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793338495, "dur": 19, + "args": { + "External id": 101741, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101741, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101741, "pid": 0, "tid": 7, "ts": 1776649793338495, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338487, "dur": 4, + "args": { + "External id": 101741, + "cbid": 211, "correlation": 101741 + } + }, + { + "ph": "s", "id": 101741, "pid": 435, "tid": 435, "ts": 1776649793338487, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793338530, "dur": 1, + "args": { + "External id": 13627, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101759, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101759, "pid": 0, "tid": 7, "ts": 1776649793338530, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338521, "dur": 6, + "args": { + "External id": 13627, + "cbid": 211, "correlation": 101759 + } + }, + { + "ph": "s", "id": 101759, "pid": 435, "tid": 435, "ts": 1776649793338521, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338533, "dur": 0, + "args": { + "External id": 101765, + "cbid": 200, "correlation": 101765 + } + }, + { + "ph": "f", "id": 101765, "pid": 435, "tid": 435, "ts": 1776649793338533, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338534, "dur": 0, + "args": { + "External id": 101766, + "cbid": 273, "correlation": 101766 + } + }, + { + "ph": "f", "id": 101766, "pid": 435, "tid": 435, "ts": 1776649793338534, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793338543, "dur": 19, + "args": { + "External id": 101767, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101767, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101767, "pid": 0, "tid": 7, "ts": 1776649793338543, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338535, "dur": 5, + "args": { + "External id": 101767, + "cbid": 211, "correlation": 101767 + } + }, + { + "ph": "s", "id": 101767, "pid": 435, "tid": 435, "ts": 1776649793338535, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793338578, "dur": 1, + "args": { + "External id": 13632, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101785, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101785, "pid": 0, "tid": 7, "ts": 1776649793338578, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338569, "dur": 6, + "args": { + "External id": 13632, + "cbid": 211, "correlation": 101785 + } + }, + { + "ph": "s", "id": 101785, "pid": 435, "tid": 435, "ts": 1776649793338569, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338581, "dur": 0, + "args": { + "External id": 101791, + "cbid": 200, "correlation": 101791 + } + }, + { + "ph": "f", "id": 101791, "pid": 435, "tid": 435, "ts": 1776649793338581, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338582, "dur": 0, + "args": { + "External id": 101792, + "cbid": 273, "correlation": 101792 + } + }, + { + "ph": "f", "id": 101792, "pid": 435, "tid": 435, "ts": 1776649793338582, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793338591, "dur": 19, + "args": { + "External id": 101793, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101793, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101793, "pid": 0, "tid": 7, "ts": 1776649793338591, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338583, "dur": 4, + "args": { + "External id": 101793, + "cbid": 211, "correlation": 101793 + } + }, + { + "ph": "s", "id": 101793, "pid": 435, "tid": 435, "ts": 1776649793338583, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793338627, "dur": 1, + "args": { + "External id": 13637, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101811, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101811, "pid": 0, "tid": 7, "ts": 1776649793338627, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338617, "dur": 6, + "args": { + "External id": 13637, + "cbid": 211, "correlation": 101811 + } + }, + { + "ph": "s", "id": 101811, "pid": 435, "tid": 435, "ts": 1776649793338617, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338630, "dur": 0, + "args": { + "External id": 101817, + "cbid": 200, "correlation": 101817 + } + }, + { + "ph": "f", "id": 101817, "pid": 435, "tid": 435, "ts": 1776649793338630, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338630, "dur": 0, + "args": { + "External id": 101818, + "cbid": 273, "correlation": 101818 + } + }, + { + "ph": "f", "id": 101818, "pid": 435, "tid": 435, "ts": 1776649793338630, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793338639, "dur": 19, + "args": { + "External id": 101819, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101819, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101819, "pid": 0, "tid": 7, "ts": 1776649793338639, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338632, "dur": 5, + "args": { + "External id": 101819, + "cbid": 211, "correlation": 101819 + } + }, + { + "ph": "s", "id": 101819, "pid": 435, "tid": 435, "ts": 1776649793338632, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793338674, "dur": 1, + "args": { + "External id": 13642, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101837, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101837, "pid": 0, "tid": 7, "ts": 1776649793338674, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338665, "dur": 6, + "args": { + "External id": 13642, + "cbid": 211, "correlation": 101837 + } + }, + { + "ph": "s", "id": 101837, "pid": 435, "tid": 435, "ts": 1776649793338665, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338678, "dur": 0, + "args": { + "External id": 101843, + "cbid": 200, "correlation": 101843 + } + }, + { + "ph": "f", "id": 101843, "pid": 435, "tid": 435, "ts": 1776649793338678, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338678, "dur": 0, + "args": { + "External id": 101844, + "cbid": 273, "correlation": 101844 + } + }, + { + "ph": "f", "id": 101844, "pid": 435, "tid": 435, "ts": 1776649793338678, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793338687, "dur": 19, + "args": { + "External id": 101845, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101845, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101845, "pid": 0, "tid": 7, "ts": 1776649793338687, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338679, "dur": 4, + "args": { + "External id": 101845, + "cbid": 211, "correlation": 101845 + } + }, + { + "ph": "s", "id": 101845, "pid": 435, "tid": 435, "ts": 1776649793338679, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793338724, "dur": 1, + "args": { + "External id": 13647, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101863, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101863, "pid": 0, "tid": 7, "ts": 1776649793338724, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338713, "dur": 7, + "args": { + "External id": 13647, + "cbid": 211, "correlation": 101863 + } + }, + { + "ph": "s", "id": 101863, "pid": 435, "tid": 435, "ts": 1776649793338713, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338726, "dur": 0, + "args": { + "External id": 101869, + "cbid": 200, "correlation": 101869 + } + }, + { + "ph": "f", "id": 101869, "pid": 435, "tid": 435, "ts": 1776649793338726, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338727, "dur": 0, + "args": { + "External id": 101870, + "cbid": 273, "correlation": 101870 + } + }, + { + "ph": "f", "id": 101870, "pid": 435, "tid": 435, "ts": 1776649793338727, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793338736, "dur": 19, + "args": { + "External id": 101871, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101871, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101871, "pid": 0, "tid": 7, "ts": 1776649793338736, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338728, "dur": 5, + "args": { + "External id": 101871, + "cbid": 211, "correlation": 101871 + } + }, + { + "ph": "s", "id": 101871, "pid": 435, "tid": 435, "ts": 1776649793338728, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793338771, "dur": 1, + "args": { + "External id": 13652, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101889, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101889, "pid": 0, "tid": 7, "ts": 1776649793338771, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338762, "dur": 6, + "args": { + "External id": 13652, + "cbid": 211, "correlation": 101889 + } + }, + { + "ph": "s", "id": 101889, "pid": 435, "tid": 435, "ts": 1776649793338762, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338774, "dur": 0, + "args": { + "External id": 101895, + "cbid": 200, "correlation": 101895 + } + }, + { + "ph": "f", "id": 101895, "pid": 435, "tid": 435, "ts": 1776649793338774, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338775, "dur": 0, + "args": { + "External id": 101896, + "cbid": 273, "correlation": 101896 + } + }, + { + "ph": "f", "id": 101896, "pid": 435, "tid": 435, "ts": 1776649793338775, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793338784, "dur": 19, + "args": { + "External id": 101897, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101897, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101897, "pid": 0, "tid": 7, "ts": 1776649793338784, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338776, "dur": 4, + "args": { + "External id": 101897, + "cbid": 211, "correlation": 101897 + } + }, + { + "ph": "s", "id": 101897, "pid": 435, "tid": 435, "ts": 1776649793338776, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793338823, "dur": 1, + "args": { + "External id": 13657, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101915, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101915, "pid": 0, "tid": 7, "ts": 1776649793338823, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338813, "dur": 6, + "args": { + "External id": 13657, + "cbid": 211, "correlation": 101915 + } + }, + { + "ph": "s", "id": 101915, "pid": 435, "tid": 435, "ts": 1776649793338813, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338825, "dur": 0, + "args": { + "External id": 101921, + "cbid": 200, "correlation": 101921 + } + }, + { + "ph": "f", "id": 101921, "pid": 435, "tid": 435, "ts": 1776649793338825, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338826, "dur": 0, + "args": { + "External id": 101922, + "cbid": 273, "correlation": 101922 + } + }, + { + "ph": "f", "id": 101922, "pid": 435, "tid": 435, "ts": 1776649793338826, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793338835, "dur": 19, + "args": { + "External id": 101923, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101923, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101923, "pid": 0, "tid": 7, "ts": 1776649793338835, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338827, "dur": 5, + "args": { + "External id": 101923, + "cbid": 211, "correlation": 101923 + } + }, + { + "ph": "s", "id": 101923, "pid": 435, "tid": 435, "ts": 1776649793338827, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649793338856, "dur": 3, + "args": { + "External id": 101931, + "cbid": 135, "correlation": 101931 + } + }, + { + "ph": "f", "id": 101931, "pid": 435, "tid": 435, "ts": 1776649793338856, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649793338871, "dur": 5, + "args": { + "External id": 101936, + "cbid": 165, "correlation": 101936 + } + }, + { + "ph": "s", "id": 101936, "pid": 435, "tid": 435, "ts": 1776649793338871, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649793338879, "dur": 4, + "args": { + "External id": 101939, + "cbid": 139, "correlation": 101939 + } + }, + { + "ph": "f", "id": 101939, "pid": 435, "tid": 435, "ts": 1776649793338879, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793338935, "dur": 1, + "args": { + "External id": 13662, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101967, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101967, "pid": 0, "tid": 7, "ts": 1776649793338935, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338924, "dur": 8, + "args": { + "External id": 13662, + "cbid": 211, "correlation": 101967 + } + }, + { + "ph": "s", "id": 101967, "pid": 435, "tid": 435, "ts": 1776649793338924, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338939, "dur": 0, + "args": { + "External id": 101973, + "cbid": 200, "correlation": 101973 + } + }, + { + "ph": "f", "id": 101973, "pid": 435, "tid": 435, "ts": 1776649793338939, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338940, "dur": 0, + "args": { + "External id": 101974, + "cbid": 273, "correlation": 101974 + } + }, + { + "ph": "f", "id": 101974, "pid": 435, "tid": 435, "ts": 1776649793338940, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793338949, "dur": 19, + "args": { + "External id": 101975, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101975, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101975, "pid": 0, "tid": 7, "ts": 1776649793338949, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338941, "dur": 5, + "args": { + "External id": 101975, + "cbid": 211, "correlation": 101975 + } + }, + { + "ph": "s", "id": 101975, "pid": 435, "tid": 435, "ts": 1776649793338941, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793338987, "dur": 1, + "args": { + "External id": 13667, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 101993, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 101993, "pid": 0, "tid": 7, "ts": 1776649793338987, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338977, "dur": 7, + "args": { + "External id": 13667, + "cbid": 211, "correlation": 101993 + } + }, + { + "ph": "s", "id": 101993, "pid": 435, "tid": 435, "ts": 1776649793338977, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338990, "dur": 0, + "args": { + "External id": 101999, + "cbid": 200, "correlation": 101999 + } + }, + { + "ph": "f", "id": 101999, "pid": 435, "tid": 435, "ts": 1776649793338990, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793338991, "dur": 0, + "args": { + "External id": 102000, + "cbid": 273, "correlation": 102000 + } + }, + { + "ph": "f", "id": 102000, "pid": 435, "tid": 435, "ts": 1776649793338991, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793339000, "dur": 19, + "args": { + "External id": 102001, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102001, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102001, "pid": 0, "tid": 7, "ts": 1776649793339000, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793338992, "dur": 4, + "args": { + "External id": 102001, + "cbid": 211, "correlation": 102001 + } + }, + { + "ph": "s", "id": 102001, "pid": 435, "tid": 435, "ts": 1776649793338992, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793339036, "dur": 1, + "args": { + "External id": 13672, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102019, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102019, "pid": 0, "tid": 7, "ts": 1776649793339036, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339026, "dur": 7, + "args": { + "External id": 13672, + "cbid": 211, "correlation": 102019 + } + }, + { + "ph": "s", "id": 102019, "pid": 435, "tid": 435, "ts": 1776649793339026, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339039, "dur": 0, + "args": { + "External id": 102025, + "cbid": 200, "correlation": 102025 + } + }, + { + "ph": "f", "id": 102025, "pid": 435, "tid": 435, "ts": 1776649793339039, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339039, "dur": 0, + "args": { + "External id": 102026, + "cbid": 273, "correlation": 102026 + } + }, + { + "ph": "f", "id": 102026, "pid": 435, "tid": 435, "ts": 1776649793339039, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793339048, "dur": 19, + "args": { + "External id": 102027, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102027, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102027, "pid": 0, "tid": 7, "ts": 1776649793339048, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339041, "dur": 4, + "args": { + "External id": 102027, + "cbid": 211, "correlation": 102027 + } + }, + { + "ph": "s", "id": 102027, "pid": 435, "tid": 435, "ts": 1776649793339041, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793339083, "dur": 1, + "args": { + "External id": 13677, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102045, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102045, "pid": 0, "tid": 7, "ts": 1776649793339083, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339074, "dur": 6, + "args": { + "External id": 13677, + "cbid": 211, "correlation": 102045 + } + }, + { + "ph": "s", "id": 102045, "pid": 435, "tid": 435, "ts": 1776649793339074, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339087, "dur": 0, + "args": { + "External id": 102051, + "cbid": 200, "correlation": 102051 + } + }, + { + "ph": "f", "id": 102051, "pid": 435, "tid": 435, "ts": 1776649793339087, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339087, "dur": 0, + "args": { + "External id": 102052, + "cbid": 273, "correlation": 102052 + } + }, + { + "ph": "f", "id": 102052, "pid": 435, "tid": 435, "ts": 1776649793339087, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793339096, "dur": 19, + "args": { + "External id": 102053, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102053, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102053, "pid": 0, "tid": 7, "ts": 1776649793339096, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339088, "dur": 4, + "args": { + "External id": 102053, + "cbid": 211, "correlation": 102053 + } + }, + { + "ph": "s", "id": 102053, "pid": 435, "tid": 435, "ts": 1776649793339088, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793339132, "dur": 1, + "args": { + "External id": 13682, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102071, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102071, "pid": 0, "tid": 7, "ts": 1776649793339132, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339122, "dur": 7, + "args": { + "External id": 13682, + "cbid": 211, "correlation": 102071 + } + }, + { + "ph": "s", "id": 102071, "pid": 435, "tid": 435, "ts": 1776649793339122, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339135, "dur": 0, + "args": { + "External id": 102077, + "cbid": 200, "correlation": 102077 + } + }, + { + "ph": "f", "id": 102077, "pid": 435, "tid": 435, "ts": 1776649793339135, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339136, "dur": 0, + "args": { + "External id": 102078, + "cbid": 273, "correlation": 102078 + } + }, + { + "ph": "f", "id": 102078, "pid": 435, "tid": 435, "ts": 1776649793339136, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793339145, "dur": 19, + "args": { + "External id": 102079, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102079, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102079, "pid": 0, "tid": 7, "ts": 1776649793339145, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339137, "dur": 4, + "args": { + "External id": 102079, + "cbid": 211, "correlation": 102079 + } + }, + { + "ph": "s", "id": 102079, "pid": 435, "tid": 435, "ts": 1776649793339137, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793339180, "dur": 1, + "args": { + "External id": 13687, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102097, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102097, "pid": 0, "tid": 7, "ts": 1776649793339180, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339171, "dur": 6, + "args": { + "External id": 13687, + "cbid": 211, "correlation": 102097 + } + }, + { + "ph": "s", "id": 102097, "pid": 435, "tid": 435, "ts": 1776649793339171, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339183, "dur": 0, + "args": { + "External id": 102103, + "cbid": 200, "correlation": 102103 + } + }, + { + "ph": "f", "id": 102103, "pid": 435, "tid": 435, "ts": 1776649793339183, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339184, "dur": 0, + "args": { + "External id": 102104, + "cbid": 273, "correlation": 102104 + } + }, + { + "ph": "f", "id": 102104, "pid": 435, "tid": 435, "ts": 1776649793339184, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793339192, "dur": 19, + "args": { + "External id": 102105, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102105, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102105, "pid": 0, "tid": 7, "ts": 1776649793339192, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339185, "dur": 4, + "args": { + "External id": 102105, + "cbid": 211, "correlation": 102105 + } + }, + { + "ph": "s", "id": 102105, "pid": 435, "tid": 435, "ts": 1776649793339185, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793339228, "dur": 1, + "args": { + "External id": 13692, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102123, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102123, "pid": 0, "tid": 7, "ts": 1776649793339228, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339219, "dur": 6, + "args": { + "External id": 13692, + "cbid": 211, "correlation": 102123 + } + }, + { + "ph": "s", "id": 102123, "pid": 435, "tid": 435, "ts": 1776649793339219, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339231, "dur": 0, + "args": { + "External id": 102129, + "cbid": 200, "correlation": 102129 + } + }, + { + "ph": "f", "id": 102129, "pid": 435, "tid": 435, "ts": 1776649793339231, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339232, "dur": 0, + "args": { + "External id": 102130, + "cbid": 273, "correlation": 102130 + } + }, + { + "ph": "f", "id": 102130, "pid": 435, "tid": 435, "ts": 1776649793339232, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793339241, "dur": 19, + "args": { + "External id": 102131, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102131, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102131, "pid": 0, "tid": 7, "ts": 1776649793339241, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339233, "dur": 4, + "args": { + "External id": 102131, + "cbid": 211, "correlation": 102131 + } + }, + { + "ph": "s", "id": 102131, "pid": 435, "tid": 435, "ts": 1776649793339233, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793339276, "dur": 1, + "args": { + "External id": 13697, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102149, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102149, "pid": 0, "tid": 7, "ts": 1776649793339276, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339267, "dur": 6, + "args": { + "External id": 13697, + "cbid": 211, "correlation": 102149 + } + }, + { + "ph": "s", "id": 102149, "pid": 435, "tid": 435, "ts": 1776649793339267, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339279, "dur": 0, + "args": { + "External id": 102155, + "cbid": 200, "correlation": 102155 + } + }, + { + "ph": "f", "id": 102155, "pid": 435, "tid": 435, "ts": 1776649793339279, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339280, "dur": 0, + "args": { + "External id": 102156, + "cbid": 273, "correlation": 102156 + } + }, + { + "ph": "f", "id": 102156, "pid": 435, "tid": 435, "ts": 1776649793339280, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793339289, "dur": 19, + "args": { + "External id": 102157, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102157, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102157, "pid": 0, "tid": 7, "ts": 1776649793339289, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339281, "dur": 4, + "args": { + "External id": 102157, + "cbid": 211, "correlation": 102157 + } + }, + { + "ph": "s", "id": 102157, "pid": 435, "tid": 435, "ts": 1776649793339281, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793339324, "dur": 1, + "args": { + "External id": 13702, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102175, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102175, "pid": 0, "tid": 7, "ts": 1776649793339324, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339315, "dur": 6, + "args": { + "External id": 13702, + "cbid": 211, "correlation": 102175 + } + }, + { + "ph": "s", "id": 102175, "pid": 435, "tid": 435, "ts": 1776649793339315, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339327, "dur": 0, + "args": { + "External id": 102181, + "cbid": 200, "correlation": 102181 + } + }, + { + "ph": "f", "id": 102181, "pid": 435, "tid": 435, "ts": 1776649793339327, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339328, "dur": 0, + "args": { + "External id": 102182, + "cbid": 273, "correlation": 102182 + } + }, + { + "ph": "f", "id": 102182, "pid": 435, "tid": 435, "ts": 1776649793339328, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793339337, "dur": 19, + "args": { + "External id": 102183, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102183, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102183, "pid": 0, "tid": 7, "ts": 1776649793339337, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339329, "dur": 4, + "args": { + "External id": 102183, + "cbid": 211, "correlation": 102183 + } + }, + { + "ph": "s", "id": 102183, "pid": 435, "tid": 435, "ts": 1776649793339329, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793339372, "dur": 1, + "args": { + "External id": 13707, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102201, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102201, "pid": 0, "tid": 7, "ts": 1776649793339372, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339363, "dur": 6, + "args": { + "External id": 13707, + "cbid": 211, "correlation": 102201 + } + }, + { + "ph": "s", "id": 102201, "pid": 435, "tid": 435, "ts": 1776649793339363, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339375, "dur": 0, + "args": { + "External id": 102207, + "cbid": 200, "correlation": 102207 + } + }, + { + "ph": "f", "id": 102207, "pid": 435, "tid": 435, "ts": 1776649793339375, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339376, "dur": 0, + "args": { + "External id": 102208, + "cbid": 273, "correlation": 102208 + } + }, + { + "ph": "f", "id": 102208, "pid": 435, "tid": 435, "ts": 1776649793339376, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793339385, "dur": 19, + "args": { + "External id": 102209, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102209, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102209, "pid": 0, "tid": 7, "ts": 1776649793339385, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339377, "dur": 5, + "args": { + "External id": 102209, + "cbid": 211, "correlation": 102209 + } + }, + { + "ph": "s", "id": 102209, "pid": 435, "tid": 435, "ts": 1776649793339377, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649793339399, "dur": 9, + "args": { + "External id": 102212, + "cbid": 165, "correlation": 102212 + } + }, + { + "ph": "s", "id": 102212, "pid": 435, "tid": 435, "ts": 1776649793339399, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649793339430, "dur": 3, + "args": { + "External id": 102221, + "cbid": 135, "correlation": 102221 + } + }, + { + "ph": "f", "id": 102221, "pid": 435, "tid": 435, "ts": 1776649793339430, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793339479, "dur": 1, + "args": { + "External id": 13712, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102241, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102241, "pid": 0, "tid": 7, "ts": 1776649793339479, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339468, "dur": 8, + "args": { + "External id": 13712, + "cbid": 211, "correlation": 102241 + } + }, + { + "ph": "s", "id": 102241, "pid": 435, "tid": 435, "ts": 1776649793339468, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339483, "dur": 0, + "args": { + "External id": 102247, + "cbid": 200, "correlation": 102247 + } + }, + { + "ph": "f", "id": 102247, "pid": 435, "tid": 435, "ts": 1776649793339483, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339484, "dur": 0, + "args": { + "External id": 102248, + "cbid": 273, "correlation": 102248 + } + }, + { + "ph": "f", "id": 102248, "pid": 435, "tid": 435, "ts": 1776649793339484, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793339493, "dur": 19, + "args": { + "External id": 102249, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102249, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102249, "pid": 0, "tid": 7, "ts": 1776649793339493, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339485, "dur": 5, + "args": { + "External id": 102249, + "cbid": 211, "correlation": 102249 + } + }, + { + "ph": "s", "id": 102249, "pid": 435, "tid": 435, "ts": 1776649793339485, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793339530, "dur": 1, + "args": { + "External id": 13717, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102267, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102267, "pid": 0, "tid": 7, "ts": 1776649793339530, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339521, "dur": 6, + "args": { + "External id": 13717, + "cbid": 211, "correlation": 102267 + } + }, + { + "ph": "s", "id": 102267, "pid": 435, "tid": 435, "ts": 1776649793339521, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339533, "dur": 0, + "args": { + "External id": 102273, + "cbid": 200, "correlation": 102273 + } + }, + { + "ph": "f", "id": 102273, "pid": 435, "tid": 435, "ts": 1776649793339533, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339534, "dur": 0, + "args": { + "External id": 102274, + "cbid": 273, "correlation": 102274 + } + }, + { + "ph": "f", "id": 102274, "pid": 435, "tid": 435, "ts": 1776649793339534, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793339543, "dur": 19, + "args": { + "External id": 102275, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102275, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102275, "pid": 0, "tid": 7, "ts": 1776649793339543, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339535, "dur": 4, + "args": { + "External id": 102275, + "cbid": 211, "correlation": 102275 + } + }, + { + "ph": "s", "id": 102275, "pid": 435, "tid": 435, "ts": 1776649793339535, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793339579, "dur": 1, + "args": { + "External id": 13722, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102293, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102293, "pid": 0, "tid": 7, "ts": 1776649793339579, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339569, "dur": 6, + "args": { + "External id": 13722, + "cbid": 211, "correlation": 102293 + } + }, + { + "ph": "s", "id": 102293, "pid": 435, "tid": 435, "ts": 1776649793339569, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339582, "dur": 0, + "args": { + "External id": 102299, + "cbid": 200, "correlation": 102299 + } + }, + { + "ph": "f", "id": 102299, "pid": 435, "tid": 435, "ts": 1776649793339582, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339583, "dur": 0, + "args": { + "External id": 102300, + "cbid": 273, "correlation": 102300 + } + }, + { + "ph": "f", "id": 102300, "pid": 435, "tid": 435, "ts": 1776649793339583, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793339591, "dur": 19, + "args": { + "External id": 102301, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102301, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102301, "pid": 0, "tid": 7, "ts": 1776649793339591, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339584, "dur": 4, + "args": { + "External id": 102301, + "cbid": 211, "correlation": 102301 + } + }, + { + "ph": "s", "id": 102301, "pid": 435, "tid": 435, "ts": 1776649793339584, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793339627, "dur": 1, + "args": { + "External id": 13727, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102319, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102319, "pid": 0, "tid": 7, "ts": 1776649793339627, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339617, "dur": 6, + "args": { + "External id": 13727, + "cbid": 211, "correlation": 102319 + } + }, + { + "ph": "s", "id": 102319, "pid": 435, "tid": 435, "ts": 1776649793339617, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339631, "dur": 0, + "args": { + "External id": 102325, + "cbid": 200, "correlation": 102325 + } + }, + { + "ph": "f", "id": 102325, "pid": 435, "tid": 435, "ts": 1776649793339631, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339632, "dur": 1, + "args": { + "External id": 102326, + "cbid": 273, "correlation": 102326 + } + }, + { + "ph": "f", "id": 102326, "pid": 435, "tid": 435, "ts": 1776649793339632, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793339642, "dur": 19, + "args": { + "External id": 102327, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102327, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102327, "pid": 0, "tid": 7, "ts": 1776649793339642, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339633, "dur": 5, + "args": { + "External id": 102327, + "cbid": 211, "correlation": 102327 + } + }, + { + "ph": "s", "id": 102327, "pid": 435, "tid": 435, "ts": 1776649793339633, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793339679, "dur": 1, + "args": { + "External id": 13732, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102345, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102345, "pid": 0, "tid": 7, "ts": 1776649793339679, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339670, "dur": 6, + "args": { + "External id": 13732, + "cbid": 211, "correlation": 102345 + } + }, + { + "ph": "s", "id": 102345, "pid": 435, "tid": 435, "ts": 1776649793339670, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339682, "dur": 0, + "args": { + "External id": 102351, + "cbid": 200, "correlation": 102351 + } + }, + { + "ph": "f", "id": 102351, "pid": 435, "tid": 435, "ts": 1776649793339682, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339683, "dur": 0, + "args": { + "External id": 102352, + "cbid": 273, "correlation": 102352 + } + }, + { + "ph": "f", "id": 102352, "pid": 435, "tid": 435, "ts": 1776649793339683, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793339692, "dur": 19, + "args": { + "External id": 102353, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102353, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102353, "pid": 0, "tid": 7, "ts": 1776649793339692, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339684, "dur": 5, + "args": { + "External id": 102353, + "cbid": 211, "correlation": 102353 + } + }, + { + "ph": "s", "id": 102353, "pid": 435, "tid": 435, "ts": 1776649793339684, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793339728, "dur": 1, + "args": { + "External id": 13737, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102371, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102371, "pid": 0, "tid": 7, "ts": 1776649793339728, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339718, "dur": 6, + "args": { + "External id": 13737, + "cbid": 211, "correlation": 102371 + } + }, + { + "ph": "s", "id": 102371, "pid": 435, "tid": 435, "ts": 1776649793339718, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339731, "dur": 0, + "args": { + "External id": 102377, + "cbid": 200, "correlation": 102377 + } + }, + { + "ph": "f", "id": 102377, "pid": 435, "tid": 435, "ts": 1776649793339731, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339732, "dur": 0, + "args": { + "External id": 102378, + "cbid": 273, "correlation": 102378 + } + }, + { + "ph": "f", "id": 102378, "pid": 435, "tid": 435, "ts": 1776649793339732, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793339741, "dur": 19, + "args": { + "External id": 102379, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102379, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102379, "pid": 0, "tid": 7, "ts": 1776649793339741, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339733, "dur": 4, + "args": { + "External id": 102379, + "cbid": 211, "correlation": 102379 + } + }, + { + "ph": "s", "id": 102379, "pid": 435, "tid": 435, "ts": 1776649793339733, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793339776, "dur": 1, + "args": { + "External id": 13742, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102397, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102397, "pid": 0, "tid": 7, "ts": 1776649793339776, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339767, "dur": 6, + "args": { + "External id": 13742, + "cbid": 211, "correlation": 102397 + } + }, + { + "ph": "s", "id": 102397, "pid": 435, "tid": 435, "ts": 1776649793339767, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339779, "dur": 0, + "args": { + "External id": 102403, + "cbid": 200, "correlation": 102403 + } + }, + { + "ph": "f", "id": 102403, "pid": 435, "tid": 435, "ts": 1776649793339779, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339780, "dur": 0, + "args": { + "External id": 102404, + "cbid": 273, "correlation": 102404 + } + }, + { + "ph": "f", "id": 102404, "pid": 435, "tid": 435, "ts": 1776649793339780, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793339789, "dur": 19, + "args": { + "External id": 102405, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102405, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102405, "pid": 0, "tid": 7, "ts": 1776649793339789, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339781, "dur": 4, + "args": { + "External id": 102405, + "cbid": 211, "correlation": 102405 + } + }, + { + "ph": "s", "id": 102405, "pid": 435, "tid": 435, "ts": 1776649793339781, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793339824, "dur": 1, + "args": { + "External id": 13747, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102423, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102423, "pid": 0, "tid": 7, "ts": 1776649793339824, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339815, "dur": 6, + "args": { + "External id": 13747, + "cbid": 211, "correlation": 102423 + } + }, + { + "ph": "s", "id": 102423, "pid": 435, "tid": 435, "ts": 1776649793339815, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339827, "dur": 0, + "args": { + "External id": 102429, + "cbid": 200, "correlation": 102429 + } + }, + { + "ph": "f", "id": 102429, "pid": 435, "tid": 435, "ts": 1776649793339827, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339828, "dur": 0, + "args": { + "External id": 102430, + "cbid": 273, "correlation": 102430 + } + }, + { + "ph": "f", "id": 102430, "pid": 435, "tid": 435, "ts": 1776649793339828, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793339837, "dur": 19, + "args": { + "External id": 102431, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102431, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102431, "pid": 0, "tid": 7, "ts": 1776649793339837, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339829, "dur": 4, + "args": { + "External id": 102431, + "cbid": 211, "correlation": 102431 + } + }, + { + "ph": "s", "id": 102431, "pid": 435, "tid": 435, "ts": 1776649793339829, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793339872, "dur": 1, + "args": { + "External id": 13752, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102449, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102449, "pid": 0, "tid": 7, "ts": 1776649793339872, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339863, "dur": 6, + "args": { + "External id": 13752, + "cbid": 211, "correlation": 102449 + } + }, + { + "ph": "s", "id": 102449, "pid": 435, "tid": 435, "ts": 1776649793339863, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339875, "dur": 0, + "args": { + "External id": 102455, + "cbid": 200, "correlation": 102455 + } + }, + { + "ph": "f", "id": 102455, "pid": 435, "tid": 435, "ts": 1776649793339875, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339876, "dur": 0, + "args": { + "External id": 102456, + "cbid": 273, "correlation": 102456 + } + }, + { + "ph": "f", "id": 102456, "pid": 435, "tid": 435, "ts": 1776649793339876, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793339885, "dur": 19, + "args": { + "External id": 102457, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102457, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102457, "pid": 0, "tid": 7, "ts": 1776649793339885, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339877, "dur": 4, + "args": { + "External id": 102457, + "cbid": 211, "correlation": 102457 + } + }, + { + "ph": "s", "id": 102457, "pid": 435, "tid": 435, "ts": 1776649793339877, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793339920, "dur": 1, + "args": { + "External id": 13757, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102475, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102475, "pid": 0, "tid": 7, "ts": 1776649793339920, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339911, "dur": 6, + "args": { + "External id": 13757, + "cbid": 211, "correlation": 102475 + } + }, + { + "ph": "s", "id": 102475, "pid": 435, "tid": 435, "ts": 1776649793339911, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339923, "dur": 0, + "args": { + "External id": 102481, + "cbid": 200, "correlation": 102481 + } + }, + { + "ph": "f", "id": 102481, "pid": 435, "tid": 435, "ts": 1776649793339923, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339924, "dur": 0, + "args": { + "External id": 102482, + "cbid": 273, "correlation": 102482 + } + }, + { + "ph": "f", "id": 102482, "pid": 435, "tid": 435, "ts": 1776649793339924, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793339933, "dur": 19, + "args": { + "External id": 102483, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102483, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102483, "pid": 0, "tid": 7, "ts": 1776649793339933, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339925, "dur": 4, + "args": { + "External id": 102483, + "cbid": 211, "correlation": 102483 + } + }, + { + "ph": "s", "id": 102483, "pid": 435, "tid": 435, "ts": 1776649793339925, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793339969, "dur": 1, + "args": { + "External id": 13762, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102501, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102501, "pid": 0, "tid": 7, "ts": 1776649793339969, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339959, "dur": 6, + "args": { + "External id": 13762, + "cbid": 211, "correlation": 102501 + } + }, + { + "ph": "s", "id": 102501, "pid": 435, "tid": 435, "ts": 1776649793339959, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339972, "dur": 0, + "args": { + "External id": 102507, + "cbid": 200, "correlation": 102507 + } + }, + { + "ph": "f", "id": 102507, "pid": 435, "tid": 435, "ts": 1776649793339972, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793339972, "dur": 0, + "args": { + "External id": 102508, + "cbid": 273, "correlation": 102508 + } + }, + { + "ph": "f", "id": 102508, "pid": 435, "tid": 435, "ts": 1776649793339972, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793339981, "dur": 19, + "args": { + "External id": 102509, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102509, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102509, "pid": 0, "tid": 7, "ts": 1776649793339981, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793339973, "dur": 5, + "args": { + "External id": 102509, + "cbid": 211, "correlation": 102509 + } + }, + { + "ph": "s", "id": 102509, "pid": 435, "tid": 435, "ts": 1776649793339973, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793340017, "dur": 1, + "args": { + "External id": 13767, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102527, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102527, "pid": 0, "tid": 7, "ts": 1776649793340017, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340008, "dur": 6, + "args": { + "External id": 13767, + "cbid": 211, "correlation": 102527 + } + }, + { + "ph": "s", "id": 102527, "pid": 435, "tid": 435, "ts": 1776649793340008, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340020, "dur": 0, + "args": { + "External id": 102533, + "cbid": 200, "correlation": 102533 + } + }, + { + "ph": "f", "id": 102533, "pid": 435, "tid": 435, "ts": 1776649793340020, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340021, "dur": 0, + "args": { + "External id": 102534, + "cbid": 273, "correlation": 102534 + } + }, + { + "ph": "f", "id": 102534, "pid": 435, "tid": 435, "ts": 1776649793340021, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793340030, "dur": 19, + "args": { + "External id": 102535, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102535, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102535, "pid": 0, "tid": 7, "ts": 1776649793340030, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340022, "dur": 4, + "args": { + "External id": 102535, + "cbid": 211, "correlation": 102535 + } + }, + { + "ph": "s", "id": 102535, "pid": 435, "tid": 435, "ts": 1776649793340022, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793340065, "dur": 1, + "args": { + "External id": 13772, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102553, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102553, "pid": 0, "tid": 7, "ts": 1776649793340065, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340056, "dur": 6, + "args": { + "External id": 13772, + "cbid": 211, "correlation": 102553 + } + }, + { + "ph": "s", "id": 102553, "pid": 435, "tid": 435, "ts": 1776649793340056, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340068, "dur": 0, + "args": { + "External id": 102559, + "cbid": 200, "correlation": 102559 + } + }, + { + "ph": "f", "id": 102559, "pid": 435, "tid": 435, "ts": 1776649793340068, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340069, "dur": 0, + "args": { + "External id": 102560, + "cbid": 273, "correlation": 102560 + } + }, + { + "ph": "f", "id": 102560, "pid": 435, "tid": 435, "ts": 1776649793340069, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793340078, "dur": 19, + "args": { + "External id": 102561, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102561, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102561, "pid": 0, "tid": 7, "ts": 1776649793340078, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340070, "dur": 4, + "args": { + "External id": 102561, + "cbid": 211, "correlation": 102561 + } + }, + { + "ph": "s", "id": 102561, "pid": 435, "tid": 435, "ts": 1776649793340070, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793340113, "dur": 1, + "args": { + "External id": 13777, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102579, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102579, "pid": 0, "tid": 7, "ts": 1776649793340113, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340104, "dur": 6, + "args": { + "External id": 13777, + "cbid": 211, "correlation": 102579 + } + }, + { + "ph": "s", "id": 102579, "pid": 435, "tid": 435, "ts": 1776649793340104, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340116, "dur": 0, + "args": { + "External id": 102585, + "cbid": 200, "correlation": 102585 + } + }, + { + "ph": "f", "id": 102585, "pid": 435, "tid": 435, "ts": 1776649793340116, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340116, "dur": 0, + "args": { + "External id": 102586, + "cbid": 273, "correlation": 102586 + } + }, + { + "ph": "f", "id": 102586, "pid": 435, "tid": 435, "ts": 1776649793340116, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793340125, "dur": 19, + "args": { + "External id": 102587, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102587, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102587, "pid": 0, "tid": 7, "ts": 1776649793340125, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340118, "dur": 4, + "args": { + "External id": 102587, + "cbid": 211, "correlation": 102587 + } + }, + { + "ph": "s", "id": 102587, "pid": 435, "tid": 435, "ts": 1776649793340118, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793340161, "dur": 1, + "args": { + "External id": 13782, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102605, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102605, "pid": 0, "tid": 7, "ts": 1776649793340161, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340151, "dur": 6, + "args": { + "External id": 13782, + "cbid": 211, "correlation": 102605 + } + }, + { + "ph": "s", "id": 102605, "pid": 435, "tid": 435, "ts": 1776649793340151, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340164, "dur": 0, + "args": { + "External id": 102611, + "cbid": 200, "correlation": 102611 + } + }, + { + "ph": "f", "id": 102611, "pid": 435, "tid": 435, "ts": 1776649793340164, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340164, "dur": 0, + "args": { + "External id": 102612, + "cbid": 273, "correlation": 102612 + } + }, + { + "ph": "f", "id": 102612, "pid": 435, "tid": 435, "ts": 1776649793340164, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793340173, "dur": 19, + "args": { + "External id": 102613, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102613, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102613, "pid": 0, "tid": 7, "ts": 1776649793340173, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340165, "dur": 4, + "args": { + "External id": 102613, + "cbid": 211, "correlation": 102613 + } + }, + { + "ph": "s", "id": 102613, "pid": 435, "tid": 435, "ts": 1776649793340165, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793340209, "dur": 1, + "args": { + "External id": 13787, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102631, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102631, "pid": 0, "tid": 7, "ts": 1776649793340209, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340199, "dur": 6, + "args": { + "External id": 13787, + "cbid": 211, "correlation": 102631 + } + }, + { + "ph": "s", "id": 102631, "pid": 435, "tid": 435, "ts": 1776649793340199, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340211, "dur": 0, + "args": { + "External id": 102637, + "cbid": 200, "correlation": 102637 + } + }, + { + "ph": "f", "id": 102637, "pid": 435, "tid": 435, "ts": 1776649793340211, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340212, "dur": 0, + "args": { + "External id": 102638, + "cbid": 273, "correlation": 102638 + } + }, + { + "ph": "f", "id": 102638, "pid": 435, "tid": 435, "ts": 1776649793340212, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793340239, "dur": 19, + "args": { + "External id": 102639, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102639, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102639, "pid": 0, "tid": 7, "ts": 1776649793340239, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340213, "dur": 23, + "args": { + "External id": 102639, + "cbid": 211, "correlation": 102639 + } + }, + { + "ph": "s", "id": 102639, "pid": 435, "tid": 435, "ts": 1776649793340213, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793340280, "dur": 1, + "args": { + "External id": 13792, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102657, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102657, "pid": 0, "tid": 7, "ts": 1776649793340280, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340270, "dur": 7, + "args": { + "External id": 13792, + "cbid": 211, "correlation": 102657 + } + }, + { + "ph": "s", "id": 102657, "pid": 435, "tid": 435, "ts": 1776649793340270, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340284, "dur": 0, + "args": { + "External id": 102663, + "cbid": 200, "correlation": 102663 + } + }, + { + "ph": "f", "id": 102663, "pid": 435, "tid": 435, "ts": 1776649793340284, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340284, "dur": 0, + "args": { + "External id": 102664, + "cbid": 273, "correlation": 102664 + } + }, + { + "ph": "f", "id": 102664, "pid": 435, "tid": 435, "ts": 1776649793340284, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793340293, "dur": 19, + "args": { + "External id": 102665, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102665, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102665, "pid": 0, "tid": 7, "ts": 1776649793340293, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340286, "dur": 5, + "args": { + "External id": 102665, + "cbid": 211, "correlation": 102665 + } + }, + { + "ph": "s", "id": 102665, "pid": 435, "tid": 435, "ts": 1776649793340286, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793340329, "dur": 1, + "args": { + "External id": 13797, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102683, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102683, "pid": 0, "tid": 7, "ts": 1776649793340329, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340320, "dur": 6, + "args": { + "External id": 13797, + "cbid": 211, "correlation": 102683 + } + }, + { + "ph": "s", "id": 102683, "pid": 435, "tid": 435, "ts": 1776649793340320, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340333, "dur": 0, + "args": { + "External id": 102689, + "cbid": 200, "correlation": 102689 + } + }, + { + "ph": "f", "id": 102689, "pid": 435, "tid": 435, "ts": 1776649793340333, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340333, "dur": 0, + "args": { + "External id": 102690, + "cbid": 273, "correlation": 102690 + } + }, + { + "ph": "f", "id": 102690, "pid": 435, "tid": 435, "ts": 1776649793340333, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793340342, "dur": 19, + "args": { + "External id": 102691, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102691, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102691, "pid": 0, "tid": 7, "ts": 1776649793340342, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340335, "dur": 4, + "args": { + "External id": 102691, + "cbid": 211, "correlation": 102691 + } + }, + { + "ph": "s", "id": 102691, "pid": 435, "tid": 435, "ts": 1776649793340335, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793340379, "dur": 1, + "args": { + "External id": 13802, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102709, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102709, "pid": 0, "tid": 7, "ts": 1776649793340379, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340369, "dur": 6, + "args": { + "External id": 13802, + "cbid": 211, "correlation": 102709 + } + }, + { + "ph": "s", "id": 102709, "pid": 435, "tid": 435, "ts": 1776649793340369, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340381, "dur": 0, + "args": { + "External id": 102715, + "cbid": 200, "correlation": 102715 + } + }, + { + "ph": "f", "id": 102715, "pid": 435, "tid": 435, "ts": 1776649793340381, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340382, "dur": 0, + "args": { + "External id": 102716, + "cbid": 273, "correlation": 102716 + } + }, + { + "ph": "f", "id": 102716, "pid": 435, "tid": 435, "ts": 1776649793340382, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793340391, "dur": 19, + "args": { + "External id": 102717, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102717, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102717, "pid": 0, "tid": 7, "ts": 1776649793340391, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340383, "dur": 4, + "args": { + "External id": 102717, + "cbid": 211, "correlation": 102717 + } + }, + { + "ph": "s", "id": 102717, "pid": 435, "tid": 435, "ts": 1776649793340383, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793340426, "dur": 1, + "args": { + "External id": 13807, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102735, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102735, "pid": 0, "tid": 7, "ts": 1776649793340426, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340417, "dur": 6, + "args": { + "External id": 13807, + "cbid": 211, "correlation": 102735 + } + }, + { + "ph": "s", "id": 102735, "pid": 435, "tid": 435, "ts": 1776649793340417, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340430, "dur": 0, + "args": { + "External id": 102741, + "cbid": 200, "correlation": 102741 + } + }, + { + "ph": "f", "id": 102741, "pid": 435, "tid": 435, "ts": 1776649793340430, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340430, "dur": 0, + "args": { + "External id": 102742, + "cbid": 273, "correlation": 102742 + } + }, + { + "ph": "f", "id": 102742, "pid": 435, "tid": 435, "ts": 1776649793340430, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793340439, "dur": 19, + "args": { + "External id": 102743, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102743, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102743, "pid": 0, "tid": 7, "ts": 1776649793340439, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340431, "dur": 4, + "args": { + "External id": 102743, + "cbid": 211, "correlation": 102743 + } + }, + { + "ph": "s", "id": 102743, "pid": 435, "tid": 435, "ts": 1776649793340431, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793340475, "dur": 1, + "args": { + "External id": 13812, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102761, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102761, "pid": 0, "tid": 7, "ts": 1776649793340475, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340465, "dur": 6, + "args": { + "External id": 13812, + "cbid": 211, "correlation": 102761 + } + }, + { + "ph": "s", "id": 102761, "pid": 435, "tid": 435, "ts": 1776649793340465, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340477, "dur": 0, + "args": { + "External id": 102767, + "cbid": 200, "correlation": 102767 + } + }, + { + "ph": "f", "id": 102767, "pid": 435, "tid": 435, "ts": 1776649793340477, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340478, "dur": 0, + "args": { + "External id": 102768, + "cbid": 273, "correlation": 102768 + } + }, + { + "ph": "f", "id": 102768, "pid": 435, "tid": 435, "ts": 1776649793340478, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793340487, "dur": 19, + "args": { + "External id": 102769, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102769, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102769, "pid": 0, "tid": 7, "ts": 1776649793340487, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340479, "dur": 4, + "args": { + "External id": 102769, + "cbid": 211, "correlation": 102769 + } + }, + { + "ph": "s", "id": 102769, "pid": 435, "tid": 435, "ts": 1776649793340479, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793340522, "dur": 1, + "args": { + "External id": 13817, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102787, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102787, "pid": 0, "tid": 7, "ts": 1776649793340522, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340513, "dur": 6, + "args": { + "External id": 13817, + "cbid": 211, "correlation": 102787 + } + }, + { + "ph": "s", "id": 102787, "pid": 435, "tid": 435, "ts": 1776649793340513, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340525, "dur": 0, + "args": { + "External id": 102793, + "cbid": 200, "correlation": 102793 + } + }, + { + "ph": "f", "id": 102793, "pid": 435, "tid": 435, "ts": 1776649793340525, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340526, "dur": 0, + "args": { + "External id": 102794, + "cbid": 273, "correlation": 102794 + } + }, + { + "ph": "f", "id": 102794, "pid": 435, "tid": 435, "ts": 1776649793340526, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793340535, "dur": 19, + "args": { + "External id": 102795, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102795, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102795, "pid": 0, "tid": 7, "ts": 1776649793340535, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340527, "dur": 4, + "args": { + "External id": 102795, + "cbid": 211, "correlation": 102795 + } + }, + { + "ph": "s", "id": 102795, "pid": 435, "tid": 435, "ts": 1776649793340527, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793340570, "dur": 1, + "args": { + "External id": 13822, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102813, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102813, "pid": 0, "tid": 7, "ts": 1776649793340570, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340560, "dur": 6, + "args": { + "External id": 13822, + "cbid": 211, "correlation": 102813 + } + }, + { + "ph": "s", "id": 102813, "pid": 435, "tid": 435, "ts": 1776649793340560, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340573, "dur": 0, + "args": { + "External id": 102819, + "cbid": 200, "correlation": 102819 + } + }, + { + "ph": "f", "id": 102819, "pid": 435, "tid": 435, "ts": 1776649793340573, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340573, "dur": 0, + "args": { + "External id": 102820, + "cbid": 273, "correlation": 102820 + } + }, + { + "ph": "f", "id": 102820, "pid": 435, "tid": 435, "ts": 1776649793340573, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793340582, "dur": 19, + "args": { + "External id": 102821, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102821, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102821, "pid": 0, "tid": 7, "ts": 1776649793340582, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340574, "dur": 5, + "args": { + "External id": 102821, + "cbid": 211, "correlation": 102821 + } + }, + { + "ph": "s", "id": 102821, "pid": 435, "tid": 435, "ts": 1776649793340574, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793340628, "dur": 1, + "args": { + "External id": 13827, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102839, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102839, "pid": 0, "tid": 7, "ts": 1776649793340628, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340619, "dur": 6, + "args": { + "External id": 13827, + "cbid": 211, "correlation": 102839 + } + }, + { + "ph": "s", "id": 102839, "pid": 435, "tid": 435, "ts": 1776649793340619, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340631, "dur": 0, + "args": { + "External id": 102845, + "cbid": 200, "correlation": 102845 + } + }, + { + "ph": "f", "id": 102845, "pid": 435, "tid": 435, "ts": 1776649793340631, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340632, "dur": 0, + "args": { + "External id": 102846, + "cbid": 273, "correlation": 102846 + } + }, + { + "ph": "f", "id": 102846, "pid": 435, "tid": 435, "ts": 1776649793340632, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793340641, "dur": 19, + "args": { + "External id": 102847, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102847, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102847, "pid": 0, "tid": 7, "ts": 1776649793340641, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340633, "dur": 4, + "args": { + "External id": 102847, + "cbid": 211, "correlation": 102847 + } + }, + { + "ph": "s", "id": 102847, "pid": 435, "tid": 435, "ts": 1776649793340633, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793340676, "dur": 1, + "args": { + "External id": 13832, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102865, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102865, "pid": 0, "tid": 7, "ts": 1776649793340676, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340667, "dur": 6, + "args": { + "External id": 13832, + "cbid": 211, "correlation": 102865 + } + }, + { + "ph": "s", "id": 102865, "pid": 435, "tid": 435, "ts": 1776649793340667, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340679, "dur": 0, + "args": { + "External id": 102871, + "cbid": 200, "correlation": 102871 + } + }, + { + "ph": "f", "id": 102871, "pid": 435, "tid": 435, "ts": 1776649793340679, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340679, "dur": 0, + "args": { + "External id": 102872, + "cbid": 273, "correlation": 102872 + } + }, + { + "ph": "f", "id": 102872, "pid": 435, "tid": 435, "ts": 1776649793340679, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793340688, "dur": 19, + "args": { + "External id": 102873, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102873, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102873, "pid": 0, "tid": 7, "ts": 1776649793340688, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340680, "dur": 4, + "args": { + "External id": 102873, + "cbid": 211, "correlation": 102873 + } + }, + { + "ph": "s", "id": 102873, "pid": 435, "tid": 435, "ts": 1776649793340680, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793340724, "dur": 1, + "args": { + "External id": 13837, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102891, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102891, "pid": 0, "tid": 7, "ts": 1776649793340724, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340714, "dur": 6, + "args": { + "External id": 13837, + "cbid": 211, "correlation": 102891 + } + }, + { + "ph": "s", "id": 102891, "pid": 435, "tid": 435, "ts": 1776649793340714, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340727, "dur": 0, + "args": { + "External id": 102897, + "cbid": 200, "correlation": 102897 + } + }, + { + "ph": "f", "id": 102897, "pid": 435, "tid": 435, "ts": 1776649793340727, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340727, "dur": 0, + "args": { + "External id": 102898, + "cbid": 273, "correlation": 102898 + } + }, + { + "ph": "f", "id": 102898, "pid": 435, "tid": 435, "ts": 1776649793340727, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793340736, "dur": 19, + "args": { + "External id": 102899, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102899, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102899, "pid": 0, "tid": 7, "ts": 1776649793340736, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340729, "dur": 4, + "args": { + "External id": 102899, + "cbid": 211, "correlation": 102899 + } + }, + { + "ph": "s", "id": 102899, "pid": 435, "tid": 435, "ts": 1776649793340729, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793340771, "dur": 1, + "args": { + "External id": 13842, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102917, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102917, "pid": 0, "tid": 7, "ts": 1776649793340771, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340762, "dur": 6, + "args": { + "External id": 13842, + "cbid": 211, "correlation": 102917 + } + }, + { + "ph": "s", "id": 102917, "pid": 435, "tid": 435, "ts": 1776649793340762, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340774, "dur": 0, + "args": { + "External id": 102923, + "cbid": 200, "correlation": 102923 + } + }, + { + "ph": "f", "id": 102923, "pid": 435, "tid": 435, "ts": 1776649793340774, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340775, "dur": 0, + "args": { + "External id": 102924, + "cbid": 273, "correlation": 102924 + } + }, + { + "ph": "f", "id": 102924, "pid": 435, "tid": 435, "ts": 1776649793340775, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793340784, "dur": 19, + "args": { + "External id": 102925, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102925, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102925, "pid": 0, "tid": 7, "ts": 1776649793340784, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340776, "dur": 4, + "args": { + "External id": 102925, + "cbid": 211, "correlation": 102925 + } + }, + { + "ph": "s", "id": 102925, "pid": 435, "tid": 435, "ts": 1776649793340776, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793340819, "dur": 1, + "args": { + "External id": 13847, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102943, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102943, "pid": 0, "tid": 7, "ts": 1776649793340819, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340810, "dur": 6, + "args": { + "External id": 13847, + "cbid": 211, "correlation": 102943 + } + }, + { + "ph": "s", "id": 102943, "pid": 435, "tid": 435, "ts": 1776649793340810, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340822, "dur": 0, + "args": { + "External id": 102949, + "cbid": 200, "correlation": 102949 + } + }, + { + "ph": "f", "id": 102949, "pid": 435, "tid": 435, "ts": 1776649793340822, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340823, "dur": 0, + "args": { + "External id": 102950, + "cbid": 273, "correlation": 102950 + } + }, + { + "ph": "f", "id": 102950, "pid": 435, "tid": 435, "ts": 1776649793340823, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793340832, "dur": 19, + "args": { + "External id": 102951, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102951, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102951, "pid": 0, "tid": 7, "ts": 1776649793340832, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340824, "dur": 5, + "args": { + "External id": 102951, + "cbid": 211, "correlation": 102951 + } + }, + { + "ph": "s", "id": 102951, "pid": 435, "tid": 435, "ts": 1776649793340824, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793340867, "dur": 1, + "args": { + "External id": 13852, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102969, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102969, "pid": 0, "tid": 7, "ts": 1776649793340867, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340858, "dur": 6, + "args": { + "External id": 13852, + "cbid": 211, "correlation": 102969 + } + }, + { + "ph": "s", "id": 102969, "pid": 435, "tid": 435, "ts": 1776649793340858, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340870, "dur": 0, + "args": { + "External id": 102975, + "cbid": 200, "correlation": 102975 + } + }, + { + "ph": "f", "id": 102975, "pid": 435, "tid": 435, "ts": 1776649793340870, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340870, "dur": 0, + "args": { + "External id": 102976, + "cbid": 273, "correlation": 102976 + } + }, + { + "ph": "f", "id": 102976, "pid": 435, "tid": 435, "ts": 1776649793340870, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793340879, "dur": 19, + "args": { + "External id": 102977, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102977, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102977, "pid": 0, "tid": 7, "ts": 1776649793340879, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340871, "dur": 4, + "args": { + "External id": 102977, + "cbid": 211, "correlation": 102977 + } + }, + { + "ph": "s", "id": 102977, "pid": 435, "tid": 435, "ts": 1776649793340871, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793340914, "dur": 1, + "args": { + "External id": 13857, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 102995, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 102995, "pid": 0, "tid": 7, "ts": 1776649793340914, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340905, "dur": 6, + "args": { + "External id": 13857, + "cbid": 211, "correlation": 102995 + } + }, + { + "ph": "s", "id": 102995, "pid": 435, "tid": 435, "ts": 1776649793340905, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340917, "dur": 0, + "args": { + "External id": 103001, + "cbid": 200, "correlation": 103001 + } + }, + { + "ph": "f", "id": 103001, "pid": 435, "tid": 435, "ts": 1776649793340917, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340918, "dur": 0, + "args": { + "External id": 103002, + "cbid": 273, "correlation": 103002 + } + }, + { + "ph": "f", "id": 103002, "pid": 435, "tid": 435, "ts": 1776649793340918, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793340927, "dur": 19, + "args": { + "External id": 103003, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103003, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103003, "pid": 0, "tid": 7, "ts": 1776649793340927, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340919, "dur": 4, + "args": { + "External id": 103003, + "cbid": 211, "correlation": 103003 + } + }, + { + "ph": "s", "id": 103003, "pid": 435, "tid": 435, "ts": 1776649793340919, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793340962, "dur": 1, + "args": { + "External id": 13862, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103021, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103021, "pid": 0, "tid": 7, "ts": 1776649793340962, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340953, "dur": 6, + "args": { + "External id": 13862, + "cbid": 211, "correlation": 103021 + } + }, + { + "ph": "s", "id": 103021, "pid": 435, "tid": 435, "ts": 1776649793340953, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340965, "dur": 0, + "args": { + "External id": 103027, + "cbid": 200, "correlation": 103027 + } + }, + { + "ph": "f", "id": 103027, "pid": 435, "tid": 435, "ts": 1776649793340965, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793340965, "dur": 0, + "args": { + "External id": 103028, + "cbid": 273, "correlation": 103028 + } + }, + { + "ph": "f", "id": 103028, "pid": 435, "tid": 435, "ts": 1776649793340965, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793340975, "dur": 19, + "args": { + "External id": 103029, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103029, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103029, "pid": 0, "tid": 7, "ts": 1776649793340975, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793340966, "dur": 4, + "args": { + "External id": 103029, + "cbid": 211, "correlation": 103029 + } + }, + { + "ph": "s", "id": 103029, "pid": 435, "tid": 435, "ts": 1776649793340966, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793341012, "dur": 1, + "args": { + "External id": 13867, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103047, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103047, "pid": 0, "tid": 7, "ts": 1776649793341012, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341003, "dur": 6, + "args": { + "External id": 13867, + "cbid": 211, "correlation": 103047 + } + }, + { + "ph": "s", "id": 103047, "pid": 435, "tid": 435, "ts": 1776649793341003, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341014, "dur": 0, + "args": { + "External id": 103053, + "cbid": 200, "correlation": 103053 + } + }, + { + "ph": "f", "id": 103053, "pid": 435, "tid": 435, "ts": 1776649793341014, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341015, "dur": 0, + "args": { + "External id": 103054, + "cbid": 273, "correlation": 103054 + } + }, + { + "ph": "f", "id": 103054, "pid": 435, "tid": 435, "ts": 1776649793341015, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793341024, "dur": 19, + "args": { + "External id": 103055, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103055, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103055, "pid": 0, "tid": 7, "ts": 1776649793341024, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341016, "dur": 4, + "args": { + "External id": 103055, + "cbid": 211, "correlation": 103055 + } + }, + { + "ph": "s", "id": 103055, "pid": 435, "tid": 435, "ts": 1776649793341016, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793341059, "dur": 1, + "args": { + "External id": 13872, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103073, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103073, "pid": 0, "tid": 7, "ts": 1776649793341059, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341050, "dur": 6, + "args": { + "External id": 13872, + "cbid": 211, "correlation": 103073 + } + }, + { + "ph": "s", "id": 103073, "pid": 435, "tid": 435, "ts": 1776649793341050, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341062, "dur": 0, + "args": { + "External id": 103079, + "cbid": 200, "correlation": 103079 + } + }, + { + "ph": "f", "id": 103079, "pid": 435, "tid": 435, "ts": 1776649793341062, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341063, "dur": 0, + "args": { + "External id": 103080, + "cbid": 273, "correlation": 103080 + } + }, + { + "ph": "f", "id": 103080, "pid": 435, "tid": 435, "ts": 1776649793341063, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793341072, "dur": 19, + "args": { + "External id": 103081, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103081, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103081, "pid": 0, "tid": 7, "ts": 1776649793341072, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341064, "dur": 5, + "args": { + "External id": 103081, + "cbid": 211, "correlation": 103081 + } + }, + { + "ph": "s", "id": 103081, "pid": 435, "tid": 435, "ts": 1776649793341064, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793341107, "dur": 1, + "args": { + "External id": 13877, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103099, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103099, "pid": 0, "tid": 7, "ts": 1776649793341107, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341098, "dur": 6, + "args": { + "External id": 13877, + "cbid": 211, "correlation": 103099 + } + }, + { + "ph": "s", "id": 103099, "pid": 435, "tid": 435, "ts": 1776649793341098, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341110, "dur": 0, + "args": { + "External id": 103105, + "cbid": 200, "correlation": 103105 + } + }, + { + "ph": "f", "id": 103105, "pid": 435, "tid": 435, "ts": 1776649793341110, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341110, "dur": 0, + "args": { + "External id": 103106, + "cbid": 273, "correlation": 103106 + } + }, + { + "ph": "f", "id": 103106, "pid": 435, "tid": 435, "ts": 1776649793341110, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793341119, "dur": 19, + "args": { + "External id": 103107, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103107, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103107, "pid": 0, "tid": 7, "ts": 1776649793341119, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341111, "dur": 4, + "args": { + "External id": 103107, + "cbid": 211, "correlation": 103107 + } + }, + { + "ph": "s", "id": 103107, "pid": 435, "tid": 435, "ts": 1776649793341111, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793341154, "dur": 1, + "args": { + "External id": 13882, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103125, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103125, "pid": 0, "tid": 7, "ts": 1776649793341154, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341145, "dur": 6, + "args": { + "External id": 13882, + "cbid": 211, "correlation": 103125 + } + }, + { + "ph": "s", "id": 103125, "pid": 435, "tid": 435, "ts": 1776649793341145, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341157, "dur": 0, + "args": { + "External id": 103131, + "cbid": 200, "correlation": 103131 + } + }, + { + "ph": "f", "id": 103131, "pid": 435, "tid": 435, "ts": 1776649793341157, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341158, "dur": 0, + "args": { + "External id": 103132, + "cbid": 273, "correlation": 103132 + } + }, + { + "ph": "f", "id": 103132, "pid": 435, "tid": 435, "ts": 1776649793341158, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793341166, "dur": 19, + "args": { + "External id": 103133, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103133, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103133, "pid": 0, "tid": 7, "ts": 1776649793341166, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341159, "dur": 4, + "args": { + "External id": 103133, + "cbid": 211, "correlation": 103133 + } + }, + { + "ph": "s", "id": 103133, "pid": 435, "tid": 435, "ts": 1776649793341159, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793341201, "dur": 1, + "args": { + "External id": 13887, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103151, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103151, "pid": 0, "tid": 7, "ts": 1776649793341201, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341192, "dur": 6, + "args": { + "External id": 13887, + "cbid": 211, "correlation": 103151 + } + }, + { + "ph": "s", "id": 103151, "pid": 435, "tid": 435, "ts": 1776649793341192, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341204, "dur": 0, + "args": { + "External id": 103157, + "cbid": 200, "correlation": 103157 + } + }, + { + "ph": "f", "id": 103157, "pid": 435, "tid": 435, "ts": 1776649793341204, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341205, "dur": 0, + "args": { + "External id": 103158, + "cbid": 273, "correlation": 103158 + } + }, + { + "ph": "f", "id": 103158, "pid": 435, "tid": 435, "ts": 1776649793341205, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793341214, "dur": 19, + "args": { + "External id": 103159, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103159, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103159, "pid": 0, "tid": 7, "ts": 1776649793341214, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341206, "dur": 4, + "args": { + "External id": 103159, + "cbid": 211, "correlation": 103159 + } + }, + { + "ph": "s", "id": 103159, "pid": 435, "tid": 435, "ts": 1776649793341206, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793341249, "dur": 1, + "args": { + "External id": 13892, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103177, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103177, "pid": 0, "tid": 7, "ts": 1776649793341249, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341240, "dur": 6, + "args": { + "External id": 13892, + "cbid": 211, "correlation": 103177 + } + }, + { + "ph": "s", "id": 103177, "pid": 435, "tid": 435, "ts": 1776649793341240, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341252, "dur": 0, + "args": { + "External id": 103183, + "cbid": 200, "correlation": 103183 + } + }, + { + "ph": "f", "id": 103183, "pid": 435, "tid": 435, "ts": 1776649793341252, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341253, "dur": 0, + "args": { + "External id": 103184, + "cbid": 273, "correlation": 103184 + } + }, + { + "ph": "f", "id": 103184, "pid": 435, "tid": 435, "ts": 1776649793341253, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793341262, "dur": 19, + "args": { + "External id": 103185, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103185, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103185, "pid": 0, "tid": 7, "ts": 1776649793341262, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341254, "dur": 5, + "args": { + "External id": 103185, + "cbid": 211, "correlation": 103185 + } + }, + { + "ph": "s", "id": 103185, "pid": 435, "tid": 435, "ts": 1776649793341254, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793341297, "dur": 1, + "args": { + "External id": 13897, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103203, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103203, "pid": 0, "tid": 7, "ts": 1776649793341297, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341288, "dur": 6, + "args": { + "External id": 13897, + "cbid": 211, "correlation": 103203 + } + }, + { + "ph": "s", "id": 103203, "pid": 435, "tid": 435, "ts": 1776649793341288, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341300, "dur": 0, + "args": { + "External id": 103209, + "cbid": 200, "correlation": 103209 + } + }, + { + "ph": "f", "id": 103209, "pid": 435, "tid": 435, "ts": 1776649793341300, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341300, "dur": 0, + "args": { + "External id": 103210, + "cbid": 273, "correlation": 103210 + } + }, + { + "ph": "f", "id": 103210, "pid": 435, "tid": 435, "ts": 1776649793341300, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793341310, "dur": 19, + "args": { + "External id": 103211, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103211, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103211, "pid": 0, "tid": 7, "ts": 1776649793341310, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341301, "dur": 4, + "args": { + "External id": 103211, + "cbid": 211, "correlation": 103211 + } + }, + { + "ph": "s", "id": 103211, "pid": 435, "tid": 435, "ts": 1776649793341301, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793341344, "dur": 1, + "args": { + "External id": 13902, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103229, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103229, "pid": 0, "tid": 7, "ts": 1776649793341344, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341335, "dur": 6, + "args": { + "External id": 13902, + "cbid": 211, "correlation": 103229 + } + }, + { + "ph": "s", "id": 103229, "pid": 435, "tid": 435, "ts": 1776649793341335, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341347, "dur": 0, + "args": { + "External id": 103235, + "cbid": 200, "correlation": 103235 + } + }, + { + "ph": "f", "id": 103235, "pid": 435, "tid": 435, "ts": 1776649793341347, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341348, "dur": 0, + "args": { + "External id": 103236, + "cbid": 273, "correlation": 103236 + } + }, + { + "ph": "f", "id": 103236, "pid": 435, "tid": 435, "ts": 1776649793341348, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793341357, "dur": 19, + "args": { + "External id": 103237, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103237, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103237, "pid": 0, "tid": 7, "ts": 1776649793341357, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341349, "dur": 4, + "args": { + "External id": 103237, + "cbid": 211, "correlation": 103237 + } + }, + { + "ph": "s", "id": 103237, "pid": 435, "tid": 435, "ts": 1776649793341349, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793341392, "dur": 1, + "args": { + "External id": 13907, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103255, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103255, "pid": 0, "tid": 7, "ts": 1776649793341392, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341383, "dur": 6, + "args": { + "External id": 13907, + "cbid": 211, "correlation": 103255 + } + }, + { + "ph": "s", "id": 103255, "pid": 435, "tid": 435, "ts": 1776649793341383, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341395, "dur": 0, + "args": { + "External id": 103261, + "cbid": 200, "correlation": 103261 + } + }, + { + "ph": "f", "id": 103261, "pid": 435, "tid": 435, "ts": 1776649793341395, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341395, "dur": 0, + "args": { + "External id": 103262, + "cbid": 273, "correlation": 103262 + } + }, + { + "ph": "f", "id": 103262, "pid": 435, "tid": 435, "ts": 1776649793341395, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793341404, "dur": 19, + "args": { + "External id": 103263, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103263, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103263, "pid": 0, "tid": 7, "ts": 1776649793341404, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341397, "dur": 4, + "args": { + "External id": 103263, + "cbid": 211, "correlation": 103263 + } + }, + { + "ph": "s", "id": 103263, "pid": 435, "tid": 435, "ts": 1776649793341397, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793341439, "dur": 1, + "args": { + "External id": 13912, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103281, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103281, "pid": 0, "tid": 7, "ts": 1776649793341439, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341430, "dur": 6, + "args": { + "External id": 13912, + "cbid": 211, "correlation": 103281 + } + }, + { + "ph": "s", "id": 103281, "pid": 435, "tid": 435, "ts": 1776649793341430, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341442, "dur": 0, + "args": { + "External id": 103287, + "cbid": 200, "correlation": 103287 + } + }, + { + "ph": "f", "id": 103287, "pid": 435, "tid": 435, "ts": 1776649793341442, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341443, "dur": 0, + "args": { + "External id": 103288, + "cbid": 273, "correlation": 103288 + } + }, + { + "ph": "f", "id": 103288, "pid": 435, "tid": 435, "ts": 1776649793341443, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793341452, "dur": 19, + "args": { + "External id": 103289, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103289, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103289, "pid": 0, "tid": 7, "ts": 1776649793341452, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341444, "dur": 5, + "args": { + "External id": 103289, + "cbid": 211, "correlation": 103289 + } + }, + { + "ph": "s", "id": 103289, "pid": 435, "tid": 435, "ts": 1776649793341444, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793341486, "dur": 1, + "args": { + "External id": 13917, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103307, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103307, "pid": 0, "tid": 7, "ts": 1776649793341486, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341477, "dur": 6, + "args": { + "External id": 13917, + "cbid": 211, "correlation": 103307 + } + }, + { + "ph": "s", "id": 103307, "pid": 435, "tid": 435, "ts": 1776649793341477, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341490, "dur": 0, + "args": { + "External id": 103313, + "cbid": 200, "correlation": 103313 + } + }, + { + "ph": "f", "id": 103313, "pid": 435, "tid": 435, "ts": 1776649793341490, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341490, "dur": 0, + "args": { + "External id": 103314, + "cbid": 273, "correlation": 103314 + } + }, + { + "ph": "f", "id": 103314, "pid": 435, "tid": 435, "ts": 1776649793341490, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793341500, "dur": 19, + "args": { + "External id": 103315, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103315, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103315, "pid": 0, "tid": 7, "ts": 1776649793341500, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341491, "dur": 5, + "args": { + "External id": 103315, + "cbid": 211, "correlation": 103315 + } + }, + { + "ph": "s", "id": 103315, "pid": 435, "tid": 435, "ts": 1776649793341491, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793341535, "dur": 1, + "args": { + "External id": 13922, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103333, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103333, "pid": 0, "tid": 7, "ts": 1776649793341535, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341525, "dur": 6, + "args": { + "External id": 13922, + "cbid": 211, "correlation": 103333 + } + }, + { + "ph": "s", "id": 103333, "pid": 435, "tid": 435, "ts": 1776649793341525, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341537, "dur": 0, + "args": { + "External id": 103339, + "cbid": 200, "correlation": 103339 + } + }, + { + "ph": "f", "id": 103339, "pid": 435, "tid": 435, "ts": 1776649793341537, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341538, "dur": 0, + "args": { + "External id": 103340, + "cbid": 273, "correlation": 103340 + } + }, + { + "ph": "f", "id": 103340, "pid": 435, "tid": 435, "ts": 1776649793341538, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793341547, "dur": 19, + "args": { + "External id": 103341, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103341, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103341, "pid": 0, "tid": 7, "ts": 1776649793341547, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341539, "dur": 4, + "args": { + "External id": 103341, + "cbid": 211, "correlation": 103341 + } + }, + { + "ph": "s", "id": 103341, "pid": 435, "tid": 435, "ts": 1776649793341539, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793341582, "dur": 1, + "args": { + "External id": 13927, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103359, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103359, "pid": 0, "tid": 7, "ts": 1776649793341582, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341573, "dur": 6, + "args": { + "External id": 13927, + "cbid": 211, "correlation": 103359 + } + }, + { + "ph": "s", "id": 103359, "pid": 435, "tid": 435, "ts": 1776649793341573, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341585, "dur": 0, + "args": { + "External id": 103365, + "cbid": 200, "correlation": 103365 + } + }, + { + "ph": "f", "id": 103365, "pid": 435, "tid": 435, "ts": 1776649793341585, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341586, "dur": 0, + "args": { + "External id": 103366, + "cbid": 273, "correlation": 103366 + } + }, + { + "ph": "f", "id": 103366, "pid": 435, "tid": 435, "ts": 1776649793341586, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793341595, "dur": 19, + "args": { + "External id": 103367, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103367, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103367, "pid": 0, "tid": 7, "ts": 1776649793341595, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341587, "dur": 4, + "args": { + "External id": 103367, + "cbid": 211, "correlation": 103367 + } + }, + { + "ph": "s", "id": 103367, "pid": 435, "tid": 435, "ts": 1776649793341587, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793341630, "dur": 1, + "args": { + "External id": 13932, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103385, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103385, "pid": 0, "tid": 7, "ts": 1776649793341630, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341620, "dur": 6, + "args": { + "External id": 13932, + "cbid": 211, "correlation": 103385 + } + }, + { + "ph": "s", "id": 103385, "pid": 435, "tid": 435, "ts": 1776649793341620, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341633, "dur": 0, + "args": { + "External id": 103391, + "cbid": 200, "correlation": 103391 + } + }, + { + "ph": "f", "id": 103391, "pid": 435, "tid": 435, "ts": 1776649793341633, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341633, "dur": 0, + "args": { + "External id": 103392, + "cbid": 273, "correlation": 103392 + } + }, + { + "ph": "f", "id": 103392, "pid": 435, "tid": 435, "ts": 1776649793341633, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793341642, "dur": 19, + "args": { + "External id": 103393, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103393, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103393, "pid": 0, "tid": 7, "ts": 1776649793341642, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341634, "dur": 5, + "args": { + "External id": 103393, + "cbid": 211, "correlation": 103393 + } + }, + { + "ph": "s", "id": 103393, "pid": 435, "tid": 435, "ts": 1776649793341634, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793341677, "dur": 1, + "args": { + "External id": 13937, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103411, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103411, "pid": 0, "tid": 7, "ts": 1776649793341677, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341668, "dur": 6, + "args": { + "External id": 13937, + "cbid": 211, "correlation": 103411 + } + }, + { + "ph": "s", "id": 103411, "pid": 435, "tid": 435, "ts": 1776649793341668, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341680, "dur": 0, + "args": { + "External id": 103417, + "cbid": 200, "correlation": 103417 + } + }, + { + "ph": "f", "id": 103417, "pid": 435, "tid": 435, "ts": 1776649793341680, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341681, "dur": 0, + "args": { + "External id": 103418, + "cbid": 273, "correlation": 103418 + } + }, + { + "ph": "f", "id": 103418, "pid": 435, "tid": 435, "ts": 1776649793341681, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793341690, "dur": 19, + "args": { + "External id": 103419, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103419, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103419, "pid": 0, "tid": 7, "ts": 1776649793341690, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341682, "dur": 5, + "args": { + "External id": 103419, + "cbid": 211, "correlation": 103419 + } + }, + { + "ph": "s", "id": 103419, "pid": 435, "tid": 435, "ts": 1776649793341682, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793341725, "dur": 1, + "args": { + "External id": 13942, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103437, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103437, "pid": 0, "tid": 7, "ts": 1776649793341725, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341716, "dur": 6, + "args": { + "External id": 13942, + "cbid": 211, "correlation": 103437 + } + }, + { + "ph": "s", "id": 103437, "pid": 435, "tid": 435, "ts": 1776649793341716, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341728, "dur": 0, + "args": { + "External id": 103443, + "cbid": 200, "correlation": 103443 + } + }, + { + "ph": "f", "id": 103443, "pid": 435, "tid": 435, "ts": 1776649793341728, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341728, "dur": 0, + "args": { + "External id": 103444, + "cbid": 273, "correlation": 103444 + } + }, + { + "ph": "f", "id": 103444, "pid": 435, "tid": 435, "ts": 1776649793341728, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793341737, "dur": 19, + "args": { + "External id": 103445, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103445, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103445, "pid": 0, "tid": 7, "ts": 1776649793341737, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341730, "dur": 5, + "args": { + "External id": 103445, + "cbid": 211, "correlation": 103445 + } + }, + { + "ph": "s", "id": 103445, "pid": 435, "tid": 435, "ts": 1776649793341730, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793341773, "dur": 1, + "args": { + "External id": 13947, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103463, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103463, "pid": 0, "tid": 7, "ts": 1776649793341773, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341764, "dur": 6, + "args": { + "External id": 13947, + "cbid": 211, "correlation": 103463 + } + }, + { + "ph": "s", "id": 103463, "pid": 435, "tid": 435, "ts": 1776649793341764, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341776, "dur": 0, + "args": { + "External id": 103469, + "cbid": 200, "correlation": 103469 + } + }, + { + "ph": "f", "id": 103469, "pid": 435, "tid": 435, "ts": 1776649793341776, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341776, "dur": 0, + "args": { + "External id": 103470, + "cbid": 273, "correlation": 103470 + } + }, + { + "ph": "f", "id": 103470, "pid": 435, "tid": 435, "ts": 1776649793341776, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793341785, "dur": 19, + "args": { + "External id": 103471, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103471, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103471, "pid": 0, "tid": 7, "ts": 1776649793341785, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341777, "dur": 5, + "args": { + "External id": 103471, + "cbid": 211, "correlation": 103471 + } + }, + { + "ph": "s", "id": 103471, "pid": 435, "tid": 435, "ts": 1776649793341777, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793341820, "dur": 1, + "args": { + "External id": 13952, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103489, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103489, "pid": 0, "tid": 7, "ts": 1776649793341820, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341811, "dur": 6, + "args": { + "External id": 13952, + "cbid": 211, "correlation": 103489 + } + }, + { + "ph": "s", "id": 103489, "pid": 435, "tid": 435, "ts": 1776649793341811, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341823, "dur": 0, + "args": { + "External id": 103495, + "cbid": 200, "correlation": 103495 + } + }, + { + "ph": "f", "id": 103495, "pid": 435, "tid": 435, "ts": 1776649793341823, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341824, "dur": 0, + "args": { + "External id": 103496, + "cbid": 273, "correlation": 103496 + } + }, + { + "ph": "f", "id": 103496, "pid": 435, "tid": 435, "ts": 1776649793341824, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793341832, "dur": 19, + "args": { + "External id": 103497, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103497, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103497, "pid": 0, "tid": 7, "ts": 1776649793341832, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341825, "dur": 4, + "args": { + "External id": 103497, + "cbid": 211, "correlation": 103497 + } + }, + { + "ph": "s", "id": 103497, "pid": 435, "tid": 435, "ts": 1776649793341825, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793341867, "dur": 1, + "args": { + "External id": 13957, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103515, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103515, "pid": 0, "tid": 7, "ts": 1776649793341867, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341858, "dur": 6, + "args": { + "External id": 13957, + "cbid": 211, "correlation": 103515 + } + }, + { + "ph": "s", "id": 103515, "pid": 435, "tid": 435, "ts": 1776649793341858, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341870, "dur": 0, + "args": { + "External id": 103521, + "cbid": 200, "correlation": 103521 + } + }, + { + "ph": "f", "id": 103521, "pid": 435, "tid": 435, "ts": 1776649793341870, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341871, "dur": 0, + "args": { + "External id": 103522, + "cbid": 273, "correlation": 103522 + } + }, + { + "ph": "f", "id": 103522, "pid": 435, "tid": 435, "ts": 1776649793341871, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793341880, "dur": 19, + "args": { + "External id": 103523, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103523, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103523, "pid": 0, "tid": 7, "ts": 1776649793341880, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341872, "dur": 4, + "args": { + "External id": 103523, + "cbid": 211, "correlation": 103523 + } + }, + { + "ph": "s", "id": 103523, "pid": 435, "tid": 435, "ts": 1776649793341872, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649793341901, "dur": 4, + "args": { + "External id": 103531, + "cbid": 135, "correlation": 103531 + } + }, + { + "ph": "f", "id": 103531, "pid": 435, "tid": 435, "ts": 1776649793341901, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649793341916, "dur": 5, + "args": { + "External id": 103536, + "cbid": 165, "correlation": 103536 + } + }, + { + "ph": "s", "id": 103536, "pid": 435, "tid": 435, "ts": 1776649793341916, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649793341925, "dur": 5, + "args": { + "External id": 103539, + "cbid": 139, "correlation": 103539 + } + }, + { + "ph": "f", "id": 103539, "pid": 435, "tid": 435, "ts": 1776649793341925, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793341981, "dur": 1, + "args": { + "External id": 13962, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103567, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103567, "pid": 0, "tid": 7, "ts": 1776649793341981, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341970, "dur": 8, + "args": { + "External id": 13962, + "cbid": 211, "correlation": 103567 + } + }, + { + "ph": "s", "id": 103567, "pid": 435, "tid": 435, "ts": 1776649793341970, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341985, "dur": 0, + "args": { + "External id": 103573, + "cbid": 200, "correlation": 103573 + } + }, + { + "ph": "f", "id": 103573, "pid": 435, "tid": 435, "ts": 1776649793341985, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793341986, "dur": 0, + "args": { + "External id": 103574, + "cbid": 273, "correlation": 103574 + } + }, + { + "ph": "f", "id": 103574, "pid": 435, "tid": 435, "ts": 1776649793341986, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793341996, "dur": 19, + "args": { + "External id": 103575, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103575, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103575, "pid": 0, "tid": 7, "ts": 1776649793341996, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793341987, "dur": 5, + "args": { + "External id": 103575, + "cbid": 211, "correlation": 103575 + } + }, + { + "ph": "s", "id": 103575, "pid": 435, "tid": 435, "ts": 1776649793341987, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793342033, "dur": 1, + "args": { + "External id": 13967, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103593, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103593, "pid": 0, "tid": 7, "ts": 1776649793342033, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342023, "dur": 6, + "args": { + "External id": 13967, + "cbid": 211, "correlation": 103593 + } + }, + { + "ph": "s", "id": 103593, "pid": 435, "tid": 435, "ts": 1776649793342023, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342036, "dur": 0, + "args": { + "External id": 103599, + "cbid": 200, "correlation": 103599 + } + }, + { + "ph": "f", "id": 103599, "pid": 435, "tid": 435, "ts": 1776649793342036, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342037, "dur": 0, + "args": { + "External id": 103600, + "cbid": 273, "correlation": 103600 + } + }, + { + "ph": "f", "id": 103600, "pid": 435, "tid": 435, "ts": 1776649793342037, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793342046, "dur": 19, + "args": { + "External id": 103601, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103601, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103601, "pid": 0, "tid": 7, "ts": 1776649793342046, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342038, "dur": 4, + "args": { + "External id": 103601, + "cbid": 211, "correlation": 103601 + } + }, + { + "ph": "s", "id": 103601, "pid": 435, "tid": 435, "ts": 1776649793342038, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793342082, "dur": 1, + "args": { + "External id": 13972, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103619, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103619, "pid": 0, "tid": 7, "ts": 1776649793342082, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342073, "dur": 6, + "args": { + "External id": 13972, + "cbid": 211, "correlation": 103619 + } + }, + { + "ph": "s", "id": 103619, "pid": 435, "tid": 435, "ts": 1776649793342073, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342086, "dur": 0, + "args": { + "External id": 103625, + "cbid": 200, "correlation": 103625 + } + }, + { + "ph": "f", "id": 103625, "pid": 435, "tid": 435, "ts": 1776649793342086, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342086, "dur": 0, + "args": { + "External id": 103626, + "cbid": 273, "correlation": 103626 + } + }, + { + "ph": "f", "id": 103626, "pid": 435, "tid": 435, "ts": 1776649793342086, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793342095, "dur": 19, + "args": { + "External id": 103627, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103627, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103627, "pid": 0, "tid": 7, "ts": 1776649793342095, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342087, "dur": 4, + "args": { + "External id": 103627, + "cbid": 211, "correlation": 103627 + } + }, + { + "ph": "s", "id": 103627, "pid": 435, "tid": 435, "ts": 1776649793342087, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793342130, "dur": 1, + "args": { + "External id": 13977, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103645, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103645, "pid": 0, "tid": 7, "ts": 1776649793342130, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342121, "dur": 6, + "args": { + "External id": 13977, + "cbid": 211, "correlation": 103645 + } + }, + { + "ph": "s", "id": 103645, "pid": 435, "tid": 435, "ts": 1776649793342121, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342133, "dur": 0, + "args": { + "External id": 103651, + "cbid": 200, "correlation": 103651 + } + }, + { + "ph": "f", "id": 103651, "pid": 435, "tid": 435, "ts": 1776649793342133, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342134, "dur": 0, + "args": { + "External id": 103652, + "cbid": 273, "correlation": 103652 + } + }, + { + "ph": "f", "id": 103652, "pid": 435, "tid": 435, "ts": 1776649793342134, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793342143, "dur": 19, + "args": { + "External id": 103653, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103653, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103653, "pid": 0, "tid": 7, "ts": 1776649793342143, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342135, "dur": 4, + "args": { + "External id": 103653, + "cbid": 211, "correlation": 103653 + } + }, + { + "ph": "s", "id": 103653, "pid": 435, "tid": 435, "ts": 1776649793342135, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793342178, "dur": 1, + "args": { + "External id": 13982, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103671, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103671, "pid": 0, "tid": 7, "ts": 1776649793342178, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342169, "dur": 6, + "args": { + "External id": 13982, + "cbid": 211, "correlation": 103671 + } + }, + { + "ph": "s", "id": 103671, "pid": 435, "tid": 435, "ts": 1776649793342169, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342182, "dur": 0, + "args": { + "External id": 103677, + "cbid": 200, "correlation": 103677 + } + }, + { + "ph": "f", "id": 103677, "pid": 435, "tid": 435, "ts": 1776649793342182, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342182, "dur": 0, + "args": { + "External id": 103678, + "cbid": 273, "correlation": 103678 + } + }, + { + "ph": "f", "id": 103678, "pid": 435, "tid": 435, "ts": 1776649793342182, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793342192, "dur": 19, + "args": { + "External id": 103679, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103679, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103679, "pid": 0, "tid": 7, "ts": 1776649793342192, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342184, "dur": 5, + "args": { + "External id": 103679, + "cbid": 211, "correlation": 103679 + } + }, + { + "ph": "s", "id": 103679, "pid": 435, "tid": 435, "ts": 1776649793342184, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793342227, "dur": 1, + "args": { + "External id": 13987, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103697, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103697, "pid": 0, "tid": 7, "ts": 1776649793342227, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342218, "dur": 6, + "args": { + "External id": 13987, + "cbid": 211, "correlation": 103697 + } + }, + { + "ph": "s", "id": 103697, "pid": 435, "tid": 435, "ts": 1776649793342218, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342230, "dur": 0, + "args": { + "External id": 103703, + "cbid": 200, "correlation": 103703 + } + }, + { + "ph": "f", "id": 103703, "pid": 435, "tid": 435, "ts": 1776649793342230, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342231, "dur": 0, + "args": { + "External id": 103704, + "cbid": 273, "correlation": 103704 + } + }, + { + "ph": "f", "id": 103704, "pid": 435, "tid": 435, "ts": 1776649793342231, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793342240, "dur": 19, + "args": { + "External id": 103705, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103705, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103705, "pid": 0, "tid": 7, "ts": 1776649793342240, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342232, "dur": 4, + "args": { + "External id": 103705, + "cbid": 211, "correlation": 103705 + } + }, + { + "ph": "s", "id": 103705, "pid": 435, "tid": 435, "ts": 1776649793342232, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793342278, "dur": 1, + "args": { + "External id": 13992, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103723, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103723, "pid": 0, "tid": 7, "ts": 1776649793342278, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342269, "dur": 7, + "args": { + "External id": 13992, + "cbid": 211, "correlation": 103723 + } + }, + { + "ph": "s", "id": 103723, "pid": 435, "tid": 435, "ts": 1776649793342269, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342283, "dur": 0, + "args": { + "External id": 103729, + "cbid": 200, "correlation": 103729 + } + }, + { + "ph": "f", "id": 103729, "pid": 435, "tid": 435, "ts": 1776649793342283, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342283, "dur": 0, + "args": { + "External id": 103730, + "cbid": 273, "correlation": 103730 + } + }, + { + "ph": "f", "id": 103730, "pid": 435, "tid": 435, "ts": 1776649793342283, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793342293, "dur": 19, + "args": { + "External id": 103731, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103731, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103731, "pid": 0, "tid": 7, "ts": 1776649793342293, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342285, "dur": 5, + "args": { + "External id": 103731, + "cbid": 211, "correlation": 103731 + } + }, + { + "ph": "s", "id": 103731, "pid": 435, "tid": 435, "ts": 1776649793342285, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793342328, "dur": 1, + "args": { + "External id": 13997, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103749, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103749, "pid": 0, "tid": 7, "ts": 1776649793342328, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342319, "dur": 6, + "args": { + "External id": 13997, + "cbid": 211, "correlation": 103749 + } + }, + { + "ph": "s", "id": 103749, "pid": 435, "tid": 435, "ts": 1776649793342319, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342331, "dur": 0, + "args": { + "External id": 103755, + "cbid": 200, "correlation": 103755 + } + }, + { + "ph": "f", "id": 103755, "pid": 435, "tid": 435, "ts": 1776649793342331, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342332, "dur": 0, + "args": { + "External id": 103756, + "cbid": 273, "correlation": 103756 + } + }, + { + "ph": "f", "id": 103756, "pid": 435, "tid": 435, "ts": 1776649793342332, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793342341, "dur": 19, + "args": { + "External id": 103757, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103757, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103757, "pid": 0, "tid": 7, "ts": 1776649793342341, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342333, "dur": 5, + "args": { + "External id": 103757, + "cbid": 211, "correlation": 103757 + } + }, + { + "ph": "s", "id": 103757, "pid": 435, "tid": 435, "ts": 1776649793342333, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793342376, "dur": 1, + "args": { + "External id": 14002, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103775, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103775, "pid": 0, "tid": 7, "ts": 1776649793342376, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342367, "dur": 6, + "args": { + "External id": 14002, + "cbid": 211, "correlation": 103775 + } + }, + { + "ph": "s", "id": 103775, "pid": 435, "tid": 435, "ts": 1776649793342367, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342379, "dur": 0, + "args": { + "External id": 103781, + "cbid": 200, "correlation": 103781 + } + }, + { + "ph": "f", "id": 103781, "pid": 435, "tid": 435, "ts": 1776649793342379, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342380, "dur": 0, + "args": { + "External id": 103782, + "cbid": 273, "correlation": 103782 + } + }, + { + "ph": "f", "id": 103782, "pid": 435, "tid": 435, "ts": 1776649793342380, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793342389, "dur": 19, + "args": { + "External id": 103783, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103783, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103783, "pid": 0, "tid": 7, "ts": 1776649793342389, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342381, "dur": 5, + "args": { + "External id": 103783, + "cbid": 211, "correlation": 103783 + } + }, + { + "ph": "s", "id": 103783, "pid": 435, "tid": 435, "ts": 1776649793342381, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793342424, "dur": 1, + "args": { + "External id": 14007, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103801, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103801, "pid": 0, "tid": 7, "ts": 1776649793342424, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342415, "dur": 6, + "args": { + "External id": 14007, + "cbid": 211, "correlation": 103801 + } + }, + { + "ph": "s", "id": 103801, "pid": 435, "tid": 435, "ts": 1776649793342415, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342427, "dur": 0, + "args": { + "External id": 103807, + "cbid": 200, "correlation": 103807 + } + }, + { + "ph": "f", "id": 103807, "pid": 435, "tid": 435, "ts": 1776649793342427, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342428, "dur": 0, + "args": { + "External id": 103808, + "cbid": 273, "correlation": 103808 + } + }, + { + "ph": "f", "id": 103808, "pid": 435, "tid": 435, "ts": 1776649793342428, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793342437, "dur": 19, + "args": { + "External id": 103809, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103809, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103809, "pid": 0, "tid": 7, "ts": 1776649793342437, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342429, "dur": 4, + "args": { + "External id": 103809, + "cbid": 211, "correlation": 103809 + } + }, + { + "ph": "s", "id": 103809, "pid": 435, "tid": 435, "ts": 1776649793342429, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649793342450, "dur": 10, + "args": { + "External id": 103812, + "cbid": 165, "correlation": 103812 + } + }, + { + "ph": "s", "id": 103812, "pid": 435, "tid": 435, "ts": 1776649793342450, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649793342481, "dur": 3, + "args": { + "External id": 103821, + "cbid": 135, "correlation": 103821 + } + }, + { + "ph": "f", "id": 103821, "pid": 435, "tid": 435, "ts": 1776649793342481, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793342535, "dur": 1, + "args": { + "External id": 14012, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103841, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103841, "pid": 0, "tid": 7, "ts": 1776649793342535, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342524, "dur": 9, + "args": { + "External id": 14012, + "cbid": 211, "correlation": 103841 + } + }, + { + "ph": "s", "id": 103841, "pid": 435, "tid": 435, "ts": 1776649793342524, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342539, "dur": 0, + "args": { + "External id": 103847, + "cbid": 200, "correlation": 103847 + } + }, + { + "ph": "f", "id": 103847, "pid": 435, "tid": 435, "ts": 1776649793342539, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342540, "dur": 1, + "args": { + "External id": 103848, + "cbid": 273, "correlation": 103848 + } + }, + { + "ph": "f", "id": 103848, "pid": 435, "tid": 435, "ts": 1776649793342540, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793342550, "dur": 19, + "args": { + "External id": 103849, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103849, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103849, "pid": 0, "tid": 7, "ts": 1776649793342550, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342542, "dur": 5, + "args": { + "External id": 103849, + "cbid": 211, "correlation": 103849 + } + }, + { + "ph": "s", "id": 103849, "pid": 435, "tid": 435, "ts": 1776649793342542, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793342588, "dur": 1, + "args": { + "External id": 14017, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103867, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103867, "pid": 0, "tid": 7, "ts": 1776649793342588, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342578, "dur": 6, + "args": { + "External id": 14017, + "cbid": 211, "correlation": 103867 + } + }, + { + "ph": "s", "id": 103867, "pid": 435, "tid": 435, "ts": 1776649793342578, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342591, "dur": 0, + "args": { + "External id": 103873, + "cbid": 200, "correlation": 103873 + } + }, + { + "ph": "f", "id": 103873, "pid": 435, "tid": 435, "ts": 1776649793342591, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342591, "dur": 0, + "args": { + "External id": 103874, + "cbid": 273, "correlation": 103874 + } + }, + { + "ph": "f", "id": 103874, "pid": 435, "tid": 435, "ts": 1776649793342591, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793342600, "dur": 19, + "args": { + "External id": 103875, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103875, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103875, "pid": 0, "tid": 7, "ts": 1776649793342600, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342593, "dur": 4, + "args": { + "External id": 103875, + "cbid": 211, "correlation": 103875 + } + }, + { + "ph": "s", "id": 103875, "pid": 435, "tid": 435, "ts": 1776649793342593, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793342636, "dur": 1, + "args": { + "External id": 14022, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103893, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103893, "pid": 0, "tid": 7, "ts": 1776649793342636, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342627, "dur": 6, + "args": { + "External id": 14022, + "cbid": 211, "correlation": 103893 + } + }, + { + "ph": "s", "id": 103893, "pid": 435, "tid": 435, "ts": 1776649793342627, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342639, "dur": 0, + "args": { + "External id": 103899, + "cbid": 200, "correlation": 103899 + } + }, + { + "ph": "f", "id": 103899, "pid": 435, "tid": 435, "ts": 1776649793342639, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342640, "dur": 0, + "args": { + "External id": 103900, + "cbid": 273, "correlation": 103900 + } + }, + { + "ph": "f", "id": 103900, "pid": 435, "tid": 435, "ts": 1776649793342640, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793342649, "dur": 19, + "args": { + "External id": 103901, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103901, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103901, "pid": 0, "tid": 7, "ts": 1776649793342649, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342641, "dur": 5, + "args": { + "External id": 103901, + "cbid": 211, "correlation": 103901 + } + }, + { + "ph": "s", "id": 103901, "pid": 435, "tid": 435, "ts": 1776649793342641, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793342685, "dur": 1, + "args": { + "External id": 14027, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103919, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103919, "pid": 0, "tid": 7, "ts": 1776649793342685, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342676, "dur": 6, + "args": { + "External id": 14027, + "cbid": 211, "correlation": 103919 + } + }, + { + "ph": "s", "id": 103919, "pid": 435, "tid": 435, "ts": 1776649793342676, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342688, "dur": 0, + "args": { + "External id": 103925, + "cbid": 200, "correlation": 103925 + } + }, + { + "ph": "f", "id": 103925, "pid": 435, "tid": 435, "ts": 1776649793342688, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342689, "dur": 0, + "args": { + "External id": 103926, + "cbid": 273, "correlation": 103926 + } + }, + { + "ph": "f", "id": 103926, "pid": 435, "tid": 435, "ts": 1776649793342689, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793342698, "dur": 19, + "args": { + "External id": 103927, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103927, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103927, "pid": 0, "tid": 7, "ts": 1776649793342698, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342690, "dur": 5, + "args": { + "External id": 103927, + "cbid": 211, "correlation": 103927 + } + }, + { + "ph": "s", "id": 103927, "pid": 435, "tid": 435, "ts": 1776649793342690, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793342733, "dur": 1, + "args": { + "External id": 14032, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103945, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103945, "pid": 0, "tid": 7, "ts": 1776649793342733, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342724, "dur": 6, + "args": { + "External id": 14032, + "cbid": 211, "correlation": 103945 + } + }, + { + "ph": "s", "id": 103945, "pid": 435, "tid": 435, "ts": 1776649793342724, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342736, "dur": 0, + "args": { + "External id": 103951, + "cbid": 200, "correlation": 103951 + } + }, + { + "ph": "f", "id": 103951, "pid": 435, "tid": 435, "ts": 1776649793342736, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342736, "dur": 0, + "args": { + "External id": 103952, + "cbid": 273, "correlation": 103952 + } + }, + { + "ph": "f", "id": 103952, "pid": 435, "tid": 435, "ts": 1776649793342736, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793342745, "dur": 19, + "args": { + "External id": 103953, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103953, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103953, "pid": 0, "tid": 7, "ts": 1776649793342745, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342738, "dur": 4, + "args": { + "External id": 103953, + "cbid": 211, "correlation": 103953 + } + }, + { + "ph": "s", "id": 103953, "pid": 435, "tid": 435, "ts": 1776649793342738, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793342781, "dur": 1, + "args": { + "External id": 14037, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103971, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103971, "pid": 0, "tid": 7, "ts": 1776649793342781, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342771, "dur": 6, + "args": { + "External id": 14037, + "cbid": 211, "correlation": 103971 + } + }, + { + "ph": "s", "id": 103971, "pid": 435, "tid": 435, "ts": 1776649793342771, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342783, "dur": 0, + "args": { + "External id": 103977, + "cbid": 200, "correlation": 103977 + } + }, + { + "ph": "f", "id": 103977, "pid": 435, "tid": 435, "ts": 1776649793342783, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342784, "dur": 0, + "args": { + "External id": 103978, + "cbid": 273, "correlation": 103978 + } + }, + { + "ph": "f", "id": 103978, "pid": 435, "tid": 435, "ts": 1776649793342784, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793342793, "dur": 19, + "args": { + "External id": 103979, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103979, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103979, "pid": 0, "tid": 7, "ts": 1776649793342793, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342785, "dur": 5, + "args": { + "External id": 103979, + "cbid": 211, "correlation": 103979 + } + }, + { + "ph": "s", "id": 103979, "pid": 435, "tid": 435, "ts": 1776649793342785, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793342829, "dur": 1, + "args": { + "External id": 14042, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 103997, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 103997, "pid": 0, "tid": 7, "ts": 1776649793342829, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342819, "dur": 6, + "args": { + "External id": 14042, + "cbid": 211, "correlation": 103997 + } + }, + { + "ph": "s", "id": 103997, "pid": 435, "tid": 435, "ts": 1776649793342819, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342832, "dur": 0, + "args": { + "External id": 104003, + "cbid": 200, "correlation": 104003 + } + }, + { + "ph": "f", "id": 104003, "pid": 435, "tid": 435, "ts": 1776649793342832, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342832, "dur": 0, + "args": { + "External id": 104004, + "cbid": 273, "correlation": 104004 + } + }, + { + "ph": "f", "id": 104004, "pid": 435, "tid": 435, "ts": 1776649793342832, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793342841, "dur": 19, + "args": { + "External id": 104005, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104005, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104005, "pid": 0, "tid": 7, "ts": 1776649793342841, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342833, "dur": 4, + "args": { + "External id": 104005, + "cbid": 211, "correlation": 104005 + } + }, + { + "ph": "s", "id": 104005, "pid": 435, "tid": 435, "ts": 1776649793342833, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793342878, "dur": 1, + "args": { + "External id": 14047, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104023, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104023, "pid": 0, "tid": 7, "ts": 1776649793342878, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342868, "dur": 7, + "args": { + "External id": 14047, + "cbid": 211, "correlation": 104023 + } + }, + { + "ph": "s", "id": 104023, "pid": 435, "tid": 435, "ts": 1776649793342868, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342881, "dur": 0, + "args": { + "External id": 104029, + "cbid": 200, "correlation": 104029 + } + }, + { + "ph": "f", "id": 104029, "pid": 435, "tid": 435, "ts": 1776649793342881, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342882, "dur": 0, + "args": { + "External id": 104030, + "cbid": 273, "correlation": 104030 + } + }, + { + "ph": "f", "id": 104030, "pid": 435, "tid": 435, "ts": 1776649793342882, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793342892, "dur": 19, + "args": { + "External id": 104031, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104031, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104031, "pid": 0, "tid": 7, "ts": 1776649793342892, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342883, "dur": 5, + "args": { + "External id": 104031, + "cbid": 211, "correlation": 104031 + } + }, + { + "ph": "s", "id": 104031, "pid": 435, "tid": 435, "ts": 1776649793342883, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793342928, "dur": 1, + "args": { + "External id": 14052, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104049, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104049, "pid": 0, "tid": 7, "ts": 1776649793342928, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342919, "dur": 6, + "args": { + "External id": 14052, + "cbid": 211, "correlation": 104049 + } + }, + { + "ph": "s", "id": 104049, "pid": 435, "tid": 435, "ts": 1776649793342919, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342931, "dur": 0, + "args": { + "External id": 104055, + "cbid": 200, "correlation": 104055 + } + }, + { + "ph": "f", "id": 104055, "pid": 435, "tid": 435, "ts": 1776649793342931, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342932, "dur": 0, + "args": { + "External id": 104056, + "cbid": 273, "correlation": 104056 + } + }, + { + "ph": "f", "id": 104056, "pid": 435, "tid": 435, "ts": 1776649793342932, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793342941, "dur": 19, + "args": { + "External id": 104057, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104057, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104057, "pid": 0, "tid": 7, "ts": 1776649793342941, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342933, "dur": 4, + "args": { + "External id": 104057, + "cbid": 211, "correlation": 104057 + } + }, + { + "ph": "s", "id": 104057, "pid": 435, "tid": 435, "ts": 1776649793342933, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793342977, "dur": 1, + "args": { + "External id": 14057, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104075, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104075, "pid": 0, "tid": 7, "ts": 1776649793342977, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342967, "dur": 6, + "args": { + "External id": 14057, + "cbid": 211, "correlation": 104075 + } + }, + { + "ph": "s", "id": 104075, "pid": 435, "tid": 435, "ts": 1776649793342967, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342979, "dur": 0, + "args": { + "External id": 104081, + "cbid": 200, "correlation": 104081 + } + }, + { + "ph": "f", "id": 104081, "pid": 435, "tid": 435, "ts": 1776649793342979, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793342980, "dur": 0, + "args": { + "External id": 104082, + "cbid": 273, "correlation": 104082 + } + }, + { + "ph": "f", "id": 104082, "pid": 435, "tid": 435, "ts": 1776649793342980, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793342989, "dur": 19, + "args": { + "External id": 104083, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104083, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104083, "pid": 0, "tid": 7, "ts": 1776649793342989, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793342981, "dur": 4, + "args": { + "External id": 104083, + "cbid": 211, "correlation": 104083 + } + }, + { + "ph": "s", "id": 104083, "pid": 435, "tid": 435, "ts": 1776649793342981, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793343024, "dur": 1, + "args": { + "External id": 14062, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104101, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104101, "pid": 0, "tid": 7, "ts": 1776649793343024, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343015, "dur": 6, + "args": { + "External id": 14062, + "cbid": 211, "correlation": 104101 + } + }, + { + "ph": "s", "id": 104101, "pid": 435, "tid": 435, "ts": 1776649793343015, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343027, "dur": 0, + "args": { + "External id": 104107, + "cbid": 200, "correlation": 104107 + } + }, + { + "ph": "f", "id": 104107, "pid": 435, "tid": 435, "ts": 1776649793343027, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343027, "dur": 0, + "args": { + "External id": 104108, + "cbid": 273, "correlation": 104108 + } + }, + { + "ph": "f", "id": 104108, "pid": 435, "tid": 435, "ts": 1776649793343027, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793343036, "dur": 19, + "args": { + "External id": 104109, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104109, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104109, "pid": 0, "tid": 7, "ts": 1776649793343036, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343028, "dur": 5, + "args": { + "External id": 104109, + "cbid": 211, "correlation": 104109 + } + }, + { + "ph": "s", "id": 104109, "pid": 435, "tid": 435, "ts": 1776649793343028, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793343072, "dur": 1, + "args": { + "External id": 14067, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104127, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104127, "pid": 0, "tid": 7, "ts": 1776649793343072, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343063, "dur": 6, + "args": { + "External id": 14067, + "cbid": 211, "correlation": 104127 + } + }, + { + "ph": "s", "id": 104127, "pid": 435, "tid": 435, "ts": 1776649793343063, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343075, "dur": 0, + "args": { + "External id": 104133, + "cbid": 200, "correlation": 104133 + } + }, + { + "ph": "f", "id": 104133, "pid": 435, "tid": 435, "ts": 1776649793343075, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343076, "dur": 0, + "args": { + "External id": 104134, + "cbid": 273, "correlation": 104134 + } + }, + { + "ph": "f", "id": 104134, "pid": 435, "tid": 435, "ts": 1776649793343076, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793343085, "dur": 19, + "args": { + "External id": 104135, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104135, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104135, "pid": 0, "tid": 7, "ts": 1776649793343085, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343077, "dur": 5, + "args": { + "External id": 104135, + "cbid": 211, "correlation": 104135 + } + }, + { + "ph": "s", "id": 104135, "pid": 435, "tid": 435, "ts": 1776649793343077, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793343119, "dur": 1, + "args": { + "External id": 14072, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104153, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104153, "pid": 0, "tid": 7, "ts": 1776649793343119, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343111, "dur": 6, + "args": { + "External id": 14072, + "cbid": 211, "correlation": 104153 + } + }, + { + "ph": "s", "id": 104153, "pid": 435, "tid": 435, "ts": 1776649793343111, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343122, "dur": 0, + "args": { + "External id": 104159, + "cbid": 200, "correlation": 104159 + } + }, + { + "ph": "f", "id": 104159, "pid": 435, "tid": 435, "ts": 1776649793343122, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343123, "dur": 0, + "args": { + "External id": 104160, + "cbid": 273, "correlation": 104160 + } + }, + { + "ph": "f", "id": 104160, "pid": 435, "tid": 435, "ts": 1776649793343123, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793343132, "dur": 19, + "args": { + "External id": 104161, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104161, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104161, "pid": 0, "tid": 7, "ts": 1776649793343132, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343124, "dur": 4, + "args": { + "External id": 104161, + "cbid": 211, "correlation": 104161 + } + }, + { + "ph": "s", "id": 104161, "pid": 435, "tid": 435, "ts": 1776649793343124, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793343167, "dur": 1, + "args": { + "External id": 14077, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104179, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104179, "pid": 0, "tid": 7, "ts": 1776649793343167, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343158, "dur": 6, + "args": { + "External id": 14077, + "cbid": 211, "correlation": 104179 + } + }, + { + "ph": "s", "id": 104179, "pid": 435, "tid": 435, "ts": 1776649793343158, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343170, "dur": 0, + "args": { + "External id": 104185, + "cbid": 200, "correlation": 104185 + } + }, + { + "ph": "f", "id": 104185, "pid": 435, "tid": 435, "ts": 1776649793343170, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343171, "dur": 0, + "args": { + "External id": 104186, + "cbid": 273, "correlation": 104186 + } + }, + { + "ph": "f", "id": 104186, "pid": 435, "tid": 435, "ts": 1776649793343171, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793343180, "dur": 19, + "args": { + "External id": 104187, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104187, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104187, "pid": 0, "tid": 7, "ts": 1776649793343180, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343172, "dur": 5, + "args": { + "External id": 104187, + "cbid": 211, "correlation": 104187 + } + }, + { + "ph": "s", "id": 104187, "pid": 435, "tid": 435, "ts": 1776649793343172, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793343215, "dur": 1, + "args": { + "External id": 14082, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104205, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104205, "pid": 0, "tid": 7, "ts": 1776649793343215, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343206, "dur": 6, + "args": { + "External id": 14082, + "cbid": 211, "correlation": 104205 + } + }, + { + "ph": "s", "id": 104205, "pid": 435, "tid": 435, "ts": 1776649793343206, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343218, "dur": 0, + "args": { + "External id": 104211, + "cbid": 200, "correlation": 104211 + } + }, + { + "ph": "f", "id": 104211, "pid": 435, "tid": 435, "ts": 1776649793343218, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343219, "dur": 0, + "args": { + "External id": 104212, + "cbid": 273, "correlation": 104212 + } + }, + { + "ph": "f", "id": 104212, "pid": 435, "tid": 435, "ts": 1776649793343219, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793343228, "dur": 19, + "args": { + "External id": 104213, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104213, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104213, "pid": 0, "tid": 7, "ts": 1776649793343228, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343220, "dur": 4, + "args": { + "External id": 104213, + "cbid": 211, "correlation": 104213 + } + }, + { + "ph": "s", "id": 104213, "pid": 435, "tid": 435, "ts": 1776649793343220, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793343263, "dur": 1, + "args": { + "External id": 14087, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104231, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104231, "pid": 0, "tid": 7, "ts": 1776649793343263, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343254, "dur": 6, + "args": { + "External id": 14087, + "cbid": 211, "correlation": 104231 + } + }, + { + "ph": "s", "id": 104231, "pid": 435, "tid": 435, "ts": 1776649793343254, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343266, "dur": 0, + "args": { + "External id": 104237, + "cbid": 200, "correlation": 104237 + } + }, + { + "ph": "f", "id": 104237, "pid": 435, "tid": 435, "ts": 1776649793343266, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343266, "dur": 0, + "args": { + "External id": 104238, + "cbid": 273, "correlation": 104238 + } + }, + { + "ph": "f", "id": 104238, "pid": 435, "tid": 435, "ts": 1776649793343266, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793343275, "dur": 19, + "args": { + "External id": 104239, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104239, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104239, "pid": 0, "tid": 7, "ts": 1776649793343275, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343267, "dur": 4, + "args": { + "External id": 104239, + "cbid": 211, "correlation": 104239 + } + }, + { + "ph": "s", "id": 104239, "pid": 435, "tid": 435, "ts": 1776649793343267, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793343311, "dur": 1, + "args": { + "External id": 14092, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104257, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104257, "pid": 0, "tid": 7, "ts": 1776649793343311, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343301, "dur": 6, + "args": { + "External id": 14092, + "cbid": 211, "correlation": 104257 + } + }, + { + "ph": "s", "id": 104257, "pid": 435, "tid": 435, "ts": 1776649793343301, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343314, "dur": 0, + "args": { + "External id": 104263, + "cbid": 200, "correlation": 104263 + } + }, + { + "ph": "f", "id": 104263, "pid": 435, "tid": 435, "ts": 1776649793343314, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343314, "dur": 0, + "args": { + "External id": 104264, + "cbid": 273, "correlation": 104264 + } + }, + { + "ph": "f", "id": 104264, "pid": 435, "tid": 435, "ts": 1776649793343314, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793343323, "dur": 19, + "args": { + "External id": 104265, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104265, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104265, "pid": 0, "tid": 7, "ts": 1776649793343323, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343316, "dur": 4, + "args": { + "External id": 104265, + "cbid": 211, "correlation": 104265 + } + }, + { + "ph": "s", "id": 104265, "pid": 435, "tid": 435, "ts": 1776649793343316, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793343359, "dur": 1, + "args": { + "External id": 14097, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104283, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104283, "pid": 0, "tid": 7, "ts": 1776649793343359, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343349, "dur": 6, + "args": { + "External id": 14097, + "cbid": 211, "correlation": 104283 + } + }, + { + "ph": "s", "id": 104283, "pid": 435, "tid": 435, "ts": 1776649793343349, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343362, "dur": 0, + "args": { + "External id": 104289, + "cbid": 200, "correlation": 104289 + } + }, + { + "ph": "f", "id": 104289, "pid": 435, "tid": 435, "ts": 1776649793343362, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343362, "dur": 0, + "args": { + "External id": 104290, + "cbid": 273, "correlation": 104290 + } + }, + { + "ph": "f", "id": 104290, "pid": 435, "tid": 435, "ts": 1776649793343362, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793343371, "dur": 19, + "args": { + "External id": 104291, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104291, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104291, "pid": 0, "tid": 7, "ts": 1776649793343371, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343363, "dur": 5, + "args": { + "External id": 104291, + "cbid": 211, "correlation": 104291 + } + }, + { + "ph": "s", "id": 104291, "pid": 435, "tid": 435, "ts": 1776649793343363, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793343406, "dur": 1, + "args": { + "External id": 14102, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104309, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104309, "pid": 0, "tid": 7, "ts": 1776649793343406, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343398, "dur": 6, + "args": { + "External id": 14102, + "cbid": 211, "correlation": 104309 + } + }, + { + "ph": "s", "id": 104309, "pid": 435, "tid": 435, "ts": 1776649793343398, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343409, "dur": 0, + "args": { + "External id": 104315, + "cbid": 200, "correlation": 104315 + } + }, + { + "ph": "f", "id": 104315, "pid": 435, "tid": 435, "ts": 1776649793343409, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343410, "dur": 0, + "args": { + "External id": 104316, + "cbid": 273, "correlation": 104316 + } + }, + { + "ph": "f", "id": 104316, "pid": 435, "tid": 435, "ts": 1776649793343410, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793343419, "dur": 19, + "args": { + "External id": 104317, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104317, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104317, "pid": 0, "tid": 7, "ts": 1776649793343419, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343411, "dur": 4, + "args": { + "External id": 104317, + "cbid": 211, "correlation": 104317 + } + }, + { + "ph": "s", "id": 104317, "pid": 435, "tid": 435, "ts": 1776649793343411, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793343454, "dur": 1, + "args": { + "External id": 14107, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104335, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104335, "pid": 0, "tid": 7, "ts": 1776649793343454, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343445, "dur": 6, + "args": { + "External id": 14107, + "cbid": 211, "correlation": 104335 + } + }, + { + "ph": "s", "id": 104335, "pid": 435, "tid": 435, "ts": 1776649793343445, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343457, "dur": 0, + "args": { + "External id": 104341, + "cbid": 200, "correlation": 104341 + } + }, + { + "ph": "f", "id": 104341, "pid": 435, "tid": 435, "ts": 1776649793343457, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343458, "dur": 0, + "args": { + "External id": 104342, + "cbid": 273, "correlation": 104342 + } + }, + { + "ph": "f", "id": 104342, "pid": 435, "tid": 435, "ts": 1776649793343458, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793343467, "dur": 19, + "args": { + "External id": 104343, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104343, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104343, "pid": 0, "tid": 7, "ts": 1776649793343467, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343459, "dur": 4, + "args": { + "External id": 104343, + "cbid": 211, "correlation": 104343 + } + }, + { + "ph": "s", "id": 104343, "pid": 435, "tid": 435, "ts": 1776649793343459, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793343502, "dur": 1, + "args": { + "External id": 14112, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104361, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104361, "pid": 0, "tid": 7, "ts": 1776649793343502, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343493, "dur": 6, + "args": { + "External id": 14112, + "cbid": 211, "correlation": 104361 + } + }, + { + "ph": "s", "id": 104361, "pid": 435, "tid": 435, "ts": 1776649793343493, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343505, "dur": 0, + "args": { + "External id": 104367, + "cbid": 200, "correlation": 104367 + } + }, + { + "ph": "f", "id": 104367, "pid": 435, "tid": 435, "ts": 1776649793343505, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343506, "dur": 0, + "args": { + "External id": 104368, + "cbid": 273, "correlation": 104368 + } + }, + { + "ph": "f", "id": 104368, "pid": 435, "tid": 435, "ts": 1776649793343506, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793343515, "dur": 19, + "args": { + "External id": 104369, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104369, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104369, "pid": 0, "tid": 7, "ts": 1776649793343515, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343507, "dur": 4, + "args": { + "External id": 104369, + "cbid": 211, "correlation": 104369 + } + }, + { + "ph": "s", "id": 104369, "pid": 435, "tid": 435, "ts": 1776649793343507, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793343550, "dur": 1, + "args": { + "External id": 14117, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104387, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104387, "pid": 0, "tid": 7, "ts": 1776649793343550, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343541, "dur": 6, + "args": { + "External id": 14117, + "cbid": 211, "correlation": 104387 + } + }, + { + "ph": "s", "id": 104387, "pid": 435, "tid": 435, "ts": 1776649793343541, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343553, "dur": 0, + "args": { + "External id": 104393, + "cbid": 200, "correlation": 104393 + } + }, + { + "ph": "f", "id": 104393, "pid": 435, "tid": 435, "ts": 1776649793343553, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343553, "dur": 0, + "args": { + "External id": 104394, + "cbid": 273, "correlation": 104394 + } + }, + { + "ph": "f", "id": 104394, "pid": 435, "tid": 435, "ts": 1776649793343553, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793343562, "dur": 19, + "args": { + "External id": 104395, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104395, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104395, "pid": 0, "tid": 7, "ts": 1776649793343562, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343555, "dur": 4, + "args": { + "External id": 104395, + "cbid": 211, "correlation": 104395 + } + }, + { + "ph": "s", "id": 104395, "pid": 435, "tid": 435, "ts": 1776649793343555, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793343598, "dur": 1, + "args": { + "External id": 14122, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104413, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104413, "pid": 0, "tid": 7, "ts": 1776649793343598, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343588, "dur": 6, + "args": { + "External id": 14122, + "cbid": 211, "correlation": 104413 + } + }, + { + "ph": "s", "id": 104413, "pid": 435, "tid": 435, "ts": 1776649793343588, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343600, "dur": 0, + "args": { + "External id": 104419, + "cbid": 200, "correlation": 104419 + } + }, + { + "ph": "f", "id": 104419, "pid": 435, "tid": 435, "ts": 1776649793343600, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343601, "dur": 0, + "args": { + "External id": 104420, + "cbid": 273, "correlation": 104420 + } + }, + { + "ph": "f", "id": 104420, "pid": 435, "tid": 435, "ts": 1776649793343601, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793343610, "dur": 19, + "args": { + "External id": 104421, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104421, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104421, "pid": 0, "tid": 7, "ts": 1776649793343610, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343602, "dur": 4, + "args": { + "External id": 104421, + "cbid": 211, "correlation": 104421 + } + }, + { + "ph": "s", "id": 104421, "pid": 435, "tid": 435, "ts": 1776649793343602, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793343646, "dur": 1, + "args": { + "External id": 14127, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104439, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104439, "pid": 0, "tid": 7, "ts": 1776649793343646, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343637, "dur": 6, + "args": { + "External id": 14127, + "cbid": 211, "correlation": 104439 + } + }, + { + "ph": "s", "id": 104439, "pid": 435, "tid": 435, "ts": 1776649793343637, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343649, "dur": 0, + "args": { + "External id": 104445, + "cbid": 200, "correlation": 104445 + } + }, + { + "ph": "f", "id": 104445, "pid": 435, "tid": 435, "ts": 1776649793343649, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343649, "dur": 0, + "args": { + "External id": 104446, + "cbid": 273, "correlation": 104446 + } + }, + { + "ph": "f", "id": 104446, "pid": 435, "tid": 435, "ts": 1776649793343649, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793343658, "dur": 19, + "args": { + "External id": 104447, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104447, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104447, "pid": 0, "tid": 7, "ts": 1776649793343658, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343650, "dur": 5, + "args": { + "External id": 104447, + "cbid": 211, "correlation": 104447 + } + }, + { + "ph": "s", "id": 104447, "pid": 435, "tid": 435, "ts": 1776649793343650, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793343693, "dur": 1, + "args": { + "External id": 14132, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104465, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104465, "pid": 0, "tid": 7, "ts": 1776649793343693, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343684, "dur": 6, + "args": { + "External id": 14132, + "cbid": 211, "correlation": 104465 + } + }, + { + "ph": "s", "id": 104465, "pid": 435, "tid": 435, "ts": 1776649793343684, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343696, "dur": 0, + "args": { + "External id": 104471, + "cbid": 200, "correlation": 104471 + } + }, + { + "ph": "f", "id": 104471, "pid": 435, "tid": 435, "ts": 1776649793343696, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343697, "dur": 0, + "args": { + "External id": 104472, + "cbid": 273, "correlation": 104472 + } + }, + { + "ph": "f", "id": 104472, "pid": 435, "tid": 435, "ts": 1776649793343697, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793343706, "dur": 19, + "args": { + "External id": 104473, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104473, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104473, "pid": 0, "tid": 7, "ts": 1776649793343706, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343698, "dur": 4, + "args": { + "External id": 104473, + "cbid": 211, "correlation": 104473 + } + }, + { + "ph": "s", "id": 104473, "pid": 435, "tid": 435, "ts": 1776649793343698, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793343741, "dur": 1, + "args": { + "External id": 14137, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104491, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104491, "pid": 0, "tid": 7, "ts": 1776649793343741, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343732, "dur": 6, + "args": { + "External id": 14137, + "cbid": 211, "correlation": 104491 + } + }, + { + "ph": "s", "id": 104491, "pid": 435, "tid": 435, "ts": 1776649793343732, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343744, "dur": 0, + "args": { + "External id": 104497, + "cbid": 200, "correlation": 104497 + } + }, + { + "ph": "f", "id": 104497, "pid": 435, "tid": 435, "ts": 1776649793343744, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343745, "dur": 0, + "args": { + "External id": 104498, + "cbid": 273, "correlation": 104498 + } + }, + { + "ph": "f", "id": 104498, "pid": 435, "tid": 435, "ts": 1776649793343745, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793343753, "dur": 19, + "args": { + "External id": 104499, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104499, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104499, "pid": 0, "tid": 7, "ts": 1776649793343753, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343746, "dur": 5, + "args": { + "External id": 104499, + "cbid": 211, "correlation": 104499 + } + }, + { + "ph": "s", "id": 104499, "pid": 435, "tid": 435, "ts": 1776649793343746, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793343789, "dur": 1, + "args": { + "External id": 14142, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104517, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104517, "pid": 0, "tid": 7, "ts": 1776649793343789, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343780, "dur": 6, + "args": { + "External id": 14142, + "cbid": 211, "correlation": 104517 + } + }, + { + "ph": "s", "id": 104517, "pid": 435, "tid": 435, "ts": 1776649793343780, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343792, "dur": 0, + "args": { + "External id": 104523, + "cbid": 200, "correlation": 104523 + } + }, + { + "ph": "f", "id": 104523, "pid": 435, "tid": 435, "ts": 1776649793343792, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343793, "dur": 0, + "args": { + "External id": 104524, + "cbid": 273, "correlation": 104524 + } + }, + { + "ph": "f", "id": 104524, "pid": 435, "tid": 435, "ts": 1776649793343793, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793343802, "dur": 19, + "args": { + "External id": 104525, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104525, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104525, "pid": 0, "tid": 7, "ts": 1776649793343802, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343794, "dur": 4, + "args": { + "External id": 104525, + "cbid": 211, "correlation": 104525 + } + }, + { + "ph": "s", "id": 104525, "pid": 435, "tid": 435, "ts": 1776649793343794, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793343837, "dur": 1, + "args": { + "External id": 14147, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104543, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104543, "pid": 0, "tid": 7, "ts": 1776649793343837, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343828, "dur": 6, + "args": { + "External id": 14147, + "cbid": 211, "correlation": 104543 + } + }, + { + "ph": "s", "id": 104543, "pid": 435, "tid": 435, "ts": 1776649793343828, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343840, "dur": 0, + "args": { + "External id": 104549, + "cbid": 200, "correlation": 104549 + } + }, + { + "ph": "f", "id": 104549, "pid": 435, "tid": 435, "ts": 1776649793343840, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343841, "dur": 0, + "args": { + "External id": 104550, + "cbid": 273, "correlation": 104550 + } + }, + { + "ph": "f", "id": 104550, "pid": 435, "tid": 435, "ts": 1776649793343841, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793343850, "dur": 19, + "args": { + "External id": 104551, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104551, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104551, "pid": 0, "tid": 7, "ts": 1776649793343850, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343842, "dur": 4, + "args": { + "External id": 104551, + "cbid": 211, "correlation": 104551 + } + }, + { + "ph": "s", "id": 104551, "pid": 435, "tid": 435, "ts": 1776649793343842, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793343885, "dur": 1, + "args": { + "External id": 14152, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104569, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104569, "pid": 0, "tid": 7, "ts": 1776649793343885, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343876, "dur": 6, + "args": { + "External id": 14152, + "cbid": 211, "correlation": 104569 + } + }, + { + "ph": "s", "id": 104569, "pid": 435, "tid": 435, "ts": 1776649793343876, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343888, "dur": 0, + "args": { + "External id": 104575, + "cbid": 200, "correlation": 104575 + } + }, + { + "ph": "f", "id": 104575, "pid": 435, "tid": 435, "ts": 1776649793343888, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343888, "dur": 0, + "args": { + "External id": 104576, + "cbid": 273, "correlation": 104576 + } + }, + { + "ph": "f", "id": 104576, "pid": 435, "tid": 435, "ts": 1776649793343888, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793343897, "dur": 19, + "args": { + "External id": 104577, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104577, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104577, "pid": 0, "tid": 7, "ts": 1776649793343897, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343889, "dur": 4, + "args": { + "External id": 104577, + "cbid": 211, "correlation": 104577 + } + }, + { + "ph": "s", "id": 104577, "pid": 435, "tid": 435, "ts": 1776649793343889, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793343933, "dur": 1, + "args": { + "External id": 14157, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104595, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104595, "pid": 0, "tid": 7, "ts": 1776649793343933, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343923, "dur": 6, + "args": { + "External id": 14157, + "cbid": 211, "correlation": 104595 + } + }, + { + "ph": "s", "id": 104595, "pid": 435, "tid": 435, "ts": 1776649793343923, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343936, "dur": 0, + "args": { + "External id": 104601, + "cbid": 200, "correlation": 104601 + } + }, + { + "ph": "f", "id": 104601, "pid": 435, "tid": 435, "ts": 1776649793343936, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343936, "dur": 0, + "args": { + "External id": 104602, + "cbid": 273, "correlation": 104602 + } + }, + { + "ph": "f", "id": 104602, "pid": 435, "tid": 435, "ts": 1776649793343936, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793343945, "dur": 19, + "args": { + "External id": 104603, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104603, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104603, "pid": 0, "tid": 7, "ts": 1776649793343945, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343937, "dur": 4, + "args": { + "External id": 104603, + "cbid": 211, "correlation": 104603 + } + }, + { + "ph": "s", "id": 104603, "pid": 435, "tid": 435, "ts": 1776649793343937, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793343981, "dur": 1, + "args": { + "External id": 14162, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104621, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104621, "pid": 0, "tid": 7, "ts": 1776649793343981, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343972, "dur": 6, + "args": { + "External id": 14162, + "cbid": 211, "correlation": 104621 + } + }, + { + "ph": "s", "id": 104621, "pid": 435, "tid": 435, "ts": 1776649793343972, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343984, "dur": 0, + "args": { + "External id": 104627, + "cbid": 200, "correlation": 104627 + } + }, + { + "ph": "f", "id": 104627, "pid": 435, "tid": 435, "ts": 1776649793343984, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793343985, "dur": 0, + "args": { + "External id": 104628, + "cbid": 273, "correlation": 104628 + } + }, + { + "ph": "f", "id": 104628, "pid": 435, "tid": 435, "ts": 1776649793343985, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793343994, "dur": 19, + "args": { + "External id": 104629, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104629, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104629, "pid": 0, "tid": 7, "ts": 1776649793343994, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793343986, "dur": 4, + "args": { + "External id": 104629, + "cbid": 211, "correlation": 104629 + } + }, + { + "ph": "s", "id": 104629, "pid": 435, "tid": 435, "ts": 1776649793343986, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793344029, "dur": 1, + "args": { + "External id": 14167, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104647, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104647, "pid": 0, "tid": 7, "ts": 1776649793344029, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344020, "dur": 6, + "args": { + "External id": 14167, + "cbid": 211, "correlation": 104647 + } + }, + { + "ph": "s", "id": 104647, "pid": 435, "tid": 435, "ts": 1776649793344020, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344032, "dur": 0, + "args": { + "External id": 104653, + "cbid": 200, "correlation": 104653 + } + }, + { + "ph": "f", "id": 104653, "pid": 435, "tid": 435, "ts": 1776649793344032, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344033, "dur": 0, + "args": { + "External id": 104654, + "cbid": 273, "correlation": 104654 + } + }, + { + "ph": "f", "id": 104654, "pid": 435, "tid": 435, "ts": 1776649793344033, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793344042, "dur": 19, + "args": { + "External id": 104655, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104655, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104655, "pid": 0, "tid": 7, "ts": 1776649793344042, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344034, "dur": 5, + "args": { + "External id": 104655, + "cbid": 211, "correlation": 104655 + } + }, + { + "ph": "s", "id": 104655, "pid": 435, "tid": 435, "ts": 1776649793344034, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793344077, "dur": 1, + "args": { + "External id": 14172, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104673, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104673, "pid": 0, "tid": 7, "ts": 1776649793344077, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344068, "dur": 6, + "args": { + "External id": 14172, + "cbid": 211, "correlation": 104673 + } + }, + { + "ph": "s", "id": 104673, "pid": 435, "tid": 435, "ts": 1776649793344068, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344080, "dur": 0, + "args": { + "External id": 104679, + "cbid": 200, "correlation": 104679 + } + }, + { + "ph": "f", "id": 104679, "pid": 435, "tid": 435, "ts": 1776649793344080, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344081, "dur": 0, + "args": { + "External id": 104680, + "cbid": 273, "correlation": 104680 + } + }, + { + "ph": "f", "id": 104680, "pid": 435, "tid": 435, "ts": 1776649793344081, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793344090, "dur": 19, + "args": { + "External id": 104681, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104681, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104681, "pid": 0, "tid": 7, "ts": 1776649793344090, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344082, "dur": 4, + "args": { + "External id": 104681, + "cbid": 211, "correlation": 104681 + } + }, + { + "ph": "s", "id": 104681, "pid": 435, "tid": 435, "ts": 1776649793344082, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793344126, "dur": 1, + "args": { + "External id": 14177, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104699, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104699, "pid": 0, "tid": 7, "ts": 1776649793344126, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344116, "dur": 6, + "args": { + "External id": 14177, + "cbid": 211, "correlation": 104699 + } + }, + { + "ph": "s", "id": 104699, "pid": 435, "tid": 435, "ts": 1776649793344116, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344129, "dur": 0, + "args": { + "External id": 104705, + "cbid": 200, "correlation": 104705 + } + }, + { + "ph": "f", "id": 104705, "pid": 435, "tid": 435, "ts": 1776649793344129, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344129, "dur": 0, + "args": { + "External id": 104706, + "cbid": 273, "correlation": 104706 + } + }, + { + "ph": "f", "id": 104706, "pid": 435, "tid": 435, "ts": 1776649793344129, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793344139, "dur": 19, + "args": { + "External id": 104707, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104707, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104707, "pid": 0, "tid": 7, "ts": 1776649793344139, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344130, "dur": 4, + "args": { + "External id": 104707, + "cbid": 211, "correlation": 104707 + } + }, + { + "ph": "s", "id": 104707, "pid": 435, "tid": 435, "ts": 1776649793344130, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793344173, "dur": 1, + "args": { + "External id": 14182, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104725, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104725, "pid": 0, "tid": 7, "ts": 1776649793344173, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344164, "dur": 6, + "args": { + "External id": 14182, + "cbid": 211, "correlation": 104725 + } + }, + { + "ph": "s", "id": 104725, "pid": 435, "tid": 435, "ts": 1776649793344164, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344176, "dur": 0, + "args": { + "External id": 104731, + "cbid": 200, "correlation": 104731 + } + }, + { + "ph": "f", "id": 104731, "pid": 435, "tid": 435, "ts": 1776649793344176, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344177, "dur": 0, + "args": { + "External id": 104732, + "cbid": 273, "correlation": 104732 + } + }, + { + "ph": "f", "id": 104732, "pid": 435, "tid": 435, "ts": 1776649793344177, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793344186, "dur": 19, + "args": { + "External id": 104733, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104733, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104733, "pid": 0, "tid": 7, "ts": 1776649793344186, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344178, "dur": 5, + "args": { + "External id": 104733, + "cbid": 211, "correlation": 104733 + } + }, + { + "ph": "s", "id": 104733, "pid": 435, "tid": 435, "ts": 1776649793344178, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793344239, "dur": 1, + "args": { + "External id": 14187, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104751, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104751, "pid": 0, "tid": 7, "ts": 1776649793344239, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344212, "dur": 24, + "args": { + "External id": 14187, + "cbid": 211, "correlation": 104751 + } + }, + { + "ph": "s", "id": 104751, "pid": 435, "tid": 435, "ts": 1776649793344212, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344243, "dur": 0, + "args": { + "External id": 104757, + "cbid": 200, "correlation": 104757 + } + }, + { + "ph": "f", "id": 104757, "pid": 435, "tid": 435, "ts": 1776649793344243, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344244, "dur": 1, + "args": { + "External id": 104758, + "cbid": 273, "correlation": 104758 + } + }, + { + "ph": "f", "id": 104758, "pid": 435, "tid": 435, "ts": 1776649793344244, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793344255, "dur": 19, + "args": { + "External id": 104759, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104759, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104759, "pid": 0, "tid": 7, "ts": 1776649793344255, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344246, "dur": 5, + "args": { + "External id": 104759, + "cbid": 211, "correlation": 104759 + } + }, + { + "ph": "s", "id": 104759, "pid": 435, "tid": 435, "ts": 1776649793344246, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793344293, "dur": 1, + "args": { + "External id": 14192, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104777, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104777, "pid": 0, "tid": 7, "ts": 1776649793344293, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344283, "dur": 6, + "args": { + "External id": 14192, + "cbid": 211, "correlation": 104777 + } + }, + { + "ph": "s", "id": 104777, "pid": 435, "tid": 435, "ts": 1776649793344283, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344296, "dur": 0, + "args": { + "External id": 104783, + "cbid": 200, "correlation": 104783 + } + }, + { + "ph": "f", "id": 104783, "pid": 435, "tid": 435, "ts": 1776649793344296, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344297, "dur": 0, + "args": { + "External id": 104784, + "cbid": 273, "correlation": 104784 + } + }, + { + "ph": "f", "id": 104784, "pid": 435, "tid": 435, "ts": 1776649793344297, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793344305, "dur": 19, + "args": { + "External id": 104785, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104785, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104785, "pid": 0, "tid": 7, "ts": 1776649793344305, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344298, "dur": 4, + "args": { + "External id": 104785, + "cbid": 211, "correlation": 104785 + } + }, + { + "ph": "s", "id": 104785, "pid": 435, "tid": 435, "ts": 1776649793344298, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793344341, "dur": 1, + "args": { + "External id": 14197, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104803, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104803, "pid": 0, "tid": 7, "ts": 1776649793344341, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344331, "dur": 7, + "args": { + "External id": 14197, + "cbid": 211, "correlation": 104803 + } + }, + { + "ph": "s", "id": 104803, "pid": 435, "tid": 435, "ts": 1776649793344331, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344344, "dur": 0, + "args": { + "External id": 104809, + "cbid": 200, "correlation": 104809 + } + }, + { + "ph": "f", "id": 104809, "pid": 435, "tid": 435, "ts": 1776649793344344, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344345, "dur": 0, + "args": { + "External id": 104810, + "cbid": 273, "correlation": 104810 + } + }, + { + "ph": "f", "id": 104810, "pid": 435, "tid": 435, "ts": 1776649793344345, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793344354, "dur": 19, + "args": { + "External id": 104811, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104811, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104811, "pid": 0, "tid": 7, "ts": 1776649793344354, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344346, "dur": 4, + "args": { + "External id": 104811, + "cbid": 211, "correlation": 104811 + } + }, + { + "ph": "s", "id": 104811, "pid": 435, "tid": 435, "ts": 1776649793344346, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793344390, "dur": 1, + "args": { + "External id": 14202, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104829, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104829, "pid": 0, "tid": 7, "ts": 1776649793344390, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344380, "dur": 6, + "args": { + "External id": 14202, + "cbid": 211, "correlation": 104829 + } + }, + { + "ph": "s", "id": 104829, "pid": 435, "tid": 435, "ts": 1776649793344380, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344393, "dur": 0, + "args": { + "External id": 104835, + "cbid": 200, "correlation": 104835 + } + }, + { + "ph": "f", "id": 104835, "pid": 435, "tid": 435, "ts": 1776649793344393, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344393, "dur": 0, + "args": { + "External id": 104836, + "cbid": 273, "correlation": 104836 + } + }, + { + "ph": "f", "id": 104836, "pid": 435, "tid": 435, "ts": 1776649793344393, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793344403, "dur": 19, + "args": { + "External id": 104837, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104837, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104837, "pid": 0, "tid": 7, "ts": 1776649793344403, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344395, "dur": 5, + "args": { + "External id": 104837, + "cbid": 211, "correlation": 104837 + } + }, + { + "ph": "s", "id": 104837, "pid": 435, "tid": 435, "ts": 1776649793344395, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793344438, "dur": 1, + "args": { + "External id": 14207, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104855, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104855, "pid": 0, "tid": 7, "ts": 1776649793344438, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344429, "dur": 6, + "args": { + "External id": 14207, + "cbid": 211, "correlation": 104855 + } + }, + { + "ph": "s", "id": 104855, "pid": 435, "tid": 435, "ts": 1776649793344429, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344441, "dur": 0, + "args": { + "External id": 104861, + "cbid": 200, "correlation": 104861 + } + }, + { + "ph": "f", "id": 104861, "pid": 435, "tid": 435, "ts": 1776649793344441, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344442, "dur": 0, + "args": { + "External id": 104862, + "cbid": 273, "correlation": 104862 + } + }, + { + "ph": "f", "id": 104862, "pid": 435, "tid": 435, "ts": 1776649793344442, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793344451, "dur": 19, + "args": { + "External id": 104863, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104863, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104863, "pid": 0, "tid": 7, "ts": 1776649793344451, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344443, "dur": 5, + "args": { + "External id": 104863, + "cbid": 211, "correlation": 104863 + } + }, + { + "ph": "s", "id": 104863, "pid": 435, "tid": 435, "ts": 1776649793344443, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793344486, "dur": 1, + "args": { + "External id": 14212, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104881, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104881, "pid": 0, "tid": 7, "ts": 1776649793344486, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344477, "dur": 6, + "args": { + "External id": 14212, + "cbid": 211, "correlation": 104881 + } + }, + { + "ph": "s", "id": 104881, "pid": 435, "tid": 435, "ts": 1776649793344477, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344489, "dur": 0, + "args": { + "External id": 104887, + "cbid": 200, "correlation": 104887 + } + }, + { + "ph": "f", "id": 104887, "pid": 435, "tid": 435, "ts": 1776649793344489, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344490, "dur": 0, + "args": { + "External id": 104888, + "cbid": 273, "correlation": 104888 + } + }, + { + "ph": "f", "id": 104888, "pid": 435, "tid": 435, "ts": 1776649793344490, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793344499, "dur": 19, + "args": { + "External id": 104889, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104889, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104889, "pid": 0, "tid": 7, "ts": 1776649793344499, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344491, "dur": 5, + "args": { + "External id": 104889, + "cbid": 211, "correlation": 104889 + } + }, + { + "ph": "s", "id": 104889, "pid": 435, "tid": 435, "ts": 1776649793344491, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793344534, "dur": 1, + "args": { + "External id": 14217, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104907, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104907, "pid": 0, "tid": 7, "ts": 1776649793344534, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344525, "dur": 6, + "args": { + "External id": 14217, + "cbid": 211, "correlation": 104907 + } + }, + { + "ph": "s", "id": 104907, "pid": 435, "tid": 435, "ts": 1776649793344525, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344537, "dur": 0, + "args": { + "External id": 104913, + "cbid": 200, "correlation": 104913 + } + }, + { + "ph": "f", "id": 104913, "pid": 435, "tid": 435, "ts": 1776649793344537, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344537, "dur": 0, + "args": { + "External id": 104914, + "cbid": 273, "correlation": 104914 + } + }, + { + "ph": "f", "id": 104914, "pid": 435, "tid": 435, "ts": 1776649793344537, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793344546, "dur": 19, + "args": { + "External id": 104915, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104915, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104915, "pid": 0, "tid": 7, "ts": 1776649793344546, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344539, "dur": 5, + "args": { + "External id": 104915, + "cbid": 211, "correlation": 104915 + } + }, + { + "ph": "s", "id": 104915, "pid": 435, "tid": 435, "ts": 1776649793344539, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793344582, "dur": 1, + "args": { + "External id": 14222, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104933, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104933, "pid": 0, "tid": 7, "ts": 1776649793344582, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344572, "dur": 6, + "args": { + "External id": 14222, + "cbid": 211, "correlation": 104933 + } + }, + { + "ph": "s", "id": 104933, "pid": 435, "tid": 435, "ts": 1776649793344572, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344585, "dur": 0, + "args": { + "External id": 104939, + "cbid": 200, "correlation": 104939 + } + }, + { + "ph": "f", "id": 104939, "pid": 435, "tid": 435, "ts": 1776649793344585, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344585, "dur": 0, + "args": { + "External id": 104940, + "cbid": 273, "correlation": 104940 + } + }, + { + "ph": "f", "id": 104940, "pid": 435, "tid": 435, "ts": 1776649793344585, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793344594, "dur": 19, + "args": { + "External id": 104941, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104941, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104941, "pid": 0, "tid": 7, "ts": 1776649793344594, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344586, "dur": 5, + "args": { + "External id": 104941, + "cbid": 211, "correlation": 104941 + } + }, + { + "ph": "s", "id": 104941, "pid": 435, "tid": 435, "ts": 1776649793344586, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793344630, "dur": 1, + "args": { + "External id": 14227, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104959, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104959, "pid": 0, "tid": 7, "ts": 1776649793344630, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344621, "dur": 6, + "args": { + "External id": 14227, + "cbid": 211, "correlation": 104959 + } + }, + { + "ph": "s", "id": 104959, "pid": 435, "tid": 435, "ts": 1776649793344621, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344633, "dur": 0, + "args": { + "External id": 104965, + "cbid": 200, "correlation": 104965 + } + }, + { + "ph": "f", "id": 104965, "pid": 435, "tid": 435, "ts": 1776649793344633, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344634, "dur": 0, + "args": { + "External id": 104966, + "cbid": 273, "correlation": 104966 + } + }, + { + "ph": "f", "id": 104966, "pid": 435, "tid": 435, "ts": 1776649793344634, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793344643, "dur": 19, + "args": { + "External id": 104967, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104967, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104967, "pid": 0, "tid": 7, "ts": 1776649793344643, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344635, "dur": 4, + "args": { + "External id": 104967, + "cbid": 211, "correlation": 104967 + } + }, + { + "ph": "s", "id": 104967, "pid": 435, "tid": 435, "ts": 1776649793344635, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793344678, "dur": 1, + "args": { + "External id": 14232, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104985, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104985, "pid": 0, "tid": 7, "ts": 1776649793344678, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344668, "dur": 6, + "args": { + "External id": 14232, + "cbid": 211, "correlation": 104985 + } + }, + { + "ph": "s", "id": 104985, "pid": 435, "tid": 435, "ts": 1776649793344668, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344681, "dur": 0, + "args": { + "External id": 104991, + "cbid": 200, "correlation": 104991 + } + }, + { + "ph": "f", "id": 104991, "pid": 435, "tid": 435, "ts": 1776649793344681, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344682, "dur": 0, + "args": { + "External id": 104992, + "cbid": 273, "correlation": 104992 + } + }, + { + "ph": "f", "id": 104992, "pid": 435, "tid": 435, "ts": 1776649793344682, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793344691, "dur": 19, + "args": { + "External id": 104993, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 104993, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 104993, "pid": 0, "tid": 7, "ts": 1776649793344691, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344683, "dur": 4, + "args": { + "External id": 104993, + "cbid": 211, "correlation": 104993 + } + }, + { + "ph": "s", "id": 104993, "pid": 435, "tid": 435, "ts": 1776649793344683, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793344726, "dur": 1, + "args": { + "External id": 14237, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 105011, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 105011, "pid": 0, "tid": 7, "ts": 1776649793344726, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344717, "dur": 6, + "args": { + "External id": 14237, + "cbid": 211, "correlation": 105011 + } + }, + { + "ph": "s", "id": 105011, "pid": 435, "tid": 435, "ts": 1776649793344717, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344729, "dur": 0, + "args": { + "External id": 105017, + "cbid": 200, "correlation": 105017 + } + }, + { + "ph": "f", "id": 105017, "pid": 435, "tid": 435, "ts": 1776649793344729, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344730, "dur": 0, + "args": { + "External id": 105018, + "cbid": 273, "correlation": 105018 + } + }, + { + "ph": "f", "id": 105018, "pid": 435, "tid": 435, "ts": 1776649793344730, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793344739, "dur": 19, + "args": { + "External id": 105019, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 105019, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 105019, "pid": 0, "tid": 7, "ts": 1776649793344739, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344731, "dur": 5, + "args": { + "External id": 105019, + "cbid": 211, "correlation": 105019 + } + }, + { + "ph": "s", "id": 105019, "pid": 435, "tid": 435, "ts": 1776649793344731, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793344774, "dur": 1, + "args": { + "External id": 14242, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 105037, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 105037, "pid": 0, "tid": 7, "ts": 1776649793344774, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344765, "dur": 6, + "args": { + "External id": 14242, + "cbid": 211, "correlation": 105037 + } + }, + { + "ph": "s", "id": 105037, "pid": 435, "tid": 435, "ts": 1776649793344765, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344777, "dur": 0, + "args": { + "External id": 105043, + "cbid": 200, "correlation": 105043 + } + }, + { + "ph": "f", "id": 105043, "pid": 435, "tid": 435, "ts": 1776649793344777, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344777, "dur": 0, + "args": { + "External id": 105044, + "cbid": 273, "correlation": 105044 + } + }, + { + "ph": "f", "id": 105044, "pid": 435, "tid": 435, "ts": 1776649793344777, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793344787, "dur": 19, + "args": { + "External id": 105045, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 105045, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 105045, "pid": 0, "tid": 7, "ts": 1776649793344787, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344779, "dur": 4, + "args": { + "External id": 105045, + "cbid": 211, "correlation": 105045 + } + }, + { + "ph": "s", "id": 105045, "pid": 435, "tid": 435, "ts": 1776649793344779, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793344823, "dur": 1, + "args": { + "External id": 14247, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 105063, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 105063, "pid": 0, "tid": 7, "ts": 1776649793344823, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344813, "dur": 7, + "args": { + "External id": 14247, + "cbid": 211, "correlation": 105063 + } + }, + { + "ph": "s", "id": 105063, "pid": 435, "tid": 435, "ts": 1776649793344813, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344827, "dur": 0, + "args": { + "External id": 105069, + "cbid": 200, "correlation": 105069 + } + }, + { + "ph": "f", "id": 105069, "pid": 435, "tid": 435, "ts": 1776649793344827, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344827, "dur": 0, + "args": { + "External id": 105070, + "cbid": 273, "correlation": 105070 + } + }, + { + "ph": "f", "id": 105070, "pid": 435, "tid": 435, "ts": 1776649793344827, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793344837, "dur": 19, + "args": { + "External id": 105071, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 105071, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 105071, "pid": 0, "tid": 7, "ts": 1776649793344837, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344828, "dur": 5, + "args": { + "External id": 105071, + "cbid": 211, "correlation": 105071 + } + }, + { + "ph": "s", "id": 105071, "pid": 435, "tid": 435, "ts": 1776649793344828, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793344873, "dur": 1, + "args": { + "External id": 14252, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 105089, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 105089, "pid": 0, "tid": 7, "ts": 1776649793344873, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344864, "dur": 6, + "args": { + "External id": 14252, + "cbid": 211, "correlation": 105089 + } + }, + { + "ph": "s", "id": 105089, "pid": 435, "tid": 435, "ts": 1776649793344864, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344877, "dur": 0, + "args": { + "External id": 105095, + "cbid": 200, "correlation": 105095 + } + }, + { + "ph": "f", "id": 105095, "pid": 435, "tid": 435, "ts": 1776649793344877, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344877, "dur": 0, + "args": { + "External id": 105096, + "cbid": 273, "correlation": 105096 + } + }, + { + "ph": "f", "id": 105096, "pid": 435, "tid": 435, "ts": 1776649793344877, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793344886, "dur": 19, + "args": { + "External id": 105097, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 105097, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 105097, "pid": 0, "tid": 7, "ts": 1776649793344886, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344878, "dur": 4, + "args": { + "External id": 105097, + "cbid": 211, "correlation": 105097 + } + }, + { + "ph": "s", "id": 105097, "pid": 435, "tid": 435, "ts": 1776649793344878, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649793344921, "dur": 1, + "args": { + "External id": 14257, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 105115, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.018519, + "warps per SM": 0.074074, + "grid": [2, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 105115, "pid": 0, "tid": 7, "ts": 1776649793344921, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344912, "dur": 6, + "args": { + "External id": 14257, + "cbid": 211, "correlation": 105115 + } + }, + { + "ph": "s", "id": 105115, "pid": 435, "tid": 435, "ts": 1776649793344912, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344924, "dur": 0, + "args": { + "External id": 105121, + "cbid": 200, "correlation": 105121 + } + }, + { + "ph": "f", "id": 105121, "pid": 435, "tid": 435, "ts": 1776649793344924, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649793344924, "dur": 0, + "args": { + "External id": 105122, + "cbid": 273, "correlation": 105122 + } + }, + { + "ph": "f", "id": 105122, "pid": 435, "tid": 435, "ts": 1776649793344924, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649793344933, "dur": 19, + "args": { + "External id": 105123, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 105123, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 105123, "pid": 0, "tid": 7, "ts": 1776649793344933, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649793344926, "dur": 4, + "args": { + "External id": 105123, + "cbid": 211, "correlation": 105123 + } + }, + { + "ph": "s", "id": 105123, "pid": 435, "tid": 435, "ts": 1776649793344926, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649793344955, "dur": 3, + "args": { + "External id": 105131, + "cbid": 135, "correlation": 105131 + } + }, + { + "ph": "f", "id": 105131, "pid": 435, "tid": 435, "ts": 1776649793344955, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649793344970, "dur": 5, + "args": { + "External id": 105136, + "cbid": 165, "correlation": 105136 + } + }, + { + "ph": "s", "id": 105136, "pid": 435, "tid": 435, "ts": 1776649793344970, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649793344978, "dur": 4, + "args": { + "External id": 105139, + "cbid": 139, "correlation": 105139 + } + }, + { + "ph": "f", "id": 105139, "pid": 435, "tid": 435, "ts": 1776649793344978, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649793345188, "dur": 5, + "args": { + "External id": 105152, + "cbid": 165, "correlation": 105152 + } + }, + { + "ph": "s", "id": 105152, "pid": 435, "tid": 435, "ts": 1776649793345188, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetStreamPriorityRange", "pid": 435, "tid": 435, + "ts": 1776649789514098, "dur": 180293, + "args": { + "External id": 2, + "cbid": 205, "correlation": 4 + } + }, + { + "ph": "f", "id": 4, "pid": 435, "tid": 435, "ts": 1776649789514098, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649789694435, "dur": 15, + "args": { + "External id": 2, + "cbid": 317, "correlation": 7 + } + }, + { + "ph": "f", "id": 7, "pid": 435, "tid": 435, "ts": 1776649789694435, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 435, "tid": 435, + "ts": 1776649789694466, "dur": 390, + "args": { + "External id": 2, + "cbid": 20, "correlation": 8 + } + }, + { + "ph": "f", "id": 8, "pid": 435, "tid": 435, "ts": 1776649789694466, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649789695118, "dur": 1, + "args": { + "External id": 3, + "cbid": 317, "correlation": 20 + } + }, + { + "ph": "f", "id": 20, "pid": 435, "tid": 435, "ts": 1776649789695118, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649789695122, "dur": 0, + "args": { + "External id": 3, + "cbid": 317, "correlation": 24 + } + }, + { + "ph": "f", "id": 24, "pid": 435, "tid": 435, "ts": 1776649789695122, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::distribution_elementwise_grid_stride_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::(anonymous namespace)::distribution_nullary_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2} const&, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(int, float)#1}>(int, at::PhiloxCudaState, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::(anonymous namespace)::distribution_nullary_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2} const&, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(int, float)#1})", "pid": 0, "tid": 7, + "ts": 1776649789696747, "dur": 3, + "args": { + "External id": 3, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26, + "registers per thread": 38, + "shared memory": 0, + "blocks per SM": 0.148148, + "warps per SM": 1.185185, + "grid": [16, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 2 + } + }, + { + "ph": "f", "id": 26, "pid": 0, "tid": 7, "ts": 1776649789696747, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789695128, "dur": 1621, + "args": { + "External id": 3, + "cbid": 211, "correlation": 26 + } + }, + { + "ph": "s", "id": 26, "pid": 435, "tid": 435, "ts": 1776649789695128, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649789696837, "dur": 1, + "args": { + "External id": 33, + "cbid": 317, "correlation": 33 + } + }, + { + "ph": "f", "id": 33, "pid": 435, "tid": 435, "ts": 1776649789696837, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649789696848, "dur": 0, + "args": { + "External id": 37, + "cbid": 317, "correlation": 37 + } + }, + { + "ph": "f", "id": 37, "pid": 435, "tid": 435, "ts": 1776649789696848, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649789696930, "dur": 0, + "args": { + "External id": 5, + "cbid": 317, "correlation": 43 + } + }, + { + "ph": "f", "id": 43, "pid": 435, "tid": 435, "ts": 1776649789696930, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 435, "tid": 435, + "ts": 1776649789696932, "dur": 259, + "args": { + "External id": 5, + "cbid": 20, "correlation": 44 + } + }, + { + "ph": "f", "id": 44, "pid": 435, "tid": 435, "ts": 1776649789696932, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649789697214, "dur": 1, + "args": { + "External id": 6, + "cbid": 317, "correlation": 56 + } + }, + { + "ph": "f", "id": 56, "pid": 435, "tid": 435, "ts": 1776649789697214, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649789697217, "dur": 0, + "args": { + "External id": 6, + "cbid": 317, "correlation": 60 + } + }, + { + "ph": "f", "id": 60, "pid": 435, "tid": 435, "ts": 1776649789697217, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::distribution_elementwise_grid_stride_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::(anonymous namespace)::distribution_nullary_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2} const&, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(int, float)#1}>(int, at::PhiloxCudaState, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::(anonymous namespace)::distribution_nullary_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2} const&, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(int, float)#1})", "pid": 0, "tid": 7, + "ts": 1776649789697235, "dur": 85, + "args": { + "External id": 6, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 62, + "registers per thread": 38, + "shared memory": 0, + "blocks per SM": 8.000000, + "warps per SM": 64.000000, + "grid": [864, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 75 + } + }, + { + "ph": "f", "id": 62, "pid": 0, "tid": 7, "ts": 1776649789697235, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789697220, "dur": 13, + "args": { + "External id": 6, + "cbid": 211, "correlation": 62 + } + }, + { + "ph": "s", "id": 62, "pid": 435, "tid": 435, "ts": 1776649789697220, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649789697472, "dur": 1, + "args": { + "External id": 7, + "cbid": 317, "correlation": 71 + } + }, + { + "ph": "f", "id": 71, "pid": 435, "tid": 435, "ts": 1776649789697472, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 435, "tid": 435, + "ts": 1776649789697473, "dur": 233, + "args": { + "External id": 7, + "cbid": 20, "correlation": 72 + } + }, + { + "ph": "f", "id": 72, "pid": 435, "tid": 435, "ts": 1776649789697473, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor >, at::detail::Array >(int, at::native::AUnaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649789708249, "dur": 40, + "args": { + "External id": 7, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 76, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 76, "pid": 0, "tid": 7, "ts": 1776649789708249, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789697748, "dur": 10506, + "args": { + "External id": 7, + "cbid": 211, "correlation": 76 + } + }, + { + "ph": "s", "id": 76, "pid": 435, "tid": 435, "ts": 1776649789697748, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AbsFunctor, at::detail::Array >(int, at::native::AbsFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649789711459, "dur": 41, + "args": { + "External id": 12, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 96, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 96, "pid": 0, "tid": 7, "ts": 1776649789711459, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789708798, "dur": 2659, + "args": { + "External id": 12, + "cbid": 211, "correlation": 96 + } + }, + { + "ph": "s", "id": 96, "pid": 435, "tid": 435, "ts": 1776649789708798, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<128, 4, at::native::ReduceOp >, unsigned int, c10::Half, 4> >(at::native::ReduceOp >, unsigned int, c10::Half, 4>)", "pid": 0, "tid": 7, + "ts": 1776649789727049, "dur": 47, + "args": { + "External id": 14, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 109, + "registers per thread": 49, + "shared memory": 1040, + "blocks per SM": 9.481482, + "warps per SM": 37.925926, + "grid": [1024, 1, 1], + "block": [32, 4, 1], + "est. achieved occupancy %": 56 + } + }, + { + "ph": "f", "id": 109, "pid": 0, "tid": 7, "ts": 1776649789727049, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789711660, "dur": 15396, + "args": { + "External id": 14, + "cbid": 211, "correlation": 109 + } + }, + { + "ph": "s", "id": 109, "pid": 435, "tid": 435, "ts": 1776649789711660, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BUnaryFunctor >, at::detail::Array >(int, at::native::BUnaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649789734968, "dur": 2, + "args": { + "External id": 16, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 119, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 2.370370, + "warps per SM": 9.481482, + "grid": [256, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 15 + } + }, + { + "ph": "f", "id": 119, "pid": 0, "tid": 7, "ts": 1776649789734968, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789727388, "dur": 7581, + "args": { + "External id": 16, + "cbid": 211, "correlation": 119 + } + }, + { + "ph": "s", "id": 119, "pid": 435, "tid": 435, "ts": 1776649789727388, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#8}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#8}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649789748043, "dur": 2, + "args": { + "External id": 17, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 129, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 2.370370, + "warps per SM": 9.481482, + "grid": [256, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 15 + } + }, + { + "ph": "f", "id": 129, "pid": 0, "tid": 7, "ts": 1776649789748043, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789735238, "dur": 12809, + "args": { + "External id": 17, + "cbid": 211, "correlation": 129 + } + }, + { + "ph": "s", "id": 129, "pid": 435, "tid": 435, "ts": 1776649789735238, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649789748548, "dur": 99, + "args": { + "External id": 22, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 139, + "registers per thread": 20, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 139, "pid": 0, "tid": 7, "ts": 1776649789748548, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789748491, "dur": 56, + "args": { + "External id": 22, + "cbid": 211, "correlation": 139 + } + }, + { + "ph": "s", "id": 139, "pid": 435, "tid": 435, "ts": 1776649789748491, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649789748651, "dur": 3, + "args": { + "External id": 23, + "cbid": 317, "correlation": 148 + } + }, + { + "ph": "f", "id": 148, "pid": 435, "tid": 435, "ts": 1776649789748651, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 435, "tid": 435, + "ts": 1776649789748654, "dur": 599, + "args": { + "External id": 23, + "cbid": 20, "correlation": 149 + } + }, + { + "ph": "f", "id": 149, "pid": 435, "tid": 435, "ts": 1776649789748654, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649789757534, "dur": 39, + "args": { + "External id": 23, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 153, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 153, "pid": 0, "tid": 7, "ts": 1776649789757534, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789749282, "dur": 8253, + "args": { + "External id": 23, + "cbid": 211, "correlation": 153 + } + }, + { + "ph": "s", "id": 153, "pid": 435, "tid": 435, "ts": 1776649789749282, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#8}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#8}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649789757677, "dur": 41, + "args": { + "External id": 24, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 163, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 163, "pid": 0, "tid": 7, "ts": 1776649789757677, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789757664, "dur": 11, + "args": { + "External id": 24, + "cbid": 211, "correlation": 163 + } + }, + { + "ph": "s", "id": 163, "pid": 435, "tid": 435, "ts": 1776649789757664, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649789766628, "dur": 95, + "args": { + "External id": 29, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 176, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 176, "pid": 0, "tid": 7, "ts": 1776649789766628, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789757887, "dur": 8744, + "args": { + "External id": 29, + "cbid": 211, "correlation": 176 + } + }, + { + "ph": "s", "id": 176, "pid": 435, "tid": 435, "ts": 1776649789757887, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, at::detail::Array >(int, at::native::CUDAFunctorOnSelf_add, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649789777145, "dur": 33, + "args": { + "External id": 30, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 187, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 187, "pid": 0, "tid": 7, "ts": 1776649789777145, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789766919, "dur": 10227, + "args": { + "External id": 30, + "cbid": 211, "correlation": 187 + } + }, + { + "ph": "s", "id": 187, "pid": 435, "tid": 435, "ts": 1776649789766919, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#1}::operator()() const::{lambda(unsigned char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649789777403, "dur": 57, + "args": { + "External id": 34, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 200, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 200, "pid": 0, "tid": 7, "ts": 1776649789777403, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789777338, "dur": 63, + "args": { + "External id": 34, + "cbid": 211, "correlation": 200 + } + }, + { + "ph": "s", "id": 200, "pid": 435, "tid": 435, "ts": 1776649789777338, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast >(at::TensorIteratorBase&, at::native::BUnaryFunctor const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast >(at::TensorIteratorBase&, at::native::BUnaryFunctor const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649789785935, "dur": 37, + "args": { + "External id": 45, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 220, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 220, "pid": 0, "tid": 7, "ts": 1776649789785935, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789777817, "dur": 8122, + "args": { + "External id": 45, + "cbid": 211, "correlation": 220 + } + }, + { + "ph": "s", "id": 220, "pid": 435, "tid": 435, "ts": 1776649789777817, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649789795834, "dur": 39, + "args": { + "External id": 48, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 230, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 230, "pid": 0, "tid": 7, "ts": 1776649789795834, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789786124, "dur": 9713, + "args": { + "External id": 48, + "cbid": 211, "correlation": 230 + } + }, + { + "ph": "s", "id": 230, "pid": 435, "tid": 435, "ts": 1776649789786124, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor >, at::detail::Array >(int, at::native::AUnaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649789796206, "dur": 17, + "args": { + "External id": 52, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 240, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 240, "pid": 0, "tid": 7, "ts": 1776649789796206, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789796169, "dur": 35, + "args": { + "External id": 52, + "cbid": 211, "correlation": 240 + } + }, + { + "ph": "s", "id": 240, "pid": 435, "tid": 435, "ts": 1776649789796169, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649789796318, "dur": 30, + "args": { + "External id": 56, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 253, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 253, "pid": 0, "tid": 7, "ts": 1776649789796318, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789796305, "dur": 11, + "args": { + "External id": 56, + "cbid": 211, "correlation": 253 + } + }, + { + "ph": "s", "id": 253, "pid": 435, "tid": 435, "ts": 1776649789796305, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BUnaryFunctor, at::detail::Array >(int, at::native::BUnaryFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649789796424, "dur": 17, + "args": { + "External id": 57, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 267, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 267, "pid": 0, "tid": 7, "ts": 1776649789796424, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789796392, "dur": 30, + "args": { + "External id": 57, + "cbid": 211, "correlation": 267 + } + }, + { + "ph": "s", "id": 267, "pid": 435, "tid": 435, "ts": 1776649789796392, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649789796460, "dur": 30, + "args": { + "External id": 62, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 280, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 280, "pid": 0, "tid": 7, "ts": 1776649789796460, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789796450, "dur": 7, + "args": { + "External id": 62, + "cbid": 211, "correlation": 280 + } + }, + { + "ph": "s", "id": 280, "pid": 435, "tid": 435, "ts": 1776649789796450, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy_contig, unsigned int, 3, 128, 1>(at::native::(anonymous namespace)::OpaqueType<1u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata, unsigned int, 128, 1>, at::native::(anonymous namespace)::TensorSizeStride, int, unsigned int)", "pid": 0, "tid": 7, + "ts": 1776649789800733, "dur": 64, + "args": { + "External id": 68, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 292, + "registers per thread": 21, + "shared memory": 0, + "blocks per SM": 4.000000, + "warps per SM": 64.000000, + "grid": [216, 2, 1], + "block": [512, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 292, "pid": 0, "tid": 7, "ts": 1776649789800733, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789796691, "dur": 4043, + "args": { + "External id": 68, + "cbid": 211, "correlation": 292 + } + }, + { + "ph": "s", "id": 292, "pid": 435, "tid": 435, "ts": 1776649789796691, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649789801049, "dur": 71, + "args": { + "External id": 79, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 305, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 305, "pid": 0, "tid": 7, "ts": 1776649789801049, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789800996, "dur": 55, + "args": { + "External id": 79, + "cbid": 211, "correlation": 305 + } + }, + { + "ph": "s", "id": 305, "pid": 435, "tid": 435, "ts": 1776649789800996, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, at::detail::Array >(int, at::native::CUDAFunctorOnSelf_add, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649789801168, "dur": 33, + "args": { + "External id": 82, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 316, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 316, "pid": 0, "tid": 7, "ts": 1776649789801168, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789801154, "dur": 11, + "args": { + "External id": 82, + "cbid": 211, "correlation": 316 + } + }, + { + "ph": "s", "id": 316, "pid": 435, "tid": 435, "ts": 1776649789801154, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649789801206, "dur": 2, + "args": { + "External id": 85, + "cbid": 317, "correlation": 325 + } + }, + { + "ph": "f", "id": 325, "pid": 435, "tid": 435, "ts": 1776649789801206, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 435, "tid": 435, + "ts": 1776649789801209, "dur": 418, + "args": { + "External id": 85, + "cbid": 20, "correlation": 326 + } + }, + { + "ph": "f", "id": 326, "pid": 435, "tid": 435, "ts": 1776649789801209, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649789801713, "dur": 91, + "args": { + "External id": 86, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 333, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 333, "pid": 0, "tid": 7, "ts": 1776649789801713, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789801646, "dur": 65, + "args": { + "External id": 86, + "cbid": 211, "correlation": 333 + } + }, + { + "ph": "s", "id": 333, "pid": 435, "tid": 435, "ts": 1776649789801646, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649789801749, "dur": 1, + "args": { + "External id": 87, + "cbid": 317, "correlation": 343 + } + }, + { + "ph": "f", "id": 343, "pid": 435, "tid": 435, "ts": 1776649789801749, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 435, "tid": 435, + "ts": 1776649789801750, "dur": 218, + "args": { + "External id": 87, + "cbid": 20, "correlation": 344 + } + }, + { + "ph": "f", "id": 344, "pid": 435, "tid": 435, "ts": 1776649789801750, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor >, at::detail::Array >(int, at::native::BinaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649789802020, "dur": 59, + "args": { + "External id": 87, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 348, + "registers per thread": 22, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 348, "pid": 0, "tid": 7, "ts": 1776649789802020, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789801982, "dur": 36, + "args": { + "External id": 87, + "cbid": 211, "correlation": 348 + } + }, + { + "ph": "s", "id": 348, "pid": 435, "tid": 435, "ts": 1776649789801982, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFree", "pid": 435, "tid": 435, + "ts": 1776649789804243, "dur": 2089, + "args": { + "External id": 89, + "cbid": 22, "correlation": 883 + } + }, + { + "ph": "f", "id": 883, "pid": 435, "tid": 435, "ts": 1776649789804243, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789808299, "dur": 0, + "args": { + "External id": 89, + "cbid": 200, "correlation": 1412 + } + }, + { + "ph": "f", "id": 1412, "pid": 435, "tid": 435, "ts": 1776649789808299, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789808299, "dur": 0, + "args": { + "External id": 89, + "cbid": 200, "correlation": 1413 + } + }, + { + "ph": "f", "id": 1413, "pid": 435, "tid": 435, "ts": 1776649789808299, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789808300, "dur": 0, + "args": { + "External id": 89, + "cbid": 200, "correlation": 1414 + } + }, + { + "ph": "f", "id": 1414, "pid": 435, "tid": 435, "ts": 1776649789808300, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789808300, "dur": 0, + "args": { + "External id": 89, + "cbid": 200, "correlation": 1415 + } + }, + { + "ph": "f", "id": 1415, "pid": 435, "tid": 435, "ts": 1776649789808300, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789808301, "dur": 0, + "args": { + "External id": 89, + "cbid": 200, "correlation": 1416 + } + }, + { + "ph": "f", "id": 1416, "pid": 435, "tid": 435, "ts": 1776649789808301, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789808301, "dur": 0, + "args": { + "External id": 89, + "cbid": 200, "correlation": 1417 + } + }, + { + "ph": "f", "id": 1417, "pid": 435, "tid": 435, "ts": 1776649789808301, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789808301, "dur": 0, + "args": { + "External id": 89, + "cbid": 200, "correlation": 1418 + } + }, + { + "ph": "f", "id": 1418, "pid": 435, "tid": 435, "ts": 1776649789808301, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789808302, "dur": 0, + "args": { + "External id": 89, + "cbid": 200, "correlation": 1419 + } + }, + { + "ph": "f", "id": 1419, "pid": 435, "tid": 435, "ts": 1776649789808302, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789808302, "dur": 0, + "args": { + "External id": 89, + "cbid": 200, "correlation": 1420 + } + }, + { + "ph": "f", "id": 1420, "pid": 435, "tid": 435, "ts": 1776649789808302, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789808303, "dur": 0, + "args": { + "External id": 89, + "cbid": 200, "correlation": 1421 + } + }, + { + "ph": "f", "id": 1421, "pid": 435, "tid": 435, "ts": 1776649789808303, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789808303, "dur": 0, + "args": { + "External id": 89, + "cbid": 200, "correlation": 1422 + } + }, + { + "ph": "f", "id": 1422, "pid": 435, "tid": 435, "ts": 1776649789808303, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789808303, "dur": 0, + "args": { + "External id": 89, + "cbid": 200, "correlation": 1423 + } + }, + { + "ph": "f", "id": 1423, "pid": 435, "tid": 435, "ts": 1776649789808303, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789808304, "dur": 0, + "args": { + "External id": 89, + "cbid": 200, "correlation": 1424 + } + }, + { + "ph": "f", "id": 1424, "pid": 435, "tid": 435, "ts": 1776649789808304, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789808304, "dur": 0, + "args": { + "External id": 89, + "cbid": 200, "correlation": 1425 + } + }, + { + "ph": "f", "id": 1425, "pid": 435, "tid": 435, "ts": 1776649789808304, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789808305, "dur": 0, + "args": { + "External id": 89, + "cbid": 200, "correlation": 1426 + } + }, + { + "ph": "f", "id": 1426, "pid": 435, "tid": 435, "ts": 1776649789808305, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789808305, "dur": 0, + "args": { + "External id": 89, + "cbid": 200, "correlation": 1427 + } + }, + { + "ph": "f", "id": 1427, "pid": 435, "tid": 435, "ts": 1776649789808305, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789808305, "dur": 0, + "args": { + "External id": 89, + "cbid": 200, "correlation": 1428 + } + }, + { + "ph": "f", "id": 1428, "pid": 435, "tid": 435, "ts": 1776649789808305, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPoint", "pid": 435, "tid": 435, + "ts": 1776649789808306, "dur": 0, + "args": { + "External id": 89, + "cbid": 406, "correlation": 1429 + } + }, + { + "ph": "f", "id": 1429, "pid": 435, "tid": 435, "ts": 1776649789808306, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaGetDriverEntryPoint", "pid": 435, "tid": 435, + "ts": 1776649789808307, "dur": 0, + "args": { + "External id": 89, + "cbid": 406, "correlation": 1431 + } + }, + { + "ph": "f", "id": 1431, "pid": 435, "tid": 435, "ts": 1776649789808307, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789808314, "dur": 0, + "args": { + "External id": 89, + "cbid": 200, "correlation": 1433 + } + }, + { + "ph": "f", "id": 1433, "pid": 435, "tid": 435, "ts": 1776649789808314, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaGetSymbolAddress", "pid": 435, "tid": 435, + "ts": 1776649789808330, "dur": 11069, + "args": { + "External id": 89, + "cbid": 53, "correlation": 1434 + } + }, + { + "ph": "f", "id": 1434, "pid": 435, "tid": 435, "ts": 1776649789808330, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 435, "tid": 435, + "ts": 1776649789819499, "dur": 380, + "args": { + "External id": 89, + "cbid": 20, "correlation": 1436 + } + }, + { + "ph": "f", "id": 1436, "pid": 435, "tid": 435, "ts": 1776649789819499, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 435, "tid": 435, + "ts": 1776649789819911, "dur": 4, + "args": { + "External id": 89, + "cbid": 20, "correlation": 1445 + } + }, + { + "ph": "f", "id": 1445, "pid": 435, "tid": 435, "ts": 1776649789819911, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 435, "tid": 435, + "ts": 1776649789819926, "dur": 205, + "args": { + "External id": 89, + "cbid": 20, "correlation": 1454 + } + }, + { + "ph": "f", "id": 1454, "pid": 435, "tid": 435, "ts": 1776649789819926, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789910162, "dur": 8958, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1501 + } + }, + { + "ph": "f", "id": 1501, "pid": 435, "tid": 435, "ts": 1776649789910162, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919125, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1502 + } + }, + { + "ph": "f", "id": 1502, "pid": 435, "tid": 435, "ts": 1776649789919125, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919153, "dur": 29, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1503 + } + }, + { + "ph": "f", "id": 1503, "pid": 435, "tid": 435, "ts": 1776649789919153, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919182, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1504 + } + }, + { + "ph": "f", "id": 1504, "pid": 435, "tid": 435, "ts": 1776649789919182, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919208, "dur": 31, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1505 + } + }, + { + "ph": "f", "id": 1505, "pid": 435, "tid": 435, "ts": 1776649789919208, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919240, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1506 + } + }, + { + "ph": "f", "id": 1506, "pid": 435, "tid": 435, "ts": 1776649789919240, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919264, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1507 + } + }, + { + "ph": "f", "id": 1507, "pid": 435, "tid": 435, "ts": 1776649789919264, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919291, "dur": 33, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1508 + } + }, + { + "ph": "f", "id": 1508, "pid": 435, "tid": 435, "ts": 1776649789919291, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919325, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1509 + } + }, + { + "ph": "f", "id": 1509, "pid": 435, "tid": 435, "ts": 1776649789919325, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919352, "dur": 29, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1510 + } + }, + { + "ph": "f", "id": 1510, "pid": 435, "tid": 435, "ts": 1776649789919352, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919382, "dur": 29, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1511 + } + }, + { + "ph": "f", "id": 1511, "pid": 435, "tid": 435, "ts": 1776649789919382, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919412, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1512 + } + }, + { + "ph": "f", "id": 1512, "pid": 435, "tid": 435, "ts": 1776649789919412, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919439, "dur": 30, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1513 + } + }, + { + "ph": "f", "id": 1513, "pid": 435, "tid": 435, "ts": 1776649789919439, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919470, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1514 + } + }, + { + "ph": "f", "id": 1514, "pid": 435, "tid": 435, "ts": 1776649789919470, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919498, "dur": 33, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1515 + } + }, + { + "ph": "f", "id": 1515, "pid": 435, "tid": 435, "ts": 1776649789919498, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919531, "dur": 32, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1516 + } + }, + { + "ph": "f", "id": 1516, "pid": 435, "tid": 435, "ts": 1776649789919531, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919563, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1517 + } + }, + { + "ph": "f", "id": 1517, "pid": 435, "tid": 435, "ts": 1776649789919563, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919591, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1518 + } + }, + { + "ph": "f", "id": 1518, "pid": 435, "tid": 435, "ts": 1776649789919591, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919618, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1519 + } + }, + { + "ph": "f", "id": 1519, "pid": 435, "tid": 435, "ts": 1776649789919618, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919643, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1520 + } + }, + { + "ph": "f", "id": 1520, "pid": 435, "tid": 435, "ts": 1776649789919643, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919670, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1521 + } + }, + { + "ph": "f", "id": 1521, "pid": 435, "tid": 435, "ts": 1776649789919670, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919697, "dur": 28, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1522 + } + }, + { + "ph": "f", "id": 1522, "pid": 435, "tid": 435, "ts": 1776649789919697, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919726, "dur": 30, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1523 + } + }, + { + "ph": "f", "id": 1523, "pid": 435, "tid": 435, "ts": 1776649789919726, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919756, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1524 + } + }, + { + "ph": "f", "id": 1524, "pid": 435, "tid": 435, "ts": 1776649789919756, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919784, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1525 + } + }, + { + "ph": "f", "id": 1525, "pid": 435, "tid": 435, "ts": 1776649789919784, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919811, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1526 + } + }, + { + "ph": "f", "id": 1526, "pid": 435, "tid": 435, "ts": 1776649789919811, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919835, "dur": 29, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1527 + } + }, + { + "ph": "f", "id": 1527, "pid": 435, "tid": 435, "ts": 1776649789919835, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919864, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1528 + } + }, + { + "ph": "f", "id": 1528, "pid": 435, "tid": 435, "ts": 1776649789919864, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919891, "dur": 34, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1529 + } + }, + { + "ph": "f", "id": 1529, "pid": 435, "tid": 435, "ts": 1776649789919891, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919926, "dur": 28, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1530 + } + }, + { + "ph": "f", "id": 1530, "pid": 435, "tid": 435, "ts": 1776649789919926, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919955, "dur": 28, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1531 + } + }, + { + "ph": "f", "id": 1531, "pid": 435, "tid": 435, "ts": 1776649789919955, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789919983, "dur": 32, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1532 + } + }, + { + "ph": "f", "id": 1532, "pid": 435, "tid": 435, "ts": 1776649789919983, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789920019, "dur": 8960, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1533 + } + }, + { + "ph": "f", "id": 1533, "pid": 435, "tid": 435, "ts": 1776649789920019, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789928985, "dur": 29, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1534 + } + }, + { + "ph": "f", "id": 1534, "pid": 435, "tid": 435, "ts": 1776649789928985, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929015, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1535 + } + }, + { + "ph": "f", "id": 1535, "pid": 435, "tid": 435, "ts": 1776649789929015, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929039, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1536 + } + }, + { + "ph": "f", "id": 1536, "pid": 435, "tid": 435, "ts": 1776649789929039, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929063, "dur": 29, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1537 + } + }, + { + "ph": "f", "id": 1537, "pid": 435, "tid": 435, "ts": 1776649789929063, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929092, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1538 + } + }, + { + "ph": "f", "id": 1538, "pid": 435, "tid": 435, "ts": 1776649789929092, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929119, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1539 + } + }, + { + "ph": "f", "id": 1539, "pid": 435, "tid": 435, "ts": 1776649789929119, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929143, "dur": 28, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1540 + } + }, + { + "ph": "f", "id": 1540, "pid": 435, "tid": 435, "ts": 1776649789929143, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929172, "dur": 30, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1541 + } + }, + { + "ph": "f", "id": 1541, "pid": 435, "tid": 435, "ts": 1776649789929172, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929203, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1542 + } + }, + { + "ph": "f", "id": 1542, "pid": 435, "tid": 435, "ts": 1776649789929203, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929230, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1543 + } + }, + { + "ph": "f", "id": 1543, "pid": 435, "tid": 435, "ts": 1776649789929230, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929254, "dur": 28, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1544 + } + }, + { + "ph": "f", "id": 1544, "pid": 435, "tid": 435, "ts": 1776649789929254, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929283, "dur": 28, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1545 + } + }, + { + "ph": "f", "id": 1545, "pid": 435, "tid": 435, "ts": 1776649789929283, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929312, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1546 + } + }, + { + "ph": "f", "id": 1546, "pid": 435, "tid": 435, "ts": 1776649789929312, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929337, "dur": 40, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1547 + } + }, + { + "ph": "f", "id": 1547, "pid": 435, "tid": 435, "ts": 1776649789929337, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929378, "dur": 28, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1548 + } + }, + { + "ph": "f", "id": 1548, "pid": 435, "tid": 435, "ts": 1776649789929378, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929407, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1549 + } + }, + { + "ph": "f", "id": 1549, "pid": 435, "tid": 435, "ts": 1776649789929407, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929432, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1550 + } + }, + { + "ph": "f", "id": 1550, "pid": 435, "tid": 435, "ts": 1776649789929432, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929459, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1551 + } + }, + { + "ph": "f", "id": 1551, "pid": 435, "tid": 435, "ts": 1776649789929459, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929483, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1552 + } + }, + { + "ph": "f", "id": 1552, "pid": 435, "tid": 435, "ts": 1776649789929483, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929509, "dur": 28, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1553 + } + }, + { + "ph": "f", "id": 1553, "pid": 435, "tid": 435, "ts": 1776649789929509, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929538, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1554 + } + }, + { + "ph": "f", "id": 1554, "pid": 435, "tid": 435, "ts": 1776649789929538, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929564, "dur": 29, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1555 + } + }, + { + "ph": "f", "id": 1555, "pid": 435, "tid": 435, "ts": 1776649789929564, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929594, "dur": 29, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1556 + } + }, + { + "ph": "f", "id": 1556, "pid": 435, "tid": 435, "ts": 1776649789929594, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929624, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1557 + } + }, + { + "ph": "f", "id": 1557, "pid": 435, "tid": 435, "ts": 1776649789929624, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929652, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1558 + } + }, + { + "ph": "f", "id": 1558, "pid": 435, "tid": 435, "ts": 1776649789929652, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929675, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1559 + } + }, + { + "ph": "f", "id": 1559, "pid": 435, "tid": 435, "ts": 1776649789929675, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929701, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1560 + } + }, + { + "ph": "f", "id": 1560, "pid": 435, "tid": 435, "ts": 1776649789929701, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929729, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1561 + } + }, + { + "ph": "f", "id": 1561, "pid": 435, "tid": 435, "ts": 1776649789929729, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929756, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1562 + } + }, + { + "ph": "f", "id": 1562, "pid": 435, "tid": 435, "ts": 1776649789929756, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789929781, "dur": 394, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1563 + } + }, + { + "ph": "f", "id": 1563, "pid": 435, "tid": 435, "ts": 1776649789929781, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789930176, "dur": 32, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1564 + } + }, + { + "ph": "f", "id": 1564, "pid": 435, "tid": 435, "ts": 1776649789930176, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789930219, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1565 + } + }, + { + "ph": "f", "id": 1565, "pid": 435, "tid": 435, "ts": 1776649789930219, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789930247, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1566 + } + }, + { + "ph": "f", "id": 1566, "pid": 435, "tid": 435, "ts": 1776649789930247, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789930274, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1567 + } + }, + { + "ph": "f", "id": 1567, "pid": 435, "tid": 435, "ts": 1776649789930274, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789930299, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1568 + } + }, + { + "ph": "f", "id": 1568, "pid": 435, "tid": 435, "ts": 1776649789930299, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789930324, "dur": 32, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1569 + } + }, + { + "ph": "f", "id": 1569, "pid": 435, "tid": 435, "ts": 1776649789930324, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789930357, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1570 + } + }, + { + "ph": "f", "id": 1570, "pid": 435, "tid": 435, "ts": 1776649789930357, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789930386, "dur": 29, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1571 + } + }, + { + "ph": "f", "id": 1571, "pid": 435, "tid": 435, "ts": 1776649789930386, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789930415, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1572 + } + }, + { + "ph": "f", "id": 1572, "pid": 435, "tid": 435, "ts": 1776649789930415, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789930439, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1573 + } + }, + { + "ph": "f", "id": 1573, "pid": 435, "tid": 435, "ts": 1776649789930439, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789930462, "dur": 28, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1574 + } + }, + { + "ph": "f", "id": 1574, "pid": 435, "tid": 435, "ts": 1776649789930462, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789930491, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1575 + } + }, + { + "ph": "f", "id": 1575, "pid": 435, "tid": 435, "ts": 1776649789930491, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789930514, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1576 + } + }, + { + "ph": "f", "id": 1576, "pid": 435, "tid": 435, "ts": 1776649789930514, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789930541, "dur": 10016, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1577 + } + }, + { + "ph": "f", "id": 1577, "pid": 435, "tid": 435, "ts": 1776649789930541, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789940562, "dur": 33, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1578 + } + }, + { + "ph": "f", "id": 1578, "pid": 435, "tid": 435, "ts": 1776649789940562, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789940596, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1579 + } + }, + { + "ph": "f", "id": 1579, "pid": 435, "tid": 435, "ts": 1776649789940596, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789940619, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1580 + } + }, + { + "ph": "f", "id": 1580, "pid": 435, "tid": 435, "ts": 1776649789940619, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789940645, "dur": 40, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1581 + } + }, + { + "ph": "f", "id": 1581, "pid": 435, "tid": 435, "ts": 1776649789940645, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789940685, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1582 + } + }, + { + "ph": "f", "id": 1582, "pid": 435, "tid": 435, "ts": 1776649789940685, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789940709, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1583 + } + }, + { + "ph": "f", "id": 1583, "pid": 435, "tid": 435, "ts": 1776649789940709, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789940737, "dur": 28, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1584 + } + }, + { + "ph": "f", "id": 1584, "pid": 435, "tid": 435, "ts": 1776649789940737, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789940765, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1585 + } + }, + { + "ph": "f", "id": 1585, "pid": 435, "tid": 435, "ts": 1776649789940765, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789940790, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1586 + } + }, + { + "ph": "f", "id": 1586, "pid": 435, "tid": 435, "ts": 1776649789940790, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789940818, "dur": 28, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1587 + } + }, + { + "ph": "f", "id": 1587, "pid": 435, "tid": 435, "ts": 1776649789940818, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789940847, "dur": 29, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1588 + } + }, + { + "ph": "f", "id": 1588, "pid": 435, "tid": 435, "ts": 1776649789940847, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789940876, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1589 + } + }, + { + "ph": "f", "id": 1589, "pid": 435, "tid": 435, "ts": 1776649789940876, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789940903, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1590 + } + }, + { + "ph": "f", "id": 1590, "pid": 435, "tid": 435, "ts": 1776649789940903, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789940930, "dur": 35, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1591 + } + }, + { + "ph": "f", "id": 1591, "pid": 435, "tid": 435, "ts": 1776649789940930, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789940966, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1592 + } + }, + { + "ph": "f", "id": 1592, "pid": 435, "tid": 435, "ts": 1776649789940966, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789940994, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1593 + } + }, + { + "ph": "f", "id": 1593, "pid": 435, "tid": 435, "ts": 1776649789940994, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941022, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1594 + } + }, + { + "ph": "f", "id": 1594, "pid": 435, "tid": 435, "ts": 1776649789941022, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941049, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1595 + } + }, + { + "ph": "f", "id": 1595, "pid": 435, "tid": 435, "ts": 1776649789941049, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941076, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1596 + } + }, + { + "ph": "f", "id": 1596, "pid": 435, "tid": 435, "ts": 1776649789941076, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941100, "dur": 29, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1597 + } + }, + { + "ph": "f", "id": 1597, "pid": 435, "tid": 435, "ts": 1776649789941100, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941130, "dur": 28, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1598 + } + }, + { + "ph": "f", "id": 1598, "pid": 435, "tid": 435, "ts": 1776649789941130, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941159, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1599 + } + }, + { + "ph": "f", "id": 1599, "pid": 435, "tid": 435, "ts": 1776649789941159, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941187, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1600 + } + }, + { + "ph": "f", "id": 1600, "pid": 435, "tid": 435, "ts": 1776649789941187, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941215, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1601 + } + }, + { + "ph": "f", "id": 1601, "pid": 435, "tid": 435, "ts": 1776649789941215, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941240, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1602 + } + }, + { + "ph": "f", "id": 1602, "pid": 435, "tid": 435, "ts": 1776649789941240, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941264, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1603 + } + }, + { + "ph": "f", "id": 1603, "pid": 435, "tid": 435, "ts": 1776649789941264, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941290, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1604 + } + }, + { + "ph": "f", "id": 1604, "pid": 435, "tid": 435, "ts": 1776649789941290, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941316, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1605 + } + }, + { + "ph": "f", "id": 1605, "pid": 435, "tid": 435, "ts": 1776649789941316, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941344, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1606 + } + }, + { + "ph": "f", "id": 1606, "pid": 435, "tid": 435, "ts": 1776649789941344, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941370, "dur": 32, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1607 + } + }, + { + "ph": "f", "id": 1607, "pid": 435, "tid": 435, "ts": 1776649789941370, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941403, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1608 + } + }, + { + "ph": "f", "id": 1608, "pid": 435, "tid": 435, "ts": 1776649789941403, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941427, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1609 + } + }, + { + "ph": "f", "id": 1609, "pid": 435, "tid": 435, "ts": 1776649789941427, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941450, "dur": 19, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1610 + } + }, + { + "ph": "f", "id": 1610, "pid": 435, "tid": 435, "ts": 1776649789941450, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941470, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1611 + } + }, + { + "ph": "f", "id": 1611, "pid": 435, "tid": 435, "ts": 1776649789941470, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941496, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1612 + } + }, + { + "ph": "f", "id": 1612, "pid": 435, "tid": 435, "ts": 1776649789941496, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941521, "dur": 21, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1613 + } + }, + { + "ph": "f", "id": 1613, "pid": 435, "tid": 435, "ts": 1776649789941521, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941543, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1614 + } + }, + { + "ph": "f", "id": 1614, "pid": 435, "tid": 435, "ts": 1776649789941543, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941565, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1615 + } + }, + { + "ph": "f", "id": 1615, "pid": 435, "tid": 435, "ts": 1776649789941565, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941589, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1616 + } + }, + { + "ph": "f", "id": 1616, "pid": 435, "tid": 435, "ts": 1776649789941589, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941616, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1617 + } + }, + { + "ph": "f", "id": 1617, "pid": 435, "tid": 435, "ts": 1776649789941616, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941639, "dur": 21, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1618 + } + }, + { + "ph": "f", "id": 1618, "pid": 435, "tid": 435, "ts": 1776649789941639, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941661, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1619 + } + }, + { + "ph": "f", "id": 1619, "pid": 435, "tid": 435, "ts": 1776649789941661, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941685, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1620 + } + }, + { + "ph": "f", "id": 1620, "pid": 435, "tid": 435, "ts": 1776649789941685, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941711, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1621 + } + }, + { + "ph": "f", "id": 1621, "pid": 435, "tid": 435, "ts": 1776649789941711, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941735, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1622 + } + }, + { + "ph": "f", "id": 1622, "pid": 435, "tid": 435, "ts": 1776649789941735, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941762, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1623 + } + }, + { + "ph": "f", "id": 1623, "pid": 435, "tid": 435, "ts": 1776649789941762, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941788, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1624 + } + }, + { + "ph": "f", "id": 1624, "pid": 435, "tid": 435, "ts": 1776649789941788, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941815, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1625 + } + }, + { + "ph": "f", "id": 1625, "pid": 435, "tid": 435, "ts": 1776649789941815, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941839, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1626 + } + }, + { + "ph": "f", "id": 1626, "pid": 435, "tid": 435, "ts": 1776649789941839, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941862, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1627 + } + }, + { + "ph": "f", "id": 1627, "pid": 435, "tid": 435, "ts": 1776649789941862, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941888, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1628 + } + }, + { + "ph": "f", "id": 1628, "pid": 435, "tid": 435, "ts": 1776649789941888, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941911, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1629 + } + }, + { + "ph": "f", "id": 1629, "pid": 435, "tid": 435, "ts": 1776649789941911, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941935, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1630 + } + }, + { + "ph": "f", "id": 1630, "pid": 435, "tid": 435, "ts": 1776649789941935, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941958, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1631 + } + }, + { + "ph": "f", "id": 1631, "pid": 435, "tid": 435, "ts": 1776649789941958, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789941984, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1632 + } + }, + { + "ph": "f", "id": 1632, "pid": 435, "tid": 435, "ts": 1776649789941984, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789942008, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1633 + } + }, + { + "ph": "f", "id": 1633, "pid": 435, "tid": 435, "ts": 1776649789942008, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789942033, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1634 + } + }, + { + "ph": "f", "id": 1634, "pid": 435, "tid": 435, "ts": 1776649789942033, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789942056, "dur": 9306, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1635 + } + }, + { + "ph": "f", "id": 1635, "pid": 435, "tid": 435, "ts": 1776649789942056, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789951367, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1636 + } + }, + { + "ph": "f", "id": 1636, "pid": 435, "tid": 435, "ts": 1776649789951367, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789951394, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1637 + } + }, + { + "ph": "f", "id": 1637, "pid": 435, "tid": 435, "ts": 1776649789951394, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789951419, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1638 + } + }, + { + "ph": "f", "id": 1638, "pid": 435, "tid": 435, "ts": 1776649789951419, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789951446, "dur": 29, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1639 + } + }, + { + "ph": "f", "id": 1639, "pid": 435, "tid": 435, "ts": 1776649789951446, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789951476, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1640 + } + }, + { + "ph": "f", "id": 1640, "pid": 435, "tid": 435, "ts": 1776649789951476, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789951501, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1641 + } + }, + { + "ph": "f", "id": 1641, "pid": 435, "tid": 435, "ts": 1776649789951501, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789951525, "dur": 32, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1642 + } + }, + { + "ph": "f", "id": 1642, "pid": 435, "tid": 435, "ts": 1776649789951525, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789951557, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1643 + } + }, + { + "ph": "f", "id": 1643, "pid": 435, "tid": 435, "ts": 1776649789951557, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789951582, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1644 + } + }, + { + "ph": "f", "id": 1644, "pid": 435, "tid": 435, "ts": 1776649789951582, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789951609, "dur": 29, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1645 + } + }, + { + "ph": "f", "id": 1645, "pid": 435, "tid": 435, "ts": 1776649789951609, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789951639, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1646 + } + }, + { + "ph": "f", "id": 1646, "pid": 435, "tid": 435, "ts": 1776649789951639, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789951664, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1647 + } + }, + { + "ph": "f", "id": 1647, "pid": 435, "tid": 435, "ts": 1776649789951664, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789951691, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1648 + } + }, + { + "ph": "f", "id": 1648, "pid": 435, "tid": 435, "ts": 1776649789951691, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789951717, "dur": 31, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1649 + } + }, + { + "ph": "f", "id": 1649, "pid": 435, "tid": 435, "ts": 1776649789951717, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789951749, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1650 + } + }, + { + "ph": "f", "id": 1650, "pid": 435, "tid": 435, "ts": 1776649789951749, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789951777, "dur": 30, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1651 + } + }, + { + "ph": "f", "id": 1651, "pid": 435, "tid": 435, "ts": 1776649789951777, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789951808, "dur": 29, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1652 + } + }, + { + "ph": "f", "id": 1652, "pid": 435, "tid": 435, "ts": 1776649789951808, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789951838, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1653 + } + }, + { + "ph": "f", "id": 1653, "pid": 435, "tid": 435, "ts": 1776649789951838, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789951865, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1654 + } + }, + { + "ph": "f", "id": 1654, "pid": 435, "tid": 435, "ts": 1776649789951865, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789951891, "dur": 30, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1655 + } + }, + { + "ph": "f", "id": 1655, "pid": 435, "tid": 435, "ts": 1776649789951891, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789951922, "dur": 31, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1656 + } + }, + { + "ph": "f", "id": 1656, "pid": 435, "tid": 435, "ts": 1776649789951922, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789951954, "dur": 35, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1657 + } + }, + { + "ph": "f", "id": 1657, "pid": 435, "tid": 435, "ts": 1776649789951954, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789951989, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1658 + } + }, + { + "ph": "f", "id": 1658, "pid": 435, "tid": 435, "ts": 1776649789951989, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789952014, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1659 + } + }, + { + "ph": "f", "id": 1659, "pid": 435, "tid": 435, "ts": 1776649789952014, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789952039, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1660 + } + }, + { + "ph": "f", "id": 1660, "pid": 435, "tid": 435, "ts": 1776649789952039, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789952064, "dur": 28, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1661 + } + }, + { + "ph": "f", "id": 1661, "pid": 435, "tid": 435, "ts": 1776649789952064, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789952092, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1662 + } + }, + { + "ph": "f", "id": 1662, "pid": 435, "tid": 435, "ts": 1776649789952092, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789952117, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1663 + } + }, + { + "ph": "f", "id": 1663, "pid": 435, "tid": 435, "ts": 1776649789952117, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789952144, "dur": 36, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1664 + } + }, + { + "ph": "f", "id": 1664, "pid": 435, "tid": 435, "ts": 1776649789952144, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789952181, "dur": 20, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1665 + } + }, + { + "ph": "f", "id": 1665, "pid": 435, "tid": 435, "ts": 1776649789952181, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789952201, "dur": 40, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1666 + } + }, + { + "ph": "f", "id": 1666, "pid": 435, "tid": 435, "ts": 1776649789952201, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789952242, "dur": 21, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1667 + } + }, + { + "ph": "f", "id": 1667, "pid": 435, "tid": 435, "ts": 1776649789952242, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789952264, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1668 + } + }, + { + "ph": "f", "id": 1668, "pid": 435, "tid": 435, "ts": 1776649789952264, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789952288, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1669 + } + }, + { + "ph": "f", "id": 1669, "pid": 435, "tid": 435, "ts": 1776649789952288, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789952315, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1670 + } + }, + { + "ph": "f", "id": 1670, "pid": 435, "tid": 435, "ts": 1776649789952315, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789952338, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1671 + } + }, + { + "ph": "f", "id": 1671, "pid": 435, "tid": 435, "ts": 1776649789952338, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789952361, "dur": 20, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1672 + } + }, + { + "ph": "f", "id": 1672, "pid": 435, "tid": 435, "ts": 1776649789952361, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789952382, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1673 + } + }, + { + "ph": "f", "id": 1673, "pid": 435, "tid": 435, "ts": 1776649789952382, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789952409, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1674 + } + }, + { + "ph": "f", "id": 1674, "pid": 435, "tid": 435, "ts": 1776649789952409, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789952435, "dur": 21, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1675 + } + }, + { + "ph": "f", "id": 1675, "pid": 435, "tid": 435, "ts": 1776649789952435, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789952456, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1676 + } + }, + { + "ph": "f", "id": 1676, "pid": 435, "tid": 435, "ts": 1776649789952456, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789952479, "dur": 378, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1677 + } + }, + { + "ph": "f", "id": 1677, "pid": 435, "tid": 435, "ts": 1776649789952479, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789952859, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1678 + } + }, + { + "ph": "f", "id": 1678, "pid": 435, "tid": 435, "ts": 1776649789952859, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789952886, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1679 + } + }, + { + "ph": "f", "id": 1679, "pid": 435, "tid": 435, "ts": 1776649789952886, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789952910, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1680 + } + }, + { + "ph": "f", "id": 1680, "pid": 435, "tid": 435, "ts": 1776649789952910, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789952933, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1681 + } + }, + { + "ph": "f", "id": 1681, "pid": 435, "tid": 435, "ts": 1776649789952933, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789952960, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1682 + } + }, + { + "ph": "f", "id": 1682, "pid": 435, "tid": 435, "ts": 1776649789952960, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789952985, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1683 + } + }, + { + "ph": "f", "id": 1683, "pid": 435, "tid": 435, "ts": 1776649789952985, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789953012, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1684 + } + }, + { + "ph": "f", "id": 1684, "pid": 435, "tid": 435, "ts": 1776649789953012, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789953034, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1685 + } + }, + { + "ph": "f", "id": 1685, "pid": 435, "tid": 435, "ts": 1776649789953034, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789953060, "dur": 21, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1686 + } + }, + { + "ph": "f", "id": 1686, "pid": 435, "tid": 435, "ts": 1776649789953060, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789953082, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1687 + } + }, + { + "ph": "f", "id": 1687, "pid": 435, "tid": 435, "ts": 1776649789953082, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789953105, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1688 + } + }, + { + "ph": "f", "id": 1688, "pid": 435, "tid": 435, "ts": 1776649789953105, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789953130, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1689 + } + }, + { + "ph": "f", "id": 1689, "pid": 435, "tid": 435, "ts": 1776649789953130, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789953152, "dur": 19, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1690 + } + }, + { + "ph": "f", "id": 1690, "pid": 435, "tid": 435, "ts": 1776649789953152, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789953172, "dur": 19, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1691 + } + }, + { + "ph": "f", "id": 1691, "pid": 435, "tid": 435, "ts": 1776649789953172, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789953192, "dur": 19, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1692 + } + }, + { + "ph": "f", "id": 1692, "pid": 435, "tid": 435, "ts": 1776649789953192, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789953212, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1693 + } + }, + { + "ph": "f", "id": 1693, "pid": 435, "tid": 435, "ts": 1776649789953212, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789953238, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1694 + } + }, + { + "ph": "f", "id": 1694, "pid": 435, "tid": 435, "ts": 1776649789953238, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789953261, "dur": 20, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1695 + } + }, + { + "ph": "f", "id": 1695, "pid": 435, "tid": 435, "ts": 1776649789953261, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789953282, "dur": 20, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1696 + } + }, + { + "ph": "f", "id": 1696, "pid": 435, "tid": 435, "ts": 1776649789953282, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789953302, "dur": 20, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1697 + } + }, + { + "ph": "f", "id": 1697, "pid": 435, "tid": 435, "ts": 1776649789953302, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789953323, "dur": 19, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1698 + } + }, + { + "ph": "f", "id": 1698, "pid": 435, "tid": 435, "ts": 1776649789953323, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789953342, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1699 + } + }, + { + "ph": "f", "id": 1699, "pid": 435, "tid": 435, "ts": 1776649789953342, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789953365, "dur": 19, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1700 + } + }, + { + "ph": "f", "id": 1700, "pid": 435, "tid": 435, "ts": 1776649789953365, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789953385, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1701 + } + }, + { + "ph": "f", "id": 1701, "pid": 435, "tid": 435, "ts": 1776649789953385, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789953411, "dur": 19, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1702 + } + }, + { + "ph": "f", "id": 1702, "pid": 435, "tid": 435, "ts": 1776649789953411, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789953432, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1703 + } + }, + { + "ph": "f", "id": 1703, "pid": 435, "tid": 435, "ts": 1776649789953432, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789953455, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1704 + } + }, + { + "ph": "f", "id": 1704, "pid": 435, "tid": 435, "ts": 1776649789953455, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789953480, "dur": 9307, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1705 + } + }, + { + "ph": "f", "id": 1705, "pid": 435, "tid": 435, "ts": 1776649789953480, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789962792, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1706 + } + }, + { + "ph": "f", "id": 1706, "pid": 435, "tid": 435, "ts": 1776649789962792, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789962820, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1707 + } + }, + { + "ph": "f", "id": 1707, "pid": 435, "tid": 435, "ts": 1776649789962820, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789962847, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1708 + } + }, + { + "ph": "f", "id": 1708, "pid": 435, "tid": 435, "ts": 1776649789962847, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789962872, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1709 + } + }, + { + "ph": "f", "id": 1709, "pid": 435, "tid": 435, "ts": 1776649789962872, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789962898, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1710 + } + }, + { + "ph": "f", "id": 1710, "pid": 435, "tid": 435, "ts": 1776649789962898, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789962926, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1711 + } + }, + { + "ph": "f", "id": 1711, "pid": 435, "tid": 435, "ts": 1776649789962926, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789962950, "dur": 29, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1712 + } + }, + { + "ph": "f", "id": 1712, "pid": 435, "tid": 435, "ts": 1776649789962950, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789962980, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1713 + } + }, + { + "ph": "f", "id": 1713, "pid": 435, "tid": 435, "ts": 1776649789962980, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963008, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1714 + } + }, + { + "ph": "f", "id": 1714, "pid": 435, "tid": 435, "ts": 1776649789963008, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963035, "dur": 30, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1715 + } + }, + { + "ph": "f", "id": 1715, "pid": 435, "tid": 435, "ts": 1776649789963035, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963065, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1716 + } + }, + { + "ph": "f", "id": 1716, "pid": 435, "tid": 435, "ts": 1776649789963065, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963089, "dur": 28, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1717 + } + }, + { + "ph": "f", "id": 1717, "pid": 435, "tid": 435, "ts": 1776649789963089, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963119, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1718 + } + }, + { + "ph": "f", "id": 1718, "pid": 435, "tid": 435, "ts": 1776649789963119, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963145, "dur": 29, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1719 + } + }, + { + "ph": "f", "id": 1719, "pid": 435, "tid": 435, "ts": 1776649789963145, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963175, "dur": 28, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1720 + } + }, + { + "ph": "f", "id": 1720, "pid": 435, "tid": 435, "ts": 1776649789963175, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963204, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1721 + } + }, + { + "ph": "f", "id": 1721, "pid": 435, "tid": 435, "ts": 1776649789963204, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963230, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1722 + } + }, + { + "ph": "f", "id": 1722, "pid": 435, "tid": 435, "ts": 1776649789963230, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963254, "dur": 28, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1723 + } + }, + { + "ph": "f", "id": 1723, "pid": 435, "tid": 435, "ts": 1776649789963254, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963284, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1724 + } + }, + { + "ph": "f", "id": 1724, "pid": 435, "tid": 435, "ts": 1776649789963284, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963307, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1725 + } + }, + { + "ph": "f", "id": 1725, "pid": 435, "tid": 435, "ts": 1776649789963307, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963332, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1726 + } + }, + { + "ph": "f", "id": 1726, "pid": 435, "tid": 435, "ts": 1776649789963332, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963360, "dur": 28, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1727 + } + }, + { + "ph": "f", "id": 1727, "pid": 435, "tid": 435, "ts": 1776649789963360, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963388, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1728 + } + }, + { + "ph": "f", "id": 1728, "pid": 435, "tid": 435, "ts": 1776649789963388, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963414, "dur": 31, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1729 + } + }, + { + "ph": "f", "id": 1729, "pid": 435, "tid": 435, "ts": 1776649789963414, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963446, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1730 + } + }, + { + "ph": "f", "id": 1730, "pid": 435, "tid": 435, "ts": 1776649789963446, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963469, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1731 + } + }, + { + "ph": "f", "id": 1731, "pid": 435, "tid": 435, "ts": 1776649789963469, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963492, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1732 + } + }, + { + "ph": "f", "id": 1732, "pid": 435, "tid": 435, "ts": 1776649789963492, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963515, "dur": 41, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1733 + } + }, + { + "ph": "f", "id": 1733, "pid": 435, "tid": 435, "ts": 1776649789963515, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963557, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1734 + } + }, + { + "ph": "f", "id": 1734, "pid": 435, "tid": 435, "ts": 1776649789963557, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963580, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1735 + } + }, + { + "ph": "f", "id": 1735, "pid": 435, "tid": 435, "ts": 1776649789963580, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963608, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1736 + } + }, + { + "ph": "f", "id": 1736, "pid": 435, "tid": 435, "ts": 1776649789963608, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963631, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1737 + } + }, + { + "ph": "f", "id": 1737, "pid": 435, "tid": 435, "ts": 1776649789963631, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963656, "dur": 19, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1738 + } + }, + { + "ph": "f", "id": 1738, "pid": 435, "tid": 435, "ts": 1776649789963656, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963676, "dur": 20, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1739 + } + }, + { + "ph": "f", "id": 1739, "pid": 435, "tid": 435, "ts": 1776649789963676, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963697, "dur": 20, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1740 + } + }, + { + "ph": "f", "id": 1740, "pid": 435, "tid": 435, "ts": 1776649789963697, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963717, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1741 + } + }, + { + "ph": "f", "id": 1741, "pid": 435, "tid": 435, "ts": 1776649789963717, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963742, "dur": 21, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1742 + } + }, + { + "ph": "f", "id": 1742, "pid": 435, "tid": 435, "ts": 1776649789963742, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963763, "dur": 21, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1743 + } + }, + { + "ph": "f", "id": 1743, "pid": 435, "tid": 435, "ts": 1776649789963763, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963785, "dur": 19, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1744 + } + }, + { + "ph": "f", "id": 1744, "pid": 435, "tid": 435, "ts": 1776649789963785, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963805, "dur": 28, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1745 + } + }, + { + "ph": "f", "id": 1745, "pid": 435, "tid": 435, "ts": 1776649789963805, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963834, "dur": 28, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1746 + } + }, + { + "ph": "f", "id": 1746, "pid": 435, "tid": 435, "ts": 1776649789963834, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963862, "dur": 19, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1747 + } + }, + { + "ph": "f", "id": 1747, "pid": 435, "tid": 435, "ts": 1776649789963862, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963882, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1748 + } + }, + { + "ph": "f", "id": 1748, "pid": 435, "tid": 435, "ts": 1776649789963882, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963905, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1749 + } + }, + { + "ph": "f", "id": 1749, "pid": 435, "tid": 435, "ts": 1776649789963905, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963930, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1750 + } + }, + { + "ph": "f", "id": 1750, "pid": 435, "tid": 435, "ts": 1776649789963930, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963953, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1751 + } + }, + { + "ph": "f", "id": 1751, "pid": 435, "tid": 435, "ts": 1776649789963953, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963976, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1752 + } + }, + { + "ph": "f", "id": 1752, "pid": 435, "tid": 435, "ts": 1776649789963976, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789963999, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1753 + } + }, + { + "ph": "f", "id": 1753, "pid": 435, "tid": 435, "ts": 1776649789963999, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789964025, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1754 + } + }, + { + "ph": "f", "id": 1754, "pid": 435, "tid": 435, "ts": 1776649789964025, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789964050, "dur": 20, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1755 + } + }, + { + "ph": "f", "id": 1755, "pid": 435, "tid": 435, "ts": 1776649789964050, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789964071, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1756 + } + }, + { + "ph": "f", "id": 1756, "pid": 435, "tid": 435, "ts": 1776649789964071, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789964094, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1757 + } + }, + { + "ph": "f", "id": 1757, "pid": 435, "tid": 435, "ts": 1776649789964094, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789964120, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1758 + } + }, + { + "ph": "f", "id": 1758, "pid": 435, "tid": 435, "ts": 1776649789964120, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789964142, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1759 + } + }, + { + "ph": "f", "id": 1759, "pid": 435, "tid": 435, "ts": 1776649789964142, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789964167, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1760 + } + }, + { + "ph": "f", "id": 1760, "pid": 435, "tid": 435, "ts": 1776649789964167, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789964191, "dur": 9160, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1761 + } + }, + { + "ph": "f", "id": 1761, "pid": 435, "tid": 435, "ts": 1776649789964191, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789973356, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1762 + } + }, + { + "ph": "f", "id": 1762, "pid": 435, "tid": 435, "ts": 1776649789973356, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789973381, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1763 + } + }, + { + "ph": "f", "id": 1763, "pid": 435, "tid": 435, "ts": 1776649789973381, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789973407, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1764 + } + }, + { + "ph": "f", "id": 1764, "pid": 435, "tid": 435, "ts": 1776649789973407, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789973432, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1765 + } + }, + { + "ph": "f", "id": 1765, "pid": 435, "tid": 435, "ts": 1776649789973432, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789973458, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1766 + } + }, + { + "ph": "f", "id": 1766, "pid": 435, "tid": 435, "ts": 1776649789973458, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789973483, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1767 + } + }, + { + "ph": "f", "id": 1767, "pid": 435, "tid": 435, "ts": 1776649789973483, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789973508, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1768 + } + }, + { + "ph": "f", "id": 1768, "pid": 435, "tid": 435, "ts": 1776649789973508, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789973532, "dur": 29, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1769 + } + }, + { + "ph": "f", "id": 1769, "pid": 435, "tid": 435, "ts": 1776649789973532, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789973561, "dur": 31, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1770 + } + }, + { + "ph": "f", "id": 1770, "pid": 435, "tid": 435, "ts": 1776649789973561, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789973592, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1771 + } + }, + { + "ph": "f", "id": 1771, "pid": 435, "tid": 435, "ts": 1776649789973592, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789973619, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1772 + } + }, + { + "ph": "f", "id": 1772, "pid": 435, "tid": 435, "ts": 1776649789973619, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789973643, "dur": 28, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1773 + } + }, + { + "ph": "f", "id": 1773, "pid": 435, "tid": 435, "ts": 1776649789973643, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789973671, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1774 + } + }, + { + "ph": "f", "id": 1774, "pid": 435, "tid": 435, "ts": 1776649789973671, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789973695, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1775 + } + }, + { + "ph": "f", "id": 1775, "pid": 435, "tid": 435, "ts": 1776649789973695, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789973722, "dur": 31, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1776 + } + }, + { + "ph": "f", "id": 1776, "pid": 435, "tid": 435, "ts": 1776649789973722, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789973754, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1777 + } + }, + { + "ph": "f", "id": 1777, "pid": 435, "tid": 435, "ts": 1776649789973754, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789973780, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1778 + } + }, + { + "ph": "f", "id": 1778, "pid": 435, "tid": 435, "ts": 1776649789973780, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789973806, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1779 + } + }, + { + "ph": "f", "id": 1779, "pid": 435, "tid": 435, "ts": 1776649789973806, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789973832, "dur": 29, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1780 + } + }, + { + "ph": "f", "id": 1780, "pid": 435, "tid": 435, "ts": 1776649789973832, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789973861, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1781 + } + }, + { + "ph": "f", "id": 1781, "pid": 435, "tid": 435, "ts": 1776649789973861, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789973888, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1782 + } + }, + { + "ph": "f", "id": 1782, "pid": 435, "tid": 435, "ts": 1776649789973888, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789973915, "dur": 37, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1783 + } + }, + { + "ph": "f", "id": 1783, "pid": 435, "tid": 435, "ts": 1776649789973915, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789973953, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1784 + } + }, + { + "ph": "f", "id": 1784, "pid": 435, "tid": 435, "ts": 1776649789973953, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789973981, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1785 + } + }, + { + "ph": "f", "id": 1785, "pid": 435, "tid": 435, "ts": 1776649789973981, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974004, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1786 + } + }, + { + "ph": "f", "id": 1786, "pid": 435, "tid": 435, "ts": 1776649789974004, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974031, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1787 + } + }, + { + "ph": "f", "id": 1787, "pid": 435, "tid": 435, "ts": 1776649789974031, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974057, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1788 + } + }, + { + "ph": "f", "id": 1788, "pid": 435, "tid": 435, "ts": 1776649789974057, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974081, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1789 + } + }, + { + "ph": "f", "id": 1789, "pid": 435, "tid": 435, "ts": 1776649789974081, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974107, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1790 + } + }, + { + "ph": "f", "id": 1790, "pid": 435, "tid": 435, "ts": 1776649789974107, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974133, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1791 + } + }, + { + "ph": "f", "id": 1791, "pid": 435, "tid": 435, "ts": 1776649789974133, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974159, "dur": 28, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1792 + } + }, + { + "ph": "f", "id": 1792, "pid": 435, "tid": 435, "ts": 1776649789974159, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974188, "dur": 20, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1793 + } + }, + { + "ph": "f", "id": 1793, "pid": 435, "tid": 435, "ts": 1776649789974188, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974209, "dur": 19, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1794 + } + }, + { + "ph": "f", "id": 1794, "pid": 435, "tid": 435, "ts": 1776649789974209, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974229, "dur": 21, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1795 + } + }, + { + "ph": "f", "id": 1795, "pid": 435, "tid": 435, "ts": 1776649789974229, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974251, "dur": 20, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1796 + } + }, + { + "ph": "f", "id": 1796, "pid": 435, "tid": 435, "ts": 1776649789974251, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974271, "dur": 379, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1797 + } + }, + { + "ph": "f", "id": 1797, "pid": 435, "tid": 435, "ts": 1776649789974271, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974651, "dur": 21, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1798 + } + }, + { + "ph": "f", "id": 1798, "pid": 435, "tid": 435, "ts": 1776649789974651, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974673, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1799 + } + }, + { + "ph": "f", "id": 1799, "pid": 435, "tid": 435, "ts": 1776649789974673, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974697, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1800 + } + }, + { + "ph": "f", "id": 1800, "pid": 435, "tid": 435, "ts": 1776649789974697, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974720, "dur": 30, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1801 + } + }, + { + "ph": "f", "id": 1801, "pid": 435, "tid": 435, "ts": 1776649789974720, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974750, "dur": 20, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1802 + } + }, + { + "ph": "f", "id": 1802, "pid": 435, "tid": 435, "ts": 1776649789974750, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974771, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1803 + } + }, + { + "ph": "f", "id": 1803, "pid": 435, "tid": 435, "ts": 1776649789974771, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974794, "dur": 21, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1804 + } + }, + { + "ph": "f", "id": 1804, "pid": 435, "tid": 435, "ts": 1776649789974794, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974816, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1805 + } + }, + { + "ph": "f", "id": 1805, "pid": 435, "tid": 435, "ts": 1776649789974816, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974842, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1806 + } + }, + { + "ph": "f", "id": 1806, "pid": 435, "tid": 435, "ts": 1776649789974842, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974864, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1807 + } + }, + { + "ph": "f", "id": 1807, "pid": 435, "tid": 435, "ts": 1776649789974864, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974888, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1808 + } + }, + { + "ph": "f", "id": 1808, "pid": 435, "tid": 435, "ts": 1776649789974888, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974910, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1809 + } + }, + { + "ph": "f", "id": 1809, "pid": 435, "tid": 435, "ts": 1776649789974910, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974934, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1810 + } + }, + { + "ph": "f", "id": 1810, "pid": 435, "tid": 435, "ts": 1776649789974934, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974960, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1811 + } + }, + { + "ph": "f", "id": 1811, "pid": 435, "tid": 435, "ts": 1776649789974960, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789974982, "dur": 19, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1812 + } + }, + { + "ph": "f", "id": 1812, "pid": 435, "tid": 435, "ts": 1776649789974982, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975003, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1813 + } + }, + { + "ph": "f", "id": 1813, "pid": 435, "tid": 435, "ts": 1776649789975003, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975028, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1814 + } + }, + { + "ph": "f", "id": 1814, "pid": 435, "tid": 435, "ts": 1776649789975028, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975051, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1815 + } + }, + { + "ph": "f", "id": 1815, "pid": 435, "tid": 435, "ts": 1776649789975051, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975074, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1816 + } + }, + { + "ph": "f", "id": 1816, "pid": 435, "tid": 435, "ts": 1776649789975074, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975102, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1817 + } + }, + { + "ph": "f", "id": 1817, "pid": 435, "tid": 435, "ts": 1776649789975102, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975125, "dur": 19, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1818 + } + }, + { + "ph": "f", "id": 1818, "pid": 435, "tid": 435, "ts": 1776649789975125, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975144, "dur": 19, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1819 + } + }, + { + "ph": "f", "id": 1819, "pid": 435, "tid": 435, "ts": 1776649789975144, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975164, "dur": 21, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1820 + } + }, + { + "ph": "f", "id": 1820, "pid": 435, "tid": 435, "ts": 1776649789975164, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975185, "dur": 19, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1821 + } + }, + { + "ph": "f", "id": 1821, "pid": 435, "tid": 435, "ts": 1776649789975185, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975205, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1822 + } + }, + { + "ph": "f", "id": 1822, "pid": 435, "tid": 435, "ts": 1776649789975205, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975229, "dur": 19, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1823 + } + }, + { + "ph": "f", "id": 1823, "pid": 435, "tid": 435, "ts": 1776649789975229, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975249, "dur": 21, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1824 + } + }, + { + "ph": "f", "id": 1824, "pid": 435, "tid": 435, "ts": 1776649789975249, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975271, "dur": 19, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1825 + } + }, + { + "ph": "f", "id": 1825, "pid": 435, "tid": 435, "ts": 1776649789975271, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975291, "dur": 19, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1826 + } + }, + { + "ph": "f", "id": 1826, "pid": 435, "tid": 435, "ts": 1776649789975291, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975311, "dur": 21, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1827 + } + }, + { + "ph": "f", "id": 1827, "pid": 435, "tid": 435, "ts": 1776649789975311, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975333, "dur": 19, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1828 + } + }, + { + "ph": "f", "id": 1828, "pid": 435, "tid": 435, "ts": 1776649789975333, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975353, "dur": 20, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1829 + } + }, + { + "ph": "f", "id": 1829, "pid": 435, "tid": 435, "ts": 1776649789975353, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975373, "dur": 19, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1830 + } + }, + { + "ph": "f", "id": 1830, "pid": 435, "tid": 435, "ts": 1776649789975373, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975393, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1831 + } + }, + { + "ph": "f", "id": 1831, "pid": 435, "tid": 435, "ts": 1776649789975393, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975416, "dur": 21, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1832 + } + }, + { + "ph": "f", "id": 1832, "pid": 435, "tid": 435, "ts": 1776649789975416, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975449, "dur": 28, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1833 + } + }, + { + "ph": "f", "id": 1833, "pid": 435, "tid": 435, "ts": 1776649789975449, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975478, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1834 + } + }, + { + "ph": "f", "id": 1834, "pid": 435, "tid": 435, "ts": 1776649789975478, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975505, "dur": 28, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1835 + } + }, + { + "ph": "f", "id": 1835, "pid": 435, "tid": 435, "ts": 1776649789975505, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975534, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1836 + } + }, + { + "ph": "f", "id": 1836, "pid": 435, "tid": 435, "ts": 1776649789975534, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975560, "dur": 55, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1837 + } + }, + { + "ph": "f", "id": 1837, "pid": 435, "tid": 435, "ts": 1776649789975560, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975616, "dur": 17, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1838 + } + }, + { + "ph": "f", "id": 1838, "pid": 435, "tid": 435, "ts": 1776649789975616, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975634, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1839 + } + }, + { + "ph": "f", "id": 1839, "pid": 435, "tid": 435, "ts": 1776649789975634, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975658, "dur": 28, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1840 + } + }, + { + "ph": "f", "id": 1840, "pid": 435, "tid": 435, "ts": 1776649789975658, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975687, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1841 + } + }, + { + "ph": "f", "id": 1841, "pid": 435, "tid": 435, "ts": 1776649789975687, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975714, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1842 + } + }, + { + "ph": "f", "id": 1842, "pid": 435, "tid": 435, "ts": 1776649789975714, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975739, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1843 + } + }, + { + "ph": "f", "id": 1843, "pid": 435, "tid": 435, "ts": 1776649789975739, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975763, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1844 + } + }, + { + "ph": "f", "id": 1844, "pid": 435, "tid": 435, "ts": 1776649789975763, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975787, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1845 + } + }, + { + "ph": "f", "id": 1845, "pid": 435, "tid": 435, "ts": 1776649789975787, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975814, "dur": 28, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1846 + } + }, + { + "ph": "f", "id": 1846, "pid": 435, "tid": 435, "ts": 1776649789975814, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975844, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1847 + } + }, + { + "ph": "f", "id": 1847, "pid": 435, "tid": 435, "ts": 1776649789975844, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975869, "dur": 29, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1848 + } + }, + { + "ph": "f", "id": 1848, "pid": 435, "tid": 435, "ts": 1776649789975869, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975898, "dur": 48, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1849 + } + }, + { + "ph": "f", "id": 1849, "pid": 435, "tid": 435, "ts": 1776649789975898, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975947, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1850 + } + }, + { + "ph": "f", "id": 1850, "pid": 435, "tid": 435, "ts": 1776649789975947, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975973, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1851 + } + }, + { + "ph": "f", "id": 1851, "pid": 435, "tid": 435, "ts": 1776649789975973, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789975998, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1852 + } + }, + { + "ph": "f", "id": 1852, "pid": 435, "tid": 435, "ts": 1776649789975998, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976024, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1853 + } + }, + { + "ph": "f", "id": 1853, "pid": 435, "tid": 435, "ts": 1776649789976024, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976051, "dur": 28, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1854 + } + }, + { + "ph": "f", "id": 1854, "pid": 435, "tid": 435, "ts": 1776649789976051, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976080, "dur": 29, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1855 + } + }, + { + "ph": "f", "id": 1855, "pid": 435, "tid": 435, "ts": 1776649789976080, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976109, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1856 + } + }, + { + "ph": "f", "id": 1856, "pid": 435, "tid": 435, "ts": 1776649789976109, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976133, "dur": 21, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1857 + } + }, + { + "ph": "f", "id": 1857, "pid": 435, "tid": 435, "ts": 1776649789976133, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976155, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1858 + } + }, + { + "ph": "f", "id": 1858, "pid": 435, "tid": 435, "ts": 1776649789976155, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976179, "dur": 21, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1859 + } + }, + { + "ph": "f", "id": 1859, "pid": 435, "tid": 435, "ts": 1776649789976179, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976200, "dur": 43, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1860 + } + }, + { + "ph": "f", "id": 1860, "pid": 435, "tid": 435, "ts": 1776649789976200, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976245, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1861 + } + }, + { + "ph": "f", "id": 1861, "pid": 435, "tid": 435, "ts": 1776649789976245, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976272, "dur": 20, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1862 + } + }, + { + "ph": "f", "id": 1862, "pid": 435, "tid": 435, "ts": 1776649789976272, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976293, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1863 + } + }, + { + "ph": "f", "id": 1863, "pid": 435, "tid": 435, "ts": 1776649789976293, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976320, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1864 + } + }, + { + "ph": "f", "id": 1864, "pid": 435, "tid": 435, "ts": 1776649789976320, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976346, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1865 + } + }, + { + "ph": "f", "id": 1865, "pid": 435, "tid": 435, "ts": 1776649789976346, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976372, "dur": 21, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1866 + } + }, + { + "ph": "f", "id": 1866, "pid": 435, "tid": 435, "ts": 1776649789976372, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976394, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1867 + } + }, + { + "ph": "f", "id": 1867, "pid": 435, "tid": 435, "ts": 1776649789976394, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976416, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1868 + } + }, + { + "ph": "f", "id": 1868, "pid": 435, "tid": 435, "ts": 1776649789976416, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976441, "dur": 21, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1869 + } + }, + { + "ph": "f", "id": 1869, "pid": 435, "tid": 435, "ts": 1776649789976441, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976463, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1870 + } + }, + { + "ph": "f", "id": 1870, "pid": 435, "tid": 435, "ts": 1776649789976463, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976487, "dur": 21, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1871 + } + }, + { + "ph": "f", "id": 1871, "pid": 435, "tid": 435, "ts": 1776649789976487, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976509, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1872 + } + }, + { + "ph": "f", "id": 1872, "pid": 435, "tid": 435, "ts": 1776649789976509, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976534, "dur": 32, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1873 + } + }, + { + "ph": "f", "id": 1873, "pid": 435, "tid": 435, "ts": 1776649789976534, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976567, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1874 + } + }, + { + "ph": "f", "id": 1874, "pid": 435, "tid": 435, "ts": 1776649789976567, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976592, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1875 + } + }, + { + "ph": "f", "id": 1875, "pid": 435, "tid": 435, "ts": 1776649789976592, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976619, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1876 + } + }, + { + "ph": "f", "id": 1876, "pid": 435, "tid": 435, "ts": 1776649789976619, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976645, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1877 + } + }, + { + "ph": "f", "id": 1877, "pid": 435, "tid": 435, "ts": 1776649789976645, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976673, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1878 + } + }, + { + "ph": "f", "id": 1878, "pid": 435, "tid": 435, "ts": 1776649789976673, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976699, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1879 + } + }, + { + "ph": "f", "id": 1879, "pid": 435, "tid": 435, "ts": 1776649789976699, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976723, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1880 + } + }, + { + "ph": "f", "id": 1880, "pid": 435, "tid": 435, "ts": 1776649789976723, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976752, "dur": 30, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1881 + } + }, + { + "ph": "f", "id": 1881, "pid": 435, "tid": 435, "ts": 1776649789976752, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976782, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1882 + } + }, + { + "ph": "f", "id": 1882, "pid": 435, "tid": 435, "ts": 1776649789976782, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976807, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1883 + } + }, + { + "ph": "f", "id": 1883, "pid": 435, "tid": 435, "ts": 1776649789976807, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976835, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1884 + } + }, + { + "ph": "f", "id": 1884, "pid": 435, "tid": 435, "ts": 1776649789976835, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976858, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1885 + } + }, + { + "ph": "f", "id": 1885, "pid": 435, "tid": 435, "ts": 1776649789976858, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976885, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1886 + } + }, + { + "ph": "f", "id": 1886, "pid": 435, "tid": 435, "ts": 1776649789976885, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976909, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1887 + } + }, + { + "ph": "f", "id": 1887, "pid": 435, "tid": 435, "ts": 1776649789976909, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976937, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1888 + } + }, + { + "ph": "f", "id": 1888, "pid": 435, "tid": 435, "ts": 1776649789976937, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976963, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1889 + } + }, + { + "ph": "f", "id": 1889, "pid": 435, "tid": 435, "ts": 1776649789976963, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789976990, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1890 + } + }, + { + "ph": "f", "id": 1890, "pid": 435, "tid": 435, "ts": 1776649789976990, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789977016, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1891 + } + }, + { + "ph": "f", "id": 1891, "pid": 435, "tid": 435, "ts": 1776649789977016, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789977040, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1892 + } + }, + { + "ph": "f", "id": 1892, "pid": 435, "tid": 435, "ts": 1776649789977040, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789977063, "dur": 25, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1893 + } + }, + { + "ph": "f", "id": 1893, "pid": 435, "tid": 435, "ts": 1776649789977063, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789977089, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1894 + } + }, + { + "ph": "f", "id": 1894, "pid": 435, "tid": 435, "ts": 1776649789977089, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789977116, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1895 + } + }, + { + "ph": "f", "id": 1895, "pid": 435, "tid": 435, "ts": 1776649789977116, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789977141, "dur": 30, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1896 + } + }, + { + "ph": "f", "id": 1896, "pid": 435, "tid": 435, "ts": 1776649789977141, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789977172, "dur": 26, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1897 + } + }, + { + "ph": "f", "id": 1897, "pid": 435, "tid": 435, "ts": 1776649789977172, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789977198, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1898 + } + }, + { + "ph": "f", "id": 1898, "pid": 435, "tid": 435, "ts": 1776649789977198, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789977222, "dur": 20, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1899 + } + }, + { + "ph": "f", "id": 1899, "pid": 435, "tid": 435, "ts": 1776649789977222, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789977243, "dur": 21, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1900 + } + }, + { + "ph": "f", "id": 1900, "pid": 435, "tid": 435, "ts": 1776649789977243, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789977265, "dur": 21, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1901 + } + }, + { + "ph": "f", "id": 1901, "pid": 435, "tid": 435, "ts": 1776649789977265, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789977286, "dur": 24, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1902 + } + }, + { + "ph": "f", "id": 1902, "pid": 435, "tid": 435, "ts": 1776649789977286, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789977311, "dur": 20, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1903 + } + }, + { + "ph": "f", "id": 1903, "pid": 435, "tid": 435, "ts": 1776649789977311, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789977332, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1904 + } + }, + { + "ph": "f", "id": 1904, "pid": 435, "tid": 435, "ts": 1776649789977332, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789977355, "dur": 21, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1905 + } + }, + { + "ph": "f", "id": 1905, "pid": 435, "tid": 435, "ts": 1776649789977355, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789977377, "dur": 20, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1906 + } + }, + { + "ph": "f", "id": 1906, "pid": 435, "tid": 435, "ts": 1776649789977377, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789977398, "dur": 27, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1907 + } + }, + { + "ph": "f", "id": 1907, "pid": 435, "tid": 435, "ts": 1776649789977398, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789977426, "dur": 20, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1908 + } + }, + { + "ph": "f", "id": 1908, "pid": 435, "tid": 435, "ts": 1776649789977426, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789977447, "dur": 21, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1909 + } + }, + { + "ph": "f", "id": 1909, "pid": 435, "tid": 435, "ts": 1776649789977447, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789977468, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1910 + } + }, + { + "ph": "f", "id": 1910, "pid": 435, "tid": 435, "ts": 1776649789977468, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789977491, "dur": 23, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1911 + } + }, + { + "ph": "f", "id": 1911, "pid": 435, "tid": 435, "ts": 1776649789977491, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649789977515, "dur": 22, + "args": { + "External id": 89, + "cbid": 273, "correlation": 1912 + } + }, + { + "ph": "f", "id": 1912, "pid": 435, "tid": 435, "ts": 1776649789977515, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649789977689, "dur": 4, + "args": { + "External id": 89, + "cbid": 210, "correlation": 1913 + } + }, + { + "ph": "f", "id": 1913, "pid": 435, "tid": 435, "ts": 1776649789977689, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789977775, "dur": 33, + "args": { + "External id": 89, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 1914, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 1914, "pid": 0, "tid": 7, "ts": 1776649789977775, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789977737, "dur": 39, + "args": { + "External id": 89, + "cbid": 211, "correlation": 1914 + } + }, + { + "ph": "s", "id": 1914, "pid": 435, "tid": 435, "ts": 1776649789977737, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789978034, "dur": 25, + "args": { + "External id": 91, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 1941, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 1941, "pid": 0, "tid": 7, "ts": 1776649789978034, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789978022, "dur": 9, + "args": { + "External id": 91, + "cbid": 211, "correlation": 1941 + } + }, + { + "ph": "s", "id": 1941, "pid": 435, "tid": 435, "ts": 1776649789978022, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789978075, "dur": 23, + "args": { + "External id": 93, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 1968, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 1968, "pid": 0, "tid": 7, "ts": 1776649789978075, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789978065, "dur": 6, + "args": { + "External id": 93, + "cbid": 211, "correlation": 1968 + } + }, + { + "ph": "s", "id": 1968, "pid": 435, "tid": 435, "ts": 1776649789978065, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789978107, "dur": 23, + "args": { + "External id": 95, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 1995, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 1995, "pid": 0, "tid": 7, "ts": 1776649789978107, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789978097, "dur": 6, + "args": { + "External id": 95, + "cbid": 211, "correlation": 1995 + } + }, + { + "ph": "s", "id": 1995, "pid": 435, "tid": 435, "ts": 1776649789978097, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789978138, "dur": 23, + "args": { + "External id": 97, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2022, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2022, "pid": 0, "tid": 7, "ts": 1776649789978138, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789978129, "dur": 6, + "args": { + "External id": 97, + "cbid": 211, "correlation": 2022 + } + }, + { + "ph": "s", "id": 2022, "pid": 435, "tid": 435, "ts": 1776649789978129, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789978168, "dur": 23, + "args": { + "External id": 99, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2049, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2049, "pid": 0, "tid": 7, "ts": 1776649789978168, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789978159, "dur": 7, + "args": { + "External id": 99, + "cbid": 211, "correlation": 2049 + } + }, + { + "ph": "s", "id": 2049, "pid": 435, "tid": 435, "ts": 1776649789978159, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789978199, "dur": 23, + "args": { + "External id": 101, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2076, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2076, "pid": 0, "tid": 7, "ts": 1776649789978199, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789978189, "dur": 6, + "args": { + "External id": 101, + "cbid": 211, "correlation": 2076 + } + }, + { + "ph": "s", "id": 2076, "pid": 435, "tid": 435, "ts": 1776649789978189, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789978230, "dur": 23, + "args": { + "External id": 103, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2103, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2103, "pid": 0, "tid": 7, "ts": 1776649789978230, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789978219, "dur": 6, + "args": { + "External id": 103, + "cbid": 211, "correlation": 2103 + } + }, + { + "ph": "s", "id": 2103, "pid": 435, "tid": 435, "ts": 1776649789978219, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789978260, "dur": 23, + "args": { + "External id": 105, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2130, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2130, "pid": 0, "tid": 7, "ts": 1776649789978260, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789978249, "dur": 6, + "args": { + "External id": 105, + "cbid": 211, "correlation": 2130 + } + }, + { + "ph": "s", "id": 2130, "pid": 435, "tid": 435, "ts": 1776649789978249, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789978289, "dur": 23, + "args": { + "External id": 107, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2157, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2157, "pid": 0, "tid": 7, "ts": 1776649789978289, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789978279, "dur": 6, + "args": { + "External id": 107, + "cbid": 211, "correlation": 2157 + } + }, + { + "ph": "s", "id": 2157, "pid": 435, "tid": 435, "ts": 1776649789978279, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649789978484, "dur": 18, + "args": { + "External id": 2163, + "cbid": 165, "correlation": 2163 + } + }, + { + "ph": "s", "id": 2163, "pid": 435, "tid": 435, "ts": 1776649789978484, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649789978672, "dur": 8, + "args": { + "External id": 2172, + "cbid": 135, "correlation": 2172 + } + }, + { + "ph": "f", "id": 2172, "pid": 435, "tid": 435, "ts": 1776649789978672, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789978732, "dur": 23, + "args": { + "External id": 109, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2198, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2198, "pid": 0, "tid": 7, "ts": 1776649789978732, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789978719, "dur": 10, + "args": { + "External id": 109, + "cbid": 211, "correlation": 2198 + } + }, + { + "ph": "s", "id": 2198, "pid": 435, "tid": 435, "ts": 1776649789978719, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789978768, "dur": 23, + "args": { + "External id": 111, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2225, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2225, "pid": 0, "tid": 7, "ts": 1776649789978768, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789978759, "dur": 6, + "args": { + "External id": 111, + "cbid": 211, "correlation": 2225 + } + }, + { + "ph": "s", "id": 2225, "pid": 435, "tid": 435, "ts": 1776649789978759, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789978799, "dur": 23, + "args": { + "External id": 113, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2252, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2252, "pid": 0, "tid": 7, "ts": 1776649789978799, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789978789, "dur": 6, + "args": { + "External id": 113, + "cbid": 211, "correlation": 2252 + } + }, + { + "ph": "s", "id": 2252, "pid": 435, "tid": 435, "ts": 1776649789978789, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789978829, "dur": 23, + "args": { + "External id": 115, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2279, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2279, "pid": 0, "tid": 7, "ts": 1776649789978829, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789978818, "dur": 6, + "args": { + "External id": 115, + "cbid": 211, "correlation": 2279 + } + }, + { + "ph": "s", "id": 2279, "pid": 435, "tid": 435, "ts": 1776649789978818, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789978858, "dur": 23, + "args": { + "External id": 117, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2306, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2306, "pid": 0, "tid": 7, "ts": 1776649789978858, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789978848, "dur": 6, + "args": { + "External id": 117, + "cbid": 211, "correlation": 2306 + } + }, + { + "ph": "s", "id": 2306, "pid": 435, "tid": 435, "ts": 1776649789978848, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789978887, "dur": 23, + "args": { + "External id": 119, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2333, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2333, "pid": 0, "tid": 7, "ts": 1776649789978887, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789978877, "dur": 6, + "args": { + "External id": 119, + "cbid": 211, "correlation": 2333 + } + }, + { + "ph": "s", "id": 2333, "pid": 435, "tid": 435, "ts": 1776649789978877, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789978921, "dur": 23, + "args": { + "External id": 121, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2360, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2360, "pid": 0, "tid": 7, "ts": 1776649789978921, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789978911, "dur": 6, + "args": { + "External id": 121, + "cbid": 211, "correlation": 2360 + } + }, + { + "ph": "s", "id": 2360, "pid": 435, "tid": 435, "ts": 1776649789978911, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789978951, "dur": 23, + "args": { + "External id": 123, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2387, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2387, "pid": 0, "tid": 7, "ts": 1776649789978951, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789978942, "dur": 6, + "args": { + "External id": 123, + "cbid": 211, "correlation": 2387 + } + }, + { + "ph": "s", "id": 2387, "pid": 435, "tid": 435, "ts": 1776649789978942, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789978982, "dur": 23, + "args": { + "External id": 125, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2414, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2414, "pid": 0, "tid": 7, "ts": 1776649789978982, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789978971, "dur": 6, + "args": { + "External id": 125, + "cbid": 211, "correlation": 2414 + } + }, + { + "ph": "s", "id": 2414, "pid": 435, "tid": 435, "ts": 1776649789978971, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979011, "dur": 23, + "args": { + "External id": 127, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2441, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2441, "pid": 0, "tid": 7, "ts": 1776649789979011, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979000, "dur": 6, + "args": { + "External id": 127, + "cbid": 211, "correlation": 2441 + } + }, + { + "ph": "s", "id": 2441, "pid": 435, "tid": 435, "ts": 1776649789979000, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979040, "dur": 23, + "args": { + "External id": 129, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2468, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2468, "pid": 0, "tid": 7, "ts": 1776649789979040, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979029, "dur": 6, + "args": { + "External id": 129, + "cbid": 211, "correlation": 2468 + } + }, + { + "ph": "s", "id": 2468, "pid": 435, "tid": 435, "ts": 1776649789979029, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979069, "dur": 23, + "args": { + "External id": 131, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2495, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2495, "pid": 0, "tid": 7, "ts": 1776649789979069, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979058, "dur": 6, + "args": { + "External id": 131, + "cbid": 211, "correlation": 2495 + } + }, + { + "ph": "s", "id": 2495, "pid": 435, "tid": 435, "ts": 1776649789979058, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979098, "dur": 23, + "args": { + "External id": 133, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2522, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2522, "pid": 0, "tid": 7, "ts": 1776649789979098, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979087, "dur": 6, + "args": { + "External id": 133, + "cbid": 211, "correlation": 2522 + } + }, + { + "ph": "s", "id": 2522, "pid": 435, "tid": 435, "ts": 1776649789979087, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979129, "dur": 23, + "args": { + "External id": 135, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2549, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2549, "pid": 0, "tid": 7, "ts": 1776649789979129, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979117, "dur": 10, + "args": { + "External id": 135, + "cbid": 211, "correlation": 2549 + } + }, + { + "ph": "s", "id": 2549, "pid": 435, "tid": 435, "ts": 1776649789979117, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979160, "dur": 23, + "args": { + "External id": 137, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2576, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2576, "pid": 0, "tid": 7, "ts": 1776649789979160, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979151, "dur": 6, + "args": { + "External id": 137, + "cbid": 211, "correlation": 2576 + } + }, + { + "ph": "s", "id": 2576, "pid": 435, "tid": 435, "ts": 1776649789979151, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979191, "dur": 23, + "args": { + "External id": 139, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2603, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2603, "pid": 0, "tid": 7, "ts": 1776649789979191, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979180, "dur": 6, + "args": { + "External id": 139, + "cbid": 211, "correlation": 2603 + } + }, + { + "ph": "s", "id": 2603, "pid": 435, "tid": 435, "ts": 1776649789979180, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979220, "dur": 23, + "args": { + "External id": 141, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2630, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2630, "pid": 0, "tid": 7, "ts": 1776649789979220, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979209, "dur": 6, + "args": { + "External id": 141, + "cbid": 211, "correlation": 2630 + } + }, + { + "ph": "s", "id": 2630, "pid": 435, "tid": 435, "ts": 1776649789979209, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979249, "dur": 23, + "args": { + "External id": 143, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2657, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2657, "pid": 0, "tid": 7, "ts": 1776649789979249, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979238, "dur": 6, + "args": { + "External id": 143, + "cbid": 211, "correlation": 2657 + } + }, + { + "ph": "s", "id": 2657, "pid": 435, "tid": 435, "ts": 1776649789979238, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979277, "dur": 23, + "args": { + "External id": 145, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2684, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2684, "pid": 0, "tid": 7, "ts": 1776649789979277, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979267, "dur": 5, + "args": { + "External id": 145, + "cbid": 211, "correlation": 2684 + } + }, + { + "ph": "s", "id": 2684, "pid": 435, "tid": 435, "ts": 1776649789979267, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979307, "dur": 23, + "args": { + "External id": 147, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2711, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2711, "pid": 0, "tid": 7, "ts": 1776649789979307, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979296, "dur": 6, + "args": { + "External id": 147, + "cbid": 211, "correlation": 2711 + } + }, + { + "ph": "s", "id": 2711, "pid": 435, "tid": 435, "ts": 1776649789979296, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979336, "dur": 23, + "args": { + "External id": 149, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2738, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2738, "pid": 0, "tid": 7, "ts": 1776649789979336, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979325, "dur": 6, + "args": { + "External id": 149, + "cbid": 211, "correlation": 2738 + } + }, + { + "ph": "s", "id": 2738, "pid": 435, "tid": 435, "ts": 1776649789979325, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979368, "dur": 23, + "args": { + "External id": 151, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2765, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2765, "pid": 0, "tid": 7, "ts": 1776649789979368, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979359, "dur": 6, + "args": { + "External id": 151, + "cbid": 211, "correlation": 2765 + } + }, + { + "ph": "s", "id": 2765, "pid": 435, "tid": 435, "ts": 1776649789979359, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979405, "dur": 23, + "args": { + "External id": 153, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2792, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2792, "pid": 0, "tid": 7, "ts": 1776649789979405, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979396, "dur": 6, + "args": { + "External id": 153, + "cbid": 211, "correlation": 2792 + } + }, + { + "ph": "s", "id": 2792, "pid": 435, "tid": 435, "ts": 1776649789979396, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979436, "dur": 23, + "args": { + "External id": 155, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2819, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2819, "pid": 0, "tid": 7, "ts": 1776649789979436, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979427, "dur": 6, + "args": { + "External id": 155, + "cbid": 211, "correlation": 2819 + } + }, + { + "ph": "s", "id": 2819, "pid": 435, "tid": 435, "ts": 1776649789979427, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979467, "dur": 23, + "args": { + "External id": 157, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2846, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2846, "pid": 0, "tid": 7, "ts": 1776649789979467, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979457, "dur": 6, + "args": { + "External id": 157, + "cbid": 211, "correlation": 2846 + } + }, + { + "ph": "s", "id": 2846, "pid": 435, "tid": 435, "ts": 1776649789979457, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979496, "dur": 23, + "args": { + "External id": 159, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2873, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2873, "pid": 0, "tid": 7, "ts": 1776649789979496, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979486, "dur": 6, + "args": { + "External id": 159, + "cbid": 211, "correlation": 2873 + } + }, + { + "ph": "s", "id": 2873, "pid": 435, "tid": 435, "ts": 1776649789979486, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979526, "dur": 23, + "args": { + "External id": 161, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2900, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2900, "pid": 0, "tid": 7, "ts": 1776649789979526, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979515, "dur": 6, + "args": { + "External id": 161, + "cbid": 211, "correlation": 2900 + } + }, + { + "ph": "s", "id": 2900, "pid": 435, "tid": 435, "ts": 1776649789979515, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979555, "dur": 23, + "args": { + "External id": 163, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2927, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2927, "pid": 0, "tid": 7, "ts": 1776649789979555, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979544, "dur": 6, + "args": { + "External id": 163, + "cbid": 211, "correlation": 2927 + } + }, + { + "ph": "s", "id": 2927, "pid": 435, "tid": 435, "ts": 1776649789979544, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979592, "dur": 23, + "args": { + "External id": 165, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2954, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2954, "pid": 0, "tid": 7, "ts": 1776649789979592, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979583, "dur": 6, + "args": { + "External id": 165, + "cbid": 211, "correlation": 2954 + } + }, + { + "ph": "s", "id": 2954, "pid": 435, "tid": 435, "ts": 1776649789979583, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979621, "dur": 23, + "args": { + "External id": 167, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 2981, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 2981, "pid": 0, "tid": 7, "ts": 1776649789979621, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979612, "dur": 6, + "args": { + "External id": 167, + "cbid": 211, "correlation": 2981 + } + }, + { + "ph": "s", "id": 2981, "pid": 435, "tid": 435, "ts": 1776649789979612, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979652, "dur": 23, + "args": { + "External id": 169, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3008, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3008, "pid": 0, "tid": 7, "ts": 1776649789979652, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979642, "dur": 6, + "args": { + "External id": 169, + "cbid": 211, "correlation": 3008 + } + }, + { + "ph": "s", "id": 3008, "pid": 435, "tid": 435, "ts": 1776649789979642, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979681, "dur": 23, + "args": { + "External id": 171, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3035, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3035, "pid": 0, "tid": 7, "ts": 1776649789979681, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979671, "dur": 6, + "args": { + "External id": 171, + "cbid": 211, "correlation": 3035 + } + }, + { + "ph": "s", "id": 3035, "pid": 435, "tid": 435, "ts": 1776649789979671, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979710, "dur": 23, + "args": { + "External id": 173, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3062, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3062, "pid": 0, "tid": 7, "ts": 1776649789979710, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979700, "dur": 6, + "args": { + "External id": 173, + "cbid": 211, "correlation": 3062 + } + }, + { + "ph": "s", "id": 3062, "pid": 435, "tid": 435, "ts": 1776649789979700, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979740, "dur": 23, + "args": { + "External id": 175, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3089, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3089, "pid": 0, "tid": 7, "ts": 1776649789979740, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979729, "dur": 6, + "args": { + "External id": 175, + "cbid": 211, "correlation": 3089 + } + }, + { + "ph": "s", "id": 3089, "pid": 435, "tid": 435, "ts": 1776649789979729, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979769, "dur": 23, + "args": { + "External id": 177, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3116, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3116, "pid": 0, "tid": 7, "ts": 1776649789979769, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979758, "dur": 6, + "args": { + "External id": 177, + "cbid": 211, "correlation": 3116 + } + }, + { + "ph": "s", "id": 3116, "pid": 435, "tid": 435, "ts": 1776649789979758, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979801, "dur": 23, + "args": { + "External id": 179, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3143, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3143, "pid": 0, "tid": 7, "ts": 1776649789979801, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979792, "dur": 6, + "args": { + "External id": 179, + "cbid": 211, "correlation": 3143 + } + }, + { + "ph": "s", "id": 3143, "pid": 435, "tid": 435, "ts": 1776649789979792, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979831, "dur": 23, + "args": { + "External id": 181, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3170, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3170, "pid": 0, "tid": 7, "ts": 1776649789979831, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979822, "dur": 6, + "args": { + "External id": 181, + "cbid": 211, "correlation": 3170 + } + }, + { + "ph": "s", "id": 3170, "pid": 435, "tid": 435, "ts": 1776649789979822, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979862, "dur": 23, + "args": { + "External id": 183, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3197, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3197, "pid": 0, "tid": 7, "ts": 1776649789979862, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979851, "dur": 6, + "args": { + "External id": 183, + "cbid": 211, "correlation": 3197 + } + }, + { + "ph": "s", "id": 3197, "pid": 435, "tid": 435, "ts": 1776649789979851, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979891, "dur": 23, + "args": { + "External id": 185, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3224, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3224, "pid": 0, "tid": 7, "ts": 1776649789979891, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979880, "dur": 6, + "args": { + "External id": 185, + "cbid": 211, "correlation": 3224 + } + }, + { + "ph": "s", "id": 3224, "pid": 435, "tid": 435, "ts": 1776649789979880, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979920, "dur": 23, + "args": { + "External id": 187, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3251, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3251, "pid": 0, "tid": 7, "ts": 1776649789979920, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979910, "dur": 6, + "args": { + "External id": 187, + "cbid": 211, "correlation": 3251 + } + }, + { + "ph": "s", "id": 3251, "pid": 435, "tid": 435, "ts": 1776649789979910, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979949, "dur": 23, + "args": { + "External id": 189, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3278, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3278, "pid": 0, "tid": 7, "ts": 1776649789979949, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979939, "dur": 6, + "args": { + "External id": 189, + "cbid": 211, "correlation": 3278 + } + }, + { + "ph": "s", "id": 3278, "pid": 435, "tid": 435, "ts": 1776649789979939, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789979978, "dur": 23, + "args": { + "External id": 191, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3305, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3305, "pid": 0, "tid": 7, "ts": 1776649789979978, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979968, "dur": 6, + "args": { + "External id": 191, + "cbid": 211, "correlation": 3305 + } + }, + { + "ph": "s", "id": 3305, "pid": 435, "tid": 435, "ts": 1776649789979968, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980007, "dur": 23, + "args": { + "External id": 193, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3332, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3332, "pid": 0, "tid": 7, "ts": 1776649789980007, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789979997, "dur": 6, + "args": { + "External id": 193, + "cbid": 211, "correlation": 3332 + } + }, + { + "ph": "s", "id": 3332, "pid": 435, "tid": 435, "ts": 1776649789979997, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980039, "dur": 23, + "args": { + "External id": 195, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3359, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3359, "pid": 0, "tid": 7, "ts": 1776649789980039, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980030, "dur": 6, + "args": { + "External id": 195, + "cbid": 211, "correlation": 3359 + } + }, + { + "ph": "s", "id": 3359, "pid": 435, "tid": 435, "ts": 1776649789980030, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980070, "dur": 23, + "args": { + "External id": 197, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3386, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3386, "pid": 0, "tid": 7, "ts": 1776649789980070, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980060, "dur": 6, + "args": { + "External id": 197, + "cbid": 211, "correlation": 3386 + } + }, + { + "ph": "s", "id": 3386, "pid": 435, "tid": 435, "ts": 1776649789980060, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980100, "dur": 23, + "args": { + "External id": 199, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3413, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3413, "pid": 0, "tid": 7, "ts": 1776649789980100, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980089, "dur": 6, + "args": { + "External id": 199, + "cbid": 211, "correlation": 3413 + } + }, + { + "ph": "s", "id": 3413, "pid": 435, "tid": 435, "ts": 1776649789980089, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980129, "dur": 23, + "args": { + "External id": 201, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3440, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3440, "pid": 0, "tid": 7, "ts": 1776649789980129, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980118, "dur": 6, + "args": { + "External id": 201, + "cbid": 211, "correlation": 3440 + } + }, + { + "ph": "s", "id": 3440, "pid": 435, "tid": 435, "ts": 1776649789980118, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980158, "dur": 23, + "args": { + "External id": 203, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3467, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3467, "pid": 0, "tid": 7, "ts": 1776649789980158, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980148, "dur": 6, + "args": { + "External id": 203, + "cbid": 211, "correlation": 3467 + } + }, + { + "ph": "s", "id": 3467, "pid": 435, "tid": 435, "ts": 1776649789980148, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980187, "dur": 23, + "args": { + "External id": 205, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3494, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3494, "pid": 0, "tid": 7, "ts": 1776649789980187, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980177, "dur": 6, + "args": { + "External id": 205, + "cbid": 211, "correlation": 3494 + } + }, + { + "ph": "s", "id": 3494, "pid": 435, "tid": 435, "ts": 1776649789980177, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980217, "dur": 23, + "args": { + "External id": 207, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3521, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3521, "pid": 0, "tid": 7, "ts": 1776649789980217, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980206, "dur": 6, + "args": { + "External id": 207, + "cbid": 211, "correlation": 3521 + } + }, + { + "ph": "s", "id": 3521, "pid": 435, "tid": 435, "ts": 1776649789980206, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649789980272, "dur": 4, + "args": { + "External id": 3532, + "cbid": 135, "correlation": 3532 + } + }, + { + "ph": "f", "id": 3532, "pid": 435, "tid": 435, "ts": 1776649789980272, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649789980289, "dur": 6, + "args": { + "External id": 3537, + "cbid": 165, "correlation": 3537 + } + }, + { + "ph": "s", "id": 3537, "pid": 435, "tid": 435, "ts": 1776649789980289, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649789980300, "dur": 6, + "args": { + "External id": 3540, + "cbid": 139, "correlation": 3540 + } + }, + { + "ph": "f", "id": 3540, "pid": 435, "tid": 435, "ts": 1776649789980300, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980359, "dur": 23, + "args": { + "External id": 209, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3574, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3574, "pid": 0, "tid": 7, "ts": 1776649789980359, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980347, "dur": 9, + "args": { + "External id": 209, + "cbid": 211, "correlation": 3574 + } + }, + { + "ph": "s", "id": 3574, "pid": 435, "tid": 435, "ts": 1776649789980347, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980393, "dur": 23, + "args": { + "External id": 211, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3601, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3601, "pid": 0, "tid": 7, "ts": 1776649789980393, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980384, "dur": 6, + "args": { + "External id": 211, + "cbid": 211, "correlation": 3601 + } + }, + { + "ph": "s", "id": 3601, "pid": 435, "tid": 435, "ts": 1776649789980384, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980424, "dur": 23, + "args": { + "External id": 213, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3628, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3628, "pid": 0, "tid": 7, "ts": 1776649789980424, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980414, "dur": 6, + "args": { + "External id": 213, + "cbid": 211, "correlation": 3628 + } + }, + { + "ph": "s", "id": 3628, "pid": 435, "tid": 435, "ts": 1776649789980414, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980454, "dur": 23, + "args": { + "External id": 215, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3655, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3655, "pid": 0, "tid": 7, "ts": 1776649789980454, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980444, "dur": 6, + "args": { + "External id": 215, + "cbid": 211, "correlation": 3655 + } + }, + { + "ph": "s", "id": 3655, "pid": 435, "tid": 435, "ts": 1776649789980444, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980484, "dur": 23, + "args": { + "External id": 217, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3682, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3682, "pid": 0, "tid": 7, "ts": 1776649789980484, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980473, "dur": 6, + "args": { + "External id": 217, + "cbid": 211, "correlation": 3682 + } + }, + { + "ph": "s", "id": 3682, "pid": 435, "tid": 435, "ts": 1776649789980473, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980513, "dur": 23, + "args": { + "External id": 219, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3709, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3709, "pid": 0, "tid": 7, "ts": 1776649789980513, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980503, "dur": 6, + "args": { + "External id": 219, + "cbid": 211, "correlation": 3709 + } + }, + { + "ph": "s", "id": 3709, "pid": 435, "tid": 435, "ts": 1776649789980503, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980542, "dur": 23, + "args": { + "External id": 221, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3736, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3736, "pid": 0, "tid": 7, "ts": 1776649789980542, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980532, "dur": 6, + "args": { + "External id": 221, + "cbid": 211, "correlation": 3736 + } + }, + { + "ph": "s", "id": 3736, "pid": 435, "tid": 435, "ts": 1776649789980532, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980575, "dur": 23, + "args": { + "External id": 223, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3763, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3763, "pid": 0, "tid": 7, "ts": 1776649789980575, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980565, "dur": 6, + "args": { + "External id": 223, + "cbid": 211, "correlation": 3763 + } + }, + { + "ph": "s", "id": 3763, "pid": 435, "tid": 435, "ts": 1776649789980565, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980605, "dur": 23, + "args": { + "External id": 225, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3790, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3790, "pid": 0, "tid": 7, "ts": 1776649789980605, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980596, "dur": 6, + "args": { + "External id": 225, + "cbid": 211, "correlation": 3790 + } + }, + { + "ph": "s", "id": 3790, "pid": 435, "tid": 435, "ts": 1776649789980596, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980635, "dur": 23, + "args": { + "External id": 227, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3817, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3817, "pid": 0, "tid": 7, "ts": 1776649789980635, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980625, "dur": 6, + "args": { + "External id": 227, + "cbid": 211, "correlation": 3817 + } + }, + { + "ph": "s", "id": 3817, "pid": 435, "tid": 435, "ts": 1776649789980625, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649789980651, "dur": 12, + "args": { + "External id": 3823, + "cbid": 165, "correlation": 3823 + } + }, + { + "ph": "s", "id": 3823, "pid": 435, "tid": 435, "ts": 1776649789980651, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649789980685, "dur": 3, + "args": { + "External id": 3832, + "cbid": 135, "correlation": 3832 + } + }, + { + "ph": "f", "id": 3832, "pid": 435, "tid": 435, "ts": 1776649789980685, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980727, "dur": 23, + "args": { + "External id": 229, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3858, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3858, "pid": 0, "tid": 7, "ts": 1776649789980727, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980716, "dur": 8, + "args": { + "External id": 229, + "cbid": 211, "correlation": 3858 + } + }, + { + "ph": "s", "id": 3858, "pid": 435, "tid": 435, "ts": 1776649789980716, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980759, "dur": 23, + "args": { + "External id": 231, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3885, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3885, "pid": 0, "tid": 7, "ts": 1776649789980759, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980750, "dur": 6, + "args": { + "External id": 231, + "cbid": 211, "correlation": 3885 + } + }, + { + "ph": "s", "id": 3885, "pid": 435, "tid": 435, "ts": 1776649789980750, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980790, "dur": 23, + "args": { + "External id": 233, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3912, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3912, "pid": 0, "tid": 7, "ts": 1776649789980790, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980779, "dur": 6, + "args": { + "External id": 233, + "cbid": 211, "correlation": 3912 + } + }, + { + "ph": "s", "id": 3912, "pid": 435, "tid": 435, "ts": 1776649789980779, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980824, "dur": 23, + "args": { + "External id": 235, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3939, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3939, "pid": 0, "tid": 7, "ts": 1776649789980824, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980814, "dur": 6, + "args": { + "External id": 235, + "cbid": 211, "correlation": 3939 + } + }, + { + "ph": "s", "id": 3939, "pid": 435, "tid": 435, "ts": 1776649789980814, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980855, "dur": 23, + "args": { + "External id": 237, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3966, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3966, "pid": 0, "tid": 7, "ts": 1776649789980855, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980845, "dur": 6, + "args": { + "External id": 237, + "cbid": 211, "correlation": 3966 + } + }, + { + "ph": "s", "id": 3966, "pid": 435, "tid": 435, "ts": 1776649789980845, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980884, "dur": 23, + "args": { + "External id": 239, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 3993, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 3993, "pid": 0, "tid": 7, "ts": 1776649789980884, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980874, "dur": 6, + "args": { + "External id": 239, + "cbid": 211, "correlation": 3993 + } + }, + { + "ph": "s", "id": 3993, "pid": 435, "tid": 435, "ts": 1776649789980874, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980913, "dur": 23, + "args": { + "External id": 241, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4020, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4020, "pid": 0, "tid": 7, "ts": 1776649789980913, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980903, "dur": 6, + "args": { + "External id": 241, + "cbid": 211, "correlation": 4020 + } + }, + { + "ph": "s", "id": 4020, "pid": 435, "tid": 435, "ts": 1776649789980903, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980943, "dur": 23, + "args": { + "External id": 243, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4047, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4047, "pid": 0, "tid": 7, "ts": 1776649789980943, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980932, "dur": 6, + "args": { + "External id": 243, + "cbid": 211, "correlation": 4047 + } + }, + { + "ph": "s", "id": 4047, "pid": 435, "tid": 435, "ts": 1776649789980932, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789980972, "dur": 23, + "args": { + "External id": 245, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4074, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4074, "pid": 0, "tid": 7, "ts": 1776649789980972, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980961, "dur": 6, + "args": { + "External id": 245, + "cbid": 211, "correlation": 4074 + } + }, + { + "ph": "s", "id": 4074, "pid": 435, "tid": 435, "ts": 1776649789980961, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981001, "dur": 23, + "args": { + "External id": 247, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4101, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4101, "pid": 0, "tid": 7, "ts": 1776649789981001, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789980990, "dur": 6, + "args": { + "External id": 247, + "cbid": 211, "correlation": 4101 + } + }, + { + "ph": "s", "id": 4101, "pid": 435, "tid": 435, "ts": 1776649789980990, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981032, "dur": 23, + "args": { + "External id": 249, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4128, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4128, "pid": 0, "tid": 7, "ts": 1776649789981032, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981019, "dur": 10, + "args": { + "External id": 249, + "cbid": 211, "correlation": 4128 + } + }, + { + "ph": "s", "id": 4128, "pid": 435, "tid": 435, "ts": 1776649789981019, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981063, "dur": 23, + "args": { + "External id": 251, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4155, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4155, "pid": 0, "tid": 7, "ts": 1776649789981063, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981053, "dur": 6, + "args": { + "External id": 251, + "cbid": 211, "correlation": 4155 + } + }, + { + "ph": "s", "id": 4155, "pid": 435, "tid": 435, "ts": 1776649789981053, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981093, "dur": 23, + "args": { + "External id": 253, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4182, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4182, "pid": 0, "tid": 7, "ts": 1776649789981093, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981083, "dur": 6, + "args": { + "External id": 253, + "cbid": 211, "correlation": 4182 + } + }, + { + "ph": "s", "id": 4182, "pid": 435, "tid": 435, "ts": 1776649789981083, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981123, "dur": 23, + "args": { + "External id": 255, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4209, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4209, "pid": 0, "tid": 7, "ts": 1776649789981123, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981112, "dur": 6, + "args": { + "External id": 255, + "cbid": 211, "correlation": 4209 + } + }, + { + "ph": "s", "id": 4209, "pid": 435, "tid": 435, "ts": 1776649789981112, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981152, "dur": 23, + "args": { + "External id": 257, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4236, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4236, "pid": 0, "tid": 7, "ts": 1776649789981152, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981141, "dur": 6, + "args": { + "External id": 257, + "cbid": 211, "correlation": 4236 + } + }, + { + "ph": "s", "id": 4236, "pid": 435, "tid": 435, "ts": 1776649789981141, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981181, "dur": 23, + "args": { + "External id": 259, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4263, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4263, "pid": 0, "tid": 7, "ts": 1776649789981181, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981171, "dur": 6, + "args": { + "External id": 259, + "cbid": 211, "correlation": 4263 + } + }, + { + "ph": "s", "id": 4263, "pid": 435, "tid": 435, "ts": 1776649789981171, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981211, "dur": 23, + "args": { + "External id": 261, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4290, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4290, "pid": 0, "tid": 7, "ts": 1776649789981211, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981200, "dur": 6, + "args": { + "External id": 261, + "cbid": 211, "correlation": 4290 + } + }, + { + "ph": "s", "id": 4290, "pid": 435, "tid": 435, "ts": 1776649789981200, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981240, "dur": 23, + "args": { + "External id": 263, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4317, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4317, "pid": 0, "tid": 7, "ts": 1776649789981240, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981230, "dur": 6, + "args": { + "External id": 263, + "cbid": 211, "correlation": 4317 + } + }, + { + "ph": "s", "id": 4317, "pid": 435, "tid": 435, "ts": 1776649789981230, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981278, "dur": 23, + "args": { + "External id": 265, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4344, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4344, "pid": 0, "tid": 7, "ts": 1776649789981278, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981268, "dur": 6, + "args": { + "External id": 265, + "cbid": 211, "correlation": 4344 + } + }, + { + "ph": "s", "id": 4344, "pid": 435, "tid": 435, "ts": 1776649789981268, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981308, "dur": 23, + "args": { + "External id": 267, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4371, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4371, "pid": 0, "tid": 7, "ts": 1776649789981308, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981299, "dur": 6, + "args": { + "External id": 267, + "cbid": 211, "correlation": 4371 + } + }, + { + "ph": "s", "id": 4371, "pid": 435, "tid": 435, "ts": 1776649789981299, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981339, "dur": 23, + "args": { + "External id": 269, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4398, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4398, "pid": 0, "tid": 7, "ts": 1776649789981339, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981329, "dur": 6, + "args": { + "External id": 269, + "cbid": 211, "correlation": 4398 + } + }, + { + "ph": "s", "id": 4398, "pid": 435, "tid": 435, "ts": 1776649789981329, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981369, "dur": 23, + "args": { + "External id": 271, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4425, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4425, "pid": 0, "tid": 7, "ts": 1776649789981369, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981358, "dur": 6, + "args": { + "External id": 271, + "cbid": 211, "correlation": 4425 + } + }, + { + "ph": "s", "id": 4425, "pid": 435, "tid": 435, "ts": 1776649789981358, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981397, "dur": 23, + "args": { + "External id": 273, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4452, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4452, "pid": 0, "tid": 7, "ts": 1776649789981397, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981387, "dur": 6, + "args": { + "External id": 273, + "cbid": 211, "correlation": 4452 + } + }, + { + "ph": "s", "id": 4452, "pid": 435, "tid": 435, "ts": 1776649789981387, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981427, "dur": 23, + "args": { + "External id": 275, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4479, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4479, "pid": 0, "tid": 7, "ts": 1776649789981427, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981416, "dur": 6, + "args": { + "External id": 275, + "cbid": 211, "correlation": 4479 + } + }, + { + "ph": "s", "id": 4479, "pid": 435, "tid": 435, "ts": 1776649789981416, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981455, "dur": 23, + "args": { + "External id": 277, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4506, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4506, "pid": 0, "tid": 7, "ts": 1776649789981455, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981445, "dur": 6, + "args": { + "External id": 277, + "cbid": 211, "correlation": 4506 + } + }, + { + "ph": "s", "id": 4506, "pid": 435, "tid": 435, "ts": 1776649789981445, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981487, "dur": 23, + "args": { + "External id": 279, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4533, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4533, "pid": 0, "tid": 7, "ts": 1776649789981487, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981478, "dur": 6, + "args": { + "External id": 279, + "cbid": 211, "correlation": 4533 + } + }, + { + "ph": "s", "id": 4533, "pid": 435, "tid": 435, "ts": 1776649789981478, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981517, "dur": 23, + "args": { + "External id": 281, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4560, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4560, "pid": 0, "tid": 7, "ts": 1776649789981517, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981508, "dur": 6, + "args": { + "External id": 281, + "cbid": 211, "correlation": 4560 + } + }, + { + "ph": "s", "id": 4560, "pid": 435, "tid": 435, "ts": 1776649789981508, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981548, "dur": 23, + "args": { + "External id": 283, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4587, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4587, "pid": 0, "tid": 7, "ts": 1776649789981548, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981537, "dur": 6, + "args": { + "External id": 283, + "cbid": 211, "correlation": 4587 + } + }, + { + "ph": "s", "id": 4587, "pid": 435, "tid": 435, "ts": 1776649789981537, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981576, "dur": 23, + "args": { + "External id": 285, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4614, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4614, "pid": 0, "tid": 7, "ts": 1776649789981576, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981566, "dur": 6, + "args": { + "External id": 285, + "cbid": 211, "correlation": 4614 + } + }, + { + "ph": "s", "id": 4614, "pid": 435, "tid": 435, "ts": 1776649789981566, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981606, "dur": 23, + "args": { + "External id": 287, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4641, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4641, "pid": 0, "tid": 7, "ts": 1776649789981606, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981595, "dur": 6, + "args": { + "External id": 287, + "cbid": 211, "correlation": 4641 + } + }, + { + "ph": "s", "id": 4641, "pid": 435, "tid": 435, "ts": 1776649789981595, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981638, "dur": 23, + "args": { + "External id": 289, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4668, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4668, "pid": 0, "tid": 7, "ts": 1776649789981638, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981629, "dur": 6, + "args": { + "External id": 289, + "cbid": 211, "correlation": 4668 + } + }, + { + "ph": "s", "id": 4668, "pid": 435, "tid": 435, "ts": 1776649789981629, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981669, "dur": 23, + "args": { + "External id": 291, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4695, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4695, "pid": 0, "tid": 7, "ts": 1776649789981669, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981659, "dur": 6, + "args": { + "External id": 291, + "cbid": 211, "correlation": 4695 + } + }, + { + "ph": "s", "id": 4695, "pid": 435, "tid": 435, "ts": 1776649789981659, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981707, "dur": 23, + "args": { + "External id": 293, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4722, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4722, "pid": 0, "tid": 7, "ts": 1776649789981707, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981698, "dur": 6, + "args": { + "External id": 293, + "cbid": 211, "correlation": 4722 + } + }, + { + "ph": "s", "id": 4722, "pid": 435, "tid": 435, "ts": 1776649789981698, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981737, "dur": 23, + "args": { + "External id": 295, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4749, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4749, "pid": 0, "tid": 7, "ts": 1776649789981737, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981728, "dur": 6, + "args": { + "External id": 295, + "cbid": 211, "correlation": 4749 + } + }, + { + "ph": "s", "id": 4749, "pid": 435, "tid": 435, "ts": 1776649789981728, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981768, "dur": 23, + "args": { + "External id": 297, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4776, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4776, "pid": 0, "tid": 7, "ts": 1776649789981768, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981758, "dur": 6, + "args": { + "External id": 297, + "cbid": 211, "correlation": 4776 + } + }, + { + "ph": "s", "id": 4776, "pid": 435, "tid": 435, "ts": 1776649789981758, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981798, "dur": 23, + "args": { + "External id": 299, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4803, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4803, "pid": 0, "tid": 7, "ts": 1776649789981798, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981787, "dur": 6, + "args": { + "External id": 299, + "cbid": 211, "correlation": 4803 + } + }, + { + "ph": "s", "id": 4803, "pid": 435, "tid": 435, "ts": 1776649789981787, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981827, "dur": 23, + "args": { + "External id": 301, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4830, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4830, "pid": 0, "tid": 7, "ts": 1776649789981827, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981816, "dur": 6, + "args": { + "External id": 301, + "cbid": 211, "correlation": 4830 + } + }, + { + "ph": "s", "id": 4830, "pid": 435, "tid": 435, "ts": 1776649789981816, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981856, "dur": 23, + "args": { + "External id": 303, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4857, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4857, "pid": 0, "tid": 7, "ts": 1776649789981856, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981846, "dur": 6, + "args": { + "External id": 303, + "cbid": 211, "correlation": 4857 + } + }, + { + "ph": "s", "id": 4857, "pid": 435, "tid": 435, "ts": 1776649789981846, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981885, "dur": 22, + "args": { + "External id": 305, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4884, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4884, "pid": 0, "tid": 7, "ts": 1776649789981885, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981875, "dur": 6, + "args": { + "External id": 305, + "cbid": 211, "correlation": 4884 + } + }, + { + "ph": "s", "id": 4884, "pid": 435, "tid": 435, "ts": 1776649789981875, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981915, "dur": 23, + "args": { + "External id": 307, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4911, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4911, "pid": 0, "tid": 7, "ts": 1776649789981915, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981904, "dur": 6, + "args": { + "External id": 307, + "cbid": 211, "correlation": 4911 + } + }, + { + "ph": "s", "id": 4911, "pid": 435, "tid": 435, "ts": 1776649789981904, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981947, "dur": 23, + "args": { + "External id": 309, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4938, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4938, "pid": 0, "tid": 7, "ts": 1776649789981947, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981937, "dur": 6, + "args": { + "External id": 309, + "cbid": 211, "correlation": 4938 + } + }, + { + "ph": "s", "id": 4938, "pid": 435, "tid": 435, "ts": 1776649789981937, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789981978, "dur": 23, + "args": { + "External id": 311, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4965, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4965, "pid": 0, "tid": 7, "ts": 1776649789981978, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981967, "dur": 6, + "args": { + "External id": 311, + "cbid": 211, "correlation": 4965 + } + }, + { + "ph": "s", "id": 4965, "pid": 435, "tid": 435, "ts": 1776649789981967, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982007, "dur": 23, + "args": { + "External id": 313, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 4992, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 4992, "pid": 0, "tid": 7, "ts": 1776649789982007, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789981997, "dur": 6, + "args": { + "External id": 313, + "cbid": 211, "correlation": 4992 + } + }, + { + "ph": "s", "id": 4992, "pid": 435, "tid": 435, "ts": 1776649789981997, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982036, "dur": 23, + "args": { + "External id": 315, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5019, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5019, "pid": 0, "tid": 7, "ts": 1776649789982036, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982026, "dur": 6, + "args": { + "External id": 315, + "cbid": 211, "correlation": 5019 + } + }, + { + "ph": "s", "id": 5019, "pid": 435, "tid": 435, "ts": 1776649789982026, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982066, "dur": 23, + "args": { + "External id": 317, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5046, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5046, "pid": 0, "tid": 7, "ts": 1776649789982066, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982055, "dur": 6, + "args": { + "External id": 317, + "cbid": 211, "correlation": 5046 + } + }, + { + "ph": "s", "id": 5046, "pid": 435, "tid": 435, "ts": 1776649789982055, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982095, "dur": 23, + "args": { + "External id": 319, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5073, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5073, "pid": 0, "tid": 7, "ts": 1776649789982095, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982084, "dur": 6, + "args": { + "External id": 319, + "cbid": 211, "correlation": 5073 + } + }, + { + "ph": "s", "id": 5073, "pid": 435, "tid": 435, "ts": 1776649789982084, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982124, "dur": 23, + "args": { + "External id": 321, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5100, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5100, "pid": 0, "tid": 7, "ts": 1776649789982124, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982113, "dur": 6, + "args": { + "External id": 321, + "cbid": 211, "correlation": 5100 + } + }, + { + "ph": "s", "id": 5100, "pid": 435, "tid": 435, "ts": 1776649789982113, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982156, "dur": 23, + "args": { + "External id": 323, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5127, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5127, "pid": 0, "tid": 7, "ts": 1776649789982156, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982147, "dur": 6, + "args": { + "External id": 323, + "cbid": 211, "correlation": 5127 + } + }, + { + "ph": "s", "id": 5127, "pid": 435, "tid": 435, "ts": 1776649789982147, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982187, "dur": 23, + "args": { + "External id": 325, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5154, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5154, "pid": 0, "tid": 7, "ts": 1776649789982187, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982176, "dur": 6, + "args": { + "External id": 325, + "cbid": 211, "correlation": 5154 + } + }, + { + "ph": "s", "id": 5154, "pid": 435, "tid": 435, "ts": 1776649789982176, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982217, "dur": 23, + "args": { + "External id": 327, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5181, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5181, "pid": 0, "tid": 7, "ts": 1776649789982217, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982206, "dur": 6, + "args": { + "External id": 327, + "cbid": 211, "correlation": 5181 + } + }, + { + "ph": "s", "id": 5181, "pid": 435, "tid": 435, "ts": 1776649789982206, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649789982240, "dur": 4, + "args": { + "External id": 5192, + "cbid": 135, "correlation": 5192 + } + }, + { + "ph": "f", "id": 5192, "pid": 435, "tid": 435, "ts": 1776649789982240, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649789982255, "dur": 5, + "args": { + "External id": 5197, + "cbid": 165, "correlation": 5197 + } + }, + { + "ph": "s", "id": 5197, "pid": 435, "tid": 435, "ts": 1776649789982255, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649789982264, "dur": 4, + "args": { + "External id": 5200, + "cbid": 139, "correlation": 5200 + } + }, + { + "ph": "f", "id": 5200, "pid": 435, "tid": 435, "ts": 1776649789982264, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982313, "dur": 23, + "args": { + "External id": 329, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5234, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5234, "pid": 0, "tid": 7, "ts": 1776649789982313, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982301, "dur": 8, + "args": { + "External id": 329, + "cbid": 211, "correlation": 5234 + } + }, + { + "ph": "s", "id": 5234, "pid": 435, "tid": 435, "ts": 1776649789982301, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982344, "dur": 23, + "args": { + "External id": 331, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5261, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5261, "pid": 0, "tid": 7, "ts": 1776649789982344, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982335, "dur": 6, + "args": { + "External id": 331, + "cbid": 211, "correlation": 5261 + } + }, + { + "ph": "s", "id": 5261, "pid": 435, "tid": 435, "ts": 1776649789982335, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982376, "dur": 23, + "args": { + "External id": 333, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5288, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5288, "pid": 0, "tid": 7, "ts": 1776649789982376, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982365, "dur": 6, + "args": { + "External id": 333, + "cbid": 211, "correlation": 5288 + } + }, + { + "ph": "s", "id": 5288, "pid": 435, "tid": 435, "ts": 1776649789982365, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982405, "dur": 23, + "args": { + "External id": 335, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5315, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5315, "pid": 0, "tid": 7, "ts": 1776649789982405, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982394, "dur": 6, + "args": { + "External id": 335, + "cbid": 211, "correlation": 5315 + } + }, + { + "ph": "s", "id": 5315, "pid": 435, "tid": 435, "ts": 1776649789982394, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982438, "dur": 23, + "args": { + "External id": 337, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5342, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5342, "pid": 0, "tid": 7, "ts": 1776649789982438, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982429, "dur": 6, + "args": { + "External id": 337, + "cbid": 211, "correlation": 5342 + } + }, + { + "ph": "s", "id": 5342, "pid": 435, "tid": 435, "ts": 1776649789982429, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982468, "dur": 23, + "args": { + "External id": 339, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5369, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5369, "pid": 0, "tid": 7, "ts": 1776649789982468, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982459, "dur": 6, + "args": { + "External id": 339, + "cbid": 211, "correlation": 5369 + } + }, + { + "ph": "s", "id": 5369, "pid": 435, "tid": 435, "ts": 1776649789982459, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982499, "dur": 23, + "args": { + "External id": 341, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5396, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5396, "pid": 0, "tid": 7, "ts": 1776649789982499, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982489, "dur": 6, + "args": { + "External id": 341, + "cbid": 211, "correlation": 5396 + } + }, + { + "ph": "s", "id": 5396, "pid": 435, "tid": 435, "ts": 1776649789982489, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982528, "dur": 23, + "args": { + "External id": 343, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5423, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5423, "pid": 0, "tid": 7, "ts": 1776649789982528, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982518, "dur": 6, + "args": { + "External id": 343, + "cbid": 211, "correlation": 5423 + } + }, + { + "ph": "s", "id": 5423, "pid": 435, "tid": 435, "ts": 1776649789982518, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982558, "dur": 23, + "args": { + "External id": 345, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5450, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5450, "pid": 0, "tid": 7, "ts": 1776649789982558, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982547, "dur": 6, + "args": { + "External id": 345, + "cbid": 211, "correlation": 5450 + } + }, + { + "ph": "s", "id": 5450, "pid": 435, "tid": 435, "ts": 1776649789982547, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982587, "dur": 23, + "args": { + "External id": 347, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5477, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5477, "pid": 0, "tid": 7, "ts": 1776649789982587, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982577, "dur": 6, + "args": { + "External id": 347, + "cbid": 211, "correlation": 5477 + } + }, + { + "ph": "s", "id": 5477, "pid": 435, "tid": 435, "ts": 1776649789982577, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649789982602, "dur": 11, + "args": { + "External id": 5483, + "cbid": 165, "correlation": 5483 + } + }, + { + "ph": "s", "id": 5483, "pid": 435, "tid": 435, "ts": 1776649789982602, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649789982635, "dur": 3, + "args": { + "External id": 5492, + "cbid": 135, "correlation": 5492 + } + }, + { + "ph": "f", "id": 5492, "pid": 435, "tid": 435, "ts": 1776649789982635, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982675, "dur": 23, + "args": { + "External id": 349, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5518, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5518, "pid": 0, "tid": 7, "ts": 1776649789982675, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982665, "dur": 8, + "args": { + "External id": 349, + "cbid": 211, "correlation": 5518 + } + }, + { + "ph": "s", "id": 5518, "pid": 435, "tid": 435, "ts": 1776649789982665, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982712, "dur": 23, + "args": { + "External id": 351, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5545, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5545, "pid": 0, "tid": 7, "ts": 1776649789982712, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982702, "dur": 6, + "args": { + "External id": 351, + "cbid": 211, "correlation": 5545 + } + }, + { + "ph": "s", "id": 5545, "pid": 435, "tid": 435, "ts": 1776649789982702, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982743, "dur": 23, + "args": { + "External id": 353, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5572, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5572, "pid": 0, "tid": 7, "ts": 1776649789982743, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982733, "dur": 6, + "args": { + "External id": 353, + "cbid": 211, "correlation": 5572 + } + }, + { + "ph": "s", "id": 5572, "pid": 435, "tid": 435, "ts": 1776649789982733, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982773, "dur": 23, + "args": { + "External id": 355, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5599, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5599, "pid": 0, "tid": 7, "ts": 1776649789982773, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982763, "dur": 6, + "args": { + "External id": 355, + "cbid": 211, "correlation": 5599 + } + }, + { + "ph": "s", "id": 5599, "pid": 435, "tid": 435, "ts": 1776649789982763, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982803, "dur": 23, + "args": { + "External id": 357, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5626, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5626, "pid": 0, "tid": 7, "ts": 1776649789982803, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982792, "dur": 6, + "args": { + "External id": 357, + "cbid": 211, "correlation": 5626 + } + }, + { + "ph": "s", "id": 5626, "pid": 435, "tid": 435, "ts": 1776649789982792, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982832, "dur": 23, + "args": { + "External id": 359, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5653, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5653, "pid": 0, "tid": 7, "ts": 1776649789982832, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982822, "dur": 6, + "args": { + "External id": 359, + "cbid": 211, "correlation": 5653 + } + }, + { + "ph": "s", "id": 5653, "pid": 435, "tid": 435, "ts": 1776649789982822, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982861, "dur": 23, + "args": { + "External id": 361, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5680, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5680, "pid": 0, "tid": 7, "ts": 1776649789982861, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982851, "dur": 6, + "args": { + "External id": 361, + "cbid": 211, "correlation": 5680 + } + }, + { + "ph": "s", "id": 5680, "pid": 435, "tid": 435, "ts": 1776649789982851, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982890, "dur": 23, + "args": { + "External id": 363, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5707, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5707, "pid": 0, "tid": 7, "ts": 1776649789982890, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982880, "dur": 6, + "args": { + "External id": 363, + "cbid": 211, "correlation": 5707 + } + }, + { + "ph": "s", "id": 5707, "pid": 435, "tid": 435, "ts": 1776649789982880, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982922, "dur": 23, + "args": { + "External id": 365, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5734, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5734, "pid": 0, "tid": 7, "ts": 1776649789982922, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982913, "dur": 6, + "args": { + "External id": 365, + "cbid": 211, "correlation": 5734 + } + }, + { + "ph": "s", "id": 5734, "pid": 435, "tid": 435, "ts": 1776649789982913, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982953, "dur": 23, + "args": { + "External id": 367, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5761, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5761, "pid": 0, "tid": 7, "ts": 1776649789982953, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982943, "dur": 6, + "args": { + "External id": 367, + "cbid": 211, "correlation": 5761 + } + }, + { + "ph": "s", "id": 5761, "pid": 435, "tid": 435, "ts": 1776649789982943, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789982982, "dur": 23, + "args": { + "External id": 369, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5788, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5788, "pid": 0, "tid": 7, "ts": 1776649789982982, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789982972, "dur": 6, + "args": { + "External id": 369, + "cbid": 211, "correlation": 5788 + } + }, + { + "ph": "s", "id": 5788, "pid": 435, "tid": 435, "ts": 1776649789982972, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983010, "dur": 23, + "args": { + "External id": 371, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5815, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5815, "pid": 0, "tid": 7, "ts": 1776649789983010, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983000, "dur": 6, + "args": { + "External id": 371, + "cbid": 211, "correlation": 5815 + } + }, + { + "ph": "s", "id": 5815, "pid": 435, "tid": 435, "ts": 1776649789983000, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983039, "dur": 23, + "args": { + "External id": 373, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5842, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5842, "pid": 0, "tid": 7, "ts": 1776649789983039, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983029, "dur": 6, + "args": { + "External id": 373, + "cbid": 211, "correlation": 5842 + } + }, + { + "ph": "s", "id": 5842, "pid": 435, "tid": 435, "ts": 1776649789983029, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983068, "dur": 23, + "args": { + "External id": 375, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5869, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5869, "pid": 0, "tid": 7, "ts": 1776649789983068, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983058, "dur": 6, + "args": { + "External id": 375, + "cbid": 211, "correlation": 5869 + } + }, + { + "ph": "s", "id": 5869, "pid": 435, "tid": 435, "ts": 1776649789983058, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983097, "dur": 23, + "args": { + "External id": 377, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5896, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5896, "pid": 0, "tid": 7, "ts": 1776649789983097, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983087, "dur": 6, + "args": { + "External id": 377, + "cbid": 211, "correlation": 5896 + } + }, + { + "ph": "s", "id": 5896, "pid": 435, "tid": 435, "ts": 1776649789983087, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983131, "dur": 23, + "args": { + "External id": 379, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5923, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5923, "pid": 0, "tid": 7, "ts": 1776649789983131, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983123, "dur": 6, + "args": { + "External id": 379, + "cbid": 211, "correlation": 5923 + } + }, + { + "ph": "s", "id": 5923, "pid": 435, "tid": 435, "ts": 1776649789983123, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983161, "dur": 23, + "args": { + "External id": 381, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5950, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5950, "pid": 0, "tid": 7, "ts": 1776649789983161, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983152, "dur": 6, + "args": { + "External id": 381, + "cbid": 211, "correlation": 5950 + } + }, + { + "ph": "s", "id": 5950, "pid": 435, "tid": 435, "ts": 1776649789983152, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983192, "dur": 23, + "args": { + "External id": 383, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 5977, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 5977, "pid": 0, "tid": 7, "ts": 1776649789983192, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983182, "dur": 6, + "args": { + "External id": 383, + "cbid": 211, "correlation": 5977 + } + }, + { + "ph": "s", "id": 5977, "pid": 435, "tid": 435, "ts": 1776649789983182, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983221, "dur": 23, + "args": { + "External id": 385, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6004, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6004, "pid": 0, "tid": 7, "ts": 1776649789983221, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983210, "dur": 6, + "args": { + "External id": 385, + "cbid": 211, "correlation": 6004 + } + }, + { + "ph": "s", "id": 6004, "pid": 435, "tid": 435, "ts": 1776649789983210, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983250, "dur": 23, + "args": { + "External id": 387, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6031, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6031, "pid": 0, "tid": 7, "ts": 1776649789983250, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983239, "dur": 6, + "args": { + "External id": 387, + "cbid": 211, "correlation": 6031 + } + }, + { + "ph": "s", "id": 6031, "pid": 435, "tid": 435, "ts": 1776649789983239, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983279, "dur": 23, + "args": { + "External id": 389, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6058, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6058, "pid": 0, "tid": 7, "ts": 1776649789983279, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983268, "dur": 6, + "args": { + "External id": 389, + "cbid": 211, "correlation": 6058 + } + }, + { + "ph": "s", "id": 6058, "pid": 435, "tid": 435, "ts": 1776649789983268, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983307, "dur": 23, + "args": { + "External id": 391, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6085, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6085, "pid": 0, "tid": 7, "ts": 1776649789983307, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983297, "dur": 6, + "args": { + "External id": 391, + "cbid": 211, "correlation": 6085 + } + }, + { + "ph": "s", "id": 6085, "pid": 435, "tid": 435, "ts": 1776649789983297, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983338, "dur": 23, + "args": { + "External id": 393, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6112, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6112, "pid": 0, "tid": 7, "ts": 1776649789983338, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983325, "dur": 10, + "args": { + "External id": 393, + "cbid": 211, "correlation": 6112 + } + }, + { + "ph": "s", "id": 6112, "pid": 435, "tid": 435, "ts": 1776649789983325, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983368, "dur": 23, + "args": { + "External id": 395, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6139, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6139, "pid": 0, "tid": 7, "ts": 1776649789983368, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983359, "dur": 6, + "args": { + "External id": 395, + "cbid": 211, "correlation": 6139 + } + }, + { + "ph": "s", "id": 6139, "pid": 435, "tid": 435, "ts": 1776649789983359, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983399, "dur": 23, + "args": { + "External id": 397, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6166, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6166, "pid": 0, "tid": 7, "ts": 1776649789983399, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983388, "dur": 6, + "args": { + "External id": 397, + "cbid": 211, "correlation": 6166 + } + }, + { + "ph": "s", "id": 6166, "pid": 435, "tid": 435, "ts": 1776649789983388, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983428, "dur": 23, + "args": { + "External id": 399, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6193, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6193, "pid": 0, "tid": 7, "ts": 1776649789983428, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983417, "dur": 6, + "args": { + "External id": 399, + "cbid": 211, "correlation": 6193 + } + }, + { + "ph": "s", "id": 6193, "pid": 435, "tid": 435, "ts": 1776649789983417, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983457, "dur": 23, + "args": { + "External id": 401, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6220, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6220, "pid": 0, "tid": 7, "ts": 1776649789983457, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983446, "dur": 6, + "args": { + "External id": 401, + "cbid": 211, "correlation": 6220 + } + }, + { + "ph": "s", "id": 6220, "pid": 435, "tid": 435, "ts": 1776649789983446, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983486, "dur": 23, + "args": { + "External id": 403, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6247, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6247, "pid": 0, "tid": 7, "ts": 1776649789983486, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983475, "dur": 6, + "args": { + "External id": 403, + "cbid": 211, "correlation": 6247 + } + }, + { + "ph": "s", "id": 6247, "pid": 435, "tid": 435, "ts": 1776649789983475, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983514, "dur": 23, + "args": { + "External id": 405, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6274, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6274, "pid": 0, "tid": 7, "ts": 1776649789983514, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983504, "dur": 6, + "args": { + "External id": 405, + "cbid": 211, "correlation": 6274 + } + }, + { + "ph": "s", "id": 6274, "pid": 435, "tid": 435, "ts": 1776649789983504, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983543, "dur": 23, + "args": { + "External id": 407, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6301, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6301, "pid": 0, "tid": 7, "ts": 1776649789983543, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983533, "dur": 6, + "args": { + "External id": 407, + "cbid": 211, "correlation": 6301 + } + }, + { + "ph": "s", "id": 6301, "pid": 435, "tid": 435, "ts": 1776649789983533, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983575, "dur": 23, + "args": { + "External id": 409, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6328, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6328, "pid": 0, "tid": 7, "ts": 1776649789983575, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983566, "dur": 6, + "args": { + "External id": 409, + "cbid": 211, "correlation": 6328 + } + }, + { + "ph": "s", "id": 6328, "pid": 435, "tid": 435, "ts": 1776649789983566, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983605, "dur": 23, + "args": { + "External id": 411, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6355, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6355, "pid": 0, "tid": 7, "ts": 1776649789983605, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983595, "dur": 6, + "args": { + "External id": 411, + "cbid": 211, "correlation": 6355 + } + }, + { + "ph": "s", "id": 6355, "pid": 435, "tid": 435, "ts": 1776649789983595, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983634, "dur": 23, + "args": { + "External id": 413, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6382, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6382, "pid": 0, "tid": 7, "ts": 1776649789983634, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983624, "dur": 6, + "args": { + "External id": 413, + "cbid": 211, "correlation": 6382 + } + }, + { + "ph": "s", "id": 6382, "pid": 435, "tid": 435, "ts": 1776649789983624, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983663, "dur": 23, + "args": { + "External id": 415, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6409, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6409, "pid": 0, "tid": 7, "ts": 1776649789983663, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983652, "dur": 6, + "args": { + "External id": 415, + "cbid": 211, "correlation": 6409 + } + }, + { + "ph": "s", "id": 6409, "pid": 435, "tid": 435, "ts": 1776649789983652, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983692, "dur": 23, + "args": { + "External id": 417, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6436, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6436, "pid": 0, "tid": 7, "ts": 1776649789983692, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983682, "dur": 6, + "args": { + "External id": 417, + "cbid": 211, "correlation": 6436 + } + }, + { + "ph": "s", "id": 6436, "pid": 435, "tid": 435, "ts": 1776649789983682, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983722, "dur": 23, + "args": { + "External id": 419, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6463, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6463, "pid": 0, "tid": 7, "ts": 1776649789983722, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983711, "dur": 6, + "args": { + "External id": 419, + "cbid": 211, "correlation": 6463 + } + }, + { + "ph": "s", "id": 6463, "pid": 435, "tid": 435, "ts": 1776649789983711, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983756, "dur": 23, + "args": { + "External id": 421, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6490, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6490, "pid": 0, "tid": 7, "ts": 1776649789983756, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983746, "dur": 6, + "args": { + "External id": 421, + "cbid": 211, "correlation": 6490 + } + }, + { + "ph": "s", "id": 6490, "pid": 435, "tid": 435, "ts": 1776649789983746, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983789, "dur": 23, + "args": { + "External id": 423, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6517, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6517, "pid": 0, "tid": 7, "ts": 1776649789983789, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983780, "dur": 6, + "args": { + "External id": 423, + "cbid": 211, "correlation": 6517 + } + }, + { + "ph": "s", "id": 6517, "pid": 435, "tid": 435, "ts": 1776649789983780, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983820, "dur": 23, + "args": { + "External id": 425, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6544, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6544, "pid": 0, "tid": 7, "ts": 1776649789983820, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983809, "dur": 6, + "args": { + "External id": 425, + "cbid": 211, "correlation": 6544 + } + }, + { + "ph": "s", "id": 6544, "pid": 435, "tid": 435, "ts": 1776649789983809, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983849, "dur": 23, + "args": { + "External id": 427, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6571, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6571, "pid": 0, "tid": 7, "ts": 1776649789983849, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983838, "dur": 6, + "args": { + "External id": 427, + "cbid": 211, "correlation": 6571 + } + }, + { + "ph": "s", "id": 6571, "pid": 435, "tid": 435, "ts": 1776649789983838, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983877, "dur": 23, + "args": { + "External id": 429, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6598, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6598, "pid": 0, "tid": 7, "ts": 1776649789983877, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983867, "dur": 6, + "args": { + "External id": 429, + "cbid": 211, "correlation": 6598 + } + }, + { + "ph": "s", "id": 6598, "pid": 435, "tid": 435, "ts": 1776649789983867, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983906, "dur": 23, + "args": { + "External id": 431, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6625, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6625, "pid": 0, "tid": 7, "ts": 1776649789983906, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983896, "dur": 6, + "args": { + "External id": 431, + "cbid": 211, "correlation": 6625 + } + }, + { + "ph": "s", "id": 6625, "pid": 435, "tid": 435, "ts": 1776649789983896, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983935, "dur": 23, + "args": { + "External id": 433, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6652, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6652, "pid": 0, "tid": 7, "ts": 1776649789983935, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983925, "dur": 6, + "args": { + "External id": 433, + "cbid": 211, "correlation": 6652 + } + }, + { + "ph": "s", "id": 6652, "pid": 435, "tid": 435, "ts": 1776649789983925, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983964, "dur": 23, + "args": { + "External id": 435, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6679, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6679, "pid": 0, "tid": 7, "ts": 1776649789983964, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983953, "dur": 6, + "args": { + "External id": 435, + "cbid": 211, "correlation": 6679 + } + }, + { + "ph": "s", "id": 6679, "pid": 435, "tid": 435, "ts": 1776649789983953, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789983996, "dur": 23, + "args": { + "External id": 437, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6706, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6706, "pid": 0, "tid": 7, "ts": 1776649789983996, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789983982, "dur": 10, + "args": { + "External id": 437, + "cbid": 211, "correlation": 6706 + } + }, + { + "ph": "s", "id": 6706, "pid": 435, "tid": 435, "ts": 1776649789983982, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984026, "dur": 22, + "args": { + "External id": 439, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6733, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6733, "pid": 0, "tid": 7, "ts": 1776649789984026, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984017, "dur": 6, + "args": { + "External id": 439, + "cbid": 211, "correlation": 6733 + } + }, + { + "ph": "s", "id": 6733, "pid": 435, "tid": 435, "ts": 1776649789984017, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984056, "dur": 23, + "args": { + "External id": 441, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6760, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6760, "pid": 0, "tid": 7, "ts": 1776649789984056, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984047, "dur": 6, + "args": { + "External id": 441, + "cbid": 211, "correlation": 6760 + } + }, + { + "ph": "s", "id": 6760, "pid": 435, "tid": 435, "ts": 1776649789984047, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984087, "dur": 23, + "args": { + "External id": 443, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6787, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6787, "pid": 0, "tid": 7, "ts": 1776649789984087, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984076, "dur": 6, + "args": { + "External id": 443, + "cbid": 211, "correlation": 6787 + } + }, + { + "ph": "s", "id": 6787, "pid": 435, "tid": 435, "ts": 1776649789984076, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984116, "dur": 23, + "args": { + "External id": 445, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6814, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6814, "pid": 0, "tid": 7, "ts": 1776649789984116, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984105, "dur": 6, + "args": { + "External id": 445, + "cbid": 211, "correlation": 6814 + } + }, + { + "ph": "s", "id": 6814, "pid": 435, "tid": 435, "ts": 1776649789984105, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984145, "dur": 23, + "args": { + "External id": 447, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6841, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6841, "pid": 0, "tid": 7, "ts": 1776649789984145, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984134, "dur": 6, + "args": { + "External id": 447, + "cbid": 211, "correlation": 6841 + } + }, + { + "ph": "s", "id": 6841, "pid": 435, "tid": 435, "ts": 1776649789984134, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649789984168, "dur": 4, + "args": { + "External id": 6852, + "cbid": 135, "correlation": 6852 + } + }, + { + "ph": "f", "id": 6852, "pid": 435, "tid": 435, "ts": 1776649789984168, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649789984183, "dur": 5, + "args": { + "External id": 6857, + "cbid": 165, "correlation": 6857 + } + }, + { + "ph": "s", "id": 6857, "pid": 435, "tid": 435, "ts": 1776649789984183, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649789984191, "dur": 4, + "args": { + "External id": 6860, + "cbid": 139, "correlation": 6860 + } + }, + { + "ph": "f", "id": 6860, "pid": 435, "tid": 435, "ts": 1776649789984191, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984259, "dur": 23, + "args": { + "External id": 449, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6894, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6894, "pid": 0, "tid": 7, "ts": 1776649789984259, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984248, "dur": 8, + "args": { + "External id": 449, + "cbid": 211, "correlation": 6894 + } + }, + { + "ph": "s", "id": 6894, "pid": 435, "tid": 435, "ts": 1776649789984248, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984300, "dur": 23, + "args": { + "External id": 451, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6921, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6921, "pid": 0, "tid": 7, "ts": 1776649789984300, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984290, "dur": 7, + "args": { + "External id": 451, + "cbid": 211, "correlation": 6921 + } + }, + { + "ph": "s", "id": 6921, "pid": 435, "tid": 435, "ts": 1776649789984290, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984331, "dur": 23, + "args": { + "External id": 453, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6948, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6948, "pid": 0, "tid": 7, "ts": 1776649789984331, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984322, "dur": 6, + "args": { + "External id": 453, + "cbid": 211, "correlation": 6948 + } + }, + { + "ph": "s", "id": 6948, "pid": 435, "tid": 435, "ts": 1776649789984322, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984361, "dur": 23, + "args": { + "External id": 455, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 6975, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 6975, "pid": 0, "tid": 7, "ts": 1776649789984361, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984351, "dur": 6, + "args": { + "External id": 455, + "cbid": 211, "correlation": 6975 + } + }, + { + "ph": "s", "id": 6975, "pid": 435, "tid": 435, "ts": 1776649789984351, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984392, "dur": 23, + "args": { + "External id": 457, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7002, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7002, "pid": 0, "tid": 7, "ts": 1776649789984392, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984381, "dur": 6, + "args": { + "External id": 457, + "cbid": 211, "correlation": 7002 + } + }, + { + "ph": "s", "id": 7002, "pid": 435, "tid": 435, "ts": 1776649789984381, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984421, "dur": 23, + "args": { + "External id": 459, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7029, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7029, "pid": 0, "tid": 7, "ts": 1776649789984421, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984411, "dur": 6, + "args": { + "External id": 459, + "cbid": 211, "correlation": 7029 + } + }, + { + "ph": "s", "id": 7029, "pid": 435, "tid": 435, "ts": 1776649789984411, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984451, "dur": 23, + "args": { + "External id": 461, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7056, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7056, "pid": 0, "tid": 7, "ts": 1776649789984451, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984440, "dur": 6, + "args": { + "External id": 461, + "cbid": 211, "correlation": 7056 + } + }, + { + "ph": "s", "id": 7056, "pid": 435, "tid": 435, "ts": 1776649789984440, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984480, "dur": 23, + "args": { + "External id": 463, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7083, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7083, "pid": 0, "tid": 7, "ts": 1776649789984480, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984470, "dur": 6, + "args": { + "External id": 463, + "cbid": 211, "correlation": 7083 + } + }, + { + "ph": "s", "id": 7083, "pid": 435, "tid": 435, "ts": 1776649789984470, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984509, "dur": 23, + "args": { + "External id": 465, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7110, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7110, "pid": 0, "tid": 7, "ts": 1776649789984509, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984499, "dur": 6, + "args": { + "External id": 465, + "cbid": 211, "correlation": 7110 + } + }, + { + "ph": "s", "id": 7110, "pid": 435, "tid": 435, "ts": 1776649789984499, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984542, "dur": 23, + "args": { + "External id": 467, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7137, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7137, "pid": 0, "tid": 7, "ts": 1776649789984542, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984532, "dur": 6, + "args": { + "External id": 467, + "cbid": 211, "correlation": 7137 + } + }, + { + "ph": "s", "id": 7137, "pid": 435, "tid": 435, "ts": 1776649789984532, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649789984559, "dur": 10, + "args": { + "External id": 7143, + "cbid": 165, "correlation": 7143 + } + }, + { + "ph": "s", "id": 7143, "pid": 435, "tid": 435, "ts": 1776649789984559, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649789984591, "dur": 3, + "args": { + "External id": 7152, + "cbid": 135, "correlation": 7152 + } + }, + { + "ph": "f", "id": 7152, "pid": 435, "tid": 435, "ts": 1776649789984591, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984632, "dur": 23, + "args": { + "External id": 469, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7178, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7178, "pid": 0, "tid": 7, "ts": 1776649789984632, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984621, "dur": 8, + "args": { + "External id": 469, + "cbid": 211, "correlation": 7178 + } + }, + { + "ph": "s", "id": 7178, "pid": 435, "tid": 435, "ts": 1776649789984621, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984664, "dur": 23, + "args": { + "External id": 471, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7205, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7205, "pid": 0, "tid": 7, "ts": 1776649789984664, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984655, "dur": 6, + "args": { + "External id": 471, + "cbid": 211, "correlation": 7205 + } + }, + { + "ph": "s", "id": 7205, "pid": 435, "tid": 435, "ts": 1776649789984655, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984695, "dur": 23, + "args": { + "External id": 473, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7232, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7232, "pid": 0, "tid": 7, "ts": 1776649789984695, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984684, "dur": 6, + "args": { + "External id": 473, + "cbid": 211, "correlation": 7232 + } + }, + { + "ph": "s", "id": 7232, "pid": 435, "tid": 435, "ts": 1776649789984684, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984724, "dur": 23, + "args": { + "External id": 475, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7259, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7259, "pid": 0, "tid": 7, "ts": 1776649789984724, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984713, "dur": 6, + "args": { + "External id": 475, + "cbid": 211, "correlation": 7259 + } + }, + { + "ph": "s", "id": 7259, "pid": 435, "tid": 435, "ts": 1776649789984713, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984753, "dur": 23, + "args": { + "External id": 477, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7286, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7286, "pid": 0, "tid": 7, "ts": 1776649789984753, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984743, "dur": 6, + "args": { + "External id": 477, + "cbid": 211, "correlation": 7286 + } + }, + { + "ph": "s", "id": 7286, "pid": 435, "tid": 435, "ts": 1776649789984743, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984786, "dur": 23, + "args": { + "External id": 479, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7313, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7313, "pid": 0, "tid": 7, "ts": 1776649789984786, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984777, "dur": 6, + "args": { + "External id": 479, + "cbid": 211, "correlation": 7313 + } + }, + { + "ph": "s", "id": 7313, "pid": 435, "tid": 435, "ts": 1776649789984777, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984816, "dur": 23, + "args": { + "External id": 481, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7340, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7340, "pid": 0, "tid": 7, "ts": 1776649789984816, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984807, "dur": 6, + "args": { + "External id": 481, + "cbid": 211, "correlation": 7340 + } + }, + { + "ph": "s", "id": 7340, "pid": 435, "tid": 435, "ts": 1776649789984807, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984847, "dur": 23, + "args": { + "External id": 483, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7367, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7367, "pid": 0, "tid": 7, "ts": 1776649789984847, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984836, "dur": 6, + "args": { + "External id": 483, + "cbid": 211, "correlation": 7367 + } + }, + { + "ph": "s", "id": 7367, "pid": 435, "tid": 435, "ts": 1776649789984836, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984876, "dur": 23, + "args": { + "External id": 485, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7394, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7394, "pid": 0, "tid": 7, "ts": 1776649789984876, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984865, "dur": 6, + "args": { + "External id": 485, + "cbid": 211, "correlation": 7394 + } + }, + { + "ph": "s", "id": 7394, "pid": 435, "tid": 435, "ts": 1776649789984865, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984905, "dur": 23, + "args": { + "External id": 487, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7421, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7421, "pid": 0, "tid": 7, "ts": 1776649789984905, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984894, "dur": 6, + "args": { + "External id": 487, + "cbid": 211, "correlation": 7421 + } + }, + { + "ph": "s", "id": 7421, "pid": 435, "tid": 435, "ts": 1776649789984894, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984934, "dur": 23, + "args": { + "External id": 489, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7448, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7448, "pid": 0, "tid": 7, "ts": 1776649789984934, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984923, "dur": 6, + "args": { + "External id": 489, + "cbid": 211, "correlation": 7448 + } + }, + { + "ph": "s", "id": 7448, "pid": 435, "tid": 435, "ts": 1776649789984923, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984963, "dur": 23, + "args": { + "External id": 491, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7475, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7475, "pid": 0, "tid": 7, "ts": 1776649789984963, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984952, "dur": 6, + "args": { + "External id": 491, + "cbid": 211, "correlation": 7475 + } + }, + { + "ph": "s", "id": 7475, "pid": 435, "tid": 435, "ts": 1776649789984952, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789984992, "dur": 23, + "args": { + "External id": 493, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7502, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7502, "pid": 0, "tid": 7, "ts": 1776649789984992, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789984981, "dur": 6, + "args": { + "External id": 493, + "cbid": 211, "correlation": 7502 + } + }, + { + "ph": "s", "id": 7502, "pid": 435, "tid": 435, "ts": 1776649789984981, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985024, "dur": 22, + "args": { + "External id": 495, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7529, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7529, "pid": 0, "tid": 7, "ts": 1776649789985024, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985015, "dur": 6, + "args": { + "External id": 495, + "cbid": 211, "correlation": 7529 + } + }, + { + "ph": "s", "id": 7529, "pid": 435, "tid": 435, "ts": 1776649789985015, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985055, "dur": 23, + "args": { + "External id": 497, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7556, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7556, "pid": 0, "tid": 7, "ts": 1776649789985055, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985044, "dur": 6, + "args": { + "External id": 497, + "cbid": 211, "correlation": 7556 + } + }, + { + "ph": "s", "id": 7556, "pid": 435, "tid": 435, "ts": 1776649789985044, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985084, "dur": 23, + "args": { + "External id": 499, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7583, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7583, "pid": 0, "tid": 7, "ts": 1776649789985084, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985074, "dur": 6, + "args": { + "External id": 499, + "cbid": 211, "correlation": 7583 + } + }, + { + "ph": "s", "id": 7583, "pid": 435, "tid": 435, "ts": 1776649789985074, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985114, "dur": 23, + "args": { + "External id": 501, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7610, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7610, "pid": 0, "tid": 7, "ts": 1776649789985114, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985103, "dur": 6, + "args": { + "External id": 501, + "cbid": 211, "correlation": 7610 + } + }, + { + "ph": "s", "id": 7610, "pid": 435, "tid": 435, "ts": 1776649789985103, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985143, "dur": 23, + "args": { + "External id": 503, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7637, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7637, "pid": 0, "tid": 7, "ts": 1776649789985143, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985133, "dur": 6, + "args": { + "External id": 503, + "cbid": 211, "correlation": 7637 + } + }, + { + "ph": "s", "id": 7637, "pid": 435, "tid": 435, "ts": 1776649789985133, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985172, "dur": 23, + "args": { + "External id": 505, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7664, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7664, "pid": 0, "tid": 7, "ts": 1776649789985172, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985161, "dur": 6, + "args": { + "External id": 505, + "cbid": 211, "correlation": 7664 + } + }, + { + "ph": "s", "id": 7664, "pid": 435, "tid": 435, "ts": 1776649789985161, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985201, "dur": 23, + "args": { + "External id": 507, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7691, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7691, "pid": 0, "tid": 7, "ts": 1776649789985201, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985190, "dur": 6, + "args": { + "External id": 507, + "cbid": 211, "correlation": 7691 + } + }, + { + "ph": "s", "id": 7691, "pid": 435, "tid": 435, "ts": 1776649789985190, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985233, "dur": 23, + "args": { + "External id": 509, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7718, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7718, "pid": 0, "tid": 7, "ts": 1776649789985233, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985224, "dur": 6, + "args": { + "External id": 509, + "cbid": 211, "correlation": 7718 + } + }, + { + "ph": "s", "id": 7718, "pid": 435, "tid": 435, "ts": 1776649789985224, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985264, "dur": 23, + "args": { + "External id": 511, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7745, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7745, "pid": 0, "tid": 7, "ts": 1776649789985264, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985254, "dur": 6, + "args": { + "External id": 511, + "cbid": 211, "correlation": 7745 + } + }, + { + "ph": "s", "id": 7745, "pid": 435, "tid": 435, "ts": 1776649789985254, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985367, "dur": 23, + "args": { + "External id": 513, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7772, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7772, "pid": 0, "tid": 7, "ts": 1776649789985367, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985357, "dur": 6, + "args": { + "External id": 513, + "cbid": 211, "correlation": 7772 + } + }, + { + "ph": "s", "id": 7772, "pid": 435, "tid": 435, "ts": 1776649789985357, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985398, "dur": 23, + "args": { + "External id": 515, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7799, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7799, "pid": 0, "tid": 7, "ts": 1776649789985398, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985389, "dur": 6, + "args": { + "External id": 515, + "cbid": 211, "correlation": 7799 + } + }, + { + "ph": "s", "id": 7799, "pid": 435, "tid": 435, "ts": 1776649789985389, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985428, "dur": 23, + "args": { + "External id": 517, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7826, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7826, "pid": 0, "tid": 7, "ts": 1776649789985428, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985418, "dur": 6, + "args": { + "External id": 517, + "cbid": 211, "correlation": 7826 + } + }, + { + "ph": "s", "id": 7826, "pid": 435, "tid": 435, "ts": 1776649789985418, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985457, "dur": 23, + "args": { + "External id": 519, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7853, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7853, "pid": 0, "tid": 7, "ts": 1776649789985457, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985447, "dur": 6, + "args": { + "External id": 519, + "cbid": 211, "correlation": 7853 + } + }, + { + "ph": "s", "id": 7853, "pid": 435, "tid": 435, "ts": 1776649789985447, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985486, "dur": 23, + "args": { + "External id": 521, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7880, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7880, "pid": 0, "tid": 7, "ts": 1776649789985486, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985475, "dur": 6, + "args": { + "External id": 521, + "cbid": 211, "correlation": 7880 + } + }, + { + "ph": "s", "id": 7880, "pid": 435, "tid": 435, "ts": 1776649789985475, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985518, "dur": 23, + "args": { + "External id": 523, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7907, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7907, "pid": 0, "tid": 7, "ts": 1776649789985518, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985508, "dur": 6, + "args": { + "External id": 523, + "cbid": 211, "correlation": 7907 + } + }, + { + "ph": "s", "id": 7907, "pid": 435, "tid": 435, "ts": 1776649789985508, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985549, "dur": 23, + "args": { + "External id": 525, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7934, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7934, "pid": 0, "tid": 7, "ts": 1776649789985549, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985538, "dur": 6, + "args": { + "External id": 525, + "cbid": 211, "correlation": 7934 + } + }, + { + "ph": "s", "id": 7934, "pid": 435, "tid": 435, "ts": 1776649789985538, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985578, "dur": 23, + "args": { + "External id": 527, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7961, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7961, "pid": 0, "tid": 7, "ts": 1776649789985578, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985567, "dur": 6, + "args": { + "External id": 527, + "cbid": 211, "correlation": 7961 + } + }, + { + "ph": "s", "id": 7961, "pid": 435, "tid": 435, "ts": 1776649789985567, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985607, "dur": 23, + "args": { + "External id": 529, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 7988, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 7988, "pid": 0, "tid": 7, "ts": 1776649789985607, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985596, "dur": 6, + "args": { + "External id": 529, + "cbid": 211, "correlation": 7988 + } + }, + { + "ph": "s", "id": 7988, "pid": 435, "tid": 435, "ts": 1776649789985596, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985636, "dur": 23, + "args": { + "External id": 531, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8015, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8015, "pid": 0, "tid": 7, "ts": 1776649789985636, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985625, "dur": 6, + "args": { + "External id": 531, + "cbid": 211, "correlation": 8015 + } + }, + { + "ph": "s", "id": 8015, "pid": 435, "tid": 435, "ts": 1776649789985625, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985664, "dur": 23, + "args": { + "External id": 533, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8042, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8042, "pid": 0, "tid": 7, "ts": 1776649789985664, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985654, "dur": 6, + "args": { + "External id": 533, + "cbid": 211, "correlation": 8042 + } + }, + { + "ph": "s", "id": 8042, "pid": 435, "tid": 435, "ts": 1776649789985654, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985693, "dur": 23, + "args": { + "External id": 535, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8069, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8069, "pid": 0, "tid": 7, "ts": 1776649789985693, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985683, "dur": 6, + "args": { + "External id": 535, + "cbid": 211, "correlation": 8069 + } + }, + { + "ph": "s", "id": 8069, "pid": 435, "tid": 435, "ts": 1776649789985683, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985725, "dur": 23, + "args": { + "External id": 537, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8096, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8096, "pid": 0, "tid": 7, "ts": 1776649789985725, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985712, "dur": 10, + "args": { + "External id": 537, + "cbid": 211, "correlation": 8096 + } + }, + { + "ph": "s", "id": 8096, "pid": 435, "tid": 435, "ts": 1776649789985712, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985755, "dur": 23, + "args": { + "External id": 539, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8123, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8123, "pid": 0, "tid": 7, "ts": 1776649789985755, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985746, "dur": 6, + "args": { + "External id": 539, + "cbid": 211, "correlation": 8123 + } + }, + { + "ph": "s", "id": 8123, "pid": 435, "tid": 435, "ts": 1776649789985746, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985786, "dur": 23, + "args": { + "External id": 541, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8150, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8150, "pid": 0, "tid": 7, "ts": 1776649789985786, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985776, "dur": 6, + "args": { + "External id": 541, + "cbid": 211, "correlation": 8150 + } + }, + { + "ph": "s", "id": 8150, "pid": 435, "tid": 435, "ts": 1776649789985776, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985815, "dur": 23, + "args": { + "External id": 543, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8177, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8177, "pid": 0, "tid": 7, "ts": 1776649789985815, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985804, "dur": 6, + "args": { + "External id": 543, + "cbid": 211, "correlation": 8177 + } + }, + { + "ph": "s", "id": 8177, "pid": 435, "tid": 435, "ts": 1776649789985804, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985843, "dur": 23, + "args": { + "External id": 545, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8204, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8204, "pid": 0, "tid": 7, "ts": 1776649789985843, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985833, "dur": 6, + "args": { + "External id": 545, + "cbid": 211, "correlation": 8204 + } + }, + { + "ph": "s", "id": 8204, "pid": 435, "tid": 435, "ts": 1776649789985833, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985872, "dur": 23, + "args": { + "External id": 547, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8231, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8231, "pid": 0, "tid": 7, "ts": 1776649789985872, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985862, "dur": 6, + "args": { + "External id": 547, + "cbid": 211, "correlation": 8231 + } + }, + { + "ph": "s", "id": 8231, "pid": 435, "tid": 435, "ts": 1776649789985862, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985906, "dur": 22, + "args": { + "External id": 549, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8258, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8258, "pid": 0, "tid": 7, "ts": 1776649789985906, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985897, "dur": 6, + "args": { + "External id": 549, + "cbid": 211, "correlation": 8258 + } + }, + { + "ph": "s", "id": 8258, "pid": 435, "tid": 435, "ts": 1776649789985897, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985936, "dur": 23, + "args": { + "External id": 551, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8285, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8285, "pid": 0, "tid": 7, "ts": 1776649789985936, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985927, "dur": 6, + "args": { + "External id": 551, + "cbid": 211, "correlation": 8285 + } + }, + { + "ph": "s", "id": 8285, "pid": 435, "tid": 435, "ts": 1776649789985927, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789985968, "dur": 23, + "args": { + "External id": 553, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8312, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8312, "pid": 0, "tid": 7, "ts": 1776649789985968, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789985959, "dur": 6, + "args": { + "External id": 553, + "cbid": 211, "correlation": 8312 + } + }, + { + "ph": "s", "id": 8312, "pid": 435, "tid": 435, "ts": 1776649789985959, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986023, "dur": 23, + "args": { + "External id": 555, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8339, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8339, "pid": 0, "tid": 7, "ts": 1776649789986023, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986013, "dur": 6, + "args": { + "External id": 555, + "cbid": 211, "correlation": 8339 + } + }, + { + "ph": "s", "id": 8339, "pid": 435, "tid": 435, "ts": 1776649789986013, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986052, "dur": 23, + "args": { + "External id": 557, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8366, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8366, "pid": 0, "tid": 7, "ts": 1776649789986052, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986043, "dur": 6, + "args": { + "External id": 557, + "cbid": 211, "correlation": 8366 + } + }, + { + "ph": "s", "id": 8366, "pid": 435, "tid": 435, "ts": 1776649789986043, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986083, "dur": 23, + "args": { + "External id": 559, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8393, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8393, "pid": 0, "tid": 7, "ts": 1776649789986083, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986073, "dur": 6, + "args": { + "External id": 559, + "cbid": 211, "correlation": 8393 + } + }, + { + "ph": "s", "id": 8393, "pid": 435, "tid": 435, "ts": 1776649789986073, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986112, "dur": 23, + "args": { + "External id": 561, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8420, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8420, "pid": 0, "tid": 7, "ts": 1776649789986112, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986102, "dur": 6, + "args": { + "External id": 561, + "cbid": 211, "correlation": 8420 + } + }, + { + "ph": "s", "id": 8420, "pid": 435, "tid": 435, "ts": 1776649789986102, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986142, "dur": 23, + "args": { + "External id": 563, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8447, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8447, "pid": 0, "tid": 7, "ts": 1776649789986142, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986131, "dur": 6, + "args": { + "External id": 563, + "cbid": 211, "correlation": 8447 + } + }, + { + "ph": "s", "id": 8447, "pid": 435, "tid": 435, "ts": 1776649789986131, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986171, "dur": 23, + "args": { + "External id": 565, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8474, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8474, "pid": 0, "tid": 7, "ts": 1776649789986171, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986160, "dur": 6, + "args": { + "External id": 565, + "cbid": 211, "correlation": 8474 + } + }, + { + "ph": "s", "id": 8474, "pid": 435, "tid": 435, "ts": 1776649789986160, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986202, "dur": 23, + "args": { + "External id": 567, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8501, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8501, "pid": 0, "tid": 7, "ts": 1776649789986202, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986193, "dur": 6, + "args": { + "External id": 567, + "cbid": 211, "correlation": 8501 + } + }, + { + "ph": "s", "id": 8501, "pid": 435, "tid": 435, "ts": 1776649789986193, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649789986227, "dur": 3, + "args": { + "External id": 8512, + "cbid": 135, "correlation": 8512 + } + }, + { + "ph": "f", "id": 8512, "pid": 435, "tid": 435, "ts": 1776649789986227, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649789986242, "dur": 5, + "args": { + "External id": 8517, + "cbid": 165, "correlation": 8517 + } + }, + { + "ph": "s", "id": 8517, "pid": 435, "tid": 435, "ts": 1776649789986242, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649789986251, "dur": 4, + "args": { + "External id": 8520, + "cbid": 139, "correlation": 8520 + } + }, + { + "ph": "f", "id": 8520, "pid": 435, "tid": 435, "ts": 1776649789986251, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986300, "dur": 23, + "args": { + "External id": 569, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8554, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8554, "pid": 0, "tid": 7, "ts": 1776649789986300, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986289, "dur": 8, + "args": { + "External id": 569, + "cbid": 211, "correlation": 8554 + } + }, + { + "ph": "s", "id": 8554, "pid": 435, "tid": 435, "ts": 1776649789986289, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986333, "dur": 23, + "args": { + "External id": 571, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8581, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8581, "pid": 0, "tid": 7, "ts": 1776649789986333, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986324, "dur": 6, + "args": { + "External id": 571, + "cbid": 211, "correlation": 8581 + } + }, + { + "ph": "s", "id": 8581, "pid": 435, "tid": 435, "ts": 1776649789986324, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986363, "dur": 23, + "args": { + "External id": 573, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8608, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8608, "pid": 0, "tid": 7, "ts": 1776649789986363, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986353, "dur": 6, + "args": { + "External id": 573, + "cbid": 211, "correlation": 8608 + } + }, + { + "ph": "s", "id": 8608, "pid": 435, "tid": 435, "ts": 1776649789986353, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986393, "dur": 23, + "args": { + "External id": 575, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8635, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8635, "pid": 0, "tid": 7, "ts": 1776649789986393, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986383, "dur": 6, + "args": { + "External id": 575, + "cbid": 211, "correlation": 8635 + } + }, + { + "ph": "s", "id": 8635, "pid": 435, "tid": 435, "ts": 1776649789986383, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986422, "dur": 23, + "args": { + "External id": 577, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8662, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8662, "pid": 0, "tid": 7, "ts": 1776649789986422, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986412, "dur": 6, + "args": { + "External id": 577, + "cbid": 211, "correlation": 8662 + } + }, + { + "ph": "s", "id": 8662, "pid": 435, "tid": 435, "ts": 1776649789986412, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986451, "dur": 23, + "args": { + "External id": 579, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8689, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8689, "pid": 0, "tid": 7, "ts": 1776649789986451, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986441, "dur": 6, + "args": { + "External id": 579, + "cbid": 211, "correlation": 8689 + } + }, + { + "ph": "s", "id": 8689, "pid": 435, "tid": 435, "ts": 1776649789986441, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986480, "dur": 23, + "args": { + "External id": 581, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8716, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8716, "pid": 0, "tid": 7, "ts": 1776649789986480, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986470, "dur": 6, + "args": { + "External id": 581, + "cbid": 211, "correlation": 8716 + } + }, + { + "ph": "s", "id": 8716, "pid": 435, "tid": 435, "ts": 1776649789986470, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986514, "dur": 23, + "args": { + "External id": 583, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8743, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8743, "pid": 0, "tid": 7, "ts": 1776649789986514, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986505, "dur": 6, + "args": { + "External id": 583, + "cbid": 211, "correlation": 8743 + } + }, + { + "ph": "s", "id": 8743, "pid": 435, "tid": 435, "ts": 1776649789986505, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986544, "dur": 23, + "args": { + "External id": 585, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8770, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8770, "pid": 0, "tid": 7, "ts": 1776649789986544, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986535, "dur": 6, + "args": { + "External id": 585, + "cbid": 211, "correlation": 8770 + } + }, + { + "ph": "s", "id": 8770, "pid": 435, "tid": 435, "ts": 1776649789986535, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986575, "dur": 23, + "args": { + "External id": 587, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8797, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8797, "pid": 0, "tid": 7, "ts": 1776649789986575, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986564, "dur": 6, + "args": { + "External id": 587, + "cbid": 211, "correlation": 8797 + } + }, + { + "ph": "s", "id": 8797, "pid": 435, "tid": 435, "ts": 1776649789986564, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649789986590, "dur": 12, + "args": { + "External id": 8803, + "cbid": 165, "correlation": 8803 + } + }, + { + "ph": "s", "id": 8803, "pid": 435, "tid": 435, "ts": 1776649789986590, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649789986623, "dur": 3, + "args": { + "External id": 8812, + "cbid": 135, "correlation": 8812 + } + }, + { + "ph": "f", "id": 8812, "pid": 435, "tid": 435, "ts": 1776649789986623, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986663, "dur": 23, + "args": { + "External id": 589, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8838, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8838, "pid": 0, "tid": 7, "ts": 1776649789986663, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986653, "dur": 8, + "args": { + "External id": 589, + "cbid": 211, "correlation": 8838 + } + }, + { + "ph": "s", "id": 8838, "pid": 435, "tid": 435, "ts": 1776649789986653, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986696, "dur": 23, + "args": { + "External id": 591, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8865, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8865, "pid": 0, "tid": 7, "ts": 1776649789986696, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986687, "dur": 6, + "args": { + "External id": 591, + "cbid": 211, "correlation": 8865 + } + }, + { + "ph": "s", "id": 8865, "pid": 435, "tid": 435, "ts": 1776649789986687, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986727, "dur": 23, + "args": { + "External id": 593, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8892, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8892, "pid": 0, "tid": 7, "ts": 1776649789986727, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986717, "dur": 6, + "args": { + "External id": 593, + "cbid": 211, "correlation": 8892 + } + }, + { + "ph": "s", "id": 8892, "pid": 435, "tid": 435, "ts": 1776649789986717, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986757, "dur": 23, + "args": { + "External id": 595, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8919, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8919, "pid": 0, "tid": 7, "ts": 1776649789986757, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986746, "dur": 6, + "args": { + "External id": 595, + "cbid": 211, "correlation": 8919 + } + }, + { + "ph": "s", "id": 8919, "pid": 435, "tid": 435, "ts": 1776649789986746, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986791, "dur": 23, + "args": { + "External id": 597, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8946, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8946, "pid": 0, "tid": 7, "ts": 1776649789986791, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986781, "dur": 6, + "args": { + "External id": 597, + "cbid": 211, "correlation": 8946 + } + }, + { + "ph": "s", "id": 8946, "pid": 435, "tid": 435, "ts": 1776649789986781, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986821, "dur": 23, + "args": { + "External id": 599, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 8973, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 8973, "pid": 0, "tid": 7, "ts": 1776649789986821, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986811, "dur": 6, + "args": { + "External id": 599, + "cbid": 211, "correlation": 8973 + } + }, + { + "ph": "s", "id": 8973, "pid": 435, "tid": 435, "ts": 1776649789986811, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986851, "dur": 23, + "args": { + "External id": 601, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9000, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9000, "pid": 0, "tid": 7, "ts": 1776649789986851, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986841, "dur": 6, + "args": { + "External id": 601, + "cbid": 211, "correlation": 9000 + } + }, + { + "ph": "s", "id": 9000, "pid": 435, "tid": 435, "ts": 1776649789986841, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986880, "dur": 23, + "args": { + "External id": 603, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9027, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9027, "pid": 0, "tid": 7, "ts": 1776649789986880, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986870, "dur": 6, + "args": { + "External id": 603, + "cbid": 211, "correlation": 9027 + } + }, + { + "ph": "s", "id": 9027, "pid": 435, "tid": 435, "ts": 1776649789986870, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986909, "dur": 23, + "args": { + "External id": 605, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9054, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9054, "pid": 0, "tid": 7, "ts": 1776649789986909, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986899, "dur": 6, + "args": { + "External id": 605, + "cbid": 211, "correlation": 9054 + } + }, + { + "ph": "s", "id": 9054, "pid": 435, "tid": 435, "ts": 1776649789986899, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986938, "dur": 23, + "args": { + "External id": 607, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9081, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9081, "pid": 0, "tid": 7, "ts": 1776649789986938, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986928, "dur": 6, + "args": { + "External id": 607, + "cbid": 211, "correlation": 9081 + } + }, + { + "ph": "s", "id": 9081, "pid": 435, "tid": 435, "ts": 1776649789986928, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789986967, "dur": 23, + "args": { + "External id": 609, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9108, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9108, "pid": 0, "tid": 7, "ts": 1776649789986967, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986957, "dur": 6, + "args": { + "External id": 609, + "cbid": 211, "correlation": 9108 + } + }, + { + "ph": "s", "id": 9108, "pid": 435, "tid": 435, "ts": 1776649789986957, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987004, "dur": 23, + "args": { + "External id": 611, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9135, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9135, "pid": 0, "tid": 7, "ts": 1776649789987004, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789986994, "dur": 6, + "args": { + "External id": 611, + "cbid": 211, "correlation": 9135 + } + }, + { + "ph": "s", "id": 9135, "pid": 435, "tid": 435, "ts": 1776649789986994, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987034, "dur": 23, + "args": { + "External id": 613, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9162, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9162, "pid": 0, "tid": 7, "ts": 1776649789987034, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987025, "dur": 6, + "args": { + "External id": 613, + "cbid": 211, "correlation": 9162 + } + }, + { + "ph": "s", "id": 9162, "pid": 435, "tid": 435, "ts": 1776649789987025, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987064, "dur": 23, + "args": { + "External id": 615, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9189, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9189, "pid": 0, "tid": 7, "ts": 1776649789987064, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987054, "dur": 6, + "args": { + "External id": 615, + "cbid": 211, "correlation": 9189 + } + }, + { + "ph": "s", "id": 9189, "pid": 435, "tid": 435, "ts": 1776649789987054, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987094, "dur": 23, + "args": { + "External id": 617, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9216, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9216, "pid": 0, "tid": 7, "ts": 1776649789987094, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987084, "dur": 6, + "args": { + "External id": 617, + "cbid": 211, "correlation": 9216 + } + }, + { + "ph": "s", "id": 9216, "pid": 435, "tid": 435, "ts": 1776649789987084, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987123, "dur": 23, + "args": { + "External id": 619, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9243, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9243, "pid": 0, "tid": 7, "ts": 1776649789987123, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987113, "dur": 6, + "args": { + "External id": 619, + "cbid": 211, "correlation": 9243 + } + }, + { + "ph": "s", "id": 9243, "pid": 435, "tid": 435, "ts": 1776649789987113, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987152, "dur": 23, + "args": { + "External id": 621, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9270, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9270, "pid": 0, "tid": 7, "ts": 1776649789987152, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987142, "dur": 6, + "args": { + "External id": 621, + "cbid": 211, "correlation": 9270 + } + }, + { + "ph": "s", "id": 9270, "pid": 435, "tid": 435, "ts": 1776649789987142, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987181, "dur": 23, + "args": { + "External id": 623, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9297, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9297, "pid": 0, "tid": 7, "ts": 1776649789987181, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987171, "dur": 6, + "args": { + "External id": 623, + "cbid": 211, "correlation": 9297 + } + }, + { + "ph": "s", "id": 9297, "pid": 435, "tid": 435, "ts": 1776649789987171, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987213, "dur": 23, + "args": { + "External id": 625, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9324, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9324, "pid": 0, "tid": 7, "ts": 1776649789987213, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987204, "dur": 6, + "args": { + "External id": 625, + "cbid": 211, "correlation": 9324 + } + }, + { + "ph": "s", "id": 9324, "pid": 435, "tid": 435, "ts": 1776649789987204, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987243, "dur": 23, + "args": { + "External id": 627, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9351, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9351, "pid": 0, "tid": 7, "ts": 1776649789987243, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987234, "dur": 6, + "args": { + "External id": 627, + "cbid": 211, "correlation": 9351 + } + }, + { + "ph": "s", "id": 9351, "pid": 435, "tid": 435, "ts": 1776649789987234, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987280, "dur": 23, + "args": { + "External id": 629, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9378, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9378, "pid": 0, "tid": 7, "ts": 1776649789987280, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987271, "dur": 6, + "args": { + "External id": 629, + "cbid": 211, "correlation": 9378 + } + }, + { + "ph": "s", "id": 9378, "pid": 435, "tid": 435, "ts": 1776649789987271, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987311, "dur": 23, + "args": { + "External id": 631, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9405, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9405, "pid": 0, "tid": 7, "ts": 1776649789987311, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987300, "dur": 6, + "args": { + "External id": 631, + "cbid": 211, "correlation": 9405 + } + }, + { + "ph": "s", "id": 9405, "pid": 435, "tid": 435, "ts": 1776649789987300, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987340, "dur": 23, + "args": { + "External id": 633, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9432, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9432, "pid": 0, "tid": 7, "ts": 1776649789987340, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987329, "dur": 6, + "args": { + "External id": 633, + "cbid": 211, "correlation": 9432 + } + }, + { + "ph": "s", "id": 9432, "pid": 435, "tid": 435, "ts": 1776649789987329, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987368, "dur": 23, + "args": { + "External id": 635, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9459, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9459, "pid": 0, "tid": 7, "ts": 1776649789987368, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987358, "dur": 6, + "args": { + "External id": 635, + "cbid": 211, "correlation": 9459 + } + }, + { + "ph": "s", "id": 9459, "pid": 435, "tid": 435, "ts": 1776649789987358, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987398, "dur": 23, + "args": { + "External id": 637, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9486, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9486, "pid": 0, "tid": 7, "ts": 1776649789987398, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987387, "dur": 6, + "args": { + "External id": 637, + "cbid": 211, "correlation": 9486 + } + }, + { + "ph": "s", "id": 9486, "pid": 435, "tid": 435, "ts": 1776649789987387, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987429, "dur": 22, + "args": { + "External id": 639, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9513, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9513, "pid": 0, "tid": 7, "ts": 1776649789987429, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987416, "dur": 9, + "args": { + "External id": 639, + "cbid": 211, "correlation": 9513 + } + }, + { + "ph": "s", "id": 9513, "pid": 435, "tid": 435, "ts": 1776649789987416, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987459, "dur": 23, + "args": { + "External id": 641, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9540, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9540, "pid": 0, "tid": 7, "ts": 1776649789987459, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987450, "dur": 6, + "args": { + "External id": 641, + "cbid": 211, "correlation": 9540 + } + }, + { + "ph": "s", "id": 9540, "pid": 435, "tid": 435, "ts": 1776649789987450, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987490, "dur": 23, + "args": { + "External id": 643, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9567, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9567, "pid": 0, "tid": 7, "ts": 1776649789987490, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987480, "dur": 6, + "args": { + "External id": 643, + "cbid": 211, "correlation": 9567 + } + }, + { + "ph": "s", "id": 9567, "pid": 435, "tid": 435, "ts": 1776649789987480, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987519, "dur": 23, + "args": { + "External id": 645, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9594, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9594, "pid": 0, "tid": 7, "ts": 1776649789987519, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987509, "dur": 6, + "args": { + "External id": 645, + "cbid": 211, "correlation": 9594 + } + }, + { + "ph": "s", "id": 9594, "pid": 435, "tid": 435, "ts": 1776649789987509, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987549, "dur": 23, + "args": { + "External id": 647, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9621, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9621, "pid": 0, "tid": 7, "ts": 1776649789987549, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987538, "dur": 6, + "args": { + "External id": 647, + "cbid": 211, "correlation": 9621 + } + }, + { + "ph": "s", "id": 9621, "pid": 435, "tid": 435, "ts": 1776649789987538, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987577, "dur": 23, + "args": { + "External id": 649, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9648, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9648, "pid": 0, "tid": 7, "ts": 1776649789987577, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987567, "dur": 6, + "args": { + "External id": 649, + "cbid": 211, "correlation": 9648 + } + }, + { + "ph": "s", "id": 9648, "pid": 435, "tid": 435, "ts": 1776649789987567, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987607, "dur": 23, + "args": { + "External id": 651, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9675, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9675, "pid": 0, "tid": 7, "ts": 1776649789987607, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987596, "dur": 6, + "args": { + "External id": 651, + "cbid": 211, "correlation": 9675 + } + }, + { + "ph": "s", "id": 9675, "pid": 435, "tid": 435, "ts": 1776649789987596, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987636, "dur": 23, + "args": { + "External id": 653, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9702, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9702, "pid": 0, "tid": 7, "ts": 1776649789987636, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987625, "dur": 6, + "args": { + "External id": 653, + "cbid": 211, "correlation": 9702 + } + }, + { + "ph": "s", "id": 9702, "pid": 435, "tid": 435, "ts": 1776649789987625, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987668, "dur": 23, + "args": { + "External id": 655, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9729, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9729, "pid": 0, "tid": 7, "ts": 1776649789987668, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987658, "dur": 6, + "args": { + "External id": 655, + "cbid": 211, "correlation": 9729 + } + }, + { + "ph": "s", "id": 9729, "pid": 435, "tid": 435, "ts": 1776649789987658, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987699, "dur": 23, + "args": { + "External id": 657, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9756, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9756, "pid": 0, "tid": 7, "ts": 1776649789987699, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987688, "dur": 6, + "args": { + "External id": 657, + "cbid": 211, "correlation": 9756 + } + }, + { + "ph": "s", "id": 9756, "pid": 435, "tid": 435, "ts": 1776649789987688, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987728, "dur": 23, + "args": { + "External id": 659, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9783, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9783, "pid": 0, "tid": 7, "ts": 1776649789987728, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987717, "dur": 6, + "args": { + "External id": 659, + "cbid": 211, "correlation": 9783 + } + }, + { + "ph": "s", "id": 9783, "pid": 435, "tid": 435, "ts": 1776649789987717, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987757, "dur": 23, + "args": { + "External id": 661, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9810, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9810, "pid": 0, "tid": 7, "ts": 1776649789987757, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987746, "dur": 6, + "args": { + "External id": 661, + "cbid": 211, "correlation": 9810 + } + }, + { + "ph": "s", "id": 9810, "pid": 435, "tid": 435, "ts": 1776649789987746, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987786, "dur": 23, + "args": { + "External id": 663, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9837, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9837, "pid": 0, "tid": 7, "ts": 1776649789987786, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987775, "dur": 6, + "args": { + "External id": 663, + "cbid": 211, "correlation": 9837 + } + }, + { + "ph": "s", "id": 9837, "pid": 435, "tid": 435, "ts": 1776649789987775, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987815, "dur": 23, + "args": { + "External id": 665, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9864, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9864, "pid": 0, "tid": 7, "ts": 1776649789987815, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987805, "dur": 6, + "args": { + "External id": 665, + "cbid": 211, "correlation": 9864 + } + }, + { + "ph": "s", "id": 9864, "pid": 435, "tid": 435, "ts": 1776649789987805, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987844, "dur": 23, + "args": { + "External id": 667, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9891, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9891, "pid": 0, "tid": 7, "ts": 1776649789987844, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987834, "dur": 6, + "args": { + "External id": 667, + "cbid": 211, "correlation": 9891 + } + }, + { + "ph": "s", "id": 9891, "pid": 435, "tid": 435, "ts": 1776649789987834, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987876, "dur": 23, + "args": { + "External id": 669, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9918, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9918, "pid": 0, "tid": 7, "ts": 1776649789987876, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987866, "dur": 6, + "args": { + "External id": 669, + "cbid": 211, "correlation": 9918 + } + }, + { + "ph": "s", "id": 9918, "pid": 435, "tid": 435, "ts": 1776649789987866, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987907, "dur": 23, + "args": { + "External id": 671, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9945, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9945, "pid": 0, "tid": 7, "ts": 1776649789987907, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987896, "dur": 6, + "args": { + "External id": 671, + "cbid": 211, "correlation": 9945 + } + }, + { + "ph": "s", "id": 9945, "pid": 435, "tid": 435, "ts": 1776649789987896, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987935, "dur": 23, + "args": { + "External id": 673, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9972, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9972, "pid": 0, "tid": 7, "ts": 1776649789987935, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987925, "dur": 6, + "args": { + "External id": 673, + "cbid": 211, "correlation": 9972 + } + }, + { + "ph": "s", "id": 9972, "pid": 435, "tid": 435, "ts": 1776649789987925, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987965, "dur": 23, + "args": { + "External id": 675, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 9999, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 9999, "pid": 0, "tid": 7, "ts": 1776649789987965, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987954, "dur": 6, + "args": { + "External id": 675, + "cbid": 211, "correlation": 9999 + } + }, + { + "ph": "s", "id": 9999, "pid": 435, "tid": 435, "ts": 1776649789987954, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789987999, "dur": 23, + "args": { + "External id": 677, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10026, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10026, "pid": 0, "tid": 7, "ts": 1776649789987999, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789987989, "dur": 6, + "args": { + "External id": 677, + "cbid": 211, "correlation": 10026 + } + }, + { + "ph": "s", "id": 10026, "pid": 435, "tid": 435, "ts": 1776649789987989, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789988029, "dur": 23, + "args": { + "External id": 679, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10053, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10053, "pid": 0, "tid": 7, "ts": 1776649789988029, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789988020, "dur": 6, + "args": { + "External id": 679, + "cbid": 211, "correlation": 10053 + } + }, + { + "ph": "s", "id": 10053, "pid": 435, "tid": 435, "ts": 1776649789988020, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789988059, "dur": 23, + "args": { + "External id": 681, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10080, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10080, "pid": 0, "tid": 7, "ts": 1776649789988059, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789988049, "dur": 6, + "args": { + "External id": 681, + "cbid": 211, "correlation": 10080 + } + }, + { + "ph": "s", "id": 10080, "pid": 435, "tid": 435, "ts": 1776649789988049, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789988091, "dur": 23, + "args": { + "External id": 683, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10107, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10107, "pid": 0, "tid": 7, "ts": 1776649789988091, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789988078, "dur": 9, + "args": { + "External id": 683, + "cbid": 211, "correlation": 10107 + } + }, + { + "ph": "s", "id": 10107, "pid": 435, "tid": 435, "ts": 1776649789988078, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789988121, "dur": 23, + "args": { + "External id": 685, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10134, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10134, "pid": 0, "tid": 7, "ts": 1776649789988121, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789988112, "dur": 6, + "args": { + "External id": 685, + "cbid": 211, "correlation": 10134 + } + }, + { + "ph": "s", "id": 10134, "pid": 435, "tid": 435, "ts": 1776649789988112, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x5_nn", "pid": 0, "tid": 7, + "ts": 1776649789988152, "dur": 23, + "args": { + "External id": 687, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10161, + "registers per thread": 146, + "shared memory": 81920, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10161, "pid": 0, "tid": 7, "ts": 1776649789988152, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789988141, "dur": 6, + "args": { + "External id": 687, + "cbid": 211, "correlation": 10161 + } + }, + { + "ph": "s", "id": 10161, "pid": 435, "tid": 435, "ts": 1776649789988141, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649789988175, "dur": 4, + "args": { + "External id": 10172, + "cbid": 135, "correlation": 10172 + } + }, + { + "ph": "f", "id": 10172, "pid": 435, "tid": 435, "ts": 1776649789988175, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649789988190, "dur": 5, + "args": { + "External id": 10177, + "cbid": 165, "correlation": 10177 + } + }, + { + "ph": "s", "id": 10177, "pid": 435, "tid": 435, "ts": 1776649789988190, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649789988198, "dur": 4, + "args": { + "External id": 10180, + "cbid": 139, "correlation": 10180 + } + }, + { + "ph": "f", "id": 10180, "pid": 435, "tid": 435, "ts": 1776649789988198, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor >, at::detail::Array >(int, at::native::AUnaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649789991069, "dur": 17, + "args": { + "External id": 691, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10197, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 10197, "pid": 0, "tid": 7, "ts": 1776649789991069, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789991040, "dur": 28, + "args": { + "External id": 691, + "cbid": 211, "correlation": 10197 + } + }, + { + "ph": "s", "id": 10197, "pid": 435, "tid": 435, "ts": 1776649789991040, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649789991151, "dur": 30, + "args": { + "External id": 695, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10210, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 10210, "pid": 0, "tid": 7, "ts": 1776649789991151, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789991138, "dur": 11, + "args": { + "External id": 695, + "cbid": 211, "correlation": 10210 + } + }, + { + "ph": "s", "id": 10210, "pid": 435, "tid": 435, "ts": 1776649789991138, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BUnaryFunctor, at::detail::Array >(int, at::native::BUnaryFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649789991200, "dur": 17, + "args": { + "External id": 696, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10224, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 10224, "pid": 0, "tid": 7, "ts": 1776649789991200, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789991189, "dur": 9, + "args": { + "External id": 696, + "cbid": 211, "correlation": 10224 + } + }, + { + "ph": "s", "id": 10224, "pid": 435, "tid": 435, "ts": 1776649789991189, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649789991229, "dur": 29, + "args": { + "External id": 701, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10237, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 10237, "pid": 0, "tid": 7, "ts": 1776649789991229, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789991219, "dur": 7, + "args": { + "External id": 701, + "cbid": 211, "correlation": 10237 + } + }, + { + "ph": "s", "id": 10237, "pid": 435, "tid": 435, "ts": 1776649789991219, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy_contig, unsigned int, 3, 128, 1>(at::native::(anonymous namespace)::OpaqueType<1u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata, unsigned int, 128, 1>, at::native::(anonymous namespace)::TensorSizeStride, int, unsigned int)", "pid": 0, "tid": 7, + "ts": 1776649789991306, "dur": 63, + "args": { + "External id": 707, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10249, + "registers per thread": 21, + "shared memory": 0, + "blocks per SM": 4.000000, + "warps per SM": 64.000000, + "grid": [216, 2, 1], + "block": [512, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 10249, "pid": 0, "tid": 7, "ts": 1776649789991306, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789991294, "dur": 10, + "args": { + "External id": 707, + "cbid": 211, "correlation": 10249 + } + }, + { + "ph": "s", "id": 10249, "pid": 435, "tid": 435, "ts": 1776649789991294, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649789991408, "dur": 71, + "args": { + "External id": 718, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10262, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 10262, "pid": 0, "tid": 7, "ts": 1776649789991408, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789991396, "dur": 10, + "args": { + "External id": 718, + "cbid": 211, "correlation": 10262 + } + }, + { + "ph": "s", "id": 10262, "pid": 435, "tid": 435, "ts": 1776649789991396, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, at::detail::Array >(int, at::native::CUDAFunctorOnSelf_add, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649789991480, "dur": 33, + "args": { + "External id": 721, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10273, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 10273, "pid": 0, "tid": 7, "ts": 1776649789991480, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789991449, "dur": 9, + "args": { + "External id": 721, + "cbid": 211, "correlation": 10273 + } + }, + { + "ph": "s", "id": 10273, "pid": 435, "tid": 435, "ts": 1776649789991449, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649789991515, "dur": 91, + "args": { + "External id": 725, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10286, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 10286, "pid": 0, "tid": 7, "ts": 1776649789991515, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789991482, "dur": 8, + "args": { + "External id": 725, + "cbid": 211, "correlation": 10286 + } + }, + { + "ph": "s", "id": 10286, "pid": 435, "tid": 435, "ts": 1776649789991482, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649789991515, "dur": 2, + "args": { + "External id": 726, + "cbid": 317, "correlation": 10296 + } + }, + { + "ph": "f", "id": 10296, "pid": 435, "tid": 435, "ts": 1776649789991515, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 435, "tid": 435, + "ts": 1776649789991518, "dur": 387, + "args": { + "External id": 726, + "cbid": 20, "correlation": 10297 + } + }, + { + "ph": "f", "id": 10297, "pid": 435, "tid": 435, "ts": 1776649789991518, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor >, at::detail::Array >(int, at::native::BinaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649789991934, "dur": 60, + "args": { + "External id": 726, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10301, + "registers per thread": 22, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 10301, "pid": 0, "tid": 7, "ts": 1776649789991934, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789991919, "dur": 13, + "args": { + "External id": 726, + "cbid": 211, "correlation": 10301 + } + }, + { + "ph": "s", "id": 10301, "pid": 435, "tid": 435, "ts": 1776649789991919, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649789992064, "dur": 221, + "args": { + "External id": 734, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10315, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 10315, "pid": 0, "tid": 7, "ts": 1776649789992064, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649789992051, "dur": 10, + "args": { + "External id": 734, + "cbid": 211, "correlation": 10315 + } + }, + { + "ph": "s", "id": 10315, "pid": 435, "tid": 435, "ts": 1776649789992051, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649790285205, "dur": 91532, + "args": { + "External id": 750, + "device": 0, "context": 1, + "stream": 7, "correlation": 10331, + "bytes": 33554432, "memory bandwidth (GB/s)": 0.3665842753059004 + } + }, + { + "ph": "f", "id": 10331, "pid": 0, "tid": 7, "ts": 1776649790285205, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649790285046, "dur": 91650, + "args": { + "External id": 750, + "cbid": 41, "correlation": 10331 + } + }, + { + "ph": "s", "id": 10331, "pid": 435, "tid": 435, "ts": 1776649790285046, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790376709, "dur": 41, + "args": { + "External id": 750, + "cbid": 131, "correlation": 10332 + } + }, + { + "ph": "s", "id": 10332, "pid": 435, "tid": 435, "ts": 1776649790376709, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7, + "ts": 1776649790388505, "dur": 39, + "args": { + "External id": 753, + "device": 0, "context": 1, + "stream": 7, "correlation": 10337, + "bytes": 33554432, "memory bandwidth (GB/s)": 843.5848753016894 + } + }, + { + "ph": "f", "id": 10337, "pid": 0, "tid": 7, "ts": 1776649790388505, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649790388453, "dur": 50, + "args": { + "External id": 753, + "cbid": 41, "correlation": 10337 + } + }, + { + "ph": "s", "id": 10337, "pid": 435, "tid": 435, "ts": 1776649790388453, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649790389235, "dur": 2119, + "args": { + "External id": 762, + "device": 0, "context": 1, + "stream": 7, "correlation": 10352, + "bytes": 8388608, "memory bandwidth (GB/s)": 3.9580425057670086 + } + }, + { + "ph": "f", "id": 10352, "pid": 0, "tid": 7, "ts": 1776649790389235, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649790388919, "dur": 2377, + "args": { + "External id": 762, + "cbid": 41, "correlation": 10352 + } + }, + { + "ph": "s", "id": 10352, "pid": 435, "tid": 435, "ts": 1776649790388919, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790391297, "dur": 65, + "args": { + "External id": 762, + "cbid": 131, "correlation": 10353 + } + }, + { + "ph": "s", "id": 10353, "pid": 435, "tid": 435, "ts": 1776649790391297, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649790391596, "dur": 17, + "args": { + "External id": 766, + "device": 0, "context": 1, + "stream": 7, "correlation": 10365, + "bytes": 262144, "memory bandwidth (GB/s)": 15.198515769944342 + } + }, + { + "ph": "f", "id": 10365, "pid": 0, "tid": 7, "ts": 1776649790391596, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649790391508, "dur": 86, + "args": { + "External id": 766, + "cbid": 41, "correlation": 10365 + } + }, + { + "ph": "s", "id": 10365, "pid": 435, "tid": 435, "ts": 1776649790391508, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790391595, "dur": 25, + "args": { + "External id": 766, + "cbid": 131, "correlation": 10366 + } + }, + { + "ph": "s", "id": 10366, "pid": 435, "tid": 435, "ts": 1776649790391595, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649790391655, "dur": 1, + "args": { + "External id": 770, + "device": 0, "context": 1, + "stream": 7, "correlation": 10378, + "bytes": 2048, "memory bandwidth (GB/s)": 1.4222222222222223 + } + }, + { + "ph": "f", "id": 10378, "pid": 0, "tid": 7, "ts": 1776649790391655, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649790391644, "dur": 6, + "args": { + "External id": 770, + "cbid": 41, "correlation": 10378 + } + }, + { + "ph": "s", "id": 10378, "pid": 435, "tid": 435, "ts": 1776649790391644, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790391651, "dur": 10, + "args": { + "External id": 770, + "cbid": 131, "correlation": 10379 + } + }, + { + "ph": "s", "id": 10379, "pid": 435, "tid": 435, "ts": 1776649790391651, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649790391900, "dur": 230, + "args": { + "External id": 785, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10391, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 10391, "pid": 0, "tid": 7, "ts": 1776649790391900, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790391878, "dur": 22, + "args": { + "External id": 785, + "cbid": 211, "correlation": 10391 + } + }, + { + "ph": "s", "id": 10391, "pid": 435, "tid": 435, "ts": 1776649790391878, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649790392131, "dur": 6, + "args": { + "External id": 791, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10405, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 2.370370, + "warps per SM": 9.481482, + "grid": [256, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 15 + } + }, + { + "ph": "f", "id": 10405, "pid": 0, "tid": 7, "ts": 1776649790392131, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790391954, "dur": 9, + "args": { + "External id": 791, + "cbid": 211, "correlation": 10405 + } + }, + { + "ph": "s", "id": 10405, "pid": 435, "tid": 435, "ts": 1776649790391954, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649790392139, "dur": 126, + "args": { + "External id": 793, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10416, + "registers per thread": 20, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 10416, "pid": 0, "tid": 7, "ts": 1776649790392139, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790392000, "dur": 8, + "args": { + "External id": 793, + "cbid": 211, "correlation": 10416 + } + }, + { + "ph": "s", "id": 10416, "pid": 435, "tid": 435, "ts": 1776649790392000, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649790392030, "dur": 2, + "args": { + "External id": 794, + "cbid": 317, "correlation": 10425 + } + }, + { + "ph": "f", "id": 10425, "pid": 435, "tid": 435, "ts": 1776649790392030, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 435, "tid": 435, + "ts": 1776649790392033, "dur": 444, + "args": { + "External id": 794, + "cbid": 20, "correlation": 10426 + } + }, + { + "ph": "f", "id": 10426, "pid": 435, "tid": 435, "ts": 1776649790392033, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790392505, "dur": 39, + "args": { + "External id": 794, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10430, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 10430, "pid": 0, "tid": 7, "ts": 1776649790392505, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790392490, "dur": 13, + "args": { + "External id": 794, + "cbid": 211, "correlation": 10430 + } + }, + { + "ph": "s", "id": 10430, "pid": 435, "tid": 435, "ts": 1776649790392490, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649790392555, "dur": 1, + "args": { + "External id": 797, + "cbid": 317, "correlation": 10439 + } + }, + { + "ph": "f", "id": 10439, "pid": 435, "tid": 435, "ts": 1776649790392555, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 435, "tid": 435, + "ts": 1776649790392556, "dur": 215, + "args": { + "External id": 797, + "cbid": 20, "correlation": 10440 + } + }, + { + "ph": "f", "id": 10440, "pid": 435, "tid": 435, "ts": 1776649790392556, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649790392926, "dur": 99, + "args": { + "External id": 798, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10447, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 10447, "pid": 0, "tid": 7, "ts": 1776649790392926, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790392791, "dur": 134, + "args": { + "External id": 798, + "cbid": 211, "correlation": 10447 + } + }, + { + "ph": "s", "id": 10447, "pid": 435, "tid": 435, "ts": 1776649790392791, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, at::detail::Array >(int, at::native::CUDAFunctorOnSelf_add, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790393064, "dur": 75, + "args": { + "External id": 799, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10454, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 10454, "pid": 0, "tid": 7, "ts": 1776649790393064, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790393014, "dur": 48, + "args": { + "External id": 799, + "cbid": 211, "correlation": 10454 + } + }, + { + "ph": "s", "id": 10454, "pid": 435, "tid": 435, "ts": 1776649790393014, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649790393130, "dur": 1, + "args": { + "External id": 800, + "cbid": 317, "correlation": 10463 + } + }, + { + "ph": "f", "id": 10463, "pid": 435, "tid": 435, "ts": 1776649790393130, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 435, "tid": 435, + "ts": 1776649790393131, "dur": 235, + "args": { + "External id": 800, + "cbid": 20, "correlation": 10464 + } + }, + { + "ph": "f", "id": 10464, "pid": 435, "tid": 435, "ts": 1776649790393131, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1}, at::detail::Array >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790393437, "dur": 76, + "args": { + "External id": 800, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10468, + "registers per thread": 22, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 10468, "pid": 0, "tid": 7, "ts": 1776649790393437, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790393383, "dur": 53, + "args": { + "External id": 800, + "cbid": 211, "correlation": 10468 + } + }, + { + "ph": "s", "id": 10468, "pid": 435, "tid": 435, "ts": 1776649790393383, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649790393575, "dur": 102, + "args": { + "External id": 810, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10481, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 10481, "pid": 0, "tid": 7, "ts": 1776649790393575, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790393521, "dur": 55, + "args": { + "External id": 810, + "cbid": 211, "correlation": 10481 + } + }, + { + "ph": "s", "id": 10481, "pid": 435, "tid": 435, "ts": 1776649790393521, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649790393751, "dur": 1, + "args": { + "External id": 820, + "device": 0, "context": 1, + "stream": 7, "correlation": 10496, + "bytes": 512, "memory bandwidth (GB/s)": 0.38066914498141263 + } + }, + { + "ph": "f", "id": 10496, "pid": 0, "tid": 7, "ts": 1776649790393751, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649790393736, "dur": 10, + "args": { + "External id": 820, + "cbid": 41, "correlation": 10496 + } + }, + { + "ph": "s", "id": 10496, "pid": 435, "tid": 435, "ts": 1776649790393736, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790393747, "dur": 10, + "args": { + "External id": 820, + "cbid": 131, "correlation": 10497 + } + }, + { + "ph": "s", "id": 10497, "pid": 435, "tid": 435, "ts": 1776649790393747, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649790410147, "dur": 9, + "args": { + "External id": 822, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10506, + "registers per thread": 40, + "shared memory": 0, + "blocks per SM": 2.370370, + "warps per SM": 9.481482, + "grid": [256, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 15 + } + }, + { + "ph": "f", "id": 10506, "pid": 0, "tid": 7, "ts": 1776649790410147, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790393854, "dur": 16297, + "args": { + "External id": 822, + "cbid": 211, "correlation": 10506 + } + }, + { + "ph": "s", "id": 10506, "pid": 435, "tid": 435, "ts": 1776649790393854, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649790410487, "dur": 110, + "args": { + "External id": 836, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10521, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 10521, "pid": 0, "tid": 7, "ts": 1776649790410487, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790410471, "dur": 14, + "args": { + "External id": 836, + "cbid": 211, "correlation": 10521 + } + }, + { + "ph": "s", "id": 10521, "pid": 435, "tid": 435, "ts": 1776649790410471, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649790410622, "dur": 1, + "args": { + "External id": 846, + "device": 0, "context": 1, + "stream": 7, "correlation": 10536, + "bytes": 8192, "memory bandwidth (GB/s)": 4.654545454545454 + } + }, + { + "ph": "f", "id": 10536, "pid": 0, "tid": 7, "ts": 1776649790410622, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649790410600, "dur": 18, + "args": { + "External id": 846, + "cbid": 41, "correlation": 10536 + } + }, + { + "ph": "s", "id": 10536, "pid": 435, "tid": 435, "ts": 1776649790410600, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790410620, "dur": 10, + "args": { + "External id": 846, + "cbid": 131, "correlation": 10537 + } + }, + { + "ph": "s", "id": 10537, "pid": 435, "tid": 435, "ts": 1776649790410620, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649790410747, "dur": 126, + "args": { + "External id": 842, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10546, + "registers per thread": 40, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 75 + } + }, + { + "ph": "f", "id": 10546, "pid": 0, "tid": 7, "ts": 1776649790410747, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790410671, "dur": 74, + "args": { + "External id": 842, + "cbid": 211, "correlation": 10546 + } + }, + { + "ph": "s", "id": 10546, "pid": 435, "tid": 435, "ts": 1776649790410671, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoH (Device -> Pageable)", "pid": 0, "tid": 7, + "ts": 1776649790410955, "dur": 44157, + "args": { + "External id": 855, + "device": 0, "context": 1, + "stream": 7, "correlation": 10554, + "bytes": 67108864, "memory bandwidth (GB/s)": 1.5197756138216274 + } + }, + { + "ph": "f", "id": 10554, "pid": 0, "tid": 7, "ts": 1776649790410955, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649790410934, "dur": 45429, + "args": { + "External id": 855, + "cbid": 41, "correlation": 10554 + } + }, + { + "ph": "s", "id": 10554, "pid": 435, "tid": 435, "ts": 1776649790410934, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790456365, "dur": 6, + "args": { + "External id": 855, + "cbid": 131, "correlation": 10555 + } + }, + { + "ph": "s", "id": 10555, "pid": 435, "tid": 435, "ts": 1776649790456365, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649790528127, "dur": 526, + "args": { + "External id": 865, + "device": 0, "context": 1, + "stream": 7, "correlation": 10567, + "bytes": 8388608, "memory bandwidth (GB/s)": 15.918447589444641 + } + }, + { + "ph": "f", "id": 10567, "pid": 0, "tid": 7, "ts": 1776649790528127, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649790527984, "dur": 612, + "args": { + "External id": 865, + "cbid": 41, "correlation": 10567 + } + }, + { + "ph": "s", "id": 10567, "pid": 435, "tid": 435, "ts": 1776649790527984, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790528599, "dur": 66, + "args": { + "External id": 865, + "cbid": 131, "correlation": 10568 + } + }, + { + "ph": "s", "id": 10568, "pid": 435, "tid": 435, "ts": 1776649790528599, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7, + "ts": 1776649790528845, "dur": 9, + "args": { + "External id": 871, + "device": 0, "context": 1, + "stream": 7, "correlation": 10573, + "bytes": 8388608, "memory bandwidth (GB/s)": 873.8133333333334 + } + }, + { + "ph": "f", "id": 10573, "pid": 0, "tid": 7, "ts": 1776649790528845, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649790528803, "dur": 39, + "args": { + "External id": 871, + "cbid": 41, "correlation": 10573 + } + }, + { + "ph": "s", "id": 10573, "pid": 435, "tid": 435, "ts": 1776649790528803, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7, + "ts": 1776649790528895, "dur": 3, + "args": { + "External id": 877, + "device": 0, "context": 1, + "stream": 7, "correlation": 10581, + "bytes": 262144, "memory bandwidth (GB/s)": 76.5607476635514 + } + }, + { + "ph": "f", "id": 10581, "pid": 0, "tid": 7, "ts": 1776649790528895, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649790528879, "dur": 13, + "args": { + "External id": 877, + "cbid": 41, "correlation": 10581 + } + }, + { + "ph": "s", "id": 10581, "pid": 435, "tid": 435, "ts": 1776649790528879, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790531065, "dur": 1, + "args": { + "External id": 882, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10604, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10604, "pid": 0, "tid": 7, "ts": 1776649790531065, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790529593, "dur": 1473, + "args": { + "External id": 882, + "cbid": 211, "correlation": 10604 + } + }, + { + "ph": "s", "id": 10604, "pid": 435, "tid": 435, "ts": 1776649790529593, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790531095, "dur": 2, + "args": { + "External id": 10610, + "cbid": 200, "correlation": 10610 + } + }, + { + "ph": "f", "id": 10610, "pid": 435, "tid": 435, "ts": 1776649790531095, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790531098, "dur": 222, + "args": { + "External id": 10611, + "cbid": 273, "correlation": 10611 + } + }, + { + "ph": "f", "id": 10611, "pid": 435, "tid": 435, "ts": 1776649790531098, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790531333, "dur": 15, + "args": { + "External id": 10612, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10612, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10612, "pid": 0, "tid": 7, "ts": 1776649790531333, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790531322, "dur": 8, + "args": { + "External id": 10612, + "cbid": 211, "correlation": 10612 + } + }, + { + "ph": "s", "id": 10612, "pid": 435, "tid": 435, "ts": 1776649790531322, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790531413, "dur": 1, + "args": { + "External id": 887, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10630, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10630, "pid": 0, "tid": 7, "ts": 1776649790531413, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790531402, "dur": 8, + "args": { + "External id": 887, + "cbid": 211, "correlation": 10630 + } + }, + { + "ph": "s", "id": 10630, "pid": 435, "tid": 435, "ts": 1776649790531402, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790531419, "dur": 0, + "args": { + "External id": 10636, + "cbid": 200, "correlation": 10636 + } + }, + { + "ph": "f", "id": 10636, "pid": 435, "tid": 435, "ts": 1776649790531419, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790531420, "dur": 1, + "args": { + "External id": 10637, + "cbid": 273, "correlation": 10637 + } + }, + { + "ph": "f", "id": 10637, "pid": 435, "tid": 435, "ts": 1776649790531420, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790531429, "dur": 15, + "args": { + "External id": 10638, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10638, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10638, "pid": 0, "tid": 7, "ts": 1776649790531429, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790531421, "dur": 5, + "args": { + "External id": 10638, + "cbid": 211, "correlation": 10638 + } + }, + { + "ph": "s", "id": 10638, "pid": 435, "tid": 435, "ts": 1776649790531421, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790531470, "dur": 1, + "args": { + "External id": 892, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10656, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10656, "pid": 0, "tid": 7, "ts": 1776649790531470, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790531460, "dur": 7, + "args": { + "External id": 892, + "cbid": 211, "correlation": 10656 + } + }, + { + "ph": "s", "id": 10656, "pid": 435, "tid": 435, "ts": 1776649790531460, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790531473, "dur": 0, + "args": { + "External id": 10662, + "cbid": 200, "correlation": 10662 + } + }, + { + "ph": "f", "id": 10662, "pid": 435, "tid": 435, "ts": 1776649790531473, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790531474, "dur": 0, + "args": { + "External id": 10663, + "cbid": 273, "correlation": 10663 + } + }, + { + "ph": "f", "id": 10663, "pid": 435, "tid": 435, "ts": 1776649790531474, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790531483, "dur": 15, + "args": { + "External id": 10664, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10664, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10664, "pid": 0, "tid": 7, "ts": 1776649790531483, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790531475, "dur": 5, + "args": { + "External id": 10664, + "cbid": 211, "correlation": 10664 + } + }, + { + "ph": "s", "id": 10664, "pid": 435, "tid": 435, "ts": 1776649790531475, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790531520, "dur": 1, + "args": { + "External id": 897, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10682, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10682, "pid": 0, "tid": 7, "ts": 1776649790531520, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790531511, "dur": 7, + "args": { + "External id": 897, + "cbid": 211, "correlation": 10682 + } + }, + { + "ph": "s", "id": 10682, "pid": 435, "tid": 435, "ts": 1776649790531511, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790531524, "dur": 0, + "args": { + "External id": 10688, + "cbid": 200, "correlation": 10688 + } + }, + { + "ph": "f", "id": 10688, "pid": 435, "tid": 435, "ts": 1776649790531524, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790531524, "dur": 0, + "args": { + "External id": 10689, + "cbid": 273, "correlation": 10689 + } + }, + { + "ph": "f", "id": 10689, "pid": 435, "tid": 435, "ts": 1776649790531524, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790531533, "dur": 15, + "args": { + "External id": 10690, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10690, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10690, "pid": 0, "tid": 7, "ts": 1776649790531533, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790531525, "dur": 5, + "args": { + "External id": 10690, + "cbid": 211, "correlation": 10690 + } + }, + { + "ph": "s", "id": 10690, "pid": 435, "tid": 435, "ts": 1776649790531525, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790531570, "dur": 1, + "args": { + "External id": 902, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10708, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10708, "pid": 0, "tid": 7, "ts": 1776649790531570, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790531560, "dur": 6, + "args": { + "External id": 902, + "cbid": 211, "correlation": 10708 + } + }, + { + "ph": "s", "id": 10708, "pid": 435, "tid": 435, "ts": 1776649790531560, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790531573, "dur": 0, + "args": { + "External id": 10714, + "cbid": 200, "correlation": 10714 + } + }, + { + "ph": "f", "id": 10714, "pid": 435, "tid": 435, "ts": 1776649790531573, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790531574, "dur": 0, + "args": { + "External id": 10715, + "cbid": 273, "correlation": 10715 + } + }, + { + "ph": "f", "id": 10715, "pid": 435, "tid": 435, "ts": 1776649790531574, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790531583, "dur": 15, + "args": { + "External id": 10716, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10716, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10716, "pid": 0, "tid": 7, "ts": 1776649790531583, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790531575, "dur": 4, + "args": { + "External id": 10716, + "cbid": 211, "correlation": 10716 + } + }, + { + "ph": "s", "id": 10716, "pid": 435, "tid": 435, "ts": 1776649790531575, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790531618, "dur": 1, + "args": { + "External id": 907, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10734, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10734, "pid": 0, "tid": 7, "ts": 1776649790531618, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790531609, "dur": 6, + "args": { + "External id": 907, + "cbid": 211, "correlation": 10734 + } + }, + { + "ph": "s", "id": 10734, "pid": 435, "tid": 435, "ts": 1776649790531609, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790531622, "dur": 0, + "args": { + "External id": 10740, + "cbid": 200, "correlation": 10740 + } + }, + { + "ph": "f", "id": 10740, "pid": 435, "tid": 435, "ts": 1776649790531622, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790531622, "dur": 0, + "args": { + "External id": 10741, + "cbid": 273, "correlation": 10741 + } + }, + { + "ph": "f", "id": 10741, "pid": 435, "tid": 435, "ts": 1776649790531622, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790531631, "dur": 16, + "args": { + "External id": 10742, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10742, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10742, "pid": 0, "tid": 7, "ts": 1776649790531631, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790531624, "dur": 4, + "args": { + "External id": 10742, + "cbid": 211, "correlation": 10742 + } + }, + { + "ph": "s", "id": 10742, "pid": 435, "tid": 435, "ts": 1776649790531624, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790531667, "dur": 1, + "args": { + "External id": 912, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10760, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10760, "pid": 0, "tid": 7, "ts": 1776649790531667, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790531657, "dur": 6, + "args": { + "External id": 912, + "cbid": 211, "correlation": 10760 + } + }, + { + "ph": "s", "id": 10760, "pid": 435, "tid": 435, "ts": 1776649790531657, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790531670, "dur": 0, + "args": { + "External id": 10766, + "cbid": 200, "correlation": 10766 + } + }, + { + "ph": "f", "id": 10766, "pid": 435, "tid": 435, "ts": 1776649790531670, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790531671, "dur": 0, + "args": { + "External id": 10767, + "cbid": 273, "correlation": 10767 + } + }, + { + "ph": "f", "id": 10767, "pid": 435, "tid": 435, "ts": 1776649790531671, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790531680, "dur": 16, + "args": { + "External id": 10768, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10768, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10768, "pid": 0, "tid": 7, "ts": 1776649790531680, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790531672, "dur": 4, + "args": { + "External id": 10768, + "cbid": 211, "correlation": 10768 + } + }, + { + "ph": "s", "id": 10768, "pid": 435, "tid": 435, "ts": 1776649790531672, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790531716, "dur": 1, + "args": { + "External id": 917, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10786, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10786, "pid": 0, "tid": 7, "ts": 1776649790531716, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790531706, "dur": 6, + "args": { + "External id": 917, + "cbid": 211, "correlation": 10786 + } + }, + { + "ph": "s", "id": 10786, "pid": 435, "tid": 435, "ts": 1776649790531706, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790531719, "dur": 0, + "args": { + "External id": 10792, + "cbid": 200, "correlation": 10792 + } + }, + { + "ph": "f", "id": 10792, "pid": 435, "tid": 435, "ts": 1776649790531719, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790531719, "dur": 0, + "args": { + "External id": 10793, + "cbid": 273, "correlation": 10793 + } + }, + { + "ph": "f", "id": 10793, "pid": 435, "tid": 435, "ts": 1776649790531719, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790531728, "dur": 15, + "args": { + "External id": 10794, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10794, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10794, "pid": 0, "tid": 7, "ts": 1776649790531728, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790531720, "dur": 4, + "args": { + "External id": 10794, + "cbid": 211, "correlation": 10794 + } + }, + { + "ph": "s", "id": 10794, "pid": 435, "tid": 435, "ts": 1776649790531720, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790531763, "dur": 1, + "args": { + "External id": 922, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10812, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10812, "pid": 0, "tid": 7, "ts": 1776649790531763, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790531754, "dur": 6, + "args": { + "External id": 922, + "cbid": 211, "correlation": 10812 + } + }, + { + "ph": "s", "id": 10812, "pid": 435, "tid": 435, "ts": 1776649790531754, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790531767, "dur": 0, + "args": { + "External id": 10818, + "cbid": 200, "correlation": 10818 + } + }, + { + "ph": "f", "id": 10818, "pid": 435, "tid": 435, "ts": 1776649790531767, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790531768, "dur": 0, + "args": { + "External id": 10819, + "cbid": 273, "correlation": 10819 + } + }, + { + "ph": "f", "id": 10819, "pid": 435, "tid": 435, "ts": 1776649790531768, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790531777, "dur": 15, + "args": { + "External id": 10820, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10820, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10820, "pid": 0, "tid": 7, "ts": 1776649790531777, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790531769, "dur": 4, + "args": { + "External id": 10820, + "cbid": 211, "correlation": 10820 + } + }, + { + "ph": "s", "id": 10820, "pid": 435, "tid": 435, "ts": 1776649790531769, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790531812, "dur": 1, + "args": { + "External id": 927, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10838, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10838, "pid": 0, "tid": 7, "ts": 1776649790531812, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790531803, "dur": 6, + "args": { + "External id": 927, + "cbid": 211, "correlation": 10838 + } + }, + { + "ph": "s", "id": 10838, "pid": 435, "tid": 435, "ts": 1776649790531803, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790531815, "dur": 0, + "args": { + "External id": 10844, + "cbid": 200, "correlation": 10844 + } + }, + { + "ph": "f", "id": 10844, "pid": 435, "tid": 435, "ts": 1776649790531815, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790531816, "dur": 0, + "args": { + "External id": 10845, + "cbid": 273, "correlation": 10845 + } + }, + { + "ph": "f", "id": 10845, "pid": 435, "tid": 435, "ts": 1776649790531816, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790531825, "dur": 15, + "args": { + "External id": 10846, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10846, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10846, "pid": 0, "tid": 7, "ts": 1776649790531825, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790531817, "dur": 4, + "args": { + "External id": 10846, + "cbid": 211, "correlation": 10846 + } + }, + { + "ph": "s", "id": 10846, "pid": 435, "tid": 435, "ts": 1776649790531817, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790531877, "dur": 8, + "args": { + "External id": 10849, + "cbid": 165, "correlation": 10849 + } + }, + { + "ph": "s", "id": 10849, "pid": 435, "tid": 435, "ts": 1776649790531877, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790531932, "dur": 8, + "args": { + "External id": 10858, + "cbid": 135, "correlation": 10858 + } + }, + { + "ph": "f", "id": 10858, "pid": 435, "tid": 435, "ts": 1776649790531932, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790531990, "dur": 1, + "args": { + "External id": 932, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10878, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10878, "pid": 0, "tid": 7, "ts": 1776649790531990, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790531978, "dur": 9, + "args": { + "External id": 932, + "cbid": 211, "correlation": 10878 + } + }, + { + "ph": "s", "id": 10878, "pid": 435, "tid": 435, "ts": 1776649790531978, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790531995, "dur": 0, + "args": { + "External id": 10884, + "cbid": 200, "correlation": 10884 + } + }, + { + "ph": "f", "id": 10884, "pid": 435, "tid": 435, "ts": 1776649790531995, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790531996, "dur": 1, + "args": { + "External id": 10885, + "cbid": 273, "correlation": 10885 + } + }, + { + "ph": "f", "id": 10885, "pid": 435, "tid": 435, "ts": 1776649790531996, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790532006, "dur": 15, + "args": { + "External id": 10886, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10886, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10886, "pid": 0, "tid": 7, "ts": 1776649790532006, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790531997, "dur": 5, + "args": { + "External id": 10886, + "cbid": 211, "correlation": 10886 + } + }, + { + "ph": "s", "id": 10886, "pid": 435, "tid": 435, "ts": 1776649790531997, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790532044, "dur": 1, + "args": { + "External id": 937, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10904, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10904, "pid": 0, "tid": 7, "ts": 1776649790532044, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532035, "dur": 7, + "args": { + "External id": 937, + "cbid": 211, "correlation": 10904 + } + }, + { + "ph": "s", "id": 10904, "pid": 435, "tid": 435, "ts": 1776649790532035, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532048, "dur": 0, + "args": { + "External id": 10910, + "cbid": 200, "correlation": 10910 + } + }, + { + "ph": "f", "id": 10910, "pid": 435, "tid": 435, "ts": 1776649790532048, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532048, "dur": 0, + "args": { + "External id": 10911, + "cbid": 273, "correlation": 10911 + } + }, + { + "ph": "f", "id": 10911, "pid": 435, "tid": 435, "ts": 1776649790532048, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790532057, "dur": 15, + "args": { + "External id": 10912, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10912, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10912, "pid": 0, "tid": 7, "ts": 1776649790532057, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532049, "dur": 5, + "args": { + "External id": 10912, + "cbid": 211, "correlation": 10912 + } + }, + { + "ph": "s", "id": 10912, "pid": 435, "tid": 435, "ts": 1776649790532049, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790532094, "dur": 1, + "args": { + "External id": 942, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10930, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10930, "pid": 0, "tid": 7, "ts": 1776649790532094, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532085, "dur": 6, + "args": { + "External id": 942, + "cbid": 211, "correlation": 10930 + } + }, + { + "ph": "s", "id": 10930, "pid": 435, "tid": 435, "ts": 1776649790532085, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532097, "dur": 0, + "args": { + "External id": 10936, + "cbid": 200, "correlation": 10936 + } + }, + { + "ph": "f", "id": 10936, "pid": 435, "tid": 435, "ts": 1776649790532097, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532098, "dur": 0, + "args": { + "External id": 10937, + "cbid": 273, "correlation": 10937 + } + }, + { + "ph": "f", "id": 10937, "pid": 435, "tid": 435, "ts": 1776649790532098, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790532107, "dur": 15, + "args": { + "External id": 10938, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10938, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10938, "pid": 0, "tid": 7, "ts": 1776649790532107, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532099, "dur": 4, + "args": { + "External id": 10938, + "cbid": 211, "correlation": 10938 + } + }, + { + "ph": "s", "id": 10938, "pid": 435, "tid": 435, "ts": 1776649790532099, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790532143, "dur": 1, + "args": { + "External id": 947, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10956, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10956, "pid": 0, "tid": 7, "ts": 1776649790532143, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532133, "dur": 6, + "args": { + "External id": 947, + "cbid": 211, "correlation": 10956 + } + }, + { + "ph": "s", "id": 10956, "pid": 435, "tid": 435, "ts": 1776649790532133, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532146, "dur": 0, + "args": { + "External id": 10962, + "cbid": 200, "correlation": 10962 + } + }, + { + "ph": "f", "id": 10962, "pid": 435, "tid": 435, "ts": 1776649790532146, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532146, "dur": 0, + "args": { + "External id": 10963, + "cbid": 273, "correlation": 10963 + } + }, + { + "ph": "f", "id": 10963, "pid": 435, "tid": 435, "ts": 1776649790532146, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790532155, "dur": 15, + "args": { + "External id": 10964, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10964, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10964, "pid": 0, "tid": 7, "ts": 1776649790532155, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532148, "dur": 4, + "args": { + "External id": 10964, + "cbid": 211, "correlation": 10964 + } + }, + { + "ph": "s", "id": 10964, "pid": 435, "tid": 435, "ts": 1776649790532148, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790532191, "dur": 1, + "args": { + "External id": 952, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10982, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10982, "pid": 0, "tid": 7, "ts": 1776649790532191, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532182, "dur": 6, + "args": { + "External id": 952, + "cbid": 211, "correlation": 10982 + } + }, + { + "ph": "s", "id": 10982, "pid": 435, "tid": 435, "ts": 1776649790532182, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532194, "dur": 0, + "args": { + "External id": 10988, + "cbid": 200, "correlation": 10988 + } + }, + { + "ph": "f", "id": 10988, "pid": 435, "tid": 435, "ts": 1776649790532194, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532195, "dur": 0, + "args": { + "External id": 10989, + "cbid": 273, "correlation": 10989 + } + }, + { + "ph": "f", "id": 10989, "pid": 435, "tid": 435, "ts": 1776649790532195, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790532204, "dur": 16, + "args": { + "External id": 10990, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 10990, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 10990, "pid": 0, "tid": 7, "ts": 1776649790532204, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532196, "dur": 5, + "args": { + "External id": 10990, + "cbid": 211, "correlation": 10990 + } + }, + { + "ph": "s", "id": 10990, "pid": 435, "tid": 435, "ts": 1776649790532196, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790532288, "dur": 1, + "args": { + "External id": 957, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11008, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11008, "pid": 0, "tid": 7, "ts": 1776649790532288, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532277, "dur": 8, + "args": { + "External id": 957, + "cbid": 211, "correlation": 11008 + } + }, + { + "ph": "s", "id": 11008, "pid": 435, "tid": 435, "ts": 1776649790532277, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532292, "dur": 0, + "args": { + "External id": 11014, + "cbid": 200, "correlation": 11014 + } + }, + { + "ph": "f", "id": 11014, "pid": 435, "tid": 435, "ts": 1776649790532292, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532293, "dur": 0, + "args": { + "External id": 11015, + "cbid": 273, "correlation": 11015 + } + }, + { + "ph": "f", "id": 11015, "pid": 435, "tid": 435, "ts": 1776649790532293, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790532302, "dur": 15, + "args": { + "External id": 11016, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11016, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11016, "pid": 0, "tid": 7, "ts": 1776649790532302, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532294, "dur": 5, + "args": { + "External id": 11016, + "cbid": 211, "correlation": 11016 + } + }, + { + "ph": "s", "id": 11016, "pid": 435, "tid": 435, "ts": 1776649790532294, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790532339, "dur": 1, + "args": { + "External id": 962, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11034, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11034, "pid": 0, "tid": 7, "ts": 1776649790532339, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532330, "dur": 6, + "args": { + "External id": 962, + "cbid": 211, "correlation": 11034 + } + }, + { + "ph": "s", "id": 11034, "pid": 435, "tid": 435, "ts": 1776649790532330, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532343, "dur": 0, + "args": { + "External id": 11040, + "cbid": 200, "correlation": 11040 + } + }, + { + "ph": "f", "id": 11040, "pid": 435, "tid": 435, "ts": 1776649790532343, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532344, "dur": 0, + "args": { + "External id": 11041, + "cbid": 273, "correlation": 11041 + } + }, + { + "ph": "f", "id": 11041, "pid": 435, "tid": 435, "ts": 1776649790532344, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790532353, "dur": 15, + "args": { + "External id": 11042, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11042, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11042, "pid": 0, "tid": 7, "ts": 1776649790532353, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532345, "dur": 4, + "args": { + "External id": 11042, + "cbid": 211, "correlation": 11042 + } + }, + { + "ph": "s", "id": 11042, "pid": 435, "tid": 435, "ts": 1776649790532345, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790532388, "dur": 1, + "args": { + "External id": 967, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11060, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11060, "pid": 0, "tid": 7, "ts": 1776649790532388, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532379, "dur": 6, + "args": { + "External id": 967, + "cbid": 211, "correlation": 11060 + } + }, + { + "ph": "s", "id": 11060, "pid": 435, "tid": 435, "ts": 1776649790532379, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532391, "dur": 0, + "args": { + "External id": 11066, + "cbid": 200, "correlation": 11066 + } + }, + { + "ph": "f", "id": 11066, "pid": 435, "tid": 435, "ts": 1776649790532391, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532392, "dur": 0, + "args": { + "External id": 11067, + "cbid": 273, "correlation": 11067 + } + }, + { + "ph": "f", "id": 11067, "pid": 435, "tid": 435, "ts": 1776649790532392, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790532401, "dur": 15, + "args": { + "External id": 11068, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11068, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11068, "pid": 0, "tid": 7, "ts": 1776649790532401, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532393, "dur": 4, + "args": { + "External id": 11068, + "cbid": 211, "correlation": 11068 + } + }, + { + "ph": "s", "id": 11068, "pid": 435, "tid": 435, "ts": 1776649790532393, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790532437, "dur": 1, + "args": { + "External id": 972, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11086, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11086, "pid": 0, "tid": 7, "ts": 1776649790532437, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532427, "dur": 6, + "args": { + "External id": 972, + "cbid": 211, "correlation": 11086 + } + }, + { + "ph": "s", "id": 11086, "pid": 435, "tid": 435, "ts": 1776649790532427, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532440, "dur": 0, + "args": { + "External id": 11092, + "cbid": 200, "correlation": 11092 + } + }, + { + "ph": "f", "id": 11092, "pid": 435, "tid": 435, "ts": 1776649790532440, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532440, "dur": 0, + "args": { + "External id": 11093, + "cbid": 273, "correlation": 11093 + } + }, + { + "ph": "f", "id": 11093, "pid": 435, "tid": 435, "ts": 1776649790532440, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790532450, "dur": 15, + "args": { + "External id": 11094, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11094, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11094, "pid": 0, "tid": 7, "ts": 1776649790532450, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532442, "dur": 5, + "args": { + "External id": 11094, + "cbid": 211, "correlation": 11094 + } + }, + { + "ph": "s", "id": 11094, "pid": 435, "tid": 435, "ts": 1776649790532442, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790532486, "dur": 1, + "args": { + "External id": 977, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11112, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11112, "pid": 0, "tid": 7, "ts": 1776649790532486, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532476, "dur": 6, + "args": { + "External id": 977, + "cbid": 211, "correlation": 11112 + } + }, + { + "ph": "s", "id": 11112, "pid": 435, "tid": 435, "ts": 1776649790532476, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532489, "dur": 0, + "args": { + "External id": 11118, + "cbid": 200, "correlation": 11118 + } + }, + { + "ph": "f", "id": 11118, "pid": 435, "tid": 435, "ts": 1776649790532489, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532489, "dur": 0, + "args": { + "External id": 11119, + "cbid": 273, "correlation": 11119 + } + }, + { + "ph": "f", "id": 11119, "pid": 435, "tid": 435, "ts": 1776649790532489, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790532498, "dur": 15, + "args": { + "External id": 11120, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11120, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11120, "pid": 0, "tid": 7, "ts": 1776649790532498, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532490, "dur": 4, + "args": { + "External id": 11120, + "cbid": 211, "correlation": 11120 + } + }, + { + "ph": "s", "id": 11120, "pid": 435, "tid": 435, "ts": 1776649790532490, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790532534, "dur": 1, + "args": { + "External id": 982, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11138, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11138, "pid": 0, "tid": 7, "ts": 1776649790532534, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532525, "dur": 6, + "args": { + "External id": 982, + "cbid": 211, "correlation": 11138 + } + }, + { + "ph": "s", "id": 11138, "pid": 435, "tid": 435, "ts": 1776649790532525, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532537, "dur": 0, + "args": { + "External id": 11144, + "cbid": 200, "correlation": 11144 + } + }, + { + "ph": "f", "id": 11144, "pid": 435, "tid": 435, "ts": 1776649790532537, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532538, "dur": 0, + "args": { + "External id": 11145, + "cbid": 273, "correlation": 11145 + } + }, + { + "ph": "f", "id": 11145, "pid": 435, "tid": 435, "ts": 1776649790532538, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790532547, "dur": 15, + "args": { + "External id": 11146, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11146, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11146, "pid": 0, "tid": 7, "ts": 1776649790532547, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532539, "dur": 4, + "args": { + "External id": 11146, + "cbid": 211, "correlation": 11146 + } + }, + { + "ph": "s", "id": 11146, "pid": 435, "tid": 435, "ts": 1776649790532539, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790532582, "dur": 1, + "args": { + "External id": 987, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11164, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11164, "pid": 0, "tid": 7, "ts": 1776649790532582, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532573, "dur": 6, + "args": { + "External id": 987, + "cbid": 211, "correlation": 11164 + } + }, + { + "ph": "s", "id": 11164, "pid": 435, "tid": 435, "ts": 1776649790532573, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532585, "dur": 0, + "args": { + "External id": 11170, + "cbid": 200, "correlation": 11170 + } + }, + { + "ph": "f", "id": 11170, "pid": 435, "tid": 435, "ts": 1776649790532585, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532586, "dur": 0, + "args": { + "External id": 11171, + "cbid": 273, "correlation": 11171 + } + }, + { + "ph": "f", "id": 11171, "pid": 435, "tid": 435, "ts": 1776649790532586, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790532595, "dur": 15, + "args": { + "External id": 11172, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11172, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11172, "pid": 0, "tid": 7, "ts": 1776649790532595, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532587, "dur": 4, + "args": { + "External id": 11172, + "cbid": 211, "correlation": 11172 + } + }, + { + "ph": "s", "id": 11172, "pid": 435, "tid": 435, "ts": 1776649790532587, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790532630, "dur": 1, + "args": { + "External id": 992, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11190, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11190, "pid": 0, "tid": 7, "ts": 1776649790532630, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532621, "dur": 6, + "args": { + "External id": 992, + "cbid": 211, "correlation": 11190 + } + }, + { + "ph": "s", "id": 11190, "pid": 435, "tid": 435, "ts": 1776649790532621, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532633, "dur": 0, + "args": { + "External id": 11196, + "cbid": 200, "correlation": 11196 + } + }, + { + "ph": "f", "id": 11196, "pid": 435, "tid": 435, "ts": 1776649790532633, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532634, "dur": 0, + "args": { + "External id": 11197, + "cbid": 273, "correlation": 11197 + } + }, + { + "ph": "f", "id": 11197, "pid": 435, "tid": 435, "ts": 1776649790532634, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790532643, "dur": 15, + "args": { + "External id": 11198, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11198, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11198, "pid": 0, "tid": 7, "ts": 1776649790532643, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532635, "dur": 4, + "args": { + "External id": 11198, + "cbid": 211, "correlation": 11198 + } + }, + { + "ph": "s", "id": 11198, "pid": 435, "tid": 435, "ts": 1776649790532635, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790532678, "dur": 1, + "args": { + "External id": 997, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11216, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11216, "pid": 0, "tid": 7, "ts": 1776649790532678, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532669, "dur": 6, + "args": { + "External id": 997, + "cbid": 211, "correlation": 11216 + } + }, + { + "ph": "s", "id": 11216, "pid": 435, "tid": 435, "ts": 1776649790532669, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532681, "dur": 0, + "args": { + "External id": 11222, + "cbid": 200, "correlation": 11222 + } + }, + { + "ph": "f", "id": 11222, "pid": 435, "tid": 435, "ts": 1776649790532681, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532682, "dur": 0, + "args": { + "External id": 11223, + "cbid": 273, "correlation": 11223 + } + }, + { + "ph": "f", "id": 11223, "pid": 435, "tid": 435, "ts": 1776649790532682, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790532691, "dur": 17, + "args": { + "External id": 11224, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11224, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11224, "pid": 0, "tid": 7, "ts": 1776649790532691, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532683, "dur": 4, + "args": { + "External id": 11224, + "cbid": 211, "correlation": 11224 + } + }, + { + "ph": "s", "id": 11224, "pid": 435, "tid": 435, "ts": 1776649790532683, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790532726, "dur": 1, + "args": { + "External id": 1002, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11242, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11242, "pid": 0, "tid": 7, "ts": 1776649790532726, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532717, "dur": 6, + "args": { + "External id": 1002, + "cbid": 211, "correlation": 11242 + } + }, + { + "ph": "s", "id": 11242, "pid": 435, "tid": 435, "ts": 1776649790532717, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532729, "dur": 0, + "args": { + "External id": 11248, + "cbid": 200, "correlation": 11248 + } + }, + { + "ph": "f", "id": 11248, "pid": 435, "tid": 435, "ts": 1776649790532729, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532730, "dur": 0, + "args": { + "External id": 11249, + "cbid": 273, "correlation": 11249 + } + }, + { + "ph": "f", "id": 11249, "pid": 435, "tid": 435, "ts": 1776649790532730, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790532739, "dur": 15, + "args": { + "External id": 11250, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11250, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11250, "pid": 0, "tid": 7, "ts": 1776649790532739, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532731, "dur": 4, + "args": { + "External id": 11250, + "cbid": 211, "correlation": 11250 + } + }, + { + "ph": "s", "id": 11250, "pid": 435, "tid": 435, "ts": 1776649790532731, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790532775, "dur": 1, + "args": { + "External id": 1007, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11268, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11268, "pid": 0, "tid": 7, "ts": 1776649790532775, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532766, "dur": 6, + "args": { + "External id": 1007, + "cbid": 211, "correlation": 11268 + } + }, + { + "ph": "s", "id": 11268, "pid": 435, "tid": 435, "ts": 1776649790532766, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532778, "dur": 0, + "args": { + "External id": 11274, + "cbid": 200, "correlation": 11274 + } + }, + { + "ph": "f", "id": 11274, "pid": 435, "tid": 435, "ts": 1776649790532778, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532779, "dur": 0, + "args": { + "External id": 11275, + "cbid": 273, "correlation": 11275 + } + }, + { + "ph": "f", "id": 11275, "pid": 435, "tid": 435, "ts": 1776649790532779, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790532787, "dur": 15, + "args": { + "External id": 11276, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11276, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11276, "pid": 0, "tid": 7, "ts": 1776649790532787, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532780, "dur": 4, + "args": { + "External id": 11276, + "cbid": 211, "correlation": 11276 + } + }, + { + "ph": "s", "id": 11276, "pid": 435, "tid": 435, "ts": 1776649790532780, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790532824, "dur": 1, + "args": { + "External id": 1012, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11294, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11294, "pid": 0, "tid": 7, "ts": 1776649790532824, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532814, "dur": 6, + "args": { + "External id": 1012, + "cbid": 211, "correlation": 11294 + } + }, + { + "ph": "s", "id": 11294, "pid": 435, "tid": 435, "ts": 1776649790532814, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532827, "dur": 0, + "args": { + "External id": 11300, + "cbid": 200, "correlation": 11300 + } + }, + { + "ph": "f", "id": 11300, "pid": 435, "tid": 435, "ts": 1776649790532827, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532827, "dur": 0, + "args": { + "External id": 11301, + "cbid": 273, "correlation": 11301 + } + }, + { + "ph": "f", "id": 11301, "pid": 435, "tid": 435, "ts": 1776649790532827, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790532836, "dur": 15, + "args": { + "External id": 11302, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11302, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11302, "pid": 0, "tid": 7, "ts": 1776649790532836, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532828, "dur": 4, + "args": { + "External id": 11302, + "cbid": 211, "correlation": 11302 + } + }, + { + "ph": "s", "id": 11302, "pid": 435, "tid": 435, "ts": 1776649790532828, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790532872, "dur": 1, + "args": { + "External id": 1017, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11320, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11320, "pid": 0, "tid": 7, "ts": 1776649790532872, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532863, "dur": 6, + "args": { + "External id": 1017, + "cbid": 211, "correlation": 11320 + } + }, + { + "ph": "s", "id": 11320, "pid": 435, "tid": 435, "ts": 1776649790532863, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532875, "dur": 0, + "args": { + "External id": 11326, + "cbid": 200, "correlation": 11326 + } + }, + { + "ph": "f", "id": 11326, "pid": 435, "tid": 435, "ts": 1776649790532875, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532876, "dur": 0, + "args": { + "External id": 11327, + "cbid": 273, "correlation": 11327 + } + }, + { + "ph": "f", "id": 11327, "pid": 435, "tid": 435, "ts": 1776649790532876, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790532885, "dur": 15, + "args": { + "External id": 11328, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11328, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11328, "pid": 0, "tid": 7, "ts": 1776649790532885, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532877, "dur": 4, + "args": { + "External id": 11328, + "cbid": 211, "correlation": 11328 + } + }, + { + "ph": "s", "id": 11328, "pid": 435, "tid": 435, "ts": 1776649790532877, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790532921, "dur": 1, + "args": { + "External id": 1022, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11346, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11346, "pid": 0, "tid": 7, "ts": 1776649790532921, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532912, "dur": 6, + "args": { + "External id": 1022, + "cbid": 211, "correlation": 11346 + } + }, + { + "ph": "s", "id": 11346, "pid": 435, "tid": 435, "ts": 1776649790532912, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532925, "dur": 0, + "args": { + "External id": 11352, + "cbid": 200, "correlation": 11352 + } + }, + { + "ph": "f", "id": 11352, "pid": 435, "tid": 435, "ts": 1776649790532925, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532925, "dur": 0, + "args": { + "External id": 11353, + "cbid": 273, "correlation": 11353 + } + }, + { + "ph": "f", "id": 11353, "pid": 435, "tid": 435, "ts": 1776649790532925, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790532934, "dur": 15, + "args": { + "External id": 11354, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11354, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11354, "pid": 0, "tid": 7, "ts": 1776649790532934, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532926, "dur": 5, + "args": { + "External id": 11354, + "cbid": 211, "correlation": 11354 + } + }, + { + "ph": "s", "id": 11354, "pid": 435, "tid": 435, "ts": 1776649790532926, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790532980, "dur": 1, + "args": { + "External id": 1027, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11372, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11372, "pid": 0, "tid": 7, "ts": 1776649790532980, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532970, "dur": 7, + "args": { + "External id": 1027, + "cbid": 211, "correlation": 11372 + } + }, + { + "ph": "s", "id": 11372, "pid": 435, "tid": 435, "ts": 1776649790532970, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532983, "dur": 0, + "args": { + "External id": 11378, + "cbid": 200, "correlation": 11378 + } + }, + { + "ph": "f", "id": 11378, "pid": 435, "tid": 435, "ts": 1776649790532983, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790532984, "dur": 0, + "args": { + "External id": 11379, + "cbid": 273, "correlation": 11379 + } + }, + { + "ph": "f", "id": 11379, "pid": 435, "tid": 435, "ts": 1776649790532984, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790532993, "dur": 15, + "args": { + "External id": 11380, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11380, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11380, "pid": 0, "tid": 7, "ts": 1776649790532993, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790532985, "dur": 5, + "args": { + "External id": 11380, + "cbid": 211, "correlation": 11380 + } + }, + { + "ph": "s", "id": 11380, "pid": 435, "tid": 435, "ts": 1776649790532985, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790533029, "dur": 1, + "args": { + "External id": 1032, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11398, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11398, "pid": 0, "tid": 7, "ts": 1776649790533029, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533020, "dur": 6, + "args": { + "External id": 1032, + "cbid": 211, "correlation": 11398 + } + }, + { + "ph": "s", "id": 11398, "pid": 435, "tid": 435, "ts": 1776649790533020, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533032, "dur": 0, + "args": { + "External id": 11404, + "cbid": 200, "correlation": 11404 + } + }, + { + "ph": "f", "id": 11404, "pid": 435, "tid": 435, "ts": 1776649790533032, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533033, "dur": 0, + "args": { + "External id": 11405, + "cbid": 273, "correlation": 11405 + } + }, + { + "ph": "f", "id": 11405, "pid": 435, "tid": 435, "ts": 1776649790533033, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790533042, "dur": 15, + "args": { + "External id": 11406, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11406, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11406, "pid": 0, "tid": 7, "ts": 1776649790533042, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533034, "dur": 4, + "args": { + "External id": 11406, + "cbid": 211, "correlation": 11406 + } + }, + { + "ph": "s", "id": 11406, "pid": 435, "tid": 435, "ts": 1776649790533034, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790533077, "dur": 1, + "args": { + "External id": 1037, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11424, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11424, "pid": 0, "tid": 7, "ts": 1776649790533077, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533068, "dur": 6, + "args": { + "External id": 1037, + "cbid": 211, "correlation": 11424 + } + }, + { + "ph": "s", "id": 11424, "pid": 435, "tid": 435, "ts": 1776649790533068, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533080, "dur": 0, + "args": { + "External id": 11430, + "cbid": 200, "correlation": 11430 + } + }, + { + "ph": "f", "id": 11430, "pid": 435, "tid": 435, "ts": 1776649790533080, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533081, "dur": 0, + "args": { + "External id": 11431, + "cbid": 273, "correlation": 11431 + } + }, + { + "ph": "f", "id": 11431, "pid": 435, "tid": 435, "ts": 1776649790533081, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790533090, "dur": 16, + "args": { + "External id": 11432, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11432, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11432, "pid": 0, "tid": 7, "ts": 1776649790533090, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533082, "dur": 5, + "args": { + "External id": 11432, + "cbid": 211, "correlation": 11432 + } + }, + { + "ph": "s", "id": 11432, "pid": 435, "tid": 435, "ts": 1776649790533082, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790533126, "dur": 1, + "args": { + "External id": 1042, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11450, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11450, "pid": 0, "tid": 7, "ts": 1776649790533126, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533116, "dur": 6, + "args": { + "External id": 1042, + "cbid": 211, "correlation": 11450 + } + }, + { + "ph": "s", "id": 11450, "pid": 435, "tid": 435, "ts": 1776649790533116, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533129, "dur": 0, + "args": { + "External id": 11456, + "cbid": 200, "correlation": 11456 + } + }, + { + "ph": "f", "id": 11456, "pid": 435, "tid": 435, "ts": 1776649790533129, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533129, "dur": 0, + "args": { + "External id": 11457, + "cbid": 273, "correlation": 11457 + } + }, + { + "ph": "f", "id": 11457, "pid": 435, "tid": 435, "ts": 1776649790533129, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790533138, "dur": 17, + "args": { + "External id": 11458, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11458, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11458, "pid": 0, "tid": 7, "ts": 1776649790533138, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533131, "dur": 4, + "args": { + "External id": 11458, + "cbid": 211, "correlation": 11458 + } + }, + { + "ph": "s", "id": 11458, "pid": 435, "tid": 435, "ts": 1776649790533131, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790533174, "dur": 1, + "args": { + "External id": 1047, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11476, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11476, "pid": 0, "tid": 7, "ts": 1776649790533174, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533164, "dur": 6, + "args": { + "External id": 1047, + "cbid": 211, "correlation": 11476 + } + }, + { + "ph": "s", "id": 11476, "pid": 435, "tid": 435, "ts": 1776649790533164, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533177, "dur": 0, + "args": { + "External id": 11482, + "cbid": 200, "correlation": 11482 + } + }, + { + "ph": "f", "id": 11482, "pid": 435, "tid": 435, "ts": 1776649790533177, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533177, "dur": 0, + "args": { + "External id": 11483, + "cbid": 273, "correlation": 11483 + } + }, + { + "ph": "f", "id": 11483, "pid": 435, "tid": 435, "ts": 1776649790533177, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790533186, "dur": 15, + "args": { + "External id": 11484, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11484, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11484, "pid": 0, "tid": 7, "ts": 1776649790533186, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533178, "dur": 4, + "args": { + "External id": 11484, + "cbid": 211, "correlation": 11484 + } + }, + { + "ph": "s", "id": 11484, "pid": 435, "tid": 435, "ts": 1776649790533178, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790533222, "dur": 1, + "args": { + "External id": 1052, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11502, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11502, "pid": 0, "tid": 7, "ts": 1776649790533222, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533212, "dur": 6, + "args": { + "External id": 1052, + "cbid": 211, "correlation": 11502 + } + }, + { + "ph": "s", "id": 11502, "pid": 435, "tid": 435, "ts": 1776649790533212, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533225, "dur": 0, + "args": { + "External id": 11508, + "cbid": 200, "correlation": 11508 + } + }, + { + "ph": "f", "id": 11508, "pid": 435, "tid": 435, "ts": 1776649790533225, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533225, "dur": 0, + "args": { + "External id": 11509, + "cbid": 273, "correlation": 11509 + } + }, + { + "ph": "f", "id": 11509, "pid": 435, "tid": 435, "ts": 1776649790533225, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790533234, "dur": 15, + "args": { + "External id": 11510, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11510, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11510, "pid": 0, "tid": 7, "ts": 1776649790533234, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533226, "dur": 4, + "args": { + "External id": 11510, + "cbid": 211, "correlation": 11510 + } + }, + { + "ph": "s", "id": 11510, "pid": 435, "tid": 435, "ts": 1776649790533226, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790533270, "dur": 1, + "args": { + "External id": 1057, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11528, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11528, "pid": 0, "tid": 7, "ts": 1776649790533270, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533260, "dur": 6, + "args": { + "External id": 1057, + "cbid": 211, "correlation": 11528 + } + }, + { + "ph": "s", "id": 11528, "pid": 435, "tid": 435, "ts": 1776649790533260, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533273, "dur": 0, + "args": { + "External id": 11534, + "cbid": 200, "correlation": 11534 + } + }, + { + "ph": "f", "id": 11534, "pid": 435, "tid": 435, "ts": 1776649790533273, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533273, "dur": 0, + "args": { + "External id": 11535, + "cbid": 273, "correlation": 11535 + } + }, + { + "ph": "f", "id": 11535, "pid": 435, "tid": 435, "ts": 1776649790533273, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790533282, "dur": 15, + "args": { + "External id": 11536, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11536, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11536, "pid": 0, "tid": 7, "ts": 1776649790533282, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533274, "dur": 4, + "args": { + "External id": 11536, + "cbid": 211, "correlation": 11536 + } + }, + { + "ph": "s", "id": 11536, "pid": 435, "tid": 435, "ts": 1776649790533274, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790533318, "dur": 1, + "args": { + "External id": 1062, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11554, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11554, "pid": 0, "tid": 7, "ts": 1776649790533318, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533309, "dur": 6, + "args": { + "External id": 1062, + "cbid": 211, "correlation": 11554 + } + }, + { + "ph": "s", "id": 11554, "pid": 435, "tid": 435, "ts": 1776649790533309, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533321, "dur": 0, + "args": { + "External id": 11560, + "cbid": 200, "correlation": 11560 + } + }, + { + "ph": "f", "id": 11560, "pid": 435, "tid": 435, "ts": 1776649790533321, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533322, "dur": 0, + "args": { + "External id": 11561, + "cbid": 273, "correlation": 11561 + } + }, + { + "ph": "f", "id": 11561, "pid": 435, "tid": 435, "ts": 1776649790533322, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790533331, "dur": 15, + "args": { + "External id": 11562, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11562, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11562, "pid": 0, "tid": 7, "ts": 1776649790533331, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533323, "dur": 4, + "args": { + "External id": 11562, + "cbid": 211, "correlation": 11562 + } + }, + { + "ph": "s", "id": 11562, "pid": 435, "tid": 435, "ts": 1776649790533323, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790533367, "dur": 1, + "args": { + "External id": 1067, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11580, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11580, "pid": 0, "tid": 7, "ts": 1776649790533367, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533358, "dur": 6, + "args": { + "External id": 1067, + "cbid": 211, "correlation": 11580 + } + }, + { + "ph": "s", "id": 11580, "pid": 435, "tid": 435, "ts": 1776649790533358, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533370, "dur": 0, + "args": { + "External id": 11586, + "cbid": 200, "correlation": 11586 + } + }, + { + "ph": "f", "id": 11586, "pid": 435, "tid": 435, "ts": 1776649790533370, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533371, "dur": 0, + "args": { + "External id": 11587, + "cbid": 273, "correlation": 11587 + } + }, + { + "ph": "f", "id": 11587, "pid": 435, "tid": 435, "ts": 1776649790533371, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790533380, "dur": 15, + "args": { + "External id": 11588, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11588, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11588, "pid": 0, "tid": 7, "ts": 1776649790533380, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533372, "dur": 4, + "args": { + "External id": 11588, + "cbid": 211, "correlation": 11588 + } + }, + { + "ph": "s", "id": 11588, "pid": 435, "tid": 435, "ts": 1776649790533372, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790533415, "dur": 1, + "args": { + "External id": 1072, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11606, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11606, "pid": 0, "tid": 7, "ts": 1776649790533415, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533406, "dur": 6, + "args": { + "External id": 1072, + "cbid": 211, "correlation": 11606 + } + }, + { + "ph": "s", "id": 11606, "pid": 435, "tid": 435, "ts": 1776649790533406, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533418, "dur": 0, + "args": { + "External id": 11612, + "cbid": 200, "correlation": 11612 + } + }, + { + "ph": "f", "id": 11612, "pid": 435, "tid": 435, "ts": 1776649790533418, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533419, "dur": 0, + "args": { + "External id": 11613, + "cbid": 273, "correlation": 11613 + } + }, + { + "ph": "f", "id": 11613, "pid": 435, "tid": 435, "ts": 1776649790533419, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790533428, "dur": 15, + "args": { + "External id": 11614, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11614, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11614, "pid": 0, "tid": 7, "ts": 1776649790533428, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533420, "dur": 4, + "args": { + "External id": 11614, + "cbid": 211, "correlation": 11614 + } + }, + { + "ph": "s", "id": 11614, "pid": 435, "tid": 435, "ts": 1776649790533420, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790533463, "dur": 1, + "args": { + "External id": 1077, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11632, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11632, "pid": 0, "tid": 7, "ts": 1776649790533463, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533454, "dur": 6, + "args": { + "External id": 1077, + "cbid": 211, "correlation": 11632 + } + }, + { + "ph": "s", "id": 11632, "pid": 435, "tid": 435, "ts": 1776649790533454, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533467, "dur": 0, + "args": { + "External id": 11638, + "cbid": 200, "correlation": 11638 + } + }, + { + "ph": "f", "id": 11638, "pid": 435, "tid": 435, "ts": 1776649790533467, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533467, "dur": 0, + "args": { + "External id": 11639, + "cbid": 273, "correlation": 11639 + } + }, + { + "ph": "f", "id": 11639, "pid": 435, "tid": 435, "ts": 1776649790533467, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790533476, "dur": 15, + "args": { + "External id": 11640, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11640, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11640, "pid": 0, "tid": 7, "ts": 1776649790533476, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533468, "dur": 4, + "args": { + "External id": 11640, + "cbid": 211, "correlation": 11640 + } + }, + { + "ph": "s", "id": 11640, "pid": 435, "tid": 435, "ts": 1776649790533468, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790533512, "dur": 1, + "args": { + "External id": 1082, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11658, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11658, "pid": 0, "tid": 7, "ts": 1776649790533512, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533502, "dur": 6, + "args": { + "External id": 1082, + "cbid": 211, "correlation": 11658 + } + }, + { + "ph": "s", "id": 11658, "pid": 435, "tid": 435, "ts": 1776649790533502, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533515, "dur": 0, + "args": { + "External id": 11664, + "cbid": 200, "correlation": 11664 + } + }, + { + "ph": "f", "id": 11664, "pid": 435, "tid": 435, "ts": 1776649790533515, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533515, "dur": 0, + "args": { + "External id": 11665, + "cbid": 273, "correlation": 11665 + } + }, + { + "ph": "f", "id": 11665, "pid": 435, "tid": 435, "ts": 1776649790533515, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790533524, "dur": 17, + "args": { + "External id": 11666, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11666, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11666, "pid": 0, "tid": 7, "ts": 1776649790533524, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533516, "dur": 4, + "args": { + "External id": 11666, + "cbid": 211, "correlation": 11666 + } + }, + { + "ph": "s", "id": 11666, "pid": 435, "tid": 435, "ts": 1776649790533516, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790533559, "dur": 1, + "args": { + "External id": 1087, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11684, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11684, "pid": 0, "tid": 7, "ts": 1776649790533559, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533550, "dur": 7, + "args": { + "External id": 1087, + "cbid": 211, "correlation": 11684 + } + }, + { + "ph": "s", "id": 11684, "pid": 435, "tid": 435, "ts": 1776649790533550, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533563, "dur": 0, + "args": { + "External id": 11690, + "cbid": 200, "correlation": 11690 + } + }, + { + "ph": "f", "id": 11690, "pid": 435, "tid": 435, "ts": 1776649790533563, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533563, "dur": 0, + "args": { + "External id": 11691, + "cbid": 273, "correlation": 11691 + } + }, + { + "ph": "f", "id": 11691, "pid": 435, "tid": 435, "ts": 1776649790533563, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790533572, "dur": 16, + "args": { + "External id": 11692, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11692, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11692, "pid": 0, "tid": 7, "ts": 1776649790533572, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533564, "dur": 4, + "args": { + "External id": 11692, + "cbid": 211, "correlation": 11692 + } + }, + { + "ph": "s", "id": 11692, "pid": 435, "tid": 435, "ts": 1776649790533564, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790533607, "dur": 1, + "args": { + "External id": 1092, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11710, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11710, "pid": 0, "tid": 7, "ts": 1776649790533607, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533598, "dur": 6, + "args": { + "External id": 1092, + "cbid": 211, "correlation": 11710 + } + }, + { + "ph": "s", "id": 11710, "pid": 435, "tid": 435, "ts": 1776649790533598, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533611, "dur": 0, + "args": { + "External id": 11716, + "cbid": 200, "correlation": 11716 + } + }, + { + "ph": "f", "id": 11716, "pid": 435, "tid": 435, "ts": 1776649790533611, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533611, "dur": 0, + "args": { + "External id": 11717, + "cbid": 273, "correlation": 11717 + } + }, + { + "ph": "f", "id": 11717, "pid": 435, "tid": 435, "ts": 1776649790533611, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790533620, "dur": 15, + "args": { + "External id": 11718, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11718, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11718, "pid": 0, "tid": 7, "ts": 1776649790533620, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533612, "dur": 4, + "args": { + "External id": 11718, + "cbid": 211, "correlation": 11718 + } + }, + { + "ph": "s", "id": 11718, "pid": 435, "tid": 435, "ts": 1776649790533612, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790533656, "dur": 1, + "args": { + "External id": 1097, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11736, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11736, "pid": 0, "tid": 7, "ts": 1776649790533656, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533647, "dur": 6, + "args": { + "External id": 1097, + "cbid": 211, "correlation": 11736 + } + }, + { + "ph": "s", "id": 11736, "pid": 435, "tid": 435, "ts": 1776649790533647, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533659, "dur": 0, + "args": { + "External id": 11742, + "cbid": 200, "correlation": 11742 + } + }, + { + "ph": "f", "id": 11742, "pid": 435, "tid": 435, "ts": 1776649790533659, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533660, "dur": 0, + "args": { + "External id": 11743, + "cbid": 273, "correlation": 11743 + } + }, + { + "ph": "f", "id": 11743, "pid": 435, "tid": 435, "ts": 1776649790533660, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790533669, "dur": 15, + "args": { + "External id": 11744, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11744, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11744, "pid": 0, "tid": 7, "ts": 1776649790533669, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533661, "dur": 5, + "args": { + "External id": 11744, + "cbid": 211, "correlation": 11744 + } + }, + { + "ph": "s", "id": 11744, "pid": 435, "tid": 435, "ts": 1776649790533661, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790533704, "dur": 1, + "args": { + "External id": 1102, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11762, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11762, "pid": 0, "tid": 7, "ts": 1776649790533704, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533695, "dur": 6, + "args": { + "External id": 1102, + "cbid": 211, "correlation": 11762 + } + }, + { + "ph": "s", "id": 11762, "pid": 435, "tid": 435, "ts": 1776649790533695, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533707, "dur": 0, + "args": { + "External id": 11768, + "cbid": 200, "correlation": 11768 + } + }, + { + "ph": "f", "id": 11768, "pid": 435, "tid": 435, "ts": 1776649790533707, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533708, "dur": 0, + "args": { + "External id": 11769, + "cbid": 273, "correlation": 11769 + } + }, + { + "ph": "f", "id": 11769, "pid": 435, "tid": 435, "ts": 1776649790533708, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790533717, "dur": 15, + "args": { + "External id": 11770, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11770, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11770, "pid": 0, "tid": 7, "ts": 1776649790533717, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533709, "dur": 5, + "args": { + "External id": 11770, + "cbid": 211, "correlation": 11770 + } + }, + { + "ph": "s", "id": 11770, "pid": 435, "tid": 435, "ts": 1776649790533709, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790533752, "dur": 1, + "args": { + "External id": 1107, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11788, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11788, "pid": 0, "tid": 7, "ts": 1776649790533752, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533743, "dur": 6, + "args": { + "External id": 1107, + "cbid": 211, "correlation": 11788 + } + }, + { + "ph": "s", "id": 11788, "pid": 435, "tid": 435, "ts": 1776649790533743, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533756, "dur": 0, + "args": { + "External id": 11794, + "cbid": 200, "correlation": 11794 + } + }, + { + "ph": "f", "id": 11794, "pid": 435, "tid": 435, "ts": 1776649790533756, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533756, "dur": 0, + "args": { + "External id": 11795, + "cbid": 273, "correlation": 11795 + } + }, + { + "ph": "f", "id": 11795, "pid": 435, "tid": 435, "ts": 1776649790533756, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790533765, "dur": 15, + "args": { + "External id": 11796, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11796, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11796, "pid": 0, "tid": 7, "ts": 1776649790533765, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533757, "dur": 4, + "args": { + "External id": 11796, + "cbid": 211, "correlation": 11796 + } + }, + { + "ph": "s", "id": 11796, "pid": 435, "tid": 435, "ts": 1776649790533757, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790533800, "dur": 1, + "args": { + "External id": 1112, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11814, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11814, "pid": 0, "tid": 7, "ts": 1776649790533800, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533791, "dur": 6, + "args": { + "External id": 1112, + "cbid": 211, "correlation": 11814 + } + }, + { + "ph": "s", "id": 11814, "pid": 435, "tid": 435, "ts": 1776649790533791, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533803, "dur": 0, + "args": { + "External id": 11820, + "cbid": 200, "correlation": 11820 + } + }, + { + "ph": "f", "id": 11820, "pid": 435, "tid": 435, "ts": 1776649790533803, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533804, "dur": 0, + "args": { + "External id": 11821, + "cbid": 273, "correlation": 11821 + } + }, + { + "ph": "f", "id": 11821, "pid": 435, "tid": 435, "ts": 1776649790533804, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790533813, "dur": 15, + "args": { + "External id": 11822, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11822, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11822, "pid": 0, "tid": 7, "ts": 1776649790533813, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533805, "dur": 4, + "args": { + "External id": 11822, + "cbid": 211, "correlation": 11822 + } + }, + { + "ph": "s", "id": 11822, "pid": 435, "tid": 435, "ts": 1776649790533805, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790533848, "dur": 1, + "args": { + "External id": 1117, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11840, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11840, "pid": 0, "tid": 7, "ts": 1776649790533848, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533839, "dur": 6, + "args": { + "External id": 1117, + "cbid": 211, "correlation": 11840 + } + }, + { + "ph": "s", "id": 11840, "pid": 435, "tid": 435, "ts": 1776649790533839, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533851, "dur": 0, + "args": { + "External id": 11846, + "cbid": 200, "correlation": 11846 + } + }, + { + "ph": "f", "id": 11846, "pid": 435, "tid": 435, "ts": 1776649790533851, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533852, "dur": 0, + "args": { + "External id": 11847, + "cbid": 273, "correlation": 11847 + } + }, + { + "ph": "f", "id": 11847, "pid": 435, "tid": 435, "ts": 1776649790533852, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790533861, "dur": 15, + "args": { + "External id": 11848, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11848, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11848, "pid": 0, "tid": 7, "ts": 1776649790533861, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533853, "dur": 4, + "args": { + "External id": 11848, + "cbid": 211, "correlation": 11848 + } + }, + { + "ph": "s", "id": 11848, "pid": 435, "tid": 435, "ts": 1776649790533853, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790533899, "dur": 1, + "args": { + "External id": 1122, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11866, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11866, "pid": 0, "tid": 7, "ts": 1776649790533899, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533890, "dur": 6, + "args": { + "External id": 1122, + "cbid": 211, "correlation": 11866 + } + }, + { + "ph": "s", "id": 11866, "pid": 435, "tid": 435, "ts": 1776649790533890, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533902, "dur": 0, + "args": { + "External id": 11872, + "cbid": 200, "correlation": 11872 + } + }, + { + "ph": "f", "id": 11872, "pid": 435, "tid": 435, "ts": 1776649790533902, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533903, "dur": 0, + "args": { + "External id": 11873, + "cbid": 273, "correlation": 11873 + } + }, + { + "ph": "f", "id": 11873, "pid": 435, "tid": 435, "ts": 1776649790533903, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790533912, "dur": 15, + "args": { + "External id": 11874, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11874, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11874, "pid": 0, "tid": 7, "ts": 1776649790533912, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533904, "dur": 5, + "args": { + "External id": 11874, + "cbid": 211, "correlation": 11874 + } + }, + { + "ph": "s", "id": 11874, "pid": 435, "tid": 435, "ts": 1776649790533904, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790533947, "dur": 1, + "args": { + "External id": 1127, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11892, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11892, "pid": 0, "tid": 7, "ts": 1776649790533947, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533938, "dur": 6, + "args": { + "External id": 1127, + "cbid": 211, "correlation": 11892 + } + }, + { + "ph": "s", "id": 11892, "pid": 435, "tid": 435, "ts": 1776649790533938, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533950, "dur": 0, + "args": { + "External id": 11898, + "cbid": 200, "correlation": 11898 + } + }, + { + "ph": "f", "id": 11898, "pid": 435, "tid": 435, "ts": 1776649790533950, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533951, "dur": 0, + "args": { + "External id": 11899, + "cbid": 273, "correlation": 11899 + } + }, + { + "ph": "f", "id": 11899, "pid": 435, "tid": 435, "ts": 1776649790533951, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790533960, "dur": 16, + "args": { + "External id": 11900, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11900, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11900, "pid": 0, "tid": 7, "ts": 1776649790533960, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533952, "dur": 5, + "args": { + "External id": 11900, + "cbid": 211, "correlation": 11900 + } + }, + { + "ph": "s", "id": 11900, "pid": 435, "tid": 435, "ts": 1776649790533952, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790533995, "dur": 1, + "args": { + "External id": 1132, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11918, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11918, "pid": 0, "tid": 7, "ts": 1776649790533995, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790533986, "dur": 6, + "args": { + "External id": 1132, + "cbid": 211, "correlation": 11918 + } + }, + { + "ph": "s", "id": 11918, "pid": 435, "tid": 435, "ts": 1776649790533986, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533998, "dur": 0, + "args": { + "External id": 11924, + "cbid": 200, "correlation": 11924 + } + }, + { + "ph": "f", "id": 11924, "pid": 435, "tid": 435, "ts": 1776649790533998, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790533999, "dur": 0, + "args": { + "External id": 11925, + "cbid": 273, "correlation": 11925 + } + }, + { + "ph": "f", "id": 11925, "pid": 435, "tid": 435, "ts": 1776649790533999, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790534008, "dur": 15, + "args": { + "External id": 11926, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11926, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11926, "pid": 0, "tid": 7, "ts": 1776649790534008, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534000, "dur": 4, + "args": { + "External id": 11926, + "cbid": 211, "correlation": 11926 + } + }, + { + "ph": "s", "id": 11926, "pid": 435, "tid": 435, "ts": 1776649790534000, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790534043, "dur": 1, + "args": { + "External id": 1137, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11944, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11944, "pid": 0, "tid": 7, "ts": 1776649790534043, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534034, "dur": 6, + "args": { + "External id": 1137, + "cbid": 211, "correlation": 11944 + } + }, + { + "ph": "s", "id": 11944, "pid": 435, "tid": 435, "ts": 1776649790534034, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534046, "dur": 0, + "args": { + "External id": 11950, + "cbid": 200, "correlation": 11950 + } + }, + { + "ph": "f", "id": 11950, "pid": 435, "tid": 435, "ts": 1776649790534046, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534047, "dur": 0, + "args": { + "External id": 11951, + "cbid": 273, "correlation": 11951 + } + }, + { + "ph": "f", "id": 11951, "pid": 435, "tid": 435, "ts": 1776649790534047, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790534056, "dur": 15, + "args": { + "External id": 11952, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11952, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11952, "pid": 0, "tid": 7, "ts": 1776649790534056, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534048, "dur": 4, + "args": { + "External id": 11952, + "cbid": 211, "correlation": 11952 + } + }, + { + "ph": "s", "id": 11952, "pid": 435, "tid": 435, "ts": 1776649790534048, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790534091, "dur": 1, + "args": { + "External id": 1142, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11970, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11970, "pid": 0, "tid": 7, "ts": 1776649790534091, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534082, "dur": 6, + "args": { + "External id": 1142, + "cbid": 211, "correlation": 11970 + } + }, + { + "ph": "s", "id": 11970, "pid": 435, "tid": 435, "ts": 1776649790534082, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534094, "dur": 0, + "args": { + "External id": 11976, + "cbid": 200, "correlation": 11976 + } + }, + { + "ph": "f", "id": 11976, "pid": 435, "tid": 435, "ts": 1776649790534094, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534095, "dur": 0, + "args": { + "External id": 11977, + "cbid": 273, "correlation": 11977 + } + }, + { + "ph": "f", "id": 11977, "pid": 435, "tid": 435, "ts": 1776649790534095, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790534104, "dur": 15, + "args": { + "External id": 11978, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11978, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11978, "pid": 0, "tid": 7, "ts": 1776649790534104, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534096, "dur": 4, + "args": { + "External id": 11978, + "cbid": 211, "correlation": 11978 + } + }, + { + "ph": "s", "id": 11978, "pid": 435, "tid": 435, "ts": 1776649790534096, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790534138, "dur": 1, + "args": { + "External id": 1147, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 11996, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 11996, "pid": 0, "tid": 7, "ts": 1776649790534138, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534129, "dur": 6, + "args": { + "External id": 1147, + "cbid": 211, "correlation": 11996 + } + }, + { + "ph": "s", "id": 11996, "pid": 435, "tid": 435, "ts": 1776649790534129, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534142, "dur": 0, + "args": { + "External id": 12002, + "cbid": 200, "correlation": 12002 + } + }, + { + "ph": "f", "id": 12002, "pid": 435, "tid": 435, "ts": 1776649790534142, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534142, "dur": 0, + "args": { + "External id": 12003, + "cbid": 273, "correlation": 12003 + } + }, + { + "ph": "f", "id": 12003, "pid": 435, "tid": 435, "ts": 1776649790534142, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790534151, "dur": 15, + "args": { + "External id": 12004, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12004, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12004, "pid": 0, "tid": 7, "ts": 1776649790534151, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534143, "dur": 4, + "args": { + "External id": 12004, + "cbid": 211, "correlation": 12004 + } + }, + { + "ph": "s", "id": 12004, "pid": 435, "tid": 435, "ts": 1776649790534143, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790534186, "dur": 1, + "args": { + "External id": 1152, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12022, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12022, "pid": 0, "tid": 7, "ts": 1776649790534186, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534177, "dur": 6, + "args": { + "External id": 1152, + "cbid": 211, "correlation": 12022 + } + }, + { + "ph": "s", "id": 12022, "pid": 435, "tid": 435, "ts": 1776649790534177, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534189, "dur": 0, + "args": { + "External id": 12028, + "cbid": 200, "correlation": 12028 + } + }, + { + "ph": "f", "id": 12028, "pid": 435, "tid": 435, "ts": 1776649790534189, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534190, "dur": 0, + "args": { + "External id": 12029, + "cbid": 273, "correlation": 12029 + } + }, + { + "ph": "f", "id": 12029, "pid": 435, "tid": 435, "ts": 1776649790534190, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790534199, "dur": 15, + "args": { + "External id": 12030, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12030, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12030, "pid": 0, "tid": 7, "ts": 1776649790534199, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534191, "dur": 4, + "args": { + "External id": 12030, + "cbid": 211, "correlation": 12030 + } + }, + { + "ph": "s", "id": 12030, "pid": 435, "tid": 435, "ts": 1776649790534191, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790534234, "dur": 1, + "args": { + "External id": 1157, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12048, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12048, "pid": 0, "tid": 7, "ts": 1776649790534234, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534225, "dur": 6, + "args": { + "External id": 1157, + "cbid": 211, "correlation": 12048 + } + }, + { + "ph": "s", "id": 12048, "pid": 435, "tid": 435, "ts": 1776649790534225, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534237, "dur": 0, + "args": { + "External id": 12054, + "cbid": 200, "correlation": 12054 + } + }, + { + "ph": "f", "id": 12054, "pid": 435, "tid": 435, "ts": 1776649790534237, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534238, "dur": 0, + "args": { + "External id": 12055, + "cbid": 273, "correlation": 12055 + } + }, + { + "ph": "f", "id": 12055, "pid": 435, "tid": 435, "ts": 1776649790534238, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790534247, "dur": 15, + "args": { + "External id": 12056, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12056, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12056, "pid": 0, "tid": 7, "ts": 1776649790534247, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534239, "dur": 5, + "args": { + "External id": 12056, + "cbid": 211, "correlation": 12056 + } + }, + { + "ph": "s", "id": 12056, "pid": 435, "tid": 435, "ts": 1776649790534239, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790534282, "dur": 1, + "args": { + "External id": 1162, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12074, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12074, "pid": 0, "tid": 7, "ts": 1776649790534282, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534273, "dur": 6, + "args": { + "External id": 1162, + "cbid": 211, "correlation": 12074 + } + }, + { + "ph": "s", "id": 12074, "pid": 435, "tid": 435, "ts": 1776649790534273, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534285, "dur": 0, + "args": { + "External id": 12080, + "cbid": 200, "correlation": 12080 + } + }, + { + "ph": "f", "id": 12080, "pid": 435, "tid": 435, "ts": 1776649790534285, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534285, "dur": 0, + "args": { + "External id": 12081, + "cbid": 273, "correlation": 12081 + } + }, + { + "ph": "f", "id": 12081, "pid": 435, "tid": 435, "ts": 1776649790534285, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790534295, "dur": 15, + "args": { + "External id": 12082, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12082, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12082, "pid": 0, "tid": 7, "ts": 1776649790534295, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534287, "dur": 5, + "args": { + "External id": 12082, + "cbid": 211, "correlation": 12082 + } + }, + { + "ph": "s", "id": 12082, "pid": 435, "tid": 435, "ts": 1776649790534287, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790534330, "dur": 1, + "args": { + "External id": 1167, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12100, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12100, "pid": 0, "tid": 7, "ts": 1776649790534330, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534320, "dur": 6, + "args": { + "External id": 1167, + "cbid": 211, "correlation": 12100 + } + }, + { + "ph": "s", "id": 12100, "pid": 435, "tid": 435, "ts": 1776649790534320, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534333, "dur": 0, + "args": { + "External id": 12106, + "cbid": 200, "correlation": 12106 + } + }, + { + "ph": "f", "id": 12106, "pid": 435, "tid": 435, "ts": 1776649790534333, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534333, "dur": 0, + "args": { + "External id": 12107, + "cbid": 273, "correlation": 12107 + } + }, + { + "ph": "f", "id": 12107, "pid": 435, "tid": 435, "ts": 1776649790534333, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790534342, "dur": 15, + "args": { + "External id": 12108, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12108, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12108, "pid": 0, "tid": 7, "ts": 1776649790534342, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534335, "dur": 5, + "args": { + "External id": 12108, + "cbid": 211, "correlation": 12108 + } + }, + { + "ph": "s", "id": 12108, "pid": 435, "tid": 435, "ts": 1776649790534335, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790534378, "dur": 1, + "args": { + "External id": 1172, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12126, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12126, "pid": 0, "tid": 7, "ts": 1776649790534378, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534368, "dur": 7, + "args": { + "External id": 1172, + "cbid": 211, "correlation": 12126 + } + }, + { + "ph": "s", "id": 12126, "pid": 435, "tid": 435, "ts": 1776649790534368, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534381, "dur": 0, + "args": { + "External id": 12132, + "cbid": 200, "correlation": 12132 + } + }, + { + "ph": "f", "id": 12132, "pid": 435, "tid": 435, "ts": 1776649790534381, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534381, "dur": 0, + "args": { + "External id": 12133, + "cbid": 273, "correlation": 12133 + } + }, + { + "ph": "f", "id": 12133, "pid": 435, "tid": 435, "ts": 1776649790534381, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790534390, "dur": 16, + "args": { + "External id": 12134, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12134, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12134, "pid": 0, "tid": 7, "ts": 1776649790534390, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534382, "dur": 4, + "args": { + "External id": 12134, + "cbid": 211, "correlation": 12134 + } + }, + { + "ph": "s", "id": 12134, "pid": 435, "tid": 435, "ts": 1776649790534382, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790534426, "dur": 1, + "args": { + "External id": 1177, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12152, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12152, "pid": 0, "tid": 7, "ts": 1776649790534426, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534416, "dur": 6, + "args": { + "External id": 1177, + "cbid": 211, "correlation": 12152 + } + }, + { + "ph": "s", "id": 12152, "pid": 435, "tid": 435, "ts": 1776649790534416, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534429, "dur": 0, + "args": { + "External id": 12158, + "cbid": 200, "correlation": 12158 + } + }, + { + "ph": "f", "id": 12158, "pid": 435, "tid": 435, "ts": 1776649790534429, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534429, "dur": 0, + "args": { + "External id": 12159, + "cbid": 273, "correlation": 12159 + } + }, + { + "ph": "f", "id": 12159, "pid": 435, "tid": 435, "ts": 1776649790534429, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790534438, "dur": 15, + "args": { + "External id": 12160, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12160, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12160, "pid": 0, "tid": 7, "ts": 1776649790534438, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534430, "dur": 4, + "args": { + "External id": 12160, + "cbid": 211, "correlation": 12160 + } + }, + { + "ph": "s", "id": 12160, "pid": 435, "tid": 435, "ts": 1776649790534430, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790534464, "dur": 4, + "args": { + "External id": 12168, + "cbid": 135, "correlation": 12168 + } + }, + { + "ph": "f", "id": 12168, "pid": 435, "tid": 435, "ts": 1776649790534464, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790534480, "dur": 6, + "args": { + "External id": 12173, + "cbid": 165, "correlation": 12173 + } + }, + { + "ph": "s", "id": 12173, "pid": 435, "tid": 435, "ts": 1776649790534480, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649790534490, "dur": 6, + "args": { + "External id": 12176, + "cbid": 139, "correlation": 12176 + } + }, + { + "ph": "f", "id": 12176, "pid": 435, "tid": 435, "ts": 1776649790534490, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790534552, "dur": 1, + "args": { + "External id": 1182, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12204, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12204, "pid": 0, "tid": 7, "ts": 1776649790534552, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534541, "dur": 9, + "args": { + "External id": 1182, + "cbid": 211, "correlation": 12204 + } + }, + { + "ph": "s", "id": 12204, "pid": 435, "tid": 435, "ts": 1776649790534541, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534557, "dur": 0, + "args": { + "External id": 12210, + "cbid": 200, "correlation": 12210 + } + }, + { + "ph": "f", "id": 12210, "pid": 435, "tid": 435, "ts": 1776649790534557, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534558, "dur": 0, + "args": { + "External id": 12211, + "cbid": 273, "correlation": 12211 + } + }, + { + "ph": "f", "id": 12211, "pid": 435, "tid": 435, "ts": 1776649790534558, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790534567, "dur": 15, + "args": { + "External id": 12212, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12212, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12212, "pid": 0, "tid": 7, "ts": 1776649790534567, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534559, "dur": 5, + "args": { + "External id": 12212, + "cbid": 211, "correlation": 12212 + } + }, + { + "ph": "s", "id": 12212, "pid": 435, "tid": 435, "ts": 1776649790534559, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790534605, "dur": 1, + "args": { + "External id": 1187, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12230, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12230, "pid": 0, "tid": 7, "ts": 1776649790534605, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534596, "dur": 7, + "args": { + "External id": 1187, + "cbid": 211, "correlation": 12230 + } + }, + { + "ph": "s", "id": 12230, "pid": 435, "tid": 435, "ts": 1776649790534596, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534609, "dur": 0, + "args": { + "External id": 12236, + "cbid": 200, "correlation": 12236 + } + }, + { + "ph": "f", "id": 12236, "pid": 435, "tid": 435, "ts": 1776649790534609, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534609, "dur": 0, + "args": { + "External id": 12237, + "cbid": 273, "correlation": 12237 + } + }, + { + "ph": "f", "id": 12237, "pid": 435, "tid": 435, "ts": 1776649790534609, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790534618, "dur": 15, + "args": { + "External id": 12238, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12238, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12238, "pid": 0, "tid": 7, "ts": 1776649790534618, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534610, "dur": 5, + "args": { + "External id": 12238, + "cbid": 211, "correlation": 12238 + } + }, + { + "ph": "s", "id": 12238, "pid": 435, "tid": 435, "ts": 1776649790534610, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790534656, "dur": 1, + "args": { + "External id": 1192, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12256, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12256, "pid": 0, "tid": 7, "ts": 1776649790534656, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534646, "dur": 6, + "args": { + "External id": 1192, + "cbid": 211, "correlation": 12256 + } + }, + { + "ph": "s", "id": 12256, "pid": 435, "tid": 435, "ts": 1776649790534646, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534659, "dur": 0, + "args": { + "External id": 12262, + "cbid": 200, "correlation": 12262 + } + }, + { + "ph": "f", "id": 12262, "pid": 435, "tid": 435, "ts": 1776649790534659, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534660, "dur": 0, + "args": { + "External id": 12263, + "cbid": 273, "correlation": 12263 + } + }, + { + "ph": "f", "id": 12263, "pid": 435, "tid": 435, "ts": 1776649790534660, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790534668, "dur": 15, + "args": { + "External id": 12264, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12264, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12264, "pid": 0, "tid": 7, "ts": 1776649790534668, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534661, "dur": 4, + "args": { + "External id": 12264, + "cbid": 211, "correlation": 12264 + } + }, + { + "ph": "s", "id": 12264, "pid": 435, "tid": 435, "ts": 1776649790534661, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790534705, "dur": 1, + "args": { + "External id": 1197, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12282, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12282, "pid": 0, "tid": 7, "ts": 1776649790534705, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534695, "dur": 6, + "args": { + "External id": 1197, + "cbid": 211, "correlation": 12282 + } + }, + { + "ph": "s", "id": 12282, "pid": 435, "tid": 435, "ts": 1776649790534695, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534708, "dur": 0, + "args": { + "External id": 12288, + "cbid": 200, "correlation": 12288 + } + }, + { + "ph": "f", "id": 12288, "pid": 435, "tid": 435, "ts": 1776649790534708, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534708, "dur": 0, + "args": { + "External id": 12289, + "cbid": 273, "correlation": 12289 + } + }, + { + "ph": "f", "id": 12289, "pid": 435, "tid": 435, "ts": 1776649790534708, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790534717, "dur": 15, + "args": { + "External id": 12290, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12290, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12290, "pid": 0, "tid": 7, "ts": 1776649790534717, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534709, "dur": 4, + "args": { + "External id": 12290, + "cbid": 211, "correlation": 12290 + } + }, + { + "ph": "s", "id": 12290, "pid": 435, "tid": 435, "ts": 1776649790534709, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790534753, "dur": 1, + "args": { + "External id": 1202, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12308, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12308, "pid": 0, "tid": 7, "ts": 1776649790534753, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534744, "dur": 6, + "args": { + "External id": 1202, + "cbid": 211, "correlation": 12308 + } + }, + { + "ph": "s", "id": 12308, "pid": 435, "tid": 435, "ts": 1776649790534744, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534756, "dur": 0, + "args": { + "External id": 12314, + "cbid": 200, "correlation": 12314 + } + }, + { + "ph": "f", "id": 12314, "pid": 435, "tid": 435, "ts": 1776649790534756, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534757, "dur": 0, + "args": { + "External id": 12315, + "cbid": 273, "correlation": 12315 + } + }, + { + "ph": "f", "id": 12315, "pid": 435, "tid": 435, "ts": 1776649790534757, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790534766, "dur": 15, + "args": { + "External id": 12316, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12316, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12316, "pid": 0, "tid": 7, "ts": 1776649790534766, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534758, "dur": 4, + "args": { + "External id": 12316, + "cbid": 211, "correlation": 12316 + } + }, + { + "ph": "s", "id": 12316, "pid": 435, "tid": 435, "ts": 1776649790534758, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790534801, "dur": 1, + "args": { + "External id": 1207, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12334, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12334, "pid": 0, "tid": 7, "ts": 1776649790534801, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534792, "dur": 6, + "args": { + "External id": 1207, + "cbid": 211, "correlation": 12334 + } + }, + { + "ph": "s", "id": 12334, "pid": 435, "tid": 435, "ts": 1776649790534792, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534804, "dur": 0, + "args": { + "External id": 12340, + "cbid": 200, "correlation": 12340 + } + }, + { + "ph": "f", "id": 12340, "pid": 435, "tid": 435, "ts": 1776649790534804, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534805, "dur": 0, + "args": { + "External id": 12341, + "cbid": 273, "correlation": 12341 + } + }, + { + "ph": "f", "id": 12341, "pid": 435, "tid": 435, "ts": 1776649790534805, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790534814, "dur": 15, + "args": { + "External id": 12342, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12342, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12342, "pid": 0, "tid": 7, "ts": 1776649790534814, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534806, "dur": 5, + "args": { + "External id": 12342, + "cbid": 211, "correlation": 12342 + } + }, + { + "ph": "s", "id": 12342, "pid": 435, "tid": 435, "ts": 1776649790534806, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790534849, "dur": 1, + "args": { + "External id": 1212, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12360, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12360, "pid": 0, "tid": 7, "ts": 1776649790534849, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534840, "dur": 6, + "args": { + "External id": 1212, + "cbid": 211, "correlation": 12360 + } + }, + { + "ph": "s", "id": 12360, "pid": 435, "tid": 435, "ts": 1776649790534840, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534853, "dur": 0, + "args": { + "External id": 12366, + "cbid": 200, "correlation": 12366 + } + }, + { + "ph": "f", "id": 12366, "pid": 435, "tid": 435, "ts": 1776649790534853, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534853, "dur": 0, + "args": { + "External id": 12367, + "cbid": 273, "correlation": 12367 + } + }, + { + "ph": "f", "id": 12367, "pid": 435, "tid": 435, "ts": 1776649790534853, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790534862, "dur": 16, + "args": { + "External id": 12368, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12368, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12368, "pid": 0, "tid": 7, "ts": 1776649790534862, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534854, "dur": 4, + "args": { + "External id": 12368, + "cbid": 211, "correlation": 12368 + } + }, + { + "ph": "s", "id": 12368, "pid": 435, "tid": 435, "ts": 1776649790534854, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790534898, "dur": 1, + "args": { + "External id": 1217, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12386, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12386, "pid": 0, "tid": 7, "ts": 1776649790534898, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534888, "dur": 6, + "args": { + "External id": 1217, + "cbid": 211, "correlation": 12386 + } + }, + { + "ph": "s", "id": 12386, "pid": 435, "tid": 435, "ts": 1776649790534888, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534901, "dur": 0, + "args": { + "External id": 12392, + "cbid": 200, "correlation": 12392 + } + }, + { + "ph": "f", "id": 12392, "pid": 435, "tid": 435, "ts": 1776649790534901, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534901, "dur": 0, + "args": { + "External id": 12393, + "cbid": 273, "correlation": 12393 + } + }, + { + "ph": "f", "id": 12393, "pid": 435, "tid": 435, "ts": 1776649790534901, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790534911, "dur": 18, + "args": { + "External id": 12394, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12394, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12394, "pid": 0, "tid": 7, "ts": 1776649790534911, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534903, "dur": 5, + "args": { + "External id": 12394, + "cbid": 211, "correlation": 12394 + } + }, + { + "ph": "s", "id": 12394, "pid": 435, "tid": 435, "ts": 1776649790534903, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790534946, "dur": 1, + "args": { + "External id": 1222, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12412, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12412, "pid": 0, "tid": 7, "ts": 1776649790534946, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534937, "dur": 6, + "args": { + "External id": 1222, + "cbid": 211, "correlation": 12412 + } + }, + { + "ph": "s", "id": 12412, "pid": 435, "tid": 435, "ts": 1776649790534937, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534950, "dur": 0, + "args": { + "External id": 12418, + "cbid": 200, "correlation": 12418 + } + }, + { + "ph": "f", "id": 12418, "pid": 435, "tid": 435, "ts": 1776649790534950, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534950, "dur": 0, + "args": { + "External id": 12419, + "cbid": 273, "correlation": 12419 + } + }, + { + "ph": "f", "id": 12419, "pid": 435, "tid": 435, "ts": 1776649790534950, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790534959, "dur": 15, + "args": { + "External id": 12420, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12420, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12420, "pid": 0, "tid": 7, "ts": 1776649790534959, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534951, "dur": 4, + "args": { + "External id": 12420, + "cbid": 211, "correlation": 12420 + } + }, + { + "ph": "s", "id": 12420, "pid": 435, "tid": 435, "ts": 1776649790534951, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790534994, "dur": 1, + "args": { + "External id": 1227, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12438, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12438, "pid": 0, "tid": 7, "ts": 1776649790534994, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534985, "dur": 6, + "args": { + "External id": 1227, + "cbid": 211, "correlation": 12438 + } + }, + { + "ph": "s", "id": 12438, "pid": 435, "tid": 435, "ts": 1776649790534985, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534997, "dur": 0, + "args": { + "External id": 12444, + "cbid": 200, "correlation": 12444 + } + }, + { + "ph": "f", "id": 12444, "pid": 435, "tid": 435, "ts": 1776649790534997, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790534998, "dur": 0, + "args": { + "External id": 12445, + "cbid": 273, "correlation": 12445 + } + }, + { + "ph": "f", "id": 12445, "pid": 435, "tid": 435, "ts": 1776649790534998, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790535007, "dur": 15, + "args": { + "External id": 12446, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12446, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12446, "pid": 0, "tid": 7, "ts": 1776649790535007, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790534999, "dur": 4, + "args": { + "External id": 12446, + "cbid": 211, "correlation": 12446 + } + }, + { + "ph": "s", "id": 12446, "pid": 435, "tid": 435, "ts": 1776649790534999, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790535022, "dur": 6, + "args": { + "External id": 12449, + "cbid": 165, "correlation": 12449 + } + }, + { + "ph": "s", "id": 12449, "pid": 435, "tid": 435, "ts": 1776649790535022, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790535050, "dur": 3, + "args": { + "External id": 12458, + "cbid": 135, "correlation": 12458 + } + }, + { + "ph": "f", "id": 12458, "pid": 435, "tid": 435, "ts": 1776649790535050, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790535099, "dur": 1, + "args": { + "External id": 1232, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12478, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12478, "pid": 0, "tid": 7, "ts": 1776649790535099, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535088, "dur": 8, + "args": { + "External id": 1232, + "cbid": 211, "correlation": 12478 + } + }, + { + "ph": "s", "id": 12478, "pid": 435, "tid": 435, "ts": 1776649790535088, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535103, "dur": 0, + "args": { + "External id": 12484, + "cbid": 200, "correlation": 12484 + } + }, + { + "ph": "f", "id": 12484, "pid": 435, "tid": 435, "ts": 1776649790535103, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535104, "dur": 1, + "args": { + "External id": 12485, + "cbid": 273, "correlation": 12485 + } + }, + { + "ph": "f", "id": 12485, "pid": 435, "tid": 435, "ts": 1776649790535104, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790535114, "dur": 15, + "args": { + "External id": 12486, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12486, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12486, "pid": 0, "tid": 7, "ts": 1776649790535114, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535106, "dur": 5, + "args": { + "External id": 12486, + "cbid": 211, "correlation": 12486 + } + }, + { + "ph": "s", "id": 12486, "pid": 435, "tid": 435, "ts": 1776649790535106, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790535151, "dur": 1, + "args": { + "External id": 1237, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12504, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12504, "pid": 0, "tid": 7, "ts": 1776649790535151, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535142, "dur": 6, + "args": { + "External id": 1237, + "cbid": 211, "correlation": 12504 + } + }, + { + "ph": "s", "id": 12504, "pid": 435, "tid": 435, "ts": 1776649790535142, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535154, "dur": 0, + "args": { + "External id": 12510, + "cbid": 200, "correlation": 12510 + } + }, + { + "ph": "f", "id": 12510, "pid": 435, "tid": 435, "ts": 1776649790535154, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535155, "dur": 0, + "args": { + "External id": 12511, + "cbid": 273, "correlation": 12511 + } + }, + { + "ph": "f", "id": 12511, "pid": 435, "tid": 435, "ts": 1776649790535155, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790535164, "dur": 15, + "args": { + "External id": 12512, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12512, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12512, "pid": 0, "tid": 7, "ts": 1776649790535164, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535156, "dur": 5, + "args": { + "External id": 12512, + "cbid": 211, "correlation": 12512 + } + }, + { + "ph": "s", "id": 12512, "pid": 435, "tid": 435, "ts": 1776649790535156, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790535200, "dur": 1, + "args": { + "External id": 1242, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12530, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12530, "pid": 0, "tid": 7, "ts": 1776649790535200, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535191, "dur": 6, + "args": { + "External id": 1242, + "cbid": 211, "correlation": 12530 + } + }, + { + "ph": "s", "id": 12530, "pid": 435, "tid": 435, "ts": 1776649790535191, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535203, "dur": 0, + "args": { + "External id": 12536, + "cbid": 200, "correlation": 12536 + } + }, + { + "ph": "f", "id": 12536, "pid": 435, "tid": 435, "ts": 1776649790535203, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535204, "dur": 0, + "args": { + "External id": 12537, + "cbid": 273, "correlation": 12537 + } + }, + { + "ph": "f", "id": 12537, "pid": 435, "tid": 435, "ts": 1776649790535204, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790535213, "dur": 16, + "args": { + "External id": 12538, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12538, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12538, "pid": 0, "tid": 7, "ts": 1776649790535213, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535205, "dur": 5, + "args": { + "External id": 12538, + "cbid": 211, "correlation": 12538 + } + }, + { + "ph": "s", "id": 12538, "pid": 435, "tid": 435, "ts": 1776649790535205, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790535249, "dur": 1, + "args": { + "External id": 1247, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12556, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12556, "pid": 0, "tid": 7, "ts": 1776649790535249, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535240, "dur": 6, + "args": { + "External id": 1247, + "cbid": 211, "correlation": 12556 + } + }, + { + "ph": "s", "id": 12556, "pid": 435, "tid": 435, "ts": 1776649790535240, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535252, "dur": 0, + "args": { + "External id": 12562, + "cbid": 200, "correlation": 12562 + } + }, + { + "ph": "f", "id": 12562, "pid": 435, "tid": 435, "ts": 1776649790535252, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535253, "dur": 0, + "args": { + "External id": 12563, + "cbid": 273, "correlation": 12563 + } + }, + { + "ph": "f", "id": 12563, "pid": 435, "tid": 435, "ts": 1776649790535253, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790535262, "dur": 15, + "args": { + "External id": 12564, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12564, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12564, "pid": 0, "tid": 7, "ts": 1776649790535262, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535254, "dur": 5, + "args": { + "External id": 12564, + "cbid": 211, "correlation": 12564 + } + }, + { + "ph": "s", "id": 12564, "pid": 435, "tid": 435, "ts": 1776649790535254, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790535298, "dur": 1, + "args": { + "External id": 1252, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12582, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12582, "pid": 0, "tid": 7, "ts": 1776649790535298, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535289, "dur": 6, + "args": { + "External id": 1252, + "cbid": 211, "correlation": 12582 + } + }, + { + "ph": "s", "id": 12582, "pid": 435, "tid": 435, "ts": 1776649790535289, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535301, "dur": 0, + "args": { + "External id": 12588, + "cbid": 200, "correlation": 12588 + } + }, + { + "ph": "f", "id": 12588, "pid": 435, "tid": 435, "ts": 1776649790535301, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535302, "dur": 0, + "args": { + "External id": 12589, + "cbid": 273, "correlation": 12589 + } + }, + { + "ph": "f", "id": 12589, "pid": 435, "tid": 435, "ts": 1776649790535302, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790535311, "dur": 15, + "args": { + "External id": 12590, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12590, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12590, "pid": 0, "tid": 7, "ts": 1776649790535311, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535303, "dur": 5, + "args": { + "External id": 12590, + "cbid": 211, "correlation": 12590 + } + }, + { + "ph": "s", "id": 12590, "pid": 435, "tid": 435, "ts": 1776649790535303, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790535347, "dur": 1, + "args": { + "External id": 1257, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12608, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12608, "pid": 0, "tid": 7, "ts": 1776649790535347, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535337, "dur": 6, + "args": { + "External id": 1257, + "cbid": 211, "correlation": 12608 + } + }, + { + "ph": "s", "id": 12608, "pid": 435, "tid": 435, "ts": 1776649790535337, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535350, "dur": 0, + "args": { + "External id": 12614, + "cbid": 200, "correlation": 12614 + } + }, + { + "ph": "f", "id": 12614, "pid": 435, "tid": 435, "ts": 1776649790535350, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535350, "dur": 0, + "args": { + "External id": 12615, + "cbid": 273, "correlation": 12615 + } + }, + { + "ph": "f", "id": 12615, "pid": 435, "tid": 435, "ts": 1776649790535350, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790535360, "dur": 16, + "args": { + "External id": 12616, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12616, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12616, "pid": 0, "tid": 7, "ts": 1776649790535360, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535352, "dur": 4, + "args": { + "External id": 12616, + "cbid": 211, "correlation": 12616 + } + }, + { + "ph": "s", "id": 12616, "pid": 435, "tid": 435, "ts": 1776649790535352, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790535395, "dur": 1, + "args": { + "External id": 1262, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12634, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12634, "pid": 0, "tid": 7, "ts": 1776649790535395, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535386, "dur": 6, + "args": { + "External id": 1262, + "cbid": 211, "correlation": 12634 + } + }, + { + "ph": "s", "id": 12634, "pid": 435, "tid": 435, "ts": 1776649790535386, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535399, "dur": 0, + "args": { + "External id": 12640, + "cbid": 200, "correlation": 12640 + } + }, + { + "ph": "f", "id": 12640, "pid": 435, "tid": 435, "ts": 1776649790535399, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535399, "dur": 0, + "args": { + "External id": 12641, + "cbid": 273, "correlation": 12641 + } + }, + { + "ph": "f", "id": 12641, "pid": 435, "tid": 435, "ts": 1776649790535399, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790535408, "dur": 15, + "args": { + "External id": 12642, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12642, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12642, "pid": 0, "tid": 7, "ts": 1776649790535408, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535400, "dur": 5, + "args": { + "External id": 12642, + "cbid": 211, "correlation": 12642 + } + }, + { + "ph": "s", "id": 12642, "pid": 435, "tid": 435, "ts": 1776649790535400, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790535444, "dur": 1, + "args": { + "External id": 1267, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12660, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12660, "pid": 0, "tid": 7, "ts": 1776649790535444, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535435, "dur": 6, + "args": { + "External id": 1267, + "cbid": 211, "correlation": 12660 + } + }, + { + "ph": "s", "id": 12660, "pid": 435, "tid": 435, "ts": 1776649790535435, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535448, "dur": 0, + "args": { + "External id": 12666, + "cbid": 200, "correlation": 12666 + } + }, + { + "ph": "f", "id": 12666, "pid": 435, "tid": 435, "ts": 1776649790535448, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535448, "dur": 0, + "args": { + "External id": 12667, + "cbid": 273, "correlation": 12667 + } + }, + { + "ph": "f", "id": 12667, "pid": 435, "tid": 435, "ts": 1776649790535448, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790535457, "dur": 15, + "args": { + "External id": 12668, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12668, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12668, "pid": 0, "tid": 7, "ts": 1776649790535457, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535449, "dur": 5, + "args": { + "External id": 12668, + "cbid": 211, "correlation": 12668 + } + }, + { + "ph": "s", "id": 12668, "pid": 435, "tid": 435, "ts": 1776649790535449, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790535494, "dur": 1, + "args": { + "External id": 1272, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12686, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12686, "pid": 0, "tid": 7, "ts": 1776649790535494, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535484, "dur": 6, + "args": { + "External id": 1272, + "cbid": 211, "correlation": 12686 + } + }, + { + "ph": "s", "id": 12686, "pid": 435, "tid": 435, "ts": 1776649790535484, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535497, "dur": 0, + "args": { + "External id": 12692, + "cbid": 200, "correlation": 12692 + } + }, + { + "ph": "f", "id": 12692, "pid": 435, "tid": 435, "ts": 1776649790535497, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535498, "dur": 0, + "args": { + "External id": 12693, + "cbid": 273, "correlation": 12693 + } + }, + { + "ph": "f", "id": 12693, "pid": 435, "tid": 435, "ts": 1776649790535498, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790535507, "dur": 15, + "args": { + "External id": 12694, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12694, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12694, "pid": 0, "tid": 7, "ts": 1776649790535507, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535499, "dur": 5, + "args": { + "External id": 12694, + "cbid": 211, "correlation": 12694 + } + }, + { + "ph": "s", "id": 12694, "pid": 435, "tid": 435, "ts": 1776649790535499, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790535542, "dur": 1, + "args": { + "External id": 1277, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12712, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12712, "pid": 0, "tid": 7, "ts": 1776649790535542, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535533, "dur": 6, + "args": { + "External id": 1277, + "cbid": 211, "correlation": 12712 + } + }, + { + "ph": "s", "id": 12712, "pid": 435, "tid": 435, "ts": 1776649790535533, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535546, "dur": 0, + "args": { + "External id": 12718, + "cbid": 200, "correlation": 12718 + } + }, + { + "ph": "f", "id": 12718, "pid": 435, "tid": 435, "ts": 1776649790535546, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535546, "dur": 0, + "args": { + "External id": 12719, + "cbid": 273, "correlation": 12719 + } + }, + { + "ph": "f", "id": 12719, "pid": 435, "tid": 435, "ts": 1776649790535546, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790535555, "dur": 15, + "args": { + "External id": 12720, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12720, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12720, "pid": 0, "tid": 7, "ts": 1776649790535555, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535547, "dur": 5, + "args": { + "External id": 12720, + "cbid": 211, "correlation": 12720 + } + }, + { + "ph": "s", "id": 12720, "pid": 435, "tid": 435, "ts": 1776649790535547, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790535591, "dur": 1, + "args": { + "External id": 1282, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12738, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12738, "pid": 0, "tid": 7, "ts": 1776649790535591, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535582, "dur": 6, + "args": { + "External id": 1282, + "cbid": 211, "correlation": 12738 + } + }, + { + "ph": "s", "id": 12738, "pid": 435, "tid": 435, "ts": 1776649790535582, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535594, "dur": 0, + "args": { + "External id": 12744, + "cbid": 200, "correlation": 12744 + } + }, + { + "ph": "f", "id": 12744, "pid": 435, "tid": 435, "ts": 1776649790535594, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535595, "dur": 0, + "args": { + "External id": 12745, + "cbid": 273, "correlation": 12745 + } + }, + { + "ph": "f", "id": 12745, "pid": 435, "tid": 435, "ts": 1776649790535595, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790535604, "dur": 15, + "args": { + "External id": 12746, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12746, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12746, "pid": 0, "tid": 7, "ts": 1776649790535604, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535596, "dur": 5, + "args": { + "External id": 12746, + "cbid": 211, "correlation": 12746 + } + }, + { + "ph": "s", "id": 12746, "pid": 435, "tid": 435, "ts": 1776649790535596, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790535640, "dur": 1, + "args": { + "External id": 1287, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12764, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12764, "pid": 0, "tid": 7, "ts": 1776649790535640, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535631, "dur": 6, + "args": { + "External id": 1287, + "cbid": 211, "correlation": 12764 + } + }, + { + "ph": "s", "id": 12764, "pid": 435, "tid": 435, "ts": 1776649790535631, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535643, "dur": 0, + "args": { + "External id": 12770, + "cbid": 200, "correlation": 12770 + } + }, + { + "ph": "f", "id": 12770, "pid": 435, "tid": 435, "ts": 1776649790535643, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535644, "dur": 0, + "args": { + "External id": 12771, + "cbid": 273, "correlation": 12771 + } + }, + { + "ph": "f", "id": 12771, "pid": 435, "tid": 435, "ts": 1776649790535644, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790535653, "dur": 15, + "args": { + "External id": 12772, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12772, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12772, "pid": 0, "tid": 7, "ts": 1776649790535653, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535645, "dur": 5, + "args": { + "External id": 12772, + "cbid": 211, "correlation": 12772 + } + }, + { + "ph": "s", "id": 12772, "pid": 435, "tid": 435, "ts": 1776649790535645, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790535688, "dur": 1, + "args": { + "External id": 1292, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12790, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12790, "pid": 0, "tid": 7, "ts": 1776649790535688, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535679, "dur": 6, + "args": { + "External id": 1292, + "cbid": 211, "correlation": 12790 + } + }, + { + "ph": "s", "id": 12790, "pid": 435, "tid": 435, "ts": 1776649790535679, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535691, "dur": 0, + "args": { + "External id": 12796, + "cbid": 200, "correlation": 12796 + } + }, + { + "ph": "f", "id": 12796, "pid": 435, "tid": 435, "ts": 1776649790535691, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535692, "dur": 0, + "args": { + "External id": 12797, + "cbid": 273, "correlation": 12797 + } + }, + { + "ph": "f", "id": 12797, "pid": 435, "tid": 435, "ts": 1776649790535692, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790535701, "dur": 15, + "args": { + "External id": 12798, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12798, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12798, "pid": 0, "tid": 7, "ts": 1776649790535701, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535693, "dur": 4, + "args": { + "External id": 12798, + "cbid": 211, "correlation": 12798 + } + }, + { + "ph": "s", "id": 12798, "pid": 435, "tid": 435, "ts": 1776649790535693, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790535736, "dur": 1, + "args": { + "External id": 1297, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12816, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12816, "pid": 0, "tid": 7, "ts": 1776649790535736, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535727, "dur": 6, + "args": { + "External id": 1297, + "cbid": 211, "correlation": 12816 + } + }, + { + "ph": "s", "id": 12816, "pid": 435, "tid": 435, "ts": 1776649790535727, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535739, "dur": 0, + "args": { + "External id": 12822, + "cbid": 200, "correlation": 12822 + } + }, + { + "ph": "f", "id": 12822, "pid": 435, "tid": 435, "ts": 1776649790535739, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535740, "dur": 0, + "args": { + "External id": 12823, + "cbid": 273, "correlation": 12823 + } + }, + { + "ph": "f", "id": 12823, "pid": 435, "tid": 435, "ts": 1776649790535740, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790535749, "dur": 16, + "args": { + "External id": 12824, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12824, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12824, "pid": 0, "tid": 7, "ts": 1776649790535749, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535741, "dur": 4, + "args": { + "External id": 12824, + "cbid": 211, "correlation": 12824 + } + }, + { + "ph": "s", "id": 12824, "pid": 435, "tid": 435, "ts": 1776649790535741, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790535784, "dur": 1, + "args": { + "External id": 1302, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12842, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12842, "pid": 0, "tid": 7, "ts": 1776649790535784, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535775, "dur": 6, + "args": { + "External id": 1302, + "cbid": 211, "correlation": 12842 + } + }, + { + "ph": "s", "id": 12842, "pid": 435, "tid": 435, "ts": 1776649790535775, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535787, "dur": 0, + "args": { + "External id": 12848, + "cbid": 200, "correlation": 12848 + } + }, + { + "ph": "f", "id": 12848, "pid": 435, "tid": 435, "ts": 1776649790535787, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535788, "dur": 0, + "args": { + "External id": 12849, + "cbid": 273, "correlation": 12849 + } + }, + { + "ph": "f", "id": 12849, "pid": 435, "tid": 435, "ts": 1776649790535788, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790535797, "dur": 16, + "args": { + "External id": 12850, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12850, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12850, "pid": 0, "tid": 7, "ts": 1776649790535797, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535789, "dur": 5, + "args": { + "External id": 12850, + "cbid": 211, "correlation": 12850 + } + }, + { + "ph": "s", "id": 12850, "pid": 435, "tid": 435, "ts": 1776649790535789, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790535833, "dur": 1, + "args": { + "External id": 1307, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12868, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12868, "pid": 0, "tid": 7, "ts": 1776649790535833, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535823, "dur": 6, + "args": { + "External id": 1307, + "cbid": 211, "correlation": 12868 + } + }, + { + "ph": "s", "id": 12868, "pid": 435, "tid": 435, "ts": 1776649790535823, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535836, "dur": 0, + "args": { + "External id": 12874, + "cbid": 200, "correlation": 12874 + } + }, + { + "ph": "f", "id": 12874, "pid": 435, "tid": 435, "ts": 1776649790535836, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535837, "dur": 0, + "args": { + "External id": 12875, + "cbid": 273, "correlation": 12875 + } + }, + { + "ph": "f", "id": 12875, "pid": 435, "tid": 435, "ts": 1776649790535837, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790535846, "dur": 15, + "args": { + "External id": 12876, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12876, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12876, "pid": 0, "tid": 7, "ts": 1776649790535846, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535838, "dur": 4, + "args": { + "External id": 12876, + "cbid": 211, "correlation": 12876 + } + }, + { + "ph": "s", "id": 12876, "pid": 435, "tid": 435, "ts": 1776649790535838, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790535881, "dur": 1, + "args": { + "External id": 1312, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12894, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12894, "pid": 0, "tid": 7, "ts": 1776649790535881, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535872, "dur": 6, + "args": { + "External id": 1312, + "cbid": 211, "correlation": 12894 + } + }, + { + "ph": "s", "id": 12894, "pid": 435, "tid": 435, "ts": 1776649790535872, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535884, "dur": 0, + "args": { + "External id": 12900, + "cbid": 200, "correlation": 12900 + } + }, + { + "ph": "f", "id": 12900, "pid": 435, "tid": 435, "ts": 1776649790535884, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535885, "dur": 0, + "args": { + "External id": 12901, + "cbid": 273, "correlation": 12901 + } + }, + { + "ph": "f", "id": 12901, "pid": 435, "tid": 435, "ts": 1776649790535885, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790535894, "dur": 15, + "args": { + "External id": 12902, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12902, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12902, "pid": 0, "tid": 7, "ts": 1776649790535894, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535886, "dur": 4, + "args": { + "External id": 12902, + "cbid": 211, "correlation": 12902 + } + }, + { + "ph": "s", "id": 12902, "pid": 435, "tid": 435, "ts": 1776649790535886, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790535929, "dur": 1, + "args": { + "External id": 1317, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12920, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12920, "pid": 0, "tid": 7, "ts": 1776649790535929, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535920, "dur": 6, + "args": { + "External id": 1317, + "cbid": 211, "correlation": 12920 + } + }, + { + "ph": "s", "id": 12920, "pid": 435, "tid": 435, "ts": 1776649790535920, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535932, "dur": 0, + "args": { + "External id": 12926, + "cbid": 200, "correlation": 12926 + } + }, + { + "ph": "f", "id": 12926, "pid": 435, "tid": 435, "ts": 1776649790535932, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535933, "dur": 0, + "args": { + "External id": 12927, + "cbid": 273, "correlation": 12927 + } + }, + { + "ph": "f", "id": 12927, "pid": 435, "tid": 435, "ts": 1776649790535933, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790535942, "dur": 16, + "args": { + "External id": 12928, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12928, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12928, "pid": 0, "tid": 7, "ts": 1776649790535942, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535934, "dur": 4, + "args": { + "External id": 12928, + "cbid": 211, "correlation": 12928 + } + }, + { + "ph": "s", "id": 12928, "pid": 435, "tid": 435, "ts": 1776649790535934, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790535977, "dur": 1, + "args": { + "External id": 1322, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12946, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12946, "pid": 0, "tid": 7, "ts": 1776649790535977, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535968, "dur": 6, + "args": { + "External id": 1322, + "cbid": 211, "correlation": 12946 + } + }, + { + "ph": "s", "id": 12946, "pid": 435, "tid": 435, "ts": 1776649790535968, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535980, "dur": 0, + "args": { + "External id": 12952, + "cbid": 200, "correlation": 12952 + } + }, + { + "ph": "f", "id": 12952, "pid": 435, "tid": 435, "ts": 1776649790535980, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790535981, "dur": 0, + "args": { + "External id": 12953, + "cbid": 273, "correlation": 12953 + } + }, + { + "ph": "f", "id": 12953, "pid": 435, "tid": 435, "ts": 1776649790535981, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790535990, "dur": 15, + "args": { + "External id": 12954, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12954, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12954, "pid": 0, "tid": 7, "ts": 1776649790535990, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790535982, "dur": 4, + "args": { + "External id": 12954, + "cbid": 211, "correlation": 12954 + } + }, + { + "ph": "s", "id": 12954, "pid": 435, "tid": 435, "ts": 1776649790535982, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790536025, "dur": 1, + "args": { + "External id": 1327, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12972, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12972, "pid": 0, "tid": 7, "ts": 1776649790536025, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536016, "dur": 6, + "args": { + "External id": 1327, + "cbid": 211, "correlation": 12972 + } + }, + { + "ph": "s", "id": 12972, "pid": 435, "tid": 435, "ts": 1776649790536016, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536028, "dur": 0, + "args": { + "External id": 12978, + "cbid": 200, "correlation": 12978 + } + }, + { + "ph": "f", "id": 12978, "pid": 435, "tid": 435, "ts": 1776649790536028, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536029, "dur": 0, + "args": { + "External id": 12979, + "cbid": 273, "correlation": 12979 + } + }, + { + "ph": "f", "id": 12979, "pid": 435, "tid": 435, "ts": 1776649790536029, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790536038, "dur": 15, + "args": { + "External id": 12980, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12980, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12980, "pid": 0, "tid": 7, "ts": 1776649790536038, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536030, "dur": 4, + "args": { + "External id": 12980, + "cbid": 211, "correlation": 12980 + } + }, + { + "ph": "s", "id": 12980, "pid": 435, "tid": 435, "ts": 1776649790536030, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790536073, "dur": 1, + "args": { + "External id": 1332, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 12998, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 12998, "pid": 0, "tid": 7, "ts": 1776649790536073, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536064, "dur": 6, + "args": { + "External id": 1332, + "cbid": 211, "correlation": 12998 + } + }, + { + "ph": "s", "id": 12998, "pid": 435, "tid": 435, "ts": 1776649790536064, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536076, "dur": 0, + "args": { + "External id": 13004, + "cbid": 200, "correlation": 13004 + } + }, + { + "ph": "f", "id": 13004, "pid": 435, "tid": 435, "ts": 1776649790536076, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536077, "dur": 0, + "args": { + "External id": 13005, + "cbid": 273, "correlation": 13005 + } + }, + { + "ph": "f", "id": 13005, "pid": 435, "tid": 435, "ts": 1776649790536077, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790536086, "dur": 15, + "args": { + "External id": 13006, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13006, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13006, "pid": 0, "tid": 7, "ts": 1776649790536086, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536078, "dur": 5, + "args": { + "External id": 13006, + "cbid": 211, "correlation": 13006 + } + }, + { + "ph": "s", "id": 13006, "pid": 435, "tid": 435, "ts": 1776649790536078, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790536121, "dur": 1, + "args": { + "External id": 1337, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13024, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13024, "pid": 0, "tid": 7, "ts": 1776649790536121, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536112, "dur": 6, + "args": { + "External id": 1337, + "cbid": 211, "correlation": 13024 + } + }, + { + "ph": "s", "id": 13024, "pid": 435, "tid": 435, "ts": 1776649790536112, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536124, "dur": 0, + "args": { + "External id": 13030, + "cbid": 200, "correlation": 13030 + } + }, + { + "ph": "f", "id": 13030, "pid": 435, "tid": 435, "ts": 1776649790536124, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536124, "dur": 0, + "args": { + "External id": 13031, + "cbid": 273, "correlation": 13031 + } + }, + { + "ph": "f", "id": 13031, "pid": 435, "tid": 435, "ts": 1776649790536124, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790536134, "dur": 15, + "args": { + "External id": 13032, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13032, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13032, "pid": 0, "tid": 7, "ts": 1776649790536134, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536126, "dur": 4, + "args": { + "External id": 13032, + "cbid": 211, "correlation": 13032 + } + }, + { + "ph": "s", "id": 13032, "pid": 435, "tid": 435, "ts": 1776649790536126, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790536168, "dur": 1, + "args": { + "External id": 1342, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13050, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13050, "pid": 0, "tid": 7, "ts": 1776649790536168, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536159, "dur": 6, + "args": { + "External id": 1342, + "cbid": 211, "correlation": 13050 + } + }, + { + "ph": "s", "id": 13050, "pid": 435, "tid": 435, "ts": 1776649790536159, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536172, "dur": 0, + "args": { + "External id": 13056, + "cbid": 200, "correlation": 13056 + } + }, + { + "ph": "f", "id": 13056, "pid": 435, "tid": 435, "ts": 1776649790536172, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536172, "dur": 0, + "args": { + "External id": 13057, + "cbid": 273, "correlation": 13057 + } + }, + { + "ph": "f", "id": 13057, "pid": 435, "tid": 435, "ts": 1776649790536172, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790536181, "dur": 17, + "args": { + "External id": 13058, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13058, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13058, "pid": 0, "tid": 7, "ts": 1776649790536181, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536173, "dur": 4, + "args": { + "External id": 13058, + "cbid": 211, "correlation": 13058 + } + }, + { + "ph": "s", "id": 13058, "pid": 435, "tid": 435, "ts": 1776649790536173, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790536217, "dur": 1, + "args": { + "External id": 1347, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13076, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13076, "pid": 0, "tid": 7, "ts": 1776649790536217, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536207, "dur": 6, + "args": { + "External id": 1347, + "cbid": 211, "correlation": 13076 + } + }, + { + "ph": "s", "id": 13076, "pid": 435, "tid": 435, "ts": 1776649790536207, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536240, "dur": 0, + "args": { + "External id": 13082, + "cbid": 200, "correlation": 13082 + } + }, + { + "ph": "f", "id": 13082, "pid": 435, "tid": 435, "ts": 1776649790536240, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536241, "dur": 0, + "args": { + "External id": 13083, + "cbid": 273, "correlation": 13083 + } + }, + { + "ph": "f", "id": 13083, "pid": 435, "tid": 435, "ts": 1776649790536241, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790536251, "dur": 16, + "args": { + "External id": 13084, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13084, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13084, "pid": 0, "tid": 7, "ts": 1776649790536251, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536242, "dur": 5, + "args": { + "External id": 13084, + "cbid": 211, "correlation": 13084 + } + }, + { + "ph": "s", "id": 13084, "pid": 435, "tid": 435, "ts": 1776649790536242, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790536291, "dur": 1, + "args": { + "External id": 1352, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13102, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13102, "pid": 0, "tid": 7, "ts": 1776649790536291, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536281, "dur": 7, + "args": { + "External id": 1352, + "cbid": 211, "correlation": 13102 + } + }, + { + "ph": "s", "id": 13102, "pid": 435, "tid": 435, "ts": 1776649790536281, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536294, "dur": 0, + "args": { + "External id": 13108, + "cbid": 200, "correlation": 13108 + } + }, + { + "ph": "f", "id": 13108, "pid": 435, "tid": 435, "ts": 1776649790536294, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536295, "dur": 0, + "args": { + "External id": 13109, + "cbid": 273, "correlation": 13109 + } + }, + { + "ph": "f", "id": 13109, "pid": 435, "tid": 435, "ts": 1776649790536295, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790536304, "dur": 15, + "args": { + "External id": 13110, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13110, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13110, "pid": 0, "tid": 7, "ts": 1776649790536304, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536296, "dur": 5, + "args": { + "External id": 13110, + "cbid": 211, "correlation": 13110 + } + }, + { + "ph": "s", "id": 13110, "pid": 435, "tid": 435, "ts": 1776649790536296, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790536339, "dur": 1, + "args": { + "External id": 1357, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13128, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13128, "pid": 0, "tid": 7, "ts": 1776649790536339, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536330, "dur": 6, + "args": { + "External id": 1357, + "cbid": 211, "correlation": 13128 + } + }, + { + "ph": "s", "id": 13128, "pid": 435, "tid": 435, "ts": 1776649790536330, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536343, "dur": 0, + "args": { + "External id": 13134, + "cbid": 200, "correlation": 13134 + } + }, + { + "ph": "f", "id": 13134, "pid": 435, "tid": 435, "ts": 1776649790536343, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536343, "dur": 0, + "args": { + "External id": 13135, + "cbid": 273, "correlation": 13135 + } + }, + { + "ph": "f", "id": 13135, "pid": 435, "tid": 435, "ts": 1776649790536343, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790536352, "dur": 15, + "args": { + "External id": 13136, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13136, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13136, "pid": 0, "tid": 7, "ts": 1776649790536352, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536344, "dur": 4, + "args": { + "External id": 13136, + "cbid": 211, "correlation": 13136 + } + }, + { + "ph": "s", "id": 13136, "pid": 435, "tid": 435, "ts": 1776649790536344, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790536388, "dur": 1, + "args": { + "External id": 1362, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13154, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13154, "pid": 0, "tid": 7, "ts": 1776649790536388, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536378, "dur": 6, + "args": { + "External id": 1362, + "cbid": 211, "correlation": 13154 + } + }, + { + "ph": "s", "id": 13154, "pid": 435, "tid": 435, "ts": 1776649790536378, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536391, "dur": 0, + "args": { + "External id": 13160, + "cbid": 200, "correlation": 13160 + } + }, + { + "ph": "f", "id": 13160, "pid": 435, "tid": 435, "ts": 1776649790536391, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536391, "dur": 0, + "args": { + "External id": 13161, + "cbid": 273, "correlation": 13161 + } + }, + { + "ph": "f", "id": 13161, "pid": 435, "tid": 435, "ts": 1776649790536391, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790536400, "dur": 15, + "args": { + "External id": 13162, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13162, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13162, "pid": 0, "tid": 7, "ts": 1776649790536400, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536392, "dur": 5, + "args": { + "External id": 13162, + "cbid": 211, "correlation": 13162 + } + }, + { + "ph": "s", "id": 13162, "pid": 435, "tid": 435, "ts": 1776649790536392, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790536436, "dur": 1, + "args": { + "External id": 1367, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13180, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13180, "pid": 0, "tid": 7, "ts": 1776649790536436, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536427, "dur": 6, + "args": { + "External id": 1367, + "cbid": 211, "correlation": 13180 + } + }, + { + "ph": "s", "id": 13180, "pid": 435, "tid": 435, "ts": 1776649790536427, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536439, "dur": 0, + "args": { + "External id": 13186, + "cbid": 200, "correlation": 13186 + } + }, + { + "ph": "f", "id": 13186, "pid": 435, "tid": 435, "ts": 1776649790536439, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536440, "dur": 0, + "args": { + "External id": 13187, + "cbid": 273, "correlation": 13187 + } + }, + { + "ph": "f", "id": 13187, "pid": 435, "tid": 435, "ts": 1776649790536440, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790536449, "dur": 15, + "args": { + "External id": 13188, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13188, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13188, "pid": 0, "tid": 7, "ts": 1776649790536449, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536441, "dur": 4, + "args": { + "External id": 13188, + "cbid": 211, "correlation": 13188 + } + }, + { + "ph": "s", "id": 13188, "pid": 435, "tid": 435, "ts": 1776649790536441, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790536484, "dur": 1, + "args": { + "External id": 1372, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13206, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13206, "pid": 0, "tid": 7, "ts": 1776649790536484, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536475, "dur": 6, + "args": { + "External id": 1372, + "cbid": 211, "correlation": 13206 + } + }, + { + "ph": "s", "id": 13206, "pid": 435, "tid": 435, "ts": 1776649790536475, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536487, "dur": 0, + "args": { + "External id": 13212, + "cbid": 200, "correlation": 13212 + } + }, + { + "ph": "f", "id": 13212, "pid": 435, "tid": 435, "ts": 1776649790536487, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536488, "dur": 0, + "args": { + "External id": 13213, + "cbid": 273, "correlation": 13213 + } + }, + { + "ph": "f", "id": 13213, "pid": 435, "tid": 435, "ts": 1776649790536488, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790536497, "dur": 15, + "args": { + "External id": 13214, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13214, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13214, "pid": 0, "tid": 7, "ts": 1776649790536497, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536489, "dur": 5, + "args": { + "External id": 13214, + "cbid": 211, "correlation": 13214 + } + }, + { + "ph": "s", "id": 13214, "pid": 435, "tid": 435, "ts": 1776649790536489, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790536532, "dur": 1, + "args": { + "External id": 1377, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13232, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13232, "pid": 0, "tid": 7, "ts": 1776649790536532, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536523, "dur": 6, + "args": { + "External id": 1377, + "cbid": 211, "correlation": 13232 + } + }, + { + "ph": "s", "id": 13232, "pid": 435, "tid": 435, "ts": 1776649790536523, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536535, "dur": 0, + "args": { + "External id": 13238, + "cbid": 200, "correlation": 13238 + } + }, + { + "ph": "f", "id": 13238, "pid": 435, "tid": 435, "ts": 1776649790536535, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536536, "dur": 0, + "args": { + "External id": 13239, + "cbid": 273, "correlation": 13239 + } + }, + { + "ph": "f", "id": 13239, "pid": 435, "tid": 435, "ts": 1776649790536536, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790536545, "dur": 15, + "args": { + "External id": 13240, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13240, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13240, "pid": 0, "tid": 7, "ts": 1776649790536545, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536537, "dur": 5, + "args": { + "External id": 13240, + "cbid": 211, "correlation": 13240 + } + }, + { + "ph": "s", "id": 13240, "pid": 435, "tid": 435, "ts": 1776649790536537, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790536580, "dur": 1, + "args": { + "External id": 1382, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13258, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13258, "pid": 0, "tid": 7, "ts": 1776649790536580, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536571, "dur": 6, + "args": { + "External id": 1382, + "cbid": 211, "correlation": 13258 + } + }, + { + "ph": "s", "id": 13258, "pid": 435, "tid": 435, "ts": 1776649790536571, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536583, "dur": 0, + "args": { + "External id": 13264, + "cbid": 200, "correlation": 13264 + } + }, + { + "ph": "f", "id": 13264, "pid": 435, "tid": 435, "ts": 1776649790536583, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536584, "dur": 0, + "args": { + "External id": 13265, + "cbid": 273, "correlation": 13265 + } + }, + { + "ph": "f", "id": 13265, "pid": 435, "tid": 435, "ts": 1776649790536584, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790536593, "dur": 15, + "args": { + "External id": 13266, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13266, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13266, "pid": 0, "tid": 7, "ts": 1776649790536593, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536585, "dur": 4, + "args": { + "External id": 13266, + "cbid": 211, "correlation": 13266 + } + }, + { + "ph": "s", "id": 13266, "pid": 435, "tid": 435, "ts": 1776649790536585, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790536629, "dur": 1, + "args": { + "External id": 1387, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13284, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13284, "pid": 0, "tid": 7, "ts": 1776649790536629, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536619, "dur": 6, + "args": { + "External id": 1387, + "cbid": 211, "correlation": 13284 + } + }, + { + "ph": "s", "id": 13284, "pid": 435, "tid": 435, "ts": 1776649790536619, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536632, "dur": 0, + "args": { + "External id": 13290, + "cbid": 200, "correlation": 13290 + } + }, + { + "ph": "f", "id": 13290, "pid": 435, "tid": 435, "ts": 1776649790536632, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536632, "dur": 0, + "args": { + "External id": 13291, + "cbid": 273, "correlation": 13291 + } + }, + { + "ph": "f", "id": 13291, "pid": 435, "tid": 435, "ts": 1776649790536632, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790536641, "dur": 16, + "args": { + "External id": 13292, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13292, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13292, "pid": 0, "tid": 7, "ts": 1776649790536641, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536633, "dur": 4, + "args": { + "External id": 13292, + "cbid": 211, "correlation": 13292 + } + }, + { + "ph": "s", "id": 13292, "pid": 435, "tid": 435, "ts": 1776649790536633, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790536677, "dur": 1, + "args": { + "External id": 1392, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13310, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13310, "pid": 0, "tid": 7, "ts": 1776649790536677, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536668, "dur": 6, + "args": { + "External id": 1392, + "cbid": 211, "correlation": 13310 + } + }, + { + "ph": "s", "id": 13310, "pid": 435, "tid": 435, "ts": 1776649790536668, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536680, "dur": 0, + "args": { + "External id": 13316, + "cbid": 200, "correlation": 13316 + } + }, + { + "ph": "f", "id": 13316, "pid": 435, "tid": 435, "ts": 1776649790536680, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536681, "dur": 0, + "args": { + "External id": 13317, + "cbid": 273, "correlation": 13317 + } + }, + { + "ph": "f", "id": 13317, "pid": 435, "tid": 435, "ts": 1776649790536681, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790536689, "dur": 15, + "args": { + "External id": 13318, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13318, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13318, "pid": 0, "tid": 7, "ts": 1776649790536689, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536682, "dur": 5, + "args": { + "External id": 13318, + "cbid": 211, "correlation": 13318 + } + }, + { + "ph": "s", "id": 13318, "pid": 435, "tid": 435, "ts": 1776649790536682, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790536725, "dur": 1, + "args": { + "External id": 1397, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13336, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13336, "pid": 0, "tid": 7, "ts": 1776649790536725, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536716, "dur": 6, + "args": { + "External id": 1397, + "cbid": 211, "correlation": 13336 + } + }, + { + "ph": "s", "id": 13336, "pid": 435, "tid": 435, "ts": 1776649790536716, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536728, "dur": 0, + "args": { + "External id": 13342, + "cbid": 200, "correlation": 13342 + } + }, + { + "ph": "f", "id": 13342, "pid": 435, "tid": 435, "ts": 1776649790536728, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536728, "dur": 0, + "args": { + "External id": 13343, + "cbid": 273, "correlation": 13343 + } + }, + { + "ph": "f", "id": 13343, "pid": 435, "tid": 435, "ts": 1776649790536728, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790536738, "dur": 15, + "args": { + "External id": 13344, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13344, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13344, "pid": 0, "tid": 7, "ts": 1776649790536738, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536730, "dur": 5, + "args": { + "External id": 13344, + "cbid": 211, "correlation": 13344 + } + }, + { + "ph": "s", "id": 13344, "pid": 435, "tid": 435, "ts": 1776649790536730, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790536773, "dur": 1, + "args": { + "External id": 1402, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13362, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13362, "pid": 0, "tid": 7, "ts": 1776649790536773, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536763, "dur": 6, + "args": { + "External id": 1402, + "cbid": 211, "correlation": 13362 + } + }, + { + "ph": "s", "id": 13362, "pid": 435, "tid": 435, "ts": 1776649790536763, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536776, "dur": 0, + "args": { + "External id": 13368, + "cbid": 200, "correlation": 13368 + } + }, + { + "ph": "f", "id": 13368, "pid": 435, "tid": 435, "ts": 1776649790536776, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536776, "dur": 0, + "args": { + "External id": 13369, + "cbid": 273, "correlation": 13369 + } + }, + { + "ph": "f", "id": 13369, "pid": 435, "tid": 435, "ts": 1776649790536776, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790536785, "dur": 15, + "args": { + "External id": 13370, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13370, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13370, "pid": 0, "tid": 7, "ts": 1776649790536785, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536778, "dur": 4, + "args": { + "External id": 13370, + "cbid": 211, "correlation": 13370 + } + }, + { + "ph": "s", "id": 13370, "pid": 435, "tid": 435, "ts": 1776649790536778, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790536820, "dur": 1, + "args": { + "External id": 1407, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13388, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13388, "pid": 0, "tid": 7, "ts": 1776649790536820, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536811, "dur": 6, + "args": { + "External id": 1407, + "cbid": 211, "correlation": 13388 + } + }, + { + "ph": "s", "id": 13388, "pid": 435, "tid": 435, "ts": 1776649790536811, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536823, "dur": 0, + "args": { + "External id": 13394, + "cbid": 200, "correlation": 13394 + } + }, + { + "ph": "f", "id": 13394, "pid": 435, "tid": 435, "ts": 1776649790536823, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536824, "dur": 0, + "args": { + "External id": 13395, + "cbid": 273, "correlation": 13395 + } + }, + { + "ph": "f", "id": 13395, "pid": 435, "tid": 435, "ts": 1776649790536824, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790536833, "dur": 15, + "args": { + "External id": 13396, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13396, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13396, "pid": 0, "tid": 7, "ts": 1776649790536833, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536825, "dur": 4, + "args": { + "External id": 13396, + "cbid": 211, "correlation": 13396 + } + }, + { + "ph": "s", "id": 13396, "pid": 435, "tid": 435, "ts": 1776649790536825, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790536869, "dur": 1, + "args": { + "External id": 1412, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13414, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13414, "pid": 0, "tid": 7, "ts": 1776649790536869, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536859, "dur": 6, + "args": { + "External id": 1412, + "cbid": 211, "correlation": 13414 + } + }, + { + "ph": "s", "id": 13414, "pid": 435, "tid": 435, "ts": 1776649790536859, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536872, "dur": 0, + "args": { + "External id": 13420, + "cbid": 200, "correlation": 13420 + } + }, + { + "ph": "f", "id": 13420, "pid": 435, "tid": 435, "ts": 1776649790536872, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536872, "dur": 0, + "args": { + "External id": 13421, + "cbid": 273, "correlation": 13421 + } + }, + { + "ph": "f", "id": 13421, "pid": 435, "tid": 435, "ts": 1776649790536872, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790536881, "dur": 15, + "args": { + "External id": 13422, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13422, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13422, "pid": 0, "tid": 7, "ts": 1776649790536881, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536873, "dur": 4, + "args": { + "External id": 13422, + "cbid": 211, "correlation": 13422 + } + }, + { + "ph": "s", "id": 13422, "pid": 435, "tid": 435, "ts": 1776649790536873, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790536917, "dur": 1, + "args": { + "External id": 1417, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13440, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13440, "pid": 0, "tid": 7, "ts": 1776649790536917, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536908, "dur": 6, + "args": { + "External id": 1417, + "cbid": 211, "correlation": 13440 + } + }, + { + "ph": "s", "id": 13440, "pid": 435, "tid": 435, "ts": 1776649790536908, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536920, "dur": 0, + "args": { + "External id": 13446, + "cbid": 200, "correlation": 13446 + } + }, + { + "ph": "f", "id": 13446, "pid": 435, "tid": 435, "ts": 1776649790536920, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536921, "dur": 0, + "args": { + "External id": 13447, + "cbid": 273, "correlation": 13447 + } + }, + { + "ph": "f", "id": 13447, "pid": 435, "tid": 435, "ts": 1776649790536921, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790536930, "dur": 15, + "args": { + "External id": 13448, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13448, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13448, "pid": 0, "tid": 7, "ts": 1776649790536930, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536922, "dur": 4, + "args": { + "External id": 13448, + "cbid": 211, "correlation": 13448 + } + }, + { + "ph": "s", "id": 13448, "pid": 435, "tid": 435, "ts": 1776649790536922, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790536965, "dur": 1, + "args": { + "External id": 1422, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13466, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13466, "pid": 0, "tid": 7, "ts": 1776649790536965, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536956, "dur": 6, + "args": { + "External id": 1422, + "cbid": 211, "correlation": 13466 + } + }, + { + "ph": "s", "id": 13466, "pid": 435, "tid": 435, "ts": 1776649790536956, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536968, "dur": 0, + "args": { + "External id": 13472, + "cbid": 200, "correlation": 13472 + } + }, + { + "ph": "f", "id": 13472, "pid": 435, "tid": 435, "ts": 1776649790536968, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790536969, "dur": 0, + "args": { + "External id": 13473, + "cbid": 273, "correlation": 13473 + } + }, + { + "ph": "f", "id": 13473, "pid": 435, "tid": 435, "ts": 1776649790536969, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790536978, "dur": 15, + "args": { + "External id": 13474, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13474, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13474, "pid": 0, "tid": 7, "ts": 1776649790536978, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790536970, "dur": 5, + "args": { + "External id": 13474, + "cbid": 211, "correlation": 13474 + } + }, + { + "ph": "s", "id": 13474, "pid": 435, "tid": 435, "ts": 1776649790536970, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790537013, "dur": 1, + "args": { + "External id": 1427, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13492, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13492, "pid": 0, "tid": 7, "ts": 1776649790537013, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537004, "dur": 6, + "args": { + "External id": 1427, + "cbid": 211, "correlation": 13492 + } + }, + { + "ph": "s", "id": 13492, "pid": 435, "tid": 435, "ts": 1776649790537004, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537016, "dur": 0, + "args": { + "External id": 13498, + "cbid": 200, "correlation": 13498 + } + }, + { + "ph": "f", "id": 13498, "pid": 435, "tid": 435, "ts": 1776649790537016, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537017, "dur": 0, + "args": { + "External id": 13499, + "cbid": 273, "correlation": 13499 + } + }, + { + "ph": "f", "id": 13499, "pid": 435, "tid": 435, "ts": 1776649790537017, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790537026, "dur": 15, + "args": { + "External id": 13500, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13500, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13500, "pid": 0, "tid": 7, "ts": 1776649790537026, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537018, "dur": 4, + "args": { + "External id": 13500, + "cbid": 211, "correlation": 13500 + } + }, + { + "ph": "s", "id": 13500, "pid": 435, "tid": 435, "ts": 1776649790537018, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790537061, "dur": 1, + "args": { + "External id": 1432, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13518, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13518, "pid": 0, "tid": 7, "ts": 1776649790537061, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537052, "dur": 6, + "args": { + "External id": 1432, + "cbid": 211, "correlation": 13518 + } + }, + { + "ph": "s", "id": 13518, "pid": 435, "tid": 435, "ts": 1776649790537052, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537064, "dur": 0, + "args": { + "External id": 13524, + "cbid": 200, "correlation": 13524 + } + }, + { + "ph": "f", "id": 13524, "pid": 435, "tid": 435, "ts": 1776649790537064, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537065, "dur": 0, + "args": { + "External id": 13525, + "cbid": 273, "correlation": 13525 + } + }, + { + "ph": "f", "id": 13525, "pid": 435, "tid": 435, "ts": 1776649790537065, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790537074, "dur": 17, + "args": { + "External id": 13526, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13526, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13526, "pid": 0, "tid": 7, "ts": 1776649790537074, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537066, "dur": 4, + "args": { + "External id": 13526, + "cbid": 211, "correlation": 13526 + } + }, + { + "ph": "s", "id": 13526, "pid": 435, "tid": 435, "ts": 1776649790537066, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790537109, "dur": 1, + "args": { + "External id": 1437, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13544, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13544, "pid": 0, "tid": 7, "ts": 1776649790537109, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537100, "dur": 6, + "args": { + "External id": 1437, + "cbid": 211, "correlation": 13544 + } + }, + { + "ph": "s", "id": 13544, "pid": 435, "tid": 435, "ts": 1776649790537100, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537112, "dur": 0, + "args": { + "External id": 13550, + "cbid": 200, "correlation": 13550 + } + }, + { + "ph": "f", "id": 13550, "pid": 435, "tid": 435, "ts": 1776649790537112, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537113, "dur": 0, + "args": { + "External id": 13551, + "cbid": 273, "correlation": 13551 + } + }, + { + "ph": "f", "id": 13551, "pid": 435, "tid": 435, "ts": 1776649790537113, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790537122, "dur": 15, + "args": { + "External id": 13552, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13552, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13552, "pid": 0, "tid": 7, "ts": 1776649790537122, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537114, "dur": 4, + "args": { + "External id": 13552, + "cbid": 211, "correlation": 13552 + } + }, + { + "ph": "s", "id": 13552, "pid": 435, "tid": 435, "ts": 1776649790537114, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790537157, "dur": 1, + "args": { + "External id": 1442, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13570, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13570, "pid": 0, "tid": 7, "ts": 1776649790537157, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537148, "dur": 6, + "args": { + "External id": 1442, + "cbid": 211, "correlation": 13570 + } + }, + { + "ph": "s", "id": 13570, "pid": 435, "tid": 435, "ts": 1776649790537148, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537160, "dur": 0, + "args": { + "External id": 13576, + "cbid": 200, "correlation": 13576 + } + }, + { + "ph": "f", "id": 13576, "pid": 435, "tid": 435, "ts": 1776649790537160, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537161, "dur": 0, + "args": { + "External id": 13577, + "cbid": 273, "correlation": 13577 + } + }, + { + "ph": "f", "id": 13577, "pid": 435, "tid": 435, "ts": 1776649790537161, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790537170, "dur": 15, + "args": { + "External id": 13578, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13578, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13578, "pid": 0, "tid": 7, "ts": 1776649790537170, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537162, "dur": 5, + "args": { + "External id": 13578, + "cbid": 211, "correlation": 13578 + } + }, + { + "ph": "s", "id": 13578, "pid": 435, "tid": 435, "ts": 1776649790537162, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790537205, "dur": 1, + "args": { + "External id": 1447, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13596, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13596, "pid": 0, "tid": 7, "ts": 1776649790537205, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537196, "dur": 6, + "args": { + "External id": 1447, + "cbid": 211, "correlation": 13596 + } + }, + { + "ph": "s", "id": 13596, "pid": 435, "tid": 435, "ts": 1776649790537196, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537208, "dur": 0, + "args": { + "External id": 13602, + "cbid": 200, "correlation": 13602 + } + }, + { + "ph": "f", "id": 13602, "pid": 435, "tid": 435, "ts": 1776649790537208, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537209, "dur": 0, + "args": { + "External id": 13603, + "cbid": 273, "correlation": 13603 + } + }, + { + "ph": "f", "id": 13603, "pid": 435, "tid": 435, "ts": 1776649790537209, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790537218, "dur": 15, + "args": { + "External id": 13604, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13604, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13604, "pid": 0, "tid": 7, "ts": 1776649790537218, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537210, "dur": 5, + "args": { + "External id": 13604, + "cbid": 211, "correlation": 13604 + } + }, + { + "ph": "s", "id": 13604, "pid": 435, "tid": 435, "ts": 1776649790537210, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790537253, "dur": 1, + "args": { + "External id": 1452, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13622, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13622, "pid": 0, "tid": 7, "ts": 1776649790537253, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537244, "dur": 6, + "args": { + "External id": 1452, + "cbid": 211, "correlation": 13622 + } + }, + { + "ph": "s", "id": 13622, "pid": 435, "tid": 435, "ts": 1776649790537244, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537256, "dur": 0, + "args": { + "External id": 13628, + "cbid": 200, "correlation": 13628 + } + }, + { + "ph": "f", "id": 13628, "pid": 435, "tid": 435, "ts": 1776649790537256, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537257, "dur": 0, + "args": { + "External id": 13629, + "cbid": 273, "correlation": 13629 + } + }, + { + "ph": "f", "id": 13629, "pid": 435, "tid": 435, "ts": 1776649790537257, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790537266, "dur": 15, + "args": { + "External id": 13630, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13630, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13630, "pid": 0, "tid": 7, "ts": 1776649790537266, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537258, "dur": 4, + "args": { + "External id": 13630, + "cbid": 211, "correlation": 13630 + } + }, + { + "ph": "s", "id": 13630, "pid": 435, "tid": 435, "ts": 1776649790537258, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790537301, "dur": 1, + "args": { + "External id": 1457, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13648, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13648, "pid": 0, "tid": 7, "ts": 1776649790537301, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537292, "dur": 6, + "args": { + "External id": 1457, + "cbid": 211, "correlation": 13648 + } + }, + { + "ph": "s", "id": 13648, "pid": 435, "tid": 435, "ts": 1776649790537292, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537304, "dur": 0, + "args": { + "External id": 13654, + "cbid": 200, "correlation": 13654 + } + }, + { + "ph": "f", "id": 13654, "pid": 435, "tid": 435, "ts": 1776649790537304, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537305, "dur": 0, + "args": { + "External id": 13655, + "cbid": 273, "correlation": 13655 + } + }, + { + "ph": "f", "id": 13655, "pid": 435, "tid": 435, "ts": 1776649790537305, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790537314, "dur": 15, + "args": { + "External id": 13656, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13656, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13656, "pid": 0, "tid": 7, "ts": 1776649790537314, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537306, "dur": 4, + "args": { + "External id": 13656, + "cbid": 211, "correlation": 13656 + } + }, + { + "ph": "s", "id": 13656, "pid": 435, "tid": 435, "ts": 1776649790537306, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790537349, "dur": 1, + "args": { + "External id": 1462, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13674, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13674, "pid": 0, "tid": 7, "ts": 1776649790537349, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537340, "dur": 6, + "args": { + "External id": 1462, + "cbid": 211, "correlation": 13674 + } + }, + { + "ph": "s", "id": 13674, "pid": 435, "tid": 435, "ts": 1776649790537340, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537352, "dur": 0, + "args": { + "External id": 13680, + "cbid": 200, "correlation": 13680 + } + }, + { + "ph": "f", "id": 13680, "pid": 435, "tid": 435, "ts": 1776649790537352, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537353, "dur": 0, + "args": { + "External id": 13681, + "cbid": 273, "correlation": 13681 + } + }, + { + "ph": "f", "id": 13681, "pid": 435, "tid": 435, "ts": 1776649790537353, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790537362, "dur": 15, + "args": { + "External id": 13682, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13682, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13682, "pid": 0, "tid": 7, "ts": 1776649790537362, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537354, "dur": 4, + "args": { + "External id": 13682, + "cbid": 211, "correlation": 13682 + } + }, + { + "ph": "s", "id": 13682, "pid": 435, "tid": 435, "ts": 1776649790537354, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790537397, "dur": 1, + "args": { + "External id": 1467, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13700, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13700, "pid": 0, "tid": 7, "ts": 1776649790537397, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537388, "dur": 6, + "args": { + "External id": 1467, + "cbid": 211, "correlation": 13700 + } + }, + { + "ph": "s", "id": 13700, "pid": 435, "tid": 435, "ts": 1776649790537388, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537401, "dur": 0, + "args": { + "External id": 13706, + "cbid": 200, "correlation": 13706 + } + }, + { + "ph": "f", "id": 13706, "pid": 435, "tid": 435, "ts": 1776649790537401, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537401, "dur": 0, + "args": { + "External id": 13707, + "cbid": 273, "correlation": 13707 + } + }, + { + "ph": "f", "id": 13707, "pid": 435, "tid": 435, "ts": 1776649790537401, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790537410, "dur": 15, + "args": { + "External id": 13708, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13708, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13708, "pid": 0, "tid": 7, "ts": 1776649790537410, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537402, "dur": 4, + "args": { + "External id": 13708, + "cbid": 211, "correlation": 13708 + } + }, + { + "ph": "s", "id": 13708, "pid": 435, "tid": 435, "ts": 1776649790537402, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790537445, "dur": 1, + "args": { + "External id": 1472, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13726, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13726, "pid": 0, "tid": 7, "ts": 1776649790537445, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537436, "dur": 6, + "args": { + "External id": 1472, + "cbid": 211, "correlation": 13726 + } + }, + { + "ph": "s", "id": 13726, "pid": 435, "tid": 435, "ts": 1776649790537436, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537448, "dur": 0, + "args": { + "External id": 13732, + "cbid": 200, "correlation": 13732 + } + }, + { + "ph": "f", "id": 13732, "pid": 435, "tid": 435, "ts": 1776649790537448, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537448, "dur": 0, + "args": { + "External id": 13733, + "cbid": 273, "correlation": 13733 + } + }, + { + "ph": "f", "id": 13733, "pid": 435, "tid": 435, "ts": 1776649790537448, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790537457, "dur": 16, + "args": { + "External id": 13734, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13734, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13734, "pid": 0, "tid": 7, "ts": 1776649790537457, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537449, "dur": 4, + "args": { + "External id": 13734, + "cbid": 211, "correlation": 13734 + } + }, + { + "ph": "s", "id": 13734, "pid": 435, "tid": 435, "ts": 1776649790537449, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790537493, "dur": 1, + "args": { + "External id": 1477, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13752, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13752, "pid": 0, "tid": 7, "ts": 1776649790537493, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537483, "dur": 6, + "args": { + "External id": 1477, + "cbid": 211, "correlation": 13752 + } + }, + { + "ph": "s", "id": 13752, "pid": 435, "tid": 435, "ts": 1776649790537483, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537496, "dur": 0, + "args": { + "External id": 13758, + "cbid": 200, "correlation": 13758 + } + }, + { + "ph": "f", "id": 13758, "pid": 435, "tid": 435, "ts": 1776649790537496, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537496, "dur": 0, + "args": { + "External id": 13759, + "cbid": 273, "correlation": 13759 + } + }, + { + "ph": "f", "id": 13759, "pid": 435, "tid": 435, "ts": 1776649790537496, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790537505, "dur": 16, + "args": { + "External id": 13760, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13760, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13760, "pid": 0, "tid": 7, "ts": 1776649790537505, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537497, "dur": 4, + "args": { + "External id": 13760, + "cbid": 211, "correlation": 13760 + } + }, + { + "ph": "s", "id": 13760, "pid": 435, "tid": 435, "ts": 1776649790537497, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790537528, "dur": 3, + "args": { + "External id": 13768, + "cbid": 135, "correlation": 13768 + } + }, + { + "ph": "f", "id": 13768, "pid": 435, "tid": 435, "ts": 1776649790537528, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790537543, "dur": 5, + "args": { + "External id": 13773, + "cbid": 165, "correlation": 13773 + } + }, + { + "ph": "s", "id": 13773, "pid": 435, "tid": 435, "ts": 1776649790537543, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649790537552, "dur": 4, + "args": { + "External id": 13776, + "cbid": 139, "correlation": 13776 + } + }, + { + "ph": "f", "id": 13776, "pid": 435, "tid": 435, "ts": 1776649790537552, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790537608, "dur": 1, + "args": { + "External id": 1482, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13804, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13804, "pid": 0, "tid": 7, "ts": 1776649790537608, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537597, "dur": 8, + "args": { + "External id": 1482, + "cbid": 211, "correlation": 13804 + } + }, + { + "ph": "s", "id": 13804, "pid": 435, "tid": 435, "ts": 1776649790537597, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537612, "dur": 0, + "args": { + "External id": 13810, + "cbid": 200, "correlation": 13810 + } + }, + { + "ph": "f", "id": 13810, "pid": 435, "tid": 435, "ts": 1776649790537612, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537613, "dur": 0, + "args": { + "External id": 13811, + "cbid": 273, "correlation": 13811 + } + }, + { + "ph": "f", "id": 13811, "pid": 435, "tid": 435, "ts": 1776649790537613, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790537622, "dur": 15, + "args": { + "External id": 13812, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13812, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13812, "pid": 0, "tid": 7, "ts": 1776649790537622, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537614, "dur": 5, + "args": { + "External id": 13812, + "cbid": 211, "correlation": 13812 + } + }, + { + "ph": "s", "id": 13812, "pid": 435, "tid": 435, "ts": 1776649790537614, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790537666, "dur": 1, + "args": { + "External id": 1487, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13830, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13830, "pid": 0, "tid": 7, "ts": 1776649790537666, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537656, "dur": 7, + "args": { + "External id": 1487, + "cbid": 211, "correlation": 13830 + } + }, + { + "ph": "s", "id": 13830, "pid": 435, "tid": 435, "ts": 1776649790537656, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537670, "dur": 0, + "args": { + "External id": 13836, + "cbid": 200, "correlation": 13836 + } + }, + { + "ph": "f", "id": 13836, "pid": 435, "tid": 435, "ts": 1776649790537670, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537671, "dur": 0, + "args": { + "External id": 13837, + "cbid": 273, "correlation": 13837 + } + }, + { + "ph": "f", "id": 13837, "pid": 435, "tid": 435, "ts": 1776649790537671, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790537680, "dur": 15, + "args": { + "External id": 13838, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13838, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13838, "pid": 0, "tid": 7, "ts": 1776649790537680, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537672, "dur": 4, + "args": { + "External id": 13838, + "cbid": 211, "correlation": 13838 + } + }, + { + "ph": "s", "id": 13838, "pid": 435, "tid": 435, "ts": 1776649790537672, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790537716, "dur": 1, + "args": { + "External id": 1492, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13856, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13856, "pid": 0, "tid": 7, "ts": 1776649790537716, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537707, "dur": 6, + "args": { + "External id": 1492, + "cbid": 211, "correlation": 13856 + } + }, + { + "ph": "s", "id": 13856, "pid": 435, "tid": 435, "ts": 1776649790537707, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537719, "dur": 0, + "args": { + "External id": 13862, + "cbid": 200, "correlation": 13862 + } + }, + { + "ph": "f", "id": 13862, "pid": 435, "tid": 435, "ts": 1776649790537719, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537720, "dur": 0, + "args": { + "External id": 13863, + "cbid": 273, "correlation": 13863 + } + }, + { + "ph": "f", "id": 13863, "pid": 435, "tid": 435, "ts": 1776649790537720, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790537729, "dur": 15, + "args": { + "External id": 13864, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13864, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13864, "pid": 0, "tid": 7, "ts": 1776649790537729, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537721, "dur": 4, + "args": { + "External id": 13864, + "cbid": 211, "correlation": 13864 + } + }, + { + "ph": "s", "id": 13864, "pid": 435, "tid": 435, "ts": 1776649790537721, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790537764, "dur": 1, + "args": { + "External id": 1497, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13882, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13882, "pid": 0, "tid": 7, "ts": 1776649790537764, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537755, "dur": 6, + "args": { + "External id": 1497, + "cbid": 211, "correlation": 13882 + } + }, + { + "ph": "s", "id": 13882, "pid": 435, "tid": 435, "ts": 1776649790537755, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537767, "dur": 0, + "args": { + "External id": 13888, + "cbid": 200, "correlation": 13888 + } + }, + { + "ph": "f", "id": 13888, "pid": 435, "tid": 435, "ts": 1776649790537767, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537768, "dur": 0, + "args": { + "External id": 13889, + "cbid": 273, "correlation": 13889 + } + }, + { + "ph": "f", "id": 13889, "pid": 435, "tid": 435, "ts": 1776649790537768, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790537776, "dur": 15, + "args": { + "External id": 13890, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13890, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13890, "pid": 0, "tid": 7, "ts": 1776649790537776, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537769, "dur": 4, + "args": { + "External id": 13890, + "cbid": 211, "correlation": 13890 + } + }, + { + "ph": "s", "id": 13890, "pid": 435, "tid": 435, "ts": 1776649790537769, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790537811, "dur": 1, + "args": { + "External id": 1502, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13908, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13908, "pid": 0, "tid": 7, "ts": 1776649790537811, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537802, "dur": 6, + "args": { + "External id": 1502, + "cbid": 211, "correlation": 13908 + } + }, + { + "ph": "s", "id": 13908, "pid": 435, "tid": 435, "ts": 1776649790537802, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537814, "dur": 0, + "args": { + "External id": 13914, + "cbid": 200, "correlation": 13914 + } + }, + { + "ph": "f", "id": 13914, "pid": 435, "tid": 435, "ts": 1776649790537814, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537815, "dur": 0, + "args": { + "External id": 13915, + "cbid": 273, "correlation": 13915 + } + }, + { + "ph": "f", "id": 13915, "pid": 435, "tid": 435, "ts": 1776649790537815, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790537824, "dur": 15, + "args": { + "External id": 13916, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13916, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13916, "pid": 0, "tid": 7, "ts": 1776649790537824, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537816, "dur": 4, + "args": { + "External id": 13916, + "cbid": 211, "correlation": 13916 + } + }, + { + "ph": "s", "id": 13916, "pid": 435, "tid": 435, "ts": 1776649790537816, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790537859, "dur": 1, + "args": { + "External id": 1507, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13934, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13934, "pid": 0, "tid": 7, "ts": 1776649790537859, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537850, "dur": 6, + "args": { + "External id": 1507, + "cbid": 211, "correlation": 13934 + } + }, + { + "ph": "s", "id": 13934, "pid": 435, "tid": 435, "ts": 1776649790537850, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537862, "dur": 0, + "args": { + "External id": 13940, + "cbid": 200, "correlation": 13940 + } + }, + { + "ph": "f", "id": 13940, "pid": 435, "tid": 435, "ts": 1776649790537862, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537862, "dur": 0, + "args": { + "External id": 13941, + "cbid": 273, "correlation": 13941 + } + }, + { + "ph": "f", "id": 13941, "pid": 435, "tid": 435, "ts": 1776649790537862, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790537871, "dur": 15, + "args": { + "External id": 13942, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13942, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13942, "pid": 0, "tid": 7, "ts": 1776649790537871, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537863, "dur": 4, + "args": { + "External id": 13942, + "cbid": 211, "correlation": 13942 + } + }, + { + "ph": "s", "id": 13942, "pid": 435, "tid": 435, "ts": 1776649790537863, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790537906, "dur": 1, + "args": { + "External id": 1512, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13960, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13960, "pid": 0, "tid": 7, "ts": 1776649790537906, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537897, "dur": 6, + "args": { + "External id": 1512, + "cbid": 211, "correlation": 13960 + } + }, + { + "ph": "s", "id": 13960, "pid": 435, "tid": 435, "ts": 1776649790537897, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537909, "dur": 0, + "args": { + "External id": 13966, + "cbid": 200, "correlation": 13966 + } + }, + { + "ph": "f", "id": 13966, "pid": 435, "tid": 435, "ts": 1776649790537909, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537910, "dur": 0, + "args": { + "External id": 13967, + "cbid": 273, "correlation": 13967 + } + }, + { + "ph": "f", "id": 13967, "pid": 435, "tid": 435, "ts": 1776649790537910, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790537919, "dur": 16, + "args": { + "External id": 13968, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13968, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13968, "pid": 0, "tid": 7, "ts": 1776649790537919, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537911, "dur": 5, + "args": { + "External id": 13968, + "cbid": 211, "correlation": 13968 + } + }, + { + "ph": "s", "id": 13968, "pid": 435, "tid": 435, "ts": 1776649790537911, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790537953, "dur": 1, + "args": { + "External id": 1517, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13986, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13986, "pid": 0, "tid": 7, "ts": 1776649790537953, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537944, "dur": 6, + "args": { + "External id": 1517, + "cbid": 211, "correlation": 13986 + } + }, + { + "ph": "s", "id": 13986, "pid": 435, "tid": 435, "ts": 1776649790537944, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537956, "dur": 0, + "args": { + "External id": 13992, + "cbid": 200, "correlation": 13992 + } + }, + { + "ph": "f", "id": 13992, "pid": 435, "tid": 435, "ts": 1776649790537956, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790537957, "dur": 0, + "args": { + "External id": 13993, + "cbid": 273, "correlation": 13993 + } + }, + { + "ph": "f", "id": 13993, "pid": 435, "tid": 435, "ts": 1776649790537957, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790537966, "dur": 16, + "args": { + "External id": 13994, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 13994, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 13994, "pid": 0, "tid": 7, "ts": 1776649790537966, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537958, "dur": 4, + "args": { + "External id": 13994, + "cbid": 211, "correlation": 13994 + } + }, + { + "ph": "s", "id": 13994, "pid": 435, "tid": 435, "ts": 1776649790537958, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790538001, "dur": 1, + "args": { + "External id": 1522, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14012, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14012, "pid": 0, "tid": 7, "ts": 1776649790538001, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790537992, "dur": 6, + "args": { + "External id": 1522, + "cbid": 211, "correlation": 14012 + } + }, + { + "ph": "s", "id": 14012, "pid": 435, "tid": 435, "ts": 1776649790537992, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538004, "dur": 0, + "args": { + "External id": 14018, + "cbid": 200, "correlation": 14018 + } + }, + { + "ph": "f", "id": 14018, "pid": 435, "tid": 435, "ts": 1776649790538004, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538004, "dur": 0, + "args": { + "External id": 14019, + "cbid": 273, "correlation": 14019 + } + }, + { + "ph": "f", "id": 14019, "pid": 435, "tid": 435, "ts": 1776649790538004, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790538013, "dur": 15, + "args": { + "External id": 14020, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14020, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14020, "pid": 0, "tid": 7, "ts": 1776649790538013, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538006, "dur": 5, + "args": { + "External id": 14020, + "cbid": 211, "correlation": 14020 + } + }, + { + "ph": "s", "id": 14020, "pid": 435, "tid": 435, "ts": 1776649790538006, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790538048, "dur": 1, + "args": { + "External id": 1527, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14038, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14038, "pid": 0, "tid": 7, "ts": 1776649790538048, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538039, "dur": 6, + "args": { + "External id": 1527, + "cbid": 211, "correlation": 14038 + } + }, + { + "ph": "s", "id": 14038, "pid": 435, "tid": 435, "ts": 1776649790538039, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538051, "dur": 0, + "args": { + "External id": 14044, + "cbid": 200, "correlation": 14044 + } + }, + { + "ph": "f", "id": 14044, "pid": 435, "tid": 435, "ts": 1776649790538051, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538052, "dur": 0, + "args": { + "External id": 14045, + "cbid": 273, "correlation": 14045 + } + }, + { + "ph": "f", "id": 14045, "pid": 435, "tid": 435, "ts": 1776649790538052, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790538061, "dur": 15, + "args": { + "External id": 14046, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14046, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14046, "pid": 0, "tid": 7, "ts": 1776649790538061, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538053, "dur": 5, + "args": { + "External id": 14046, + "cbid": 211, "correlation": 14046 + } + }, + { + "ph": "s", "id": 14046, "pid": 435, "tid": 435, "ts": 1776649790538053, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790538075, "dur": 5, + "args": { + "External id": 14049, + "cbid": 165, "correlation": 14049 + } + }, + { + "ph": "s", "id": 14049, "pid": 435, "tid": 435, "ts": 1776649790538075, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790538102, "dur": 3, + "args": { + "External id": 14058, + "cbid": 135, "correlation": 14058 + } + }, + { + "ph": "f", "id": 14058, "pid": 435, "tid": 435, "ts": 1776649790538102, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790538150, "dur": 1, + "args": { + "External id": 1532, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14078, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14078, "pid": 0, "tid": 7, "ts": 1776649790538150, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538140, "dur": 8, + "args": { + "External id": 1532, + "cbid": 211, "correlation": 14078 + } + }, + { + "ph": "s", "id": 14078, "pid": 435, "tid": 435, "ts": 1776649790538140, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538155, "dur": 0, + "args": { + "External id": 14084, + "cbid": 200, "correlation": 14084 + } + }, + { + "ph": "f", "id": 14084, "pid": 435, "tid": 435, "ts": 1776649790538155, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538155, "dur": 0, + "args": { + "External id": 14085, + "cbid": 273, "correlation": 14085 + } + }, + { + "ph": "f", "id": 14085, "pid": 435, "tid": 435, "ts": 1776649790538155, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790538165, "dur": 15, + "args": { + "External id": 14086, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14086, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14086, "pid": 0, "tid": 7, "ts": 1776649790538165, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538157, "dur": 5, + "args": { + "External id": 14086, + "cbid": 211, "correlation": 14086 + } + }, + { + "ph": "s", "id": 14086, "pid": 435, "tid": 435, "ts": 1776649790538157, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790538211, "dur": 1, + "args": { + "External id": 1537, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14104, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14104, "pid": 0, "tid": 7, "ts": 1776649790538211, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538201, "dur": 6, + "args": { + "External id": 1537, + "cbid": 211, "correlation": 14104 + } + }, + { + "ph": "s", "id": 14104, "pid": 435, "tid": 435, "ts": 1776649790538201, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538214, "dur": 0, + "args": { + "External id": 14110, + "cbid": 200, "correlation": 14110 + } + }, + { + "ph": "f", "id": 14110, "pid": 435, "tid": 435, "ts": 1776649790538214, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538215, "dur": 0, + "args": { + "External id": 14111, + "cbid": 273, "correlation": 14111 + } + }, + { + "ph": "f", "id": 14111, "pid": 435, "tid": 435, "ts": 1776649790538215, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790538224, "dur": 15, + "args": { + "External id": 14112, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14112, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14112, "pid": 0, "tid": 7, "ts": 1776649790538224, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538216, "dur": 4, + "args": { + "External id": 14112, + "cbid": 211, "correlation": 14112 + } + }, + { + "ph": "s", "id": 14112, "pid": 435, "tid": 435, "ts": 1776649790538216, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790538259, "dur": 1, + "args": { + "External id": 1542, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14130, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14130, "pid": 0, "tid": 7, "ts": 1776649790538259, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538250, "dur": 6, + "args": { + "External id": 1542, + "cbid": 211, "correlation": 14130 + } + }, + { + "ph": "s", "id": 14130, "pid": 435, "tid": 435, "ts": 1776649790538250, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538263, "dur": 0, + "args": { + "External id": 14136, + "cbid": 200, "correlation": 14136 + } + }, + { + "ph": "f", "id": 14136, "pid": 435, "tid": 435, "ts": 1776649790538263, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538263, "dur": 0, + "args": { + "External id": 14137, + "cbid": 273, "correlation": 14137 + } + }, + { + "ph": "f", "id": 14137, "pid": 435, "tid": 435, "ts": 1776649790538263, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790538273, "dur": 15, + "args": { + "External id": 14138, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14138, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14138, "pid": 0, "tid": 7, "ts": 1776649790538273, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538265, "dur": 5, + "args": { + "External id": 14138, + "cbid": 211, "correlation": 14138 + } + }, + { + "ph": "s", "id": 14138, "pid": 435, "tid": 435, "ts": 1776649790538265, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790538307, "dur": 1, + "args": { + "External id": 1547, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14156, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14156, "pid": 0, "tid": 7, "ts": 1776649790538307, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538298, "dur": 6, + "args": { + "External id": 1547, + "cbid": 211, "correlation": 14156 + } + }, + { + "ph": "s", "id": 14156, "pid": 435, "tid": 435, "ts": 1776649790538298, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538310, "dur": 0, + "args": { + "External id": 14162, + "cbid": 200, "correlation": 14162 + } + }, + { + "ph": "f", "id": 14162, "pid": 435, "tid": 435, "ts": 1776649790538310, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538311, "dur": 0, + "args": { + "External id": 14163, + "cbid": 273, "correlation": 14163 + } + }, + { + "ph": "f", "id": 14163, "pid": 435, "tid": 435, "ts": 1776649790538311, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790538320, "dur": 15, + "args": { + "External id": 14164, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14164, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14164, "pid": 0, "tid": 7, "ts": 1776649790538320, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538312, "dur": 4, + "args": { + "External id": 14164, + "cbid": 211, "correlation": 14164 + } + }, + { + "ph": "s", "id": 14164, "pid": 435, "tid": 435, "ts": 1776649790538312, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790538354, "dur": 1, + "args": { + "External id": 1552, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14182, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14182, "pid": 0, "tid": 7, "ts": 1776649790538354, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538345, "dur": 6, + "args": { + "External id": 1552, + "cbid": 211, "correlation": 14182 + } + }, + { + "ph": "s", "id": 14182, "pid": 435, "tid": 435, "ts": 1776649790538345, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538357, "dur": 0, + "args": { + "External id": 14188, + "cbid": 200, "correlation": 14188 + } + }, + { + "ph": "f", "id": 14188, "pid": 435, "tid": 435, "ts": 1776649790538357, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538358, "dur": 0, + "args": { + "External id": 14189, + "cbid": 273, "correlation": 14189 + } + }, + { + "ph": "f", "id": 14189, "pid": 435, "tid": 435, "ts": 1776649790538358, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790538367, "dur": 15, + "args": { + "External id": 14190, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14190, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14190, "pid": 0, "tid": 7, "ts": 1776649790538367, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538359, "dur": 4, + "args": { + "External id": 14190, + "cbid": 211, "correlation": 14190 + } + }, + { + "ph": "s", "id": 14190, "pid": 435, "tid": 435, "ts": 1776649790538359, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790538401, "dur": 1, + "args": { + "External id": 1557, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14208, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14208, "pid": 0, "tid": 7, "ts": 1776649790538401, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538392, "dur": 6, + "args": { + "External id": 1557, + "cbid": 211, "correlation": 14208 + } + }, + { + "ph": "s", "id": 14208, "pid": 435, "tid": 435, "ts": 1776649790538392, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538404, "dur": 0, + "args": { + "External id": 14214, + "cbid": 200, "correlation": 14214 + } + }, + { + "ph": "f", "id": 14214, "pid": 435, "tid": 435, "ts": 1776649790538404, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538405, "dur": 0, + "args": { + "External id": 14215, + "cbid": 273, "correlation": 14215 + } + }, + { + "ph": "f", "id": 14215, "pid": 435, "tid": 435, "ts": 1776649790538405, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790538414, "dur": 16, + "args": { + "External id": 14216, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14216, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14216, "pid": 0, "tid": 7, "ts": 1776649790538414, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538406, "dur": 4, + "args": { + "External id": 14216, + "cbid": 211, "correlation": 14216 + } + }, + { + "ph": "s", "id": 14216, "pid": 435, "tid": 435, "ts": 1776649790538406, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790538449, "dur": 1, + "args": { + "External id": 1562, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14234, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14234, "pid": 0, "tid": 7, "ts": 1776649790538449, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538440, "dur": 6, + "args": { + "External id": 1562, + "cbid": 211, "correlation": 14234 + } + }, + { + "ph": "s", "id": 14234, "pid": 435, "tid": 435, "ts": 1776649790538440, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538452, "dur": 0, + "args": { + "External id": 14240, + "cbid": 200, "correlation": 14240 + } + }, + { + "ph": "f", "id": 14240, "pid": 435, "tid": 435, "ts": 1776649790538452, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538453, "dur": 0, + "args": { + "External id": 14241, + "cbid": 273, "correlation": 14241 + } + }, + { + "ph": "f", "id": 14241, "pid": 435, "tid": 435, "ts": 1776649790538453, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790538462, "dur": 17, + "args": { + "External id": 14242, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14242, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14242, "pid": 0, "tid": 7, "ts": 1776649790538462, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538454, "dur": 5, + "args": { + "External id": 14242, + "cbid": 211, "correlation": 14242 + } + }, + { + "ph": "s", "id": 14242, "pid": 435, "tid": 435, "ts": 1776649790538454, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790538497, "dur": 1, + "args": { + "External id": 1567, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14260, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14260, "pid": 0, "tid": 7, "ts": 1776649790538497, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538488, "dur": 6, + "args": { + "External id": 1567, + "cbid": 211, "correlation": 14260 + } + }, + { + "ph": "s", "id": 14260, "pid": 435, "tid": 435, "ts": 1776649790538488, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538500, "dur": 0, + "args": { + "External id": 14266, + "cbid": 200, "correlation": 14266 + } + }, + { + "ph": "f", "id": 14266, "pid": 435, "tid": 435, "ts": 1776649790538500, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538500, "dur": 0, + "args": { + "External id": 14267, + "cbid": 273, "correlation": 14267 + } + }, + { + "ph": "f", "id": 14267, "pid": 435, "tid": 435, "ts": 1776649790538500, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790538509, "dur": 15, + "args": { + "External id": 14268, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14268, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14268, "pid": 0, "tid": 7, "ts": 1776649790538509, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538502, "dur": 5, + "args": { + "External id": 14268, + "cbid": 211, "correlation": 14268 + } + }, + { + "ph": "s", "id": 14268, "pid": 435, "tid": 435, "ts": 1776649790538502, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790538544, "dur": 1, + "args": { + "External id": 1572, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14286, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14286, "pid": 0, "tid": 7, "ts": 1776649790538544, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538535, "dur": 6, + "args": { + "External id": 1572, + "cbid": 211, "correlation": 14286 + } + }, + { + "ph": "s", "id": 14286, "pid": 435, "tid": 435, "ts": 1776649790538535, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538547, "dur": 0, + "args": { + "External id": 14292, + "cbid": 200, "correlation": 14292 + } + }, + { + "ph": "f", "id": 14292, "pid": 435, "tid": 435, "ts": 1776649790538547, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538548, "dur": 0, + "args": { + "External id": 14293, + "cbid": 273, "correlation": 14293 + } + }, + { + "ph": "f", "id": 14293, "pid": 435, "tid": 435, "ts": 1776649790538548, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790538557, "dur": 15, + "args": { + "External id": 14294, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14294, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14294, "pid": 0, "tid": 7, "ts": 1776649790538557, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538549, "dur": 4, + "args": { + "External id": 14294, + "cbid": 211, "correlation": 14294 + } + }, + { + "ph": "s", "id": 14294, "pid": 435, "tid": 435, "ts": 1776649790538549, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790538592, "dur": 1, + "args": { + "External id": 1577, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14312, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14312, "pid": 0, "tid": 7, "ts": 1776649790538592, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538582, "dur": 6, + "args": { + "External id": 1577, + "cbid": 211, "correlation": 14312 + } + }, + { + "ph": "s", "id": 14312, "pid": 435, "tid": 435, "ts": 1776649790538582, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538595, "dur": 0, + "args": { + "External id": 14318, + "cbid": 200, "correlation": 14318 + } + }, + { + "ph": "f", "id": 14318, "pid": 435, "tid": 435, "ts": 1776649790538595, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538595, "dur": 0, + "args": { + "External id": 14319, + "cbid": 273, "correlation": 14319 + } + }, + { + "ph": "f", "id": 14319, "pid": 435, "tid": 435, "ts": 1776649790538595, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790538604, "dur": 15, + "args": { + "External id": 14320, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14320, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14320, "pid": 0, "tid": 7, "ts": 1776649790538604, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538596, "dur": 5, + "args": { + "External id": 14320, + "cbid": 211, "correlation": 14320 + } + }, + { + "ph": "s", "id": 14320, "pid": 435, "tid": 435, "ts": 1776649790538596, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790538639, "dur": 1, + "args": { + "External id": 1582, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14338, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14338, "pid": 0, "tid": 7, "ts": 1776649790538639, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538630, "dur": 6, + "args": { + "External id": 1582, + "cbid": 211, "correlation": 14338 + } + }, + { + "ph": "s", "id": 14338, "pid": 435, "tid": 435, "ts": 1776649790538630, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538642, "dur": 0, + "args": { + "External id": 14344, + "cbid": 200, "correlation": 14344 + } + }, + { + "ph": "f", "id": 14344, "pid": 435, "tid": 435, "ts": 1776649790538642, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538642, "dur": 0, + "args": { + "External id": 14345, + "cbid": 273, "correlation": 14345 + } + }, + { + "ph": "f", "id": 14345, "pid": 435, "tid": 435, "ts": 1776649790538642, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790538652, "dur": 15, + "args": { + "External id": 14346, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14346, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14346, "pid": 0, "tid": 7, "ts": 1776649790538652, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538644, "dur": 4, + "args": { + "External id": 14346, + "cbid": 211, "correlation": 14346 + } + }, + { + "ph": "s", "id": 14346, "pid": 435, "tid": 435, "ts": 1776649790538644, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790538686, "dur": 1, + "args": { + "External id": 1587, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14364, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14364, "pid": 0, "tid": 7, "ts": 1776649790538686, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538677, "dur": 6, + "args": { + "External id": 1587, + "cbid": 211, "correlation": 14364 + } + }, + { + "ph": "s", "id": 14364, "pid": 435, "tid": 435, "ts": 1776649790538677, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538689, "dur": 0, + "args": { + "External id": 14370, + "cbid": 200, "correlation": 14370 + } + }, + { + "ph": "f", "id": 14370, "pid": 435, "tid": 435, "ts": 1776649790538689, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538690, "dur": 0, + "args": { + "External id": 14371, + "cbid": 273, "correlation": 14371 + } + }, + { + "ph": "f", "id": 14371, "pid": 435, "tid": 435, "ts": 1776649790538690, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790538698, "dur": 15, + "args": { + "External id": 14372, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14372, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14372, "pid": 0, "tid": 7, "ts": 1776649790538698, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538691, "dur": 4, + "args": { + "External id": 14372, + "cbid": 211, "correlation": 14372 + } + }, + { + "ph": "s", "id": 14372, "pid": 435, "tid": 435, "ts": 1776649790538691, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790538733, "dur": 1, + "args": { + "External id": 1592, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14390, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14390, "pid": 0, "tid": 7, "ts": 1776649790538733, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538724, "dur": 6, + "args": { + "External id": 1592, + "cbid": 211, "correlation": 14390 + } + }, + { + "ph": "s", "id": 14390, "pid": 435, "tid": 435, "ts": 1776649790538724, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538736, "dur": 0, + "args": { + "External id": 14396, + "cbid": 200, "correlation": 14396 + } + }, + { + "ph": "f", "id": 14396, "pid": 435, "tid": 435, "ts": 1776649790538736, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538737, "dur": 0, + "args": { + "External id": 14397, + "cbid": 273, "correlation": 14397 + } + }, + { + "ph": "f", "id": 14397, "pid": 435, "tid": 435, "ts": 1776649790538737, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790538746, "dur": 15, + "args": { + "External id": 14398, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14398, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14398, "pid": 0, "tid": 7, "ts": 1776649790538746, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538738, "dur": 4, + "args": { + "External id": 14398, + "cbid": 211, "correlation": 14398 + } + }, + { + "ph": "s", "id": 14398, "pid": 435, "tid": 435, "ts": 1776649790538738, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790538780, "dur": 1, + "args": { + "External id": 1597, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14416, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14416, "pid": 0, "tid": 7, "ts": 1776649790538780, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538771, "dur": 6, + "args": { + "External id": 1597, + "cbid": 211, "correlation": 14416 + } + }, + { + "ph": "s", "id": 14416, "pid": 435, "tid": 435, "ts": 1776649790538771, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538783, "dur": 0, + "args": { + "External id": 14422, + "cbid": 200, "correlation": 14422 + } + }, + { + "ph": "f", "id": 14422, "pid": 435, "tid": 435, "ts": 1776649790538783, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538784, "dur": 0, + "args": { + "External id": 14423, + "cbid": 273, "correlation": 14423 + } + }, + { + "ph": "f", "id": 14423, "pid": 435, "tid": 435, "ts": 1776649790538784, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790538792, "dur": 15, + "args": { + "External id": 14424, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14424, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14424, "pid": 0, "tid": 7, "ts": 1776649790538792, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538785, "dur": 4, + "args": { + "External id": 14424, + "cbid": 211, "correlation": 14424 + } + }, + { + "ph": "s", "id": 14424, "pid": 435, "tid": 435, "ts": 1776649790538785, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790538827, "dur": 1, + "args": { + "External id": 1602, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14442, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14442, "pid": 0, "tid": 7, "ts": 1776649790538827, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538818, "dur": 6, + "args": { + "External id": 1602, + "cbid": 211, "correlation": 14442 + } + }, + { + "ph": "s", "id": 14442, "pid": 435, "tid": 435, "ts": 1776649790538818, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538830, "dur": 0, + "args": { + "External id": 14448, + "cbid": 200, "correlation": 14448 + } + }, + { + "ph": "f", "id": 14448, "pid": 435, "tid": 435, "ts": 1776649790538830, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538830, "dur": 0, + "args": { + "External id": 14449, + "cbid": 273, "correlation": 14449 + } + }, + { + "ph": "f", "id": 14449, "pid": 435, "tid": 435, "ts": 1776649790538830, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790538839, "dur": 16, + "args": { + "External id": 14450, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14450, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14450, "pid": 0, "tid": 7, "ts": 1776649790538839, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538831, "dur": 4, + "args": { + "External id": 14450, + "cbid": 211, "correlation": 14450 + } + }, + { + "ph": "s", "id": 14450, "pid": 435, "tid": 435, "ts": 1776649790538831, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790538874, "dur": 1, + "args": { + "External id": 1607, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14468, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14468, "pid": 0, "tid": 7, "ts": 1776649790538874, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538865, "dur": 6, + "args": { + "External id": 1607, + "cbid": 211, "correlation": 14468 + } + }, + { + "ph": "s", "id": 14468, "pid": 435, "tid": 435, "ts": 1776649790538865, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538877, "dur": 0, + "args": { + "External id": 14474, + "cbid": 200, "correlation": 14474 + } + }, + { + "ph": "f", "id": 14474, "pid": 435, "tid": 435, "ts": 1776649790538877, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538877, "dur": 0, + "args": { + "External id": 14475, + "cbid": 273, "correlation": 14475 + } + }, + { + "ph": "f", "id": 14475, "pid": 435, "tid": 435, "ts": 1776649790538877, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790538886, "dur": 16, + "args": { + "External id": 14476, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14476, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14476, "pid": 0, "tid": 7, "ts": 1776649790538886, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538879, "dur": 4, + "args": { + "External id": 14476, + "cbid": 211, "correlation": 14476 + } + }, + { + "ph": "s", "id": 14476, "pid": 435, "tid": 435, "ts": 1776649790538879, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790538921, "dur": 1, + "args": { + "External id": 1612, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14494, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14494, "pid": 0, "tid": 7, "ts": 1776649790538921, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538912, "dur": 6, + "args": { + "External id": 1612, + "cbid": 211, "correlation": 14494 + } + }, + { + "ph": "s", "id": 14494, "pid": 435, "tid": 435, "ts": 1776649790538912, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538924, "dur": 0, + "args": { + "External id": 14500, + "cbid": 200, "correlation": 14500 + } + }, + { + "ph": "f", "id": 14500, "pid": 435, "tid": 435, "ts": 1776649790538924, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538924, "dur": 0, + "args": { + "External id": 14501, + "cbid": 273, "correlation": 14501 + } + }, + { + "ph": "f", "id": 14501, "pid": 435, "tid": 435, "ts": 1776649790538924, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790538933, "dur": 15, + "args": { + "External id": 14502, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14502, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14502, "pid": 0, "tid": 7, "ts": 1776649790538933, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538926, "dur": 4, + "args": { + "External id": 14502, + "cbid": 211, "correlation": 14502 + } + }, + { + "ph": "s", "id": 14502, "pid": 435, "tid": 435, "ts": 1776649790538926, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790538968, "dur": 1, + "args": { + "External id": 1617, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14520, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14520, "pid": 0, "tid": 7, "ts": 1776649790538968, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538959, "dur": 6, + "args": { + "External id": 1617, + "cbid": 211, "correlation": 14520 + } + }, + { + "ph": "s", "id": 14520, "pid": 435, "tid": 435, "ts": 1776649790538959, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538971, "dur": 0, + "args": { + "External id": 14526, + "cbid": 200, "correlation": 14526 + } + }, + { + "ph": "f", "id": 14526, "pid": 435, "tid": 435, "ts": 1776649790538971, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790538972, "dur": 0, + "args": { + "External id": 14527, + "cbid": 273, "correlation": 14527 + } + }, + { + "ph": "f", "id": 14527, "pid": 435, "tid": 435, "ts": 1776649790538972, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790538980, "dur": 15, + "args": { + "External id": 14528, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14528, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14528, "pid": 0, "tid": 7, "ts": 1776649790538980, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790538973, "dur": 4, + "args": { + "External id": 14528, + "cbid": 211, "correlation": 14528 + } + }, + { + "ph": "s", "id": 14528, "pid": 435, "tid": 435, "ts": 1776649790538973, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790539015, "dur": 1, + "args": { + "External id": 1622, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14546, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14546, "pid": 0, "tid": 7, "ts": 1776649790539015, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539006, "dur": 6, + "args": { + "External id": 1622, + "cbid": 211, "correlation": 14546 + } + }, + { + "ph": "s", "id": 14546, "pid": 435, "tid": 435, "ts": 1776649790539006, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539018, "dur": 0, + "args": { + "External id": 14552, + "cbid": 200, "correlation": 14552 + } + }, + { + "ph": "f", "id": 14552, "pid": 435, "tid": 435, "ts": 1776649790539018, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539018, "dur": 0, + "args": { + "External id": 14553, + "cbid": 273, "correlation": 14553 + } + }, + { + "ph": "f", "id": 14553, "pid": 435, "tid": 435, "ts": 1776649790539018, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790539027, "dur": 15, + "args": { + "External id": 14554, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14554, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14554, "pid": 0, "tid": 7, "ts": 1776649790539027, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539019, "dur": 4, + "args": { + "External id": 14554, + "cbid": 211, "correlation": 14554 + } + }, + { + "ph": "s", "id": 14554, "pid": 435, "tid": 435, "ts": 1776649790539019, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790539061, "dur": 1, + "args": { + "External id": 1627, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14572, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14572, "pid": 0, "tid": 7, "ts": 1776649790539061, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539052, "dur": 6, + "args": { + "External id": 1627, + "cbid": 211, "correlation": 14572 + } + }, + { + "ph": "s", "id": 14572, "pid": 435, "tid": 435, "ts": 1776649790539052, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539064, "dur": 0, + "args": { + "External id": 14578, + "cbid": 200, "correlation": 14578 + } + }, + { + "ph": "f", "id": 14578, "pid": 435, "tid": 435, "ts": 1776649790539064, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539065, "dur": 0, + "args": { + "External id": 14579, + "cbid": 273, "correlation": 14579 + } + }, + { + "ph": "f", "id": 14579, "pid": 435, "tid": 435, "ts": 1776649790539065, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790539074, "dur": 16, + "args": { + "External id": 14580, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14580, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14580, "pid": 0, "tid": 7, "ts": 1776649790539074, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539066, "dur": 4, + "args": { + "External id": 14580, + "cbid": 211, "correlation": 14580 + } + }, + { + "ph": "s", "id": 14580, "pid": 435, "tid": 435, "ts": 1776649790539066, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790539110, "dur": 1, + "args": { + "External id": 1632, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14598, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14598, "pid": 0, "tid": 7, "ts": 1776649790539110, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539101, "dur": 6, + "args": { + "External id": 1632, + "cbid": 211, "correlation": 14598 + } + }, + { + "ph": "s", "id": 14598, "pid": 435, "tid": 435, "ts": 1776649790539101, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539113, "dur": 0, + "args": { + "External id": 14604, + "cbid": 200, "correlation": 14604 + } + }, + { + "ph": "f", "id": 14604, "pid": 435, "tid": 435, "ts": 1776649790539113, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539114, "dur": 0, + "args": { + "External id": 14605, + "cbid": 273, "correlation": 14605 + } + }, + { + "ph": "f", "id": 14605, "pid": 435, "tid": 435, "ts": 1776649790539114, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790539123, "dur": 15, + "args": { + "External id": 14606, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14606, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14606, "pid": 0, "tid": 7, "ts": 1776649790539123, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539115, "dur": 4, + "args": { + "External id": 14606, + "cbid": 211, "correlation": 14606 + } + }, + { + "ph": "s", "id": 14606, "pid": 435, "tid": 435, "ts": 1776649790539115, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790539157, "dur": 1, + "args": { + "External id": 1637, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14624, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14624, "pid": 0, "tid": 7, "ts": 1776649790539157, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539148, "dur": 6, + "args": { + "External id": 1637, + "cbid": 211, "correlation": 14624 + } + }, + { + "ph": "s", "id": 14624, "pid": 435, "tid": 435, "ts": 1776649790539148, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539160, "dur": 0, + "args": { + "External id": 14630, + "cbid": 200, "correlation": 14630 + } + }, + { + "ph": "f", "id": 14630, "pid": 435, "tid": 435, "ts": 1776649790539160, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539161, "dur": 0, + "args": { + "External id": 14631, + "cbid": 273, "correlation": 14631 + } + }, + { + "ph": "f", "id": 14631, "pid": 435, "tid": 435, "ts": 1776649790539161, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790539170, "dur": 15, + "args": { + "External id": 14632, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14632, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14632, "pid": 0, "tid": 7, "ts": 1776649790539170, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539162, "dur": 4, + "args": { + "External id": 14632, + "cbid": 211, "correlation": 14632 + } + }, + { + "ph": "s", "id": 14632, "pid": 435, "tid": 435, "ts": 1776649790539162, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790539204, "dur": 1, + "args": { + "External id": 1642, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14650, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14650, "pid": 0, "tid": 7, "ts": 1776649790539204, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539195, "dur": 6, + "args": { + "External id": 1642, + "cbid": 211, "correlation": 14650 + } + }, + { + "ph": "s", "id": 14650, "pid": 435, "tid": 435, "ts": 1776649790539195, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539207, "dur": 0, + "args": { + "External id": 14656, + "cbid": 200, "correlation": 14656 + } + }, + { + "ph": "f", "id": 14656, "pid": 435, "tid": 435, "ts": 1776649790539207, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539207, "dur": 0, + "args": { + "External id": 14657, + "cbid": 273, "correlation": 14657 + } + }, + { + "ph": "f", "id": 14657, "pid": 435, "tid": 435, "ts": 1776649790539207, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790539216, "dur": 15, + "args": { + "External id": 14658, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14658, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14658, "pid": 0, "tid": 7, "ts": 1776649790539216, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539209, "dur": 4, + "args": { + "External id": 14658, + "cbid": 211, "correlation": 14658 + } + }, + { + "ph": "s", "id": 14658, "pid": 435, "tid": 435, "ts": 1776649790539209, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790539251, "dur": 1, + "args": { + "External id": 1647, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14676, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14676, "pid": 0, "tid": 7, "ts": 1776649790539251, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539242, "dur": 6, + "args": { + "External id": 1647, + "cbid": 211, "correlation": 14676 + } + }, + { + "ph": "s", "id": 14676, "pid": 435, "tid": 435, "ts": 1776649790539242, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539254, "dur": 0, + "args": { + "External id": 14682, + "cbid": 200, "correlation": 14682 + } + }, + { + "ph": "f", "id": 14682, "pid": 435, "tid": 435, "ts": 1776649790539254, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539254, "dur": 0, + "args": { + "External id": 14683, + "cbid": 273, "correlation": 14683 + } + }, + { + "ph": "f", "id": 14683, "pid": 435, "tid": 435, "ts": 1776649790539254, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790539264, "dur": 17, + "args": { + "External id": 14684, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14684, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14684, "pid": 0, "tid": 7, "ts": 1776649790539264, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539256, "dur": 5, + "args": { + "External id": 14684, + "cbid": 211, "correlation": 14684 + } + }, + { + "ph": "s", "id": 14684, "pid": 435, "tid": 435, "ts": 1776649790539256, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790539298, "dur": 1, + "args": { + "External id": 1652, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14702, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14702, "pid": 0, "tid": 7, "ts": 1776649790539298, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539289, "dur": 6, + "args": { + "External id": 1652, + "cbid": 211, "correlation": 14702 + } + }, + { + "ph": "s", "id": 14702, "pid": 435, "tid": 435, "ts": 1776649790539289, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539301, "dur": 0, + "args": { + "External id": 14708, + "cbid": 200, "correlation": 14708 + } + }, + { + "ph": "f", "id": 14708, "pid": 435, "tid": 435, "ts": 1776649790539301, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539302, "dur": 0, + "args": { + "External id": 14709, + "cbid": 273, "correlation": 14709 + } + }, + { + "ph": "f", "id": 14709, "pid": 435, "tid": 435, "ts": 1776649790539302, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790539311, "dur": 15, + "args": { + "External id": 14710, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14710, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14710, "pid": 0, "tid": 7, "ts": 1776649790539311, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539303, "dur": 4, + "args": { + "External id": 14710, + "cbid": 211, "correlation": 14710 + } + }, + { + "ph": "s", "id": 14710, "pid": 435, "tid": 435, "ts": 1776649790539303, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790539345, "dur": 1, + "args": { + "External id": 1657, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14728, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14728, "pid": 0, "tid": 7, "ts": 1776649790539345, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539336, "dur": 6, + "args": { + "External id": 1657, + "cbid": 211, "correlation": 14728 + } + }, + { + "ph": "s", "id": 14728, "pid": 435, "tid": 435, "ts": 1776649790539336, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539348, "dur": 0, + "args": { + "External id": 14734, + "cbid": 200, "correlation": 14734 + } + }, + { + "ph": "f", "id": 14734, "pid": 435, "tid": 435, "ts": 1776649790539348, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539349, "dur": 0, + "args": { + "External id": 14735, + "cbid": 273, "correlation": 14735 + } + }, + { + "ph": "f", "id": 14735, "pid": 435, "tid": 435, "ts": 1776649790539349, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790539358, "dur": 15, + "args": { + "External id": 14736, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14736, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14736, "pid": 0, "tid": 7, "ts": 1776649790539358, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539350, "dur": 4, + "args": { + "External id": 14736, + "cbid": 211, "correlation": 14736 + } + }, + { + "ph": "s", "id": 14736, "pid": 435, "tid": 435, "ts": 1776649790539350, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790539392, "dur": 1, + "args": { + "External id": 1662, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14754, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14754, "pid": 0, "tid": 7, "ts": 1776649790539392, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539383, "dur": 6, + "args": { + "External id": 1662, + "cbid": 211, "correlation": 14754 + } + }, + { + "ph": "s", "id": 14754, "pid": 435, "tid": 435, "ts": 1776649790539383, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539395, "dur": 0, + "args": { + "External id": 14760, + "cbid": 200, "correlation": 14760 + } + }, + { + "ph": "f", "id": 14760, "pid": 435, "tid": 435, "ts": 1776649790539395, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539396, "dur": 0, + "args": { + "External id": 14761, + "cbid": 273, "correlation": 14761 + } + }, + { + "ph": "f", "id": 14761, "pid": 435, "tid": 435, "ts": 1776649790539396, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790539405, "dur": 15, + "args": { + "External id": 14762, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14762, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14762, "pid": 0, "tid": 7, "ts": 1776649790539405, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539397, "dur": 4, + "args": { + "External id": 14762, + "cbid": 211, "correlation": 14762 + } + }, + { + "ph": "s", "id": 14762, "pid": 435, "tid": 435, "ts": 1776649790539397, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790539439, "dur": 1, + "args": { + "External id": 1667, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14780, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14780, "pid": 0, "tid": 7, "ts": 1776649790539439, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539430, "dur": 6, + "args": { + "External id": 1667, + "cbid": 211, "correlation": 14780 + } + }, + { + "ph": "s", "id": 14780, "pid": 435, "tid": 435, "ts": 1776649790539430, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539442, "dur": 0, + "args": { + "External id": 14786, + "cbid": 200, "correlation": 14786 + } + }, + { + "ph": "f", "id": 14786, "pid": 435, "tid": 435, "ts": 1776649790539442, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539443, "dur": 0, + "args": { + "External id": 14787, + "cbid": 273, "correlation": 14787 + } + }, + { + "ph": "f", "id": 14787, "pid": 435, "tid": 435, "ts": 1776649790539443, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790539452, "dur": 15, + "args": { + "External id": 14788, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14788, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14788, "pid": 0, "tid": 7, "ts": 1776649790539452, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539444, "dur": 4, + "args": { + "External id": 14788, + "cbid": 211, "correlation": 14788 + } + }, + { + "ph": "s", "id": 14788, "pid": 435, "tid": 435, "ts": 1776649790539444, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790539487, "dur": 1, + "args": { + "External id": 1672, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14806, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14806, "pid": 0, "tid": 7, "ts": 1776649790539487, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539478, "dur": 6, + "args": { + "External id": 1672, + "cbid": 211, "correlation": 14806 + } + }, + { + "ph": "s", "id": 14806, "pid": 435, "tid": 435, "ts": 1776649790539478, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539490, "dur": 0, + "args": { + "External id": 14812, + "cbid": 200, "correlation": 14812 + } + }, + { + "ph": "f", "id": 14812, "pid": 435, "tid": 435, "ts": 1776649790539490, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539490, "dur": 0, + "args": { + "External id": 14813, + "cbid": 273, "correlation": 14813 + } + }, + { + "ph": "f", "id": 14813, "pid": 435, "tid": 435, "ts": 1776649790539490, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790539499, "dur": 15, + "args": { + "External id": 14814, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14814, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14814, "pid": 0, "tid": 7, "ts": 1776649790539499, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539491, "dur": 4, + "args": { + "External id": 14814, + "cbid": 211, "correlation": 14814 + } + }, + { + "ph": "s", "id": 14814, "pid": 435, "tid": 435, "ts": 1776649790539491, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790539534, "dur": 1, + "args": { + "External id": 1677, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14832, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14832, "pid": 0, "tid": 7, "ts": 1776649790539534, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539525, "dur": 6, + "args": { + "External id": 1677, + "cbid": 211, "correlation": 14832 + } + }, + { + "ph": "s", "id": 14832, "pid": 435, "tid": 435, "ts": 1776649790539525, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539536, "dur": 0, + "args": { + "External id": 14838, + "cbid": 200, "correlation": 14838 + } + }, + { + "ph": "f", "id": 14838, "pid": 435, "tid": 435, "ts": 1776649790539536, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539537, "dur": 0, + "args": { + "External id": 14839, + "cbid": 273, "correlation": 14839 + } + }, + { + "ph": "f", "id": 14839, "pid": 435, "tid": 435, "ts": 1776649790539537, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790539546, "dur": 15, + "args": { + "External id": 14840, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14840, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14840, "pid": 0, "tid": 7, "ts": 1776649790539546, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539538, "dur": 4, + "args": { + "External id": 14840, + "cbid": 211, "correlation": 14840 + } + }, + { + "ph": "s", "id": 14840, "pid": 435, "tid": 435, "ts": 1776649790539538, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790539582, "dur": 1, + "args": { + "External id": 1682, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14858, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14858, "pid": 0, "tid": 7, "ts": 1776649790539582, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539573, "dur": 6, + "args": { + "External id": 1682, + "cbid": 211, "correlation": 14858 + } + }, + { + "ph": "s", "id": 14858, "pid": 435, "tid": 435, "ts": 1776649790539573, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539584, "dur": 0, + "args": { + "External id": 14864, + "cbid": 200, "correlation": 14864 + } + }, + { + "ph": "f", "id": 14864, "pid": 435, "tid": 435, "ts": 1776649790539584, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539585, "dur": 0, + "args": { + "External id": 14865, + "cbid": 273, "correlation": 14865 + } + }, + { + "ph": "f", "id": 14865, "pid": 435, "tid": 435, "ts": 1776649790539585, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790539594, "dur": 15, + "args": { + "External id": 14866, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14866, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14866, "pid": 0, "tid": 7, "ts": 1776649790539594, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539586, "dur": 4, + "args": { + "External id": 14866, + "cbid": 211, "correlation": 14866 + } + }, + { + "ph": "s", "id": 14866, "pid": 435, "tid": 435, "ts": 1776649790539586, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790539629, "dur": 1, + "args": { + "External id": 1687, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14884, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14884, "pid": 0, "tid": 7, "ts": 1776649790539629, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539620, "dur": 6, + "args": { + "External id": 1687, + "cbid": 211, "correlation": 14884 + } + }, + { + "ph": "s", "id": 14884, "pid": 435, "tid": 435, "ts": 1776649790539620, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539632, "dur": 0, + "args": { + "External id": 14890, + "cbid": 200, "correlation": 14890 + } + }, + { + "ph": "f", "id": 14890, "pid": 435, "tid": 435, "ts": 1776649790539632, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539633, "dur": 0, + "args": { + "External id": 14891, + "cbid": 273, "correlation": 14891 + } + }, + { + "ph": "f", "id": 14891, "pid": 435, "tid": 435, "ts": 1776649790539633, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790539642, "dur": 15, + "args": { + "External id": 14892, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14892, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14892, "pid": 0, "tid": 7, "ts": 1776649790539642, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539634, "dur": 5, + "args": { + "External id": 14892, + "cbid": 211, "correlation": 14892 + } + }, + { + "ph": "s", "id": 14892, "pid": 435, "tid": 435, "ts": 1776649790539634, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790539676, "dur": 1, + "args": { + "External id": 1692, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14910, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14910, "pid": 0, "tid": 7, "ts": 1776649790539676, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539667, "dur": 6, + "args": { + "External id": 1692, + "cbid": 211, "correlation": 14910 + } + }, + { + "ph": "s", "id": 14910, "pid": 435, "tid": 435, "ts": 1776649790539667, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539679, "dur": 0, + "args": { + "External id": 14916, + "cbid": 200, "correlation": 14916 + } + }, + { + "ph": "f", "id": 14916, "pid": 435, "tid": 435, "ts": 1776649790539679, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539680, "dur": 0, + "args": { + "External id": 14917, + "cbid": 273, "correlation": 14917 + } + }, + { + "ph": "f", "id": 14917, "pid": 435, "tid": 435, "ts": 1776649790539680, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790539689, "dur": 16, + "args": { + "External id": 14918, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14918, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14918, "pid": 0, "tid": 7, "ts": 1776649790539689, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539681, "dur": 4, + "args": { + "External id": 14918, + "cbid": 211, "correlation": 14918 + } + }, + { + "ph": "s", "id": 14918, "pid": 435, "tid": 435, "ts": 1776649790539681, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790539723, "dur": 1, + "args": { + "External id": 1697, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14936, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14936, "pid": 0, "tid": 7, "ts": 1776649790539723, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539714, "dur": 6, + "args": { + "External id": 1697, + "cbid": 211, "correlation": 14936 + } + }, + { + "ph": "s", "id": 14936, "pid": 435, "tid": 435, "ts": 1776649790539714, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539726, "dur": 0, + "args": { + "External id": 14942, + "cbid": 200, "correlation": 14942 + } + }, + { + "ph": "f", "id": 14942, "pid": 435, "tid": 435, "ts": 1776649790539726, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539727, "dur": 0, + "args": { + "External id": 14943, + "cbid": 273, "correlation": 14943 + } + }, + { + "ph": "f", "id": 14943, "pid": 435, "tid": 435, "ts": 1776649790539727, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790539736, "dur": 15, + "args": { + "External id": 14944, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14944, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14944, "pid": 0, "tid": 7, "ts": 1776649790539736, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539728, "dur": 4, + "args": { + "External id": 14944, + "cbid": 211, "correlation": 14944 + } + }, + { + "ph": "s", "id": 14944, "pid": 435, "tid": 435, "ts": 1776649790539728, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790539771, "dur": 1, + "args": { + "External id": 1702, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14962, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14962, "pid": 0, "tid": 7, "ts": 1776649790539771, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539762, "dur": 6, + "args": { + "External id": 1702, + "cbid": 211, "correlation": 14962 + } + }, + { + "ph": "s", "id": 14962, "pid": 435, "tid": 435, "ts": 1776649790539762, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539774, "dur": 0, + "args": { + "External id": 14968, + "cbid": 200, "correlation": 14968 + } + }, + { + "ph": "f", "id": 14968, "pid": 435, "tid": 435, "ts": 1776649790539774, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539774, "dur": 0, + "args": { + "External id": 14969, + "cbid": 273, "correlation": 14969 + } + }, + { + "ph": "f", "id": 14969, "pid": 435, "tid": 435, "ts": 1776649790539774, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790539783, "dur": 15, + "args": { + "External id": 14970, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14970, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14970, "pid": 0, "tid": 7, "ts": 1776649790539783, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539775, "dur": 5, + "args": { + "External id": 14970, + "cbid": 211, "correlation": 14970 + } + }, + { + "ph": "s", "id": 14970, "pid": 435, "tid": 435, "ts": 1776649790539775, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790539818, "dur": 1, + "args": { + "External id": 1707, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14988, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14988, "pid": 0, "tid": 7, "ts": 1776649790539818, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539809, "dur": 6, + "args": { + "External id": 1707, + "cbid": 211, "correlation": 14988 + } + }, + { + "ph": "s", "id": 14988, "pid": 435, "tid": 435, "ts": 1776649790539809, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539821, "dur": 0, + "args": { + "External id": 14994, + "cbid": 200, "correlation": 14994 + } + }, + { + "ph": "f", "id": 14994, "pid": 435, "tid": 435, "ts": 1776649790539821, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539821, "dur": 0, + "args": { + "External id": 14995, + "cbid": 273, "correlation": 14995 + } + }, + { + "ph": "f", "id": 14995, "pid": 435, "tid": 435, "ts": 1776649790539821, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790539831, "dur": 15, + "args": { + "External id": 14996, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 14996, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 14996, "pid": 0, "tid": 7, "ts": 1776649790539831, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539823, "dur": 4, + "args": { + "External id": 14996, + "cbid": 211, "correlation": 14996 + } + }, + { + "ph": "s", "id": 14996, "pid": 435, "tid": 435, "ts": 1776649790539823, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790539865, "dur": 1, + "args": { + "External id": 1712, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15014, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15014, "pid": 0, "tid": 7, "ts": 1776649790539865, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539856, "dur": 6, + "args": { + "External id": 1712, + "cbid": 211, "correlation": 15014 + } + }, + { + "ph": "s", "id": 15014, "pid": 435, "tid": 435, "ts": 1776649790539856, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539868, "dur": 0, + "args": { + "External id": 15020, + "cbid": 200, "correlation": 15020 + } + }, + { + "ph": "f", "id": 15020, "pid": 435, "tid": 435, "ts": 1776649790539868, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539868, "dur": 0, + "args": { + "External id": 15021, + "cbid": 273, "correlation": 15021 + } + }, + { + "ph": "f", "id": 15021, "pid": 435, "tid": 435, "ts": 1776649790539868, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790539877, "dur": 15, + "args": { + "External id": 15022, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15022, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15022, "pid": 0, "tid": 7, "ts": 1776649790539877, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539869, "dur": 4, + "args": { + "External id": 15022, + "cbid": 211, "correlation": 15022 + } + }, + { + "ph": "s", "id": 15022, "pid": 435, "tid": 435, "ts": 1776649790539869, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790539912, "dur": 1, + "args": { + "External id": 1717, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15040, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15040, "pid": 0, "tid": 7, "ts": 1776649790539912, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539902, "dur": 6, + "args": { + "External id": 1717, + "cbid": 211, "correlation": 15040 + } + }, + { + "ph": "s", "id": 15040, "pid": 435, "tid": 435, "ts": 1776649790539902, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539915, "dur": 0, + "args": { + "External id": 15046, + "cbid": 200, "correlation": 15046 + } + }, + { + "ph": "f", "id": 15046, "pid": 435, "tid": 435, "ts": 1776649790539915, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539915, "dur": 0, + "args": { + "External id": 15047, + "cbid": 273, "correlation": 15047 + } + }, + { + "ph": "f", "id": 15047, "pid": 435, "tid": 435, "ts": 1776649790539915, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790539924, "dur": 15, + "args": { + "External id": 15048, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15048, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15048, "pid": 0, "tid": 7, "ts": 1776649790539924, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539917, "dur": 5, + "args": { + "External id": 15048, + "cbid": 211, "correlation": 15048 + } + }, + { + "ph": "s", "id": 15048, "pid": 435, "tid": 435, "ts": 1776649790539917, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790539959, "dur": 1, + "args": { + "External id": 1722, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15066, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15066, "pid": 0, "tid": 7, "ts": 1776649790539959, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539950, "dur": 6, + "args": { + "External id": 1722, + "cbid": 211, "correlation": 15066 + } + }, + { + "ph": "s", "id": 15066, "pid": 435, "tid": 435, "ts": 1776649790539950, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539962, "dur": 0, + "args": { + "External id": 15072, + "cbid": 200, "correlation": 15072 + } + }, + { + "ph": "f", "id": 15072, "pid": 435, "tid": 435, "ts": 1776649790539962, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790539963, "dur": 0, + "args": { + "External id": 15073, + "cbid": 273, "correlation": 15073 + } + }, + { + "ph": "f", "id": 15073, "pid": 435, "tid": 435, "ts": 1776649790539963, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790539972, "dur": 15, + "args": { + "External id": 15074, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15074, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15074, "pid": 0, "tid": 7, "ts": 1776649790539972, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539964, "dur": 4, + "args": { + "External id": 15074, + "cbid": 211, "correlation": 15074 + } + }, + { + "ph": "s", "id": 15074, "pid": 435, "tid": 435, "ts": 1776649790539964, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790540006, "dur": 1, + "args": { + "External id": 1727, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15092, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15092, "pid": 0, "tid": 7, "ts": 1776649790540006, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790539997, "dur": 6, + "args": { + "External id": 1727, + "cbid": 211, "correlation": 15092 + } + }, + { + "ph": "s", "id": 15092, "pid": 435, "tid": 435, "ts": 1776649790539997, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540010, "dur": 0, + "args": { + "External id": 15098, + "cbid": 200, "correlation": 15098 + } + }, + { + "ph": "f", "id": 15098, "pid": 435, "tid": 435, "ts": 1776649790540010, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540010, "dur": 0, + "args": { + "External id": 15099, + "cbid": 273, "correlation": 15099 + } + }, + { + "ph": "f", "id": 15099, "pid": 435, "tid": 435, "ts": 1776649790540010, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790540019, "dur": 15, + "args": { + "External id": 15100, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15100, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15100, "pid": 0, "tid": 7, "ts": 1776649790540019, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540011, "dur": 4, + "args": { + "External id": 15100, + "cbid": 211, "correlation": 15100 + } + }, + { + "ph": "s", "id": 15100, "pid": 435, "tid": 435, "ts": 1776649790540011, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790540054, "dur": 1, + "args": { + "External id": 1732, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15118, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15118, "pid": 0, "tid": 7, "ts": 1776649790540054, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540045, "dur": 6, + "args": { + "External id": 1732, + "cbid": 211, "correlation": 15118 + } + }, + { + "ph": "s", "id": 15118, "pid": 435, "tid": 435, "ts": 1776649790540045, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540058, "dur": 0, + "args": { + "External id": 15124, + "cbid": 200, "correlation": 15124 + } + }, + { + "ph": "f", "id": 15124, "pid": 435, "tid": 435, "ts": 1776649790540058, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540058, "dur": 0, + "args": { + "External id": 15125, + "cbid": 273, "correlation": 15125 + } + }, + { + "ph": "f", "id": 15125, "pid": 435, "tid": 435, "ts": 1776649790540058, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790540068, "dur": 16, + "args": { + "External id": 15126, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15126, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15126, "pid": 0, "tid": 7, "ts": 1776649790540068, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540059, "dur": 5, + "args": { + "External id": 15126, + "cbid": 211, "correlation": 15126 + } + }, + { + "ph": "s", "id": 15126, "pid": 435, "tid": 435, "ts": 1776649790540059, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790540102, "dur": 1, + "args": { + "External id": 1737, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15144, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15144, "pid": 0, "tid": 7, "ts": 1776649790540102, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540093, "dur": 6, + "args": { + "External id": 1737, + "cbid": 211, "correlation": 15144 + } + }, + { + "ph": "s", "id": 15144, "pid": 435, "tid": 435, "ts": 1776649790540093, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540105, "dur": 0, + "args": { + "External id": 15150, + "cbid": 200, "correlation": 15150 + } + }, + { + "ph": "f", "id": 15150, "pid": 435, "tid": 435, "ts": 1776649790540105, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540106, "dur": 0, + "args": { + "External id": 15151, + "cbid": 273, "correlation": 15151 + } + }, + { + "ph": "f", "id": 15151, "pid": 435, "tid": 435, "ts": 1776649790540106, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790540115, "dur": 17, + "args": { + "External id": 15152, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15152, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15152, "pid": 0, "tid": 7, "ts": 1776649790540115, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540107, "dur": 4, + "args": { + "External id": 15152, + "cbid": 211, "correlation": 15152 + } + }, + { + "ph": "s", "id": 15152, "pid": 435, "tid": 435, "ts": 1776649790540107, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790540150, "dur": 1, + "args": { + "External id": 1742, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15170, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15170, "pid": 0, "tid": 7, "ts": 1776649790540150, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540141, "dur": 6, + "args": { + "External id": 1742, + "cbid": 211, "correlation": 15170 + } + }, + { + "ph": "s", "id": 15170, "pid": 435, "tid": 435, "ts": 1776649790540141, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540153, "dur": 0, + "args": { + "External id": 15176, + "cbid": 200, "correlation": 15176 + } + }, + { + "ph": "f", "id": 15176, "pid": 435, "tid": 435, "ts": 1776649790540153, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540154, "dur": 0, + "args": { + "External id": 15177, + "cbid": 273, "correlation": 15177 + } + }, + { + "ph": "f", "id": 15177, "pid": 435, "tid": 435, "ts": 1776649790540154, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790540163, "dur": 15, + "args": { + "External id": 15178, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15178, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15178, "pid": 0, "tid": 7, "ts": 1776649790540163, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540155, "dur": 4, + "args": { + "External id": 15178, + "cbid": 211, "correlation": 15178 + } + }, + { + "ph": "s", "id": 15178, "pid": 435, "tid": 435, "ts": 1776649790540155, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790540198, "dur": 1, + "args": { + "External id": 1747, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15196, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15196, "pid": 0, "tid": 7, "ts": 1776649790540198, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540189, "dur": 6, + "args": { + "External id": 1747, + "cbid": 211, "correlation": 15196 + } + }, + { + "ph": "s", "id": 15196, "pid": 435, "tid": 435, "ts": 1776649790540189, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540201, "dur": 0, + "args": { + "External id": 15202, + "cbid": 200, "correlation": 15202 + } + }, + { + "ph": "f", "id": 15202, "pid": 435, "tid": 435, "ts": 1776649790540201, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540202, "dur": 0, + "args": { + "External id": 15203, + "cbid": 273, "correlation": 15203 + } + }, + { + "ph": "f", "id": 15203, "pid": 435, "tid": 435, "ts": 1776649790540202, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790540210, "dur": 15, + "args": { + "External id": 15204, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15204, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15204, "pid": 0, "tid": 7, "ts": 1776649790540210, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540203, "dur": 5, + "args": { + "External id": 15204, + "cbid": 211, "correlation": 15204 + } + }, + { + "ph": "s", "id": 15204, "pid": 435, "tid": 435, "ts": 1776649790540203, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790540269, "dur": 1, + "args": { + "External id": 1752, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15222, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15222, "pid": 0, "tid": 7, "ts": 1776649790540269, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540259, "dur": 7, + "args": { + "External id": 1752, + "cbid": 211, "correlation": 15222 + } + }, + { + "ph": "s", "id": 15222, "pid": 435, "tid": 435, "ts": 1776649790540259, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540273, "dur": 0, + "args": { + "External id": 15228, + "cbid": 200, "correlation": 15228 + } + }, + { + "ph": "f", "id": 15228, "pid": 435, "tid": 435, "ts": 1776649790540273, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540274, "dur": 0, + "args": { + "External id": 15229, + "cbid": 273, "correlation": 15229 + } + }, + { + "ph": "f", "id": 15229, "pid": 435, "tid": 435, "ts": 1776649790540274, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790540283, "dur": 15, + "args": { + "External id": 15230, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15230, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15230, "pid": 0, "tid": 7, "ts": 1776649790540283, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540275, "dur": 5, + "args": { + "External id": 15230, + "cbid": 211, "correlation": 15230 + } + }, + { + "ph": "s", "id": 15230, "pid": 435, "tid": 435, "ts": 1776649790540275, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790540319, "dur": 1, + "args": { + "External id": 1757, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15248, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15248, "pid": 0, "tid": 7, "ts": 1776649790540319, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540310, "dur": 6, + "args": { + "External id": 1757, + "cbid": 211, "correlation": 15248 + } + }, + { + "ph": "s", "id": 15248, "pid": 435, "tid": 435, "ts": 1776649790540310, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540323, "dur": 0, + "args": { + "External id": 15254, + "cbid": 200, "correlation": 15254 + } + }, + { + "ph": "f", "id": 15254, "pid": 435, "tid": 435, "ts": 1776649790540323, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540323, "dur": 0, + "args": { + "External id": 15255, + "cbid": 273, "correlation": 15255 + } + }, + { + "ph": "f", "id": 15255, "pid": 435, "tid": 435, "ts": 1776649790540323, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790540332, "dur": 15, + "args": { + "External id": 15256, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15256, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15256, "pid": 0, "tid": 7, "ts": 1776649790540332, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540324, "dur": 5, + "args": { + "External id": 15256, + "cbid": 211, "correlation": 15256 + } + }, + { + "ph": "s", "id": 15256, "pid": 435, "tid": 435, "ts": 1776649790540324, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790540368, "dur": 1, + "args": { + "External id": 1762, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15274, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15274, "pid": 0, "tid": 7, "ts": 1776649790540368, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540359, "dur": 6, + "args": { + "External id": 1762, + "cbid": 211, "correlation": 15274 + } + }, + { + "ph": "s", "id": 15274, "pid": 435, "tid": 435, "ts": 1776649790540359, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540371, "dur": 0, + "args": { + "External id": 15280, + "cbid": 200, "correlation": 15280 + } + }, + { + "ph": "f", "id": 15280, "pid": 435, "tid": 435, "ts": 1776649790540371, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540372, "dur": 0, + "args": { + "External id": 15281, + "cbid": 273, "correlation": 15281 + } + }, + { + "ph": "f", "id": 15281, "pid": 435, "tid": 435, "ts": 1776649790540372, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790540381, "dur": 15, + "args": { + "External id": 15282, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15282, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15282, "pid": 0, "tid": 7, "ts": 1776649790540381, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540373, "dur": 4, + "args": { + "External id": 15282, + "cbid": 211, "correlation": 15282 + } + }, + { + "ph": "s", "id": 15282, "pid": 435, "tid": 435, "ts": 1776649790540373, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790540415, "dur": 1, + "args": { + "External id": 1767, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15300, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15300, "pid": 0, "tid": 7, "ts": 1776649790540415, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540406, "dur": 6, + "args": { + "External id": 1767, + "cbid": 211, "correlation": 15300 + } + }, + { + "ph": "s", "id": 15300, "pid": 435, "tid": 435, "ts": 1776649790540406, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540419, "dur": 0, + "args": { + "External id": 15306, + "cbid": 200, "correlation": 15306 + } + }, + { + "ph": "f", "id": 15306, "pid": 435, "tid": 435, "ts": 1776649790540419, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540419, "dur": 0, + "args": { + "External id": 15307, + "cbid": 273, "correlation": 15307 + } + }, + { + "ph": "f", "id": 15307, "pid": 435, "tid": 435, "ts": 1776649790540419, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790540428, "dur": 15, + "args": { + "External id": 15308, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15308, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15308, "pid": 0, "tid": 7, "ts": 1776649790540428, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540420, "dur": 4, + "args": { + "External id": 15308, + "cbid": 211, "correlation": 15308 + } + }, + { + "ph": "s", "id": 15308, "pid": 435, "tid": 435, "ts": 1776649790540420, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790540464, "dur": 1, + "args": { + "External id": 1772, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15326, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15326, "pid": 0, "tid": 7, "ts": 1776649790540464, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540454, "dur": 6, + "args": { + "External id": 1772, + "cbid": 211, "correlation": 15326 + } + }, + { + "ph": "s", "id": 15326, "pid": 435, "tid": 435, "ts": 1776649790540454, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540467, "dur": 0, + "args": { + "External id": 15332, + "cbid": 200, "correlation": 15332 + } + }, + { + "ph": "f", "id": 15332, "pid": 435, "tid": 435, "ts": 1776649790540467, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540467, "dur": 0, + "args": { + "External id": 15333, + "cbid": 273, "correlation": 15333 + } + }, + { + "ph": "f", "id": 15333, "pid": 435, "tid": 435, "ts": 1776649790540467, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790540477, "dur": 15, + "args": { + "External id": 15334, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15334, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15334, "pid": 0, "tid": 7, "ts": 1776649790540477, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540469, "dur": 4, + "args": { + "External id": 15334, + "cbid": 211, "correlation": 15334 + } + }, + { + "ph": "s", "id": 15334, "pid": 435, "tid": 435, "ts": 1776649790540469, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790540511, "dur": 1, + "args": { + "External id": 1777, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15352, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15352, "pid": 0, "tid": 7, "ts": 1776649790540511, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540502, "dur": 6, + "args": { + "External id": 1777, + "cbid": 211, "correlation": 15352 + } + }, + { + "ph": "s", "id": 15352, "pid": 435, "tid": 435, "ts": 1776649790540502, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540514, "dur": 0, + "args": { + "External id": 15358, + "cbid": 200, "correlation": 15358 + } + }, + { + "ph": "f", "id": 15358, "pid": 435, "tid": 435, "ts": 1776649790540514, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540515, "dur": 0, + "args": { + "External id": 15359, + "cbid": 273, "correlation": 15359 + } + }, + { + "ph": "f", "id": 15359, "pid": 435, "tid": 435, "ts": 1776649790540515, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790540524, "dur": 16, + "args": { + "External id": 15360, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15360, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15360, "pid": 0, "tid": 7, "ts": 1776649790540524, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540516, "dur": 4, + "args": { + "External id": 15360, + "cbid": 211, "correlation": 15360 + } + }, + { + "ph": "s", "id": 15360, "pid": 435, "tid": 435, "ts": 1776649790540516, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790540546, "dur": 3, + "args": { + "External id": 15368, + "cbid": 135, "correlation": 15368 + } + }, + { + "ph": "f", "id": 15368, "pid": 435, "tid": 435, "ts": 1776649790540546, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790540561, "dur": 5, + "args": { + "External id": 15373, + "cbid": 165, "correlation": 15373 + } + }, + { + "ph": "s", "id": 15373, "pid": 435, "tid": 435, "ts": 1776649790540561, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649790540570, "dur": 4, + "args": { + "External id": 15376, + "cbid": 139, "correlation": 15376 + } + }, + { + "ph": "f", "id": 15376, "pid": 435, "tid": 435, "ts": 1776649790540570, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790540625, "dur": 1, + "args": { + "External id": 1782, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15404, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15404, "pid": 0, "tid": 7, "ts": 1776649790540625, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540614, "dur": 8, + "args": { + "External id": 1782, + "cbid": 211, "correlation": 15404 + } + }, + { + "ph": "s", "id": 15404, "pid": 435, "tid": 435, "ts": 1776649790540614, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540630, "dur": 0, + "args": { + "External id": 15410, + "cbid": 200, "correlation": 15410 + } + }, + { + "ph": "f", "id": 15410, "pid": 435, "tid": 435, "ts": 1776649790540630, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540631, "dur": 0, + "args": { + "External id": 15411, + "cbid": 273, "correlation": 15411 + } + }, + { + "ph": "f", "id": 15411, "pid": 435, "tid": 435, "ts": 1776649790540631, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790540640, "dur": 15, + "args": { + "External id": 15412, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15412, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15412, "pid": 0, "tid": 7, "ts": 1776649790540640, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540632, "dur": 5, + "args": { + "External id": 15412, + "cbid": 211, "correlation": 15412 + } + }, + { + "ph": "s", "id": 15412, "pid": 435, "tid": 435, "ts": 1776649790540632, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790540678, "dur": 1, + "args": { + "External id": 1787, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15430, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15430, "pid": 0, "tid": 7, "ts": 1776649790540678, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540669, "dur": 6, + "args": { + "External id": 1787, + "cbid": 211, "correlation": 15430 + } + }, + { + "ph": "s", "id": 15430, "pid": 435, "tid": 435, "ts": 1776649790540669, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540681, "dur": 0, + "args": { + "External id": 15436, + "cbid": 200, "correlation": 15436 + } + }, + { + "ph": "f", "id": 15436, "pid": 435, "tid": 435, "ts": 1776649790540681, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540682, "dur": 0, + "args": { + "External id": 15437, + "cbid": 273, "correlation": 15437 + } + }, + { + "ph": "f", "id": 15437, "pid": 435, "tid": 435, "ts": 1776649790540682, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790540691, "dur": 15, + "args": { + "External id": 15438, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15438, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15438, "pid": 0, "tid": 7, "ts": 1776649790540691, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540683, "dur": 4, + "args": { + "External id": 15438, + "cbid": 211, "correlation": 15438 + } + }, + { + "ph": "s", "id": 15438, "pid": 435, "tid": 435, "ts": 1776649790540683, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790540728, "dur": 1, + "args": { + "External id": 1792, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15456, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15456, "pid": 0, "tid": 7, "ts": 1776649790540728, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540718, "dur": 6, + "args": { + "External id": 1792, + "cbid": 211, "correlation": 15456 + } + }, + { + "ph": "s", "id": 15456, "pid": 435, "tid": 435, "ts": 1776649790540718, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540731, "dur": 0, + "args": { + "External id": 15462, + "cbid": 200, "correlation": 15462 + } + }, + { + "ph": "f", "id": 15462, "pid": 435, "tid": 435, "ts": 1776649790540731, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540731, "dur": 0, + "args": { + "External id": 15463, + "cbid": 273, "correlation": 15463 + } + }, + { + "ph": "f", "id": 15463, "pid": 435, "tid": 435, "ts": 1776649790540731, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790540740, "dur": 15, + "args": { + "External id": 15464, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15464, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15464, "pid": 0, "tid": 7, "ts": 1776649790540740, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540733, "dur": 5, + "args": { + "External id": 15464, + "cbid": 211, "correlation": 15464 + } + }, + { + "ph": "s", "id": 15464, "pid": 435, "tid": 435, "ts": 1776649790540733, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790540777, "dur": 1, + "args": { + "External id": 1797, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15482, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15482, "pid": 0, "tid": 7, "ts": 1776649790540777, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540767, "dur": 6, + "args": { + "External id": 1797, + "cbid": 211, "correlation": 15482 + } + }, + { + "ph": "s", "id": 15482, "pid": 435, "tid": 435, "ts": 1776649790540767, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540780, "dur": 0, + "args": { + "External id": 15488, + "cbid": 200, "correlation": 15488 + } + }, + { + "ph": "f", "id": 15488, "pid": 435, "tid": 435, "ts": 1776649790540780, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540781, "dur": 0, + "args": { + "External id": 15489, + "cbid": 273, "correlation": 15489 + } + }, + { + "ph": "f", "id": 15489, "pid": 435, "tid": 435, "ts": 1776649790540781, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790540790, "dur": 15, + "args": { + "External id": 15490, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15490, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15490, "pid": 0, "tid": 7, "ts": 1776649790540790, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540782, "dur": 4, + "args": { + "External id": 15490, + "cbid": 211, "correlation": 15490 + } + }, + { + "ph": "s", "id": 15490, "pid": 435, "tid": 435, "ts": 1776649790540782, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790540825, "dur": 1, + "args": { + "External id": 1802, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15508, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15508, "pid": 0, "tid": 7, "ts": 1776649790540825, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540816, "dur": 6, + "args": { + "External id": 1802, + "cbid": 211, "correlation": 15508 + } + }, + { + "ph": "s", "id": 15508, "pid": 435, "tid": 435, "ts": 1776649790540816, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540828, "dur": 0, + "args": { + "External id": 15514, + "cbid": 200, "correlation": 15514 + } + }, + { + "ph": "f", "id": 15514, "pid": 435, "tid": 435, "ts": 1776649790540828, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540829, "dur": 0, + "args": { + "External id": 15515, + "cbid": 273, "correlation": 15515 + } + }, + { + "ph": "f", "id": 15515, "pid": 435, "tid": 435, "ts": 1776649790540829, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790540838, "dur": 15, + "args": { + "External id": 15516, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15516, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15516, "pid": 0, "tid": 7, "ts": 1776649790540838, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540830, "dur": 4, + "args": { + "External id": 15516, + "cbid": 211, "correlation": 15516 + } + }, + { + "ph": "s", "id": 15516, "pid": 435, "tid": 435, "ts": 1776649790540830, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790540873, "dur": 1, + "args": { + "External id": 1807, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15534, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15534, "pid": 0, "tid": 7, "ts": 1776649790540873, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540864, "dur": 6, + "args": { + "External id": 1807, + "cbid": 211, "correlation": 15534 + } + }, + { + "ph": "s", "id": 15534, "pid": 435, "tid": 435, "ts": 1776649790540864, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540876, "dur": 0, + "args": { + "External id": 15540, + "cbid": 200, "correlation": 15540 + } + }, + { + "ph": "f", "id": 15540, "pid": 435, "tid": 435, "ts": 1776649790540876, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540877, "dur": 0, + "args": { + "External id": 15541, + "cbid": 273, "correlation": 15541 + } + }, + { + "ph": "f", "id": 15541, "pid": 435, "tid": 435, "ts": 1776649790540877, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790540886, "dur": 15, + "args": { + "External id": 15542, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15542, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15542, "pid": 0, "tid": 7, "ts": 1776649790540886, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540878, "dur": 4, + "args": { + "External id": 15542, + "cbid": 211, "correlation": 15542 + } + }, + { + "ph": "s", "id": 15542, "pid": 435, "tid": 435, "ts": 1776649790540878, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790540922, "dur": 1, + "args": { + "External id": 1812, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15560, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15560, "pid": 0, "tid": 7, "ts": 1776649790540922, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540912, "dur": 6, + "args": { + "External id": 1812, + "cbid": 211, "correlation": 15560 + } + }, + { + "ph": "s", "id": 15560, "pid": 435, "tid": 435, "ts": 1776649790540912, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540925, "dur": 0, + "args": { + "External id": 15566, + "cbid": 200, "correlation": 15566 + } + }, + { + "ph": "f", "id": 15566, "pid": 435, "tid": 435, "ts": 1776649790540925, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540925, "dur": 0, + "args": { + "External id": 15567, + "cbid": 273, "correlation": 15567 + } + }, + { + "ph": "f", "id": 15567, "pid": 435, "tid": 435, "ts": 1776649790540925, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790540934, "dur": 15, + "args": { + "External id": 15568, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15568, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15568, "pid": 0, "tid": 7, "ts": 1776649790540934, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540926, "dur": 5, + "args": { + "External id": 15568, + "cbid": 211, "correlation": 15568 + } + }, + { + "ph": "s", "id": 15568, "pid": 435, "tid": 435, "ts": 1776649790540926, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790540970, "dur": 1, + "args": { + "External id": 1817, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15586, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15586, "pid": 0, "tid": 7, "ts": 1776649790540970, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540961, "dur": 6, + "args": { + "External id": 1817, + "cbid": 211, "correlation": 15586 + } + }, + { + "ph": "s", "id": 15586, "pid": 435, "tid": 435, "ts": 1776649790540961, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540973, "dur": 0, + "args": { + "External id": 15592, + "cbid": 200, "correlation": 15592 + } + }, + { + "ph": "f", "id": 15592, "pid": 435, "tid": 435, "ts": 1776649790540973, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790540974, "dur": 0, + "args": { + "External id": 15593, + "cbid": 273, "correlation": 15593 + } + }, + { + "ph": "f", "id": 15593, "pid": 435, "tid": 435, "ts": 1776649790540974, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790540985, "dur": 16, + "args": { + "External id": 15594, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15594, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15594, "pid": 0, "tid": 7, "ts": 1776649790540985, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790540975, "dur": 6, + "args": { + "External id": 15594, + "cbid": 211, "correlation": 15594 + } + }, + { + "ph": "s", "id": 15594, "pid": 435, "tid": 435, "ts": 1776649790540975, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790541021, "dur": 1, + "args": { + "External id": 1822, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15612, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15612, "pid": 0, "tid": 7, "ts": 1776649790541021, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541011, "dur": 6, + "args": { + "External id": 1822, + "cbid": 211, "correlation": 15612 + } + }, + { + "ph": "s", "id": 15612, "pid": 435, "tid": 435, "ts": 1776649790541011, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541024, "dur": 0, + "args": { + "External id": 15618, + "cbid": 200, "correlation": 15618 + } + }, + { + "ph": "f", "id": 15618, "pid": 435, "tid": 435, "ts": 1776649790541024, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541025, "dur": 0, + "args": { + "External id": 15619, + "cbid": 273, "correlation": 15619 + } + }, + { + "ph": "f", "id": 15619, "pid": 435, "tid": 435, "ts": 1776649790541025, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790541034, "dur": 16, + "args": { + "External id": 15620, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15620, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15620, "pid": 0, "tid": 7, "ts": 1776649790541034, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541026, "dur": 5, + "args": { + "External id": 15620, + "cbid": 211, "correlation": 15620 + } + }, + { + "ph": "s", "id": 15620, "pid": 435, "tid": 435, "ts": 1776649790541026, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790541069, "dur": 1, + "args": { + "External id": 1827, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15638, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15638, "pid": 0, "tid": 7, "ts": 1776649790541069, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541060, "dur": 6, + "args": { + "External id": 1827, + "cbid": 211, "correlation": 15638 + } + }, + { + "ph": "s", "id": 15638, "pid": 435, "tid": 435, "ts": 1776649790541060, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541072, "dur": 0, + "args": { + "External id": 15644, + "cbid": 200, "correlation": 15644 + } + }, + { + "ph": "f", "id": 15644, "pid": 435, "tid": 435, "ts": 1776649790541072, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541073, "dur": 0, + "args": { + "External id": 15645, + "cbid": 273, "correlation": 15645 + } + }, + { + "ph": "f", "id": 15645, "pid": 435, "tid": 435, "ts": 1776649790541073, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790541082, "dur": 15, + "args": { + "External id": 15646, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15646, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15646, "pid": 0, "tid": 7, "ts": 1776649790541082, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541074, "dur": 4, + "args": { + "External id": 15646, + "cbid": 211, "correlation": 15646 + } + }, + { + "ph": "s", "id": 15646, "pid": 435, "tid": 435, "ts": 1776649790541074, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790541096, "dur": 6, + "args": { + "External id": 15649, + "cbid": 165, "correlation": 15649 + } + }, + { + "ph": "s", "id": 15649, "pid": 435, "tid": 435, "ts": 1776649790541096, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790541123, "dur": 3, + "args": { + "External id": 15658, + "cbid": 135, "correlation": 15658 + } + }, + { + "ph": "f", "id": 15658, "pid": 435, "tid": 435, "ts": 1776649790541123, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790541172, "dur": 1, + "args": { + "External id": 1832, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15678, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15678, "pid": 0, "tid": 7, "ts": 1776649790541172, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541161, "dur": 8, + "args": { + "External id": 1832, + "cbid": 211, "correlation": 15678 + } + }, + { + "ph": "s", "id": 15678, "pid": 435, "tid": 435, "ts": 1776649790541161, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541176, "dur": 0, + "args": { + "External id": 15684, + "cbid": 200, "correlation": 15684 + } + }, + { + "ph": "f", "id": 15684, "pid": 435, "tid": 435, "ts": 1776649790541176, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541177, "dur": 0, + "args": { + "External id": 15685, + "cbid": 273, "correlation": 15685 + } + }, + { + "ph": "f", "id": 15685, "pid": 435, "tid": 435, "ts": 1776649790541177, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790541187, "dur": 15, + "args": { + "External id": 15686, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15686, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15686, "pid": 0, "tid": 7, "ts": 1776649790541187, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541178, "dur": 5, + "args": { + "External id": 15686, + "cbid": 211, "correlation": 15686 + } + }, + { + "ph": "s", "id": 15686, "pid": 435, "tid": 435, "ts": 1776649790541178, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790541224, "dur": 1, + "args": { + "External id": 1837, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15704, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15704, "pid": 0, "tid": 7, "ts": 1776649790541224, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541215, "dur": 7, + "args": { + "External id": 1837, + "cbid": 211, "correlation": 15704 + } + }, + { + "ph": "s", "id": 15704, "pid": 435, "tid": 435, "ts": 1776649790541215, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541227, "dur": 0, + "args": { + "External id": 15710, + "cbid": 200, "correlation": 15710 + } + }, + { + "ph": "f", "id": 15710, "pid": 435, "tid": 435, "ts": 1776649790541227, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541228, "dur": 0, + "args": { + "External id": 15711, + "cbid": 273, "correlation": 15711 + } + }, + { + "ph": "f", "id": 15711, "pid": 435, "tid": 435, "ts": 1776649790541228, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790541237, "dur": 15, + "args": { + "External id": 15712, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15712, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15712, "pid": 0, "tid": 7, "ts": 1776649790541237, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541229, "dur": 4, + "args": { + "External id": 15712, + "cbid": 211, "correlation": 15712 + } + }, + { + "ph": "s", "id": 15712, "pid": 435, "tid": 435, "ts": 1776649790541229, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790541273, "dur": 1, + "args": { + "External id": 1842, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15730, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15730, "pid": 0, "tid": 7, "ts": 1776649790541273, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541263, "dur": 7, + "args": { + "External id": 1842, + "cbid": 211, "correlation": 15730 + } + }, + { + "ph": "s", "id": 15730, "pid": 435, "tid": 435, "ts": 1776649790541263, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541276, "dur": 0, + "args": { + "External id": 15736, + "cbid": 200, "correlation": 15736 + } + }, + { + "ph": "f", "id": 15736, "pid": 435, "tid": 435, "ts": 1776649790541276, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541277, "dur": 0, + "args": { + "External id": 15737, + "cbid": 273, "correlation": 15737 + } + }, + { + "ph": "f", "id": 15737, "pid": 435, "tid": 435, "ts": 1776649790541277, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790541286, "dur": 15, + "args": { + "External id": 15738, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15738, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15738, "pid": 0, "tid": 7, "ts": 1776649790541286, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541278, "dur": 5, + "args": { + "External id": 15738, + "cbid": 211, "correlation": 15738 + } + }, + { + "ph": "s", "id": 15738, "pid": 435, "tid": 435, "ts": 1776649790541278, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790541322, "dur": 1, + "args": { + "External id": 1847, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15756, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15756, "pid": 0, "tid": 7, "ts": 1776649790541322, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541313, "dur": 7, + "args": { + "External id": 1847, + "cbid": 211, "correlation": 15756 + } + }, + { + "ph": "s", "id": 15756, "pid": 435, "tid": 435, "ts": 1776649790541313, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541325, "dur": 0, + "args": { + "External id": 15762, + "cbid": 200, "correlation": 15762 + } + }, + { + "ph": "f", "id": 15762, "pid": 435, "tid": 435, "ts": 1776649790541325, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541326, "dur": 0, + "args": { + "External id": 15763, + "cbid": 273, "correlation": 15763 + } + }, + { + "ph": "f", "id": 15763, "pid": 435, "tid": 435, "ts": 1776649790541326, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790541335, "dur": 15, + "args": { + "External id": 15764, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15764, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15764, "pid": 0, "tid": 7, "ts": 1776649790541335, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541327, "dur": 4, + "args": { + "External id": 15764, + "cbid": 211, "correlation": 15764 + } + }, + { + "ph": "s", "id": 15764, "pid": 435, "tid": 435, "ts": 1776649790541327, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790541370, "dur": 1, + "args": { + "External id": 1852, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15782, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15782, "pid": 0, "tid": 7, "ts": 1776649790541370, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541361, "dur": 6, + "args": { + "External id": 1852, + "cbid": 211, "correlation": 15782 + } + }, + { + "ph": "s", "id": 15782, "pid": 435, "tid": 435, "ts": 1776649790541361, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541373, "dur": 0, + "args": { + "External id": 15788, + "cbid": 200, "correlation": 15788 + } + }, + { + "ph": "f", "id": 15788, "pid": 435, "tid": 435, "ts": 1776649790541373, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541374, "dur": 0, + "args": { + "External id": 15789, + "cbid": 273, "correlation": 15789 + } + }, + { + "ph": "f", "id": 15789, "pid": 435, "tid": 435, "ts": 1776649790541374, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790541383, "dur": 15, + "args": { + "External id": 15790, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15790, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15790, "pid": 0, "tid": 7, "ts": 1776649790541383, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541375, "dur": 4, + "args": { + "External id": 15790, + "cbid": 211, "correlation": 15790 + } + }, + { + "ph": "s", "id": 15790, "pid": 435, "tid": 435, "ts": 1776649790541375, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790541419, "dur": 1, + "args": { + "External id": 1857, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15808, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15808, "pid": 0, "tid": 7, "ts": 1776649790541419, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541409, "dur": 6, + "args": { + "External id": 1857, + "cbid": 211, "correlation": 15808 + } + }, + { + "ph": "s", "id": 15808, "pid": 435, "tid": 435, "ts": 1776649790541409, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541422, "dur": 0, + "args": { + "External id": 15814, + "cbid": 200, "correlation": 15814 + } + }, + { + "ph": "f", "id": 15814, "pid": 435, "tid": 435, "ts": 1776649790541422, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541422, "dur": 0, + "args": { + "External id": 15815, + "cbid": 273, "correlation": 15815 + } + }, + { + "ph": "f", "id": 15815, "pid": 435, "tid": 435, "ts": 1776649790541422, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790541432, "dur": 15, + "args": { + "External id": 15816, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15816, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15816, "pid": 0, "tid": 7, "ts": 1776649790541432, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541423, "dur": 4, + "args": { + "External id": 15816, + "cbid": 211, "correlation": 15816 + } + }, + { + "ph": "s", "id": 15816, "pid": 435, "tid": 435, "ts": 1776649790541423, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790541467, "dur": 1, + "args": { + "External id": 1862, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15834, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15834, "pid": 0, "tid": 7, "ts": 1776649790541467, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541458, "dur": 6, + "args": { + "External id": 1862, + "cbid": 211, "correlation": 15834 + } + }, + { + "ph": "s", "id": 15834, "pid": 435, "tid": 435, "ts": 1776649790541458, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541471, "dur": 0, + "args": { + "External id": 15840, + "cbid": 200, "correlation": 15840 + } + }, + { + "ph": "f", "id": 15840, "pid": 435, "tid": 435, "ts": 1776649790541471, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541471, "dur": 0, + "args": { + "External id": 15841, + "cbid": 273, "correlation": 15841 + } + }, + { + "ph": "f", "id": 15841, "pid": 435, "tid": 435, "ts": 1776649790541471, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790541480, "dur": 18, + "args": { + "External id": 15842, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15842, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15842, "pid": 0, "tid": 7, "ts": 1776649790541480, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541472, "dur": 4, + "args": { + "External id": 15842, + "cbid": 211, "correlation": 15842 + } + }, + { + "ph": "s", "id": 15842, "pid": 435, "tid": 435, "ts": 1776649790541472, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790541515, "dur": 1, + "args": { + "External id": 1867, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15860, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15860, "pid": 0, "tid": 7, "ts": 1776649790541515, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541506, "dur": 6, + "args": { + "External id": 1867, + "cbid": 211, "correlation": 15860 + } + }, + { + "ph": "s", "id": 15860, "pid": 435, "tid": 435, "ts": 1776649790541506, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541519, "dur": 0, + "args": { + "External id": 15866, + "cbid": 200, "correlation": 15866 + } + }, + { + "ph": "f", "id": 15866, "pid": 435, "tid": 435, "ts": 1776649790541519, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541519, "dur": 0, + "args": { + "External id": 15867, + "cbid": 273, "correlation": 15867 + } + }, + { + "ph": "f", "id": 15867, "pid": 435, "tid": 435, "ts": 1776649790541519, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790541529, "dur": 15, + "args": { + "External id": 15868, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15868, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15868, "pid": 0, "tid": 7, "ts": 1776649790541529, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541520, "dur": 5, + "args": { + "External id": 15868, + "cbid": 211, "correlation": 15868 + } + }, + { + "ph": "s", "id": 15868, "pid": 435, "tid": 435, "ts": 1776649790541520, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790541564, "dur": 1, + "args": { + "External id": 1872, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15886, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15886, "pid": 0, "tid": 7, "ts": 1776649790541564, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541554, "dur": 6, + "args": { + "External id": 1872, + "cbid": 211, "correlation": 15886 + } + }, + { + "ph": "s", "id": 15886, "pid": 435, "tid": 435, "ts": 1776649790541554, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541567, "dur": 0, + "args": { + "External id": 15892, + "cbid": 200, "correlation": 15892 + } + }, + { + "ph": "f", "id": 15892, "pid": 435, "tid": 435, "ts": 1776649790541567, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541568, "dur": 0, + "args": { + "External id": 15893, + "cbid": 273, "correlation": 15893 + } + }, + { + "ph": "f", "id": 15893, "pid": 435, "tid": 435, "ts": 1776649790541568, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790541577, "dur": 15, + "args": { + "External id": 15894, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15894, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15894, "pid": 0, "tid": 7, "ts": 1776649790541577, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541569, "dur": 4, + "args": { + "External id": 15894, + "cbid": 211, "correlation": 15894 + } + }, + { + "ph": "s", "id": 15894, "pid": 435, "tid": 435, "ts": 1776649790541569, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790541612, "dur": 1, + "args": { + "External id": 1877, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15912, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15912, "pid": 0, "tid": 7, "ts": 1776649790541612, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541603, "dur": 6, + "args": { + "External id": 1877, + "cbid": 211, "correlation": 15912 + } + }, + { + "ph": "s", "id": 15912, "pid": 435, "tid": 435, "ts": 1776649790541603, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541615, "dur": 0, + "args": { + "External id": 15918, + "cbid": 200, "correlation": 15918 + } + }, + { + "ph": "f", "id": 15918, "pid": 435, "tid": 435, "ts": 1776649790541615, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541616, "dur": 0, + "args": { + "External id": 15919, + "cbid": 273, "correlation": 15919 + } + }, + { + "ph": "f", "id": 15919, "pid": 435, "tid": 435, "ts": 1776649790541616, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790541625, "dur": 15, + "args": { + "External id": 15920, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15920, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15920, "pid": 0, "tid": 7, "ts": 1776649790541625, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541617, "dur": 4, + "args": { + "External id": 15920, + "cbid": 211, "correlation": 15920 + } + }, + { + "ph": "s", "id": 15920, "pid": 435, "tid": 435, "ts": 1776649790541617, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790541660, "dur": 1, + "args": { + "External id": 1882, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15938, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15938, "pid": 0, "tid": 7, "ts": 1776649790541660, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541651, "dur": 6, + "args": { + "External id": 1882, + "cbid": 211, "correlation": 15938 + } + }, + { + "ph": "s", "id": 15938, "pid": 435, "tid": 435, "ts": 1776649790541651, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541663, "dur": 0, + "args": { + "External id": 15944, + "cbid": 200, "correlation": 15944 + } + }, + { + "ph": "f", "id": 15944, "pid": 435, "tid": 435, "ts": 1776649790541663, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541664, "dur": 0, + "args": { + "External id": 15945, + "cbid": 273, "correlation": 15945 + } + }, + { + "ph": "f", "id": 15945, "pid": 435, "tid": 435, "ts": 1776649790541664, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790541673, "dur": 15, + "args": { + "External id": 15946, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15946, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15946, "pid": 0, "tid": 7, "ts": 1776649790541673, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541665, "dur": 5, + "args": { + "External id": 15946, + "cbid": 211, "correlation": 15946 + } + }, + { + "ph": "s", "id": 15946, "pid": 435, "tid": 435, "ts": 1776649790541665, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790541708, "dur": 1, + "args": { + "External id": 1887, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15964, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15964, "pid": 0, "tid": 7, "ts": 1776649790541708, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541699, "dur": 6, + "args": { + "External id": 1887, + "cbid": 211, "correlation": 15964 + } + }, + { + "ph": "s", "id": 15964, "pid": 435, "tid": 435, "ts": 1776649790541699, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541712, "dur": 0, + "args": { + "External id": 15970, + "cbid": 200, "correlation": 15970 + } + }, + { + "ph": "f", "id": 15970, "pid": 435, "tid": 435, "ts": 1776649790541712, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541712, "dur": 0, + "args": { + "External id": 15971, + "cbid": 273, "correlation": 15971 + } + }, + { + "ph": "f", "id": 15971, "pid": 435, "tid": 435, "ts": 1776649790541712, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790541721, "dur": 15, + "args": { + "External id": 15972, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15972, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15972, "pid": 0, "tid": 7, "ts": 1776649790541721, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541713, "dur": 4, + "args": { + "External id": 15972, + "cbid": 211, "correlation": 15972 + } + }, + { + "ph": "s", "id": 15972, "pid": 435, "tid": 435, "ts": 1776649790541713, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790541757, "dur": 1, + "args": { + "External id": 1892, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15990, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15990, "pid": 0, "tid": 7, "ts": 1776649790541757, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541748, "dur": 6, + "args": { + "External id": 1892, + "cbid": 211, "correlation": 15990 + } + }, + { + "ph": "s", "id": 15990, "pid": 435, "tid": 435, "ts": 1776649790541748, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541760, "dur": 0, + "args": { + "External id": 15996, + "cbid": 200, "correlation": 15996 + } + }, + { + "ph": "f", "id": 15996, "pid": 435, "tid": 435, "ts": 1776649790541760, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541761, "dur": 0, + "args": { + "External id": 15997, + "cbid": 273, "correlation": 15997 + } + }, + { + "ph": "f", "id": 15997, "pid": 435, "tid": 435, "ts": 1776649790541761, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790541770, "dur": 15, + "args": { + "External id": 15998, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 15998, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 15998, "pid": 0, "tid": 7, "ts": 1776649790541770, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541762, "dur": 4, + "args": { + "External id": 15998, + "cbid": 211, "correlation": 15998 + } + }, + { + "ph": "s", "id": 15998, "pid": 435, "tid": 435, "ts": 1776649790541762, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790541805, "dur": 1, + "args": { + "External id": 1897, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16016, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16016, "pid": 0, "tid": 7, "ts": 1776649790541805, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541795, "dur": 6, + "args": { + "External id": 1897, + "cbid": 211, "correlation": 16016 + } + }, + { + "ph": "s", "id": 16016, "pid": 435, "tid": 435, "ts": 1776649790541795, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541808, "dur": 0, + "args": { + "External id": 16022, + "cbid": 200, "correlation": 16022 + } + }, + { + "ph": "f", "id": 16022, "pid": 435, "tid": 435, "ts": 1776649790541808, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541808, "dur": 0, + "args": { + "External id": 16023, + "cbid": 273, "correlation": 16023 + } + }, + { + "ph": "f", "id": 16023, "pid": 435, "tid": 435, "ts": 1776649790541808, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790541818, "dur": 15, + "args": { + "External id": 16024, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16024, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16024, "pid": 0, "tid": 7, "ts": 1776649790541818, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541810, "dur": 5, + "args": { + "External id": 16024, + "cbid": 211, "correlation": 16024 + } + }, + { + "ph": "s", "id": 16024, "pid": 435, "tid": 435, "ts": 1776649790541810, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790541853, "dur": 1, + "args": { + "External id": 1902, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16042, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16042, "pid": 0, "tid": 7, "ts": 1776649790541853, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541844, "dur": 6, + "args": { + "External id": 1902, + "cbid": 211, "correlation": 16042 + } + }, + { + "ph": "s", "id": 16042, "pid": 435, "tid": 435, "ts": 1776649790541844, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541856, "dur": 0, + "args": { + "External id": 16048, + "cbid": 200, "correlation": 16048 + } + }, + { + "ph": "f", "id": 16048, "pid": 435, "tid": 435, "ts": 1776649790541856, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541857, "dur": 0, + "args": { + "External id": 16049, + "cbid": 273, "correlation": 16049 + } + }, + { + "ph": "f", "id": 16049, "pid": 435, "tid": 435, "ts": 1776649790541857, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790541866, "dur": 15, + "args": { + "External id": 16050, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16050, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16050, "pid": 0, "tid": 7, "ts": 1776649790541866, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541858, "dur": 4, + "args": { + "External id": 16050, + "cbid": 211, "correlation": 16050 + } + }, + { + "ph": "s", "id": 16050, "pid": 435, "tid": 435, "ts": 1776649790541858, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790541901, "dur": 1, + "args": { + "External id": 1907, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16068, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16068, "pid": 0, "tid": 7, "ts": 1776649790541901, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541891, "dur": 6, + "args": { + "External id": 1907, + "cbid": 211, "correlation": 16068 + } + }, + { + "ph": "s", "id": 16068, "pid": 435, "tid": 435, "ts": 1776649790541891, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541904, "dur": 0, + "args": { + "External id": 16074, + "cbid": 200, "correlation": 16074 + } + }, + { + "ph": "f", "id": 16074, "pid": 435, "tid": 435, "ts": 1776649790541904, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541904, "dur": 0, + "args": { + "External id": 16075, + "cbid": 273, "correlation": 16075 + } + }, + { + "ph": "f", "id": 16075, "pid": 435, "tid": 435, "ts": 1776649790541904, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790541913, "dur": 16, + "args": { + "External id": 16076, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16076, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16076, "pid": 0, "tid": 7, "ts": 1776649790541913, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541906, "dur": 4, + "args": { + "External id": 16076, + "cbid": 211, "correlation": 16076 + } + }, + { + "ph": "s", "id": 16076, "pid": 435, "tid": 435, "ts": 1776649790541906, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790541948, "dur": 1, + "args": { + "External id": 1912, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16094, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16094, "pid": 0, "tid": 7, "ts": 1776649790541948, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541939, "dur": 6, + "args": { + "External id": 1912, + "cbid": 211, "correlation": 16094 + } + }, + { + "ph": "s", "id": 16094, "pid": 435, "tid": 435, "ts": 1776649790541939, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541952, "dur": 0, + "args": { + "External id": 16100, + "cbid": 200, "correlation": 16100 + } + }, + { + "ph": "f", "id": 16100, "pid": 435, "tid": 435, "ts": 1776649790541952, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541952, "dur": 0, + "args": { + "External id": 16101, + "cbid": 273, "correlation": 16101 + } + }, + { + "ph": "f", "id": 16101, "pid": 435, "tid": 435, "ts": 1776649790541952, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790541961, "dur": 15, + "args": { + "External id": 16102, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16102, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16102, "pid": 0, "tid": 7, "ts": 1776649790541961, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541953, "dur": 4, + "args": { + "External id": 16102, + "cbid": 211, "correlation": 16102 + } + }, + { + "ph": "s", "id": 16102, "pid": 435, "tid": 435, "ts": 1776649790541953, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790541996, "dur": 1, + "args": { + "External id": 1917, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16120, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16120, "pid": 0, "tid": 7, "ts": 1776649790541996, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790541987, "dur": 6, + "args": { + "External id": 1917, + "cbid": 211, "correlation": 16120 + } + }, + { + "ph": "s", "id": 16120, "pid": 435, "tid": 435, "ts": 1776649790541987, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790541999, "dur": 0, + "args": { + "External id": 16126, + "cbid": 200, "correlation": 16126 + } + }, + { + "ph": "f", "id": 16126, "pid": 435, "tid": 435, "ts": 1776649790541999, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542000, "dur": 0, + "args": { + "External id": 16127, + "cbid": 273, "correlation": 16127 + } + }, + { + "ph": "f", "id": 16127, "pid": 435, "tid": 435, "ts": 1776649790542000, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790542009, "dur": 15, + "args": { + "External id": 16128, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16128, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16128, "pid": 0, "tid": 7, "ts": 1776649790542009, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542001, "dur": 4, + "args": { + "External id": 16128, + "cbid": 211, "correlation": 16128 + } + }, + { + "ph": "s", "id": 16128, "pid": 435, "tid": 435, "ts": 1776649790542001, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790542044, "dur": 1, + "args": { + "External id": 1922, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16146, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16146, "pid": 0, "tid": 7, "ts": 1776649790542044, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542035, "dur": 6, + "args": { + "External id": 1922, + "cbid": 211, "correlation": 16146 + } + }, + { + "ph": "s", "id": 16146, "pid": 435, "tid": 435, "ts": 1776649790542035, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542047, "dur": 0, + "args": { + "External id": 16152, + "cbid": 200, "correlation": 16152 + } + }, + { + "ph": "f", "id": 16152, "pid": 435, "tid": 435, "ts": 1776649790542047, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542048, "dur": 0, + "args": { + "External id": 16153, + "cbid": 273, "correlation": 16153 + } + }, + { + "ph": "f", "id": 16153, "pid": 435, "tid": 435, "ts": 1776649790542048, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790542057, "dur": 15, + "args": { + "External id": 16154, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16154, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16154, "pid": 0, "tid": 7, "ts": 1776649790542057, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542049, "dur": 5, + "args": { + "External id": 16154, + "cbid": 211, "correlation": 16154 + } + }, + { + "ph": "s", "id": 16154, "pid": 435, "tid": 435, "ts": 1776649790542049, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790542092, "dur": 1, + "args": { + "External id": 1927, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16172, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16172, "pid": 0, "tid": 7, "ts": 1776649790542092, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542082, "dur": 6, + "args": { + "External id": 1927, + "cbid": 211, "correlation": 16172 + } + }, + { + "ph": "s", "id": 16172, "pid": 435, "tid": 435, "ts": 1776649790542082, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542095, "dur": 0, + "args": { + "External id": 16178, + "cbid": 200, "correlation": 16178 + } + }, + { + "ph": "f", "id": 16178, "pid": 435, "tid": 435, "ts": 1776649790542095, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542095, "dur": 0, + "args": { + "External id": 16179, + "cbid": 273, "correlation": 16179 + } + }, + { + "ph": "f", "id": 16179, "pid": 435, "tid": 435, "ts": 1776649790542095, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790542104, "dur": 15, + "args": { + "External id": 16180, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16180, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16180, "pid": 0, "tid": 7, "ts": 1776649790542104, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542097, "dur": 4, + "args": { + "External id": 16180, + "cbid": 211, "correlation": 16180 + } + }, + { + "ph": "s", "id": 16180, "pid": 435, "tid": 435, "ts": 1776649790542097, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790542139, "dur": 1, + "args": { + "External id": 1932, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16198, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16198, "pid": 0, "tid": 7, "ts": 1776649790542139, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542130, "dur": 6, + "args": { + "External id": 1932, + "cbid": 211, "correlation": 16198 + } + }, + { + "ph": "s", "id": 16198, "pid": 435, "tid": 435, "ts": 1776649790542130, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542143, "dur": 0, + "args": { + "External id": 16204, + "cbid": 200, "correlation": 16204 + } + }, + { + "ph": "f", "id": 16204, "pid": 435, "tid": 435, "ts": 1776649790542143, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542143, "dur": 0, + "args": { + "External id": 16205, + "cbid": 273, "correlation": 16205 + } + }, + { + "ph": "f", "id": 16205, "pid": 435, "tid": 435, "ts": 1776649790542143, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790542152, "dur": 15, + "args": { + "External id": 16206, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16206, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16206, "pid": 0, "tid": 7, "ts": 1776649790542152, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542144, "dur": 5, + "args": { + "External id": 16206, + "cbid": 211, "correlation": 16206 + } + }, + { + "ph": "s", "id": 16206, "pid": 435, "tid": 435, "ts": 1776649790542144, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790542187, "dur": 1, + "args": { + "External id": 1937, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16224, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16224, "pid": 0, "tid": 7, "ts": 1776649790542187, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542178, "dur": 6, + "args": { + "External id": 1937, + "cbid": 211, "correlation": 16224 + } + }, + { + "ph": "s", "id": 16224, "pid": 435, "tid": 435, "ts": 1776649790542178, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542190, "dur": 0, + "args": { + "External id": 16230, + "cbid": 200, "correlation": 16230 + } + }, + { + "ph": "f", "id": 16230, "pid": 435, "tid": 435, "ts": 1776649790542190, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542191, "dur": 0, + "args": { + "External id": 16231, + "cbid": 273, "correlation": 16231 + } + }, + { + "ph": "f", "id": 16231, "pid": 435, "tid": 435, "ts": 1776649790542191, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790542200, "dur": 15, + "args": { + "External id": 16232, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16232, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16232, "pid": 0, "tid": 7, "ts": 1776649790542200, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542192, "dur": 4, + "args": { + "External id": 16232, + "cbid": 211, "correlation": 16232 + } + }, + { + "ph": "s", "id": 16232, "pid": 435, "tid": 435, "ts": 1776649790542192, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790542235, "dur": 1, + "args": { + "External id": 1942, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16250, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16250, "pid": 0, "tid": 7, "ts": 1776649790542235, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542225, "dur": 6, + "args": { + "External id": 1942, + "cbid": 211, "correlation": 16250 + } + }, + { + "ph": "s", "id": 16250, "pid": 435, "tid": 435, "ts": 1776649790542225, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542238, "dur": 0, + "args": { + "External id": 16256, + "cbid": 200, "correlation": 16256 + } + }, + { + "ph": "f", "id": 16256, "pid": 435, "tid": 435, "ts": 1776649790542238, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542238, "dur": 0, + "args": { + "External id": 16257, + "cbid": 273, "correlation": 16257 + } + }, + { + "ph": "f", "id": 16257, "pid": 435, "tid": 435, "ts": 1776649790542238, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790542247, "dur": 15, + "args": { + "External id": 16258, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16258, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16258, "pid": 0, "tid": 7, "ts": 1776649790542247, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542239, "dur": 5, + "args": { + "External id": 16258, + "cbid": 211, "correlation": 16258 + } + }, + { + "ph": "s", "id": 16258, "pid": 435, "tid": 435, "ts": 1776649790542239, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790542282, "dur": 1, + "args": { + "External id": 1947, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16276, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16276, "pid": 0, "tid": 7, "ts": 1776649790542282, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542273, "dur": 6, + "args": { + "External id": 1947, + "cbid": 211, "correlation": 16276 + } + }, + { + "ph": "s", "id": 16276, "pid": 435, "tid": 435, "ts": 1776649790542273, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542285, "dur": 0, + "args": { + "External id": 16282, + "cbid": 200, "correlation": 16282 + } + }, + { + "ph": "f", "id": 16282, "pid": 435, "tid": 435, "ts": 1776649790542285, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542286, "dur": 0, + "args": { + "External id": 16283, + "cbid": 273, "correlation": 16283 + } + }, + { + "ph": "f", "id": 16283, "pid": 435, "tid": 435, "ts": 1776649790542286, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790542295, "dur": 16, + "args": { + "External id": 16284, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16284, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16284, "pid": 0, "tid": 7, "ts": 1776649790542295, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542287, "dur": 5, + "args": { + "External id": 16284, + "cbid": 211, "correlation": 16284 + } + }, + { + "ph": "s", "id": 16284, "pid": 435, "tid": 435, "ts": 1776649790542287, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790542330, "dur": 1, + "args": { + "External id": 1952, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16302, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16302, "pid": 0, "tid": 7, "ts": 1776649790542330, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542321, "dur": 6, + "args": { + "External id": 1952, + "cbid": 211, "correlation": 16302 + } + }, + { + "ph": "s", "id": 16302, "pid": 435, "tid": 435, "ts": 1776649790542321, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542333, "dur": 0, + "args": { + "External id": 16308, + "cbid": 200, "correlation": 16308 + } + }, + { + "ph": "f", "id": 16308, "pid": 435, "tid": 435, "ts": 1776649790542333, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542334, "dur": 0, + "args": { + "External id": 16309, + "cbid": 273, "correlation": 16309 + } + }, + { + "ph": "f", "id": 16309, "pid": 435, "tid": 435, "ts": 1776649790542334, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790542343, "dur": 16, + "args": { + "External id": 16310, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16310, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16310, "pid": 0, "tid": 7, "ts": 1776649790542343, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542335, "dur": 4, + "args": { + "External id": 16310, + "cbid": 211, "correlation": 16310 + } + }, + { + "ph": "s", "id": 16310, "pid": 435, "tid": 435, "ts": 1776649790542335, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790542377, "dur": 1, + "args": { + "External id": 1957, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16328, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16328, "pid": 0, "tid": 7, "ts": 1776649790542377, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542368, "dur": 6, + "args": { + "External id": 1957, + "cbid": 211, "correlation": 16328 + } + }, + { + "ph": "s", "id": 16328, "pid": 435, "tid": 435, "ts": 1776649790542368, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542380, "dur": 0, + "args": { + "External id": 16334, + "cbid": 200, "correlation": 16334 + } + }, + { + "ph": "f", "id": 16334, "pid": 435, "tid": 435, "ts": 1776649790542380, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542381, "dur": 0, + "args": { + "External id": 16335, + "cbid": 273, "correlation": 16335 + } + }, + { + "ph": "f", "id": 16335, "pid": 435, "tid": 435, "ts": 1776649790542381, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790542390, "dur": 15, + "args": { + "External id": 16336, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16336, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16336, "pid": 0, "tid": 7, "ts": 1776649790542390, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542382, "dur": 4, + "args": { + "External id": 16336, + "cbid": 211, "correlation": 16336 + } + }, + { + "ph": "s", "id": 16336, "pid": 435, "tid": 435, "ts": 1776649790542382, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790542425, "dur": 1, + "args": { + "External id": 1962, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16354, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16354, "pid": 0, "tid": 7, "ts": 1776649790542425, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542416, "dur": 6, + "args": { + "External id": 1962, + "cbid": 211, "correlation": 16354 + } + }, + { + "ph": "s", "id": 16354, "pid": 435, "tid": 435, "ts": 1776649790542416, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542428, "dur": 0, + "args": { + "External id": 16360, + "cbid": 200, "correlation": 16360 + } + }, + { + "ph": "f", "id": 16360, "pid": 435, "tid": 435, "ts": 1776649790542428, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542428, "dur": 0, + "args": { + "External id": 16361, + "cbid": 273, "correlation": 16361 + } + }, + { + "ph": "f", "id": 16361, "pid": 435, "tid": 435, "ts": 1776649790542428, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790542437, "dur": 15, + "args": { + "External id": 16362, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16362, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16362, "pid": 0, "tid": 7, "ts": 1776649790542437, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542429, "dur": 4, + "args": { + "External id": 16362, + "cbid": 211, "correlation": 16362 + } + }, + { + "ph": "s", "id": 16362, "pid": 435, "tid": 435, "ts": 1776649790542429, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790542472, "dur": 1, + "args": { + "External id": 1967, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16380, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16380, "pid": 0, "tid": 7, "ts": 1776649790542472, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542463, "dur": 6, + "args": { + "External id": 1967, + "cbid": 211, "correlation": 16380 + } + }, + { + "ph": "s", "id": 16380, "pid": 435, "tid": 435, "ts": 1776649790542463, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542475, "dur": 0, + "args": { + "External id": 16386, + "cbid": 200, "correlation": 16386 + } + }, + { + "ph": "f", "id": 16386, "pid": 435, "tid": 435, "ts": 1776649790542475, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542476, "dur": 0, + "args": { + "External id": 16387, + "cbid": 273, "correlation": 16387 + } + }, + { + "ph": "f", "id": 16387, "pid": 435, "tid": 435, "ts": 1776649790542476, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790542485, "dur": 15, + "args": { + "External id": 16388, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16388, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16388, "pid": 0, "tid": 7, "ts": 1776649790542485, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542477, "dur": 4, + "args": { + "External id": 16388, + "cbid": 211, "correlation": 16388 + } + }, + { + "ph": "s", "id": 16388, "pid": 435, "tid": 435, "ts": 1776649790542477, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790542520, "dur": 1, + "args": { + "External id": 1972, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16406, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16406, "pid": 0, "tid": 7, "ts": 1776649790542520, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542511, "dur": 6, + "args": { + "External id": 1972, + "cbid": 211, "correlation": 16406 + } + }, + { + "ph": "s", "id": 16406, "pid": 435, "tid": 435, "ts": 1776649790542511, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542523, "dur": 0, + "args": { + "External id": 16412, + "cbid": 200, "correlation": 16412 + } + }, + { + "ph": "f", "id": 16412, "pid": 435, "tid": 435, "ts": 1776649790542523, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542524, "dur": 0, + "args": { + "External id": 16413, + "cbid": 273, "correlation": 16413 + } + }, + { + "ph": "f", "id": 16413, "pid": 435, "tid": 435, "ts": 1776649790542524, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790542533, "dur": 15, + "args": { + "External id": 16414, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16414, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16414, "pid": 0, "tid": 7, "ts": 1776649790542533, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542525, "dur": 5, + "args": { + "External id": 16414, + "cbid": 211, "correlation": 16414 + } + }, + { + "ph": "s", "id": 16414, "pid": 435, "tid": 435, "ts": 1776649790542525, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790542568, "dur": 1, + "args": { + "External id": 1977, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16432, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16432, "pid": 0, "tid": 7, "ts": 1776649790542568, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542559, "dur": 6, + "args": { + "External id": 1977, + "cbid": 211, "correlation": 16432 + } + }, + { + "ph": "s", "id": 16432, "pid": 435, "tid": 435, "ts": 1776649790542559, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542571, "dur": 0, + "args": { + "External id": 16438, + "cbid": 200, "correlation": 16438 + } + }, + { + "ph": "f", "id": 16438, "pid": 435, "tid": 435, "ts": 1776649790542571, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542572, "dur": 0, + "args": { + "External id": 16439, + "cbid": 273, "correlation": 16439 + } + }, + { + "ph": "f", "id": 16439, "pid": 435, "tid": 435, "ts": 1776649790542572, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790542581, "dur": 15, + "args": { + "External id": 16440, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16440, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16440, "pid": 0, "tid": 7, "ts": 1776649790542581, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542573, "dur": 4, + "args": { + "External id": 16440, + "cbid": 211, "correlation": 16440 + } + }, + { + "ph": "s", "id": 16440, "pid": 435, "tid": 435, "ts": 1776649790542573, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790542616, "dur": 1, + "args": { + "External id": 1982, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16458, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16458, "pid": 0, "tid": 7, "ts": 1776649790542616, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542606, "dur": 6, + "args": { + "External id": 1982, + "cbid": 211, "correlation": 16458 + } + }, + { + "ph": "s", "id": 16458, "pid": 435, "tid": 435, "ts": 1776649790542606, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542619, "dur": 0, + "args": { + "External id": 16464, + "cbid": 200, "correlation": 16464 + } + }, + { + "ph": "f", "id": 16464, "pid": 435, "tid": 435, "ts": 1776649790542619, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542619, "dur": 0, + "args": { + "External id": 16465, + "cbid": 273, "correlation": 16465 + } + }, + { + "ph": "f", "id": 16465, "pid": 435, "tid": 435, "ts": 1776649790542619, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790542628, "dur": 15, + "args": { + "External id": 16466, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16466, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16466, "pid": 0, "tid": 7, "ts": 1776649790542628, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542620, "dur": 4, + "args": { + "External id": 16466, + "cbid": 211, "correlation": 16466 + } + }, + { + "ph": "s", "id": 16466, "pid": 435, "tid": 435, "ts": 1776649790542620, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790542663, "dur": 1, + "args": { + "External id": 1987, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16484, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16484, "pid": 0, "tid": 7, "ts": 1776649790542663, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542654, "dur": 6, + "args": { + "External id": 1987, + "cbid": 211, "correlation": 16484 + } + }, + { + "ph": "s", "id": 16484, "pid": 435, "tid": 435, "ts": 1776649790542654, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542666, "dur": 0, + "args": { + "External id": 16490, + "cbid": 200, "correlation": 16490 + } + }, + { + "ph": "f", "id": 16490, "pid": 435, "tid": 435, "ts": 1776649790542666, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542666, "dur": 0, + "args": { + "External id": 16491, + "cbid": 273, "correlation": 16491 + } + }, + { + "ph": "f", "id": 16491, "pid": 435, "tid": 435, "ts": 1776649790542666, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790542676, "dur": 15, + "args": { + "External id": 16492, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16492, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16492, "pid": 0, "tid": 7, "ts": 1776649790542676, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542668, "dur": 4, + "args": { + "External id": 16492, + "cbid": 211, "correlation": 16492 + } + }, + { + "ph": "s", "id": 16492, "pid": 435, "tid": 435, "ts": 1776649790542668, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790542710, "dur": 1, + "args": { + "External id": 1992, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16510, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16510, "pid": 0, "tid": 7, "ts": 1776649790542710, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542701, "dur": 6, + "args": { + "External id": 1992, + "cbid": 211, "correlation": 16510 + } + }, + { + "ph": "s", "id": 16510, "pid": 435, "tid": 435, "ts": 1776649790542701, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542713, "dur": 0, + "args": { + "External id": 16516, + "cbid": 200, "correlation": 16516 + } + }, + { + "ph": "f", "id": 16516, "pid": 435, "tid": 435, "ts": 1776649790542713, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542714, "dur": 0, + "args": { + "External id": 16517, + "cbid": 273, "correlation": 16517 + } + }, + { + "ph": "f", "id": 16517, "pid": 435, "tid": 435, "ts": 1776649790542714, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790542723, "dur": 16, + "args": { + "External id": 16518, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16518, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16518, "pid": 0, "tid": 7, "ts": 1776649790542723, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542715, "dur": 4, + "args": { + "External id": 16518, + "cbid": 211, "correlation": 16518 + } + }, + { + "ph": "s", "id": 16518, "pid": 435, "tid": 435, "ts": 1776649790542715, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790542758, "dur": 1, + "args": { + "External id": 1997, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16536, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16536, "pid": 0, "tid": 7, "ts": 1776649790542758, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542749, "dur": 6, + "args": { + "External id": 1997, + "cbid": 211, "correlation": 16536 + } + }, + { + "ph": "s", "id": 16536, "pid": 435, "tid": 435, "ts": 1776649790542749, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542761, "dur": 0, + "args": { + "External id": 16542, + "cbid": 200, "correlation": 16542 + } + }, + { + "ph": "f", "id": 16542, "pid": 435, "tid": 435, "ts": 1776649790542761, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542762, "dur": 0, + "args": { + "External id": 16543, + "cbid": 273, "correlation": 16543 + } + }, + { + "ph": "f", "id": 16543, "pid": 435, "tid": 435, "ts": 1776649790542762, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790542771, "dur": 16, + "args": { + "External id": 16544, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16544, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16544, "pid": 0, "tid": 7, "ts": 1776649790542771, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542763, "dur": 5, + "args": { + "External id": 16544, + "cbid": 211, "correlation": 16544 + } + }, + { + "ph": "s", "id": 16544, "pid": 435, "tid": 435, "ts": 1776649790542763, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790542806, "dur": 1, + "args": { + "External id": 2002, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16562, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16562, "pid": 0, "tid": 7, "ts": 1776649790542806, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542797, "dur": 6, + "args": { + "External id": 2002, + "cbid": 211, "correlation": 16562 + } + }, + { + "ph": "s", "id": 16562, "pid": 435, "tid": 435, "ts": 1776649790542797, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542809, "dur": 0, + "args": { + "External id": 16568, + "cbid": 200, "correlation": 16568 + } + }, + { + "ph": "f", "id": 16568, "pid": 435, "tid": 435, "ts": 1776649790542809, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542810, "dur": 0, + "args": { + "External id": 16569, + "cbid": 273, "correlation": 16569 + } + }, + { + "ph": "f", "id": 16569, "pid": 435, "tid": 435, "ts": 1776649790542810, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790542818, "dur": 15, + "args": { + "External id": 16570, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16570, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16570, "pid": 0, "tid": 7, "ts": 1776649790542818, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542811, "dur": 4, + "args": { + "External id": 16570, + "cbid": 211, "correlation": 16570 + } + }, + { + "ph": "s", "id": 16570, "pid": 435, "tid": 435, "ts": 1776649790542811, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790542853, "dur": 1, + "args": { + "External id": 2007, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16588, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16588, "pid": 0, "tid": 7, "ts": 1776649790542853, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542844, "dur": 6, + "args": { + "External id": 2007, + "cbid": 211, "correlation": 16588 + } + }, + { + "ph": "s", "id": 16588, "pid": 435, "tid": 435, "ts": 1776649790542844, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542857, "dur": 0, + "args": { + "External id": 16594, + "cbid": 200, "correlation": 16594 + } + }, + { + "ph": "f", "id": 16594, "pid": 435, "tid": 435, "ts": 1776649790542857, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542857, "dur": 0, + "args": { + "External id": 16595, + "cbid": 273, "correlation": 16595 + } + }, + { + "ph": "f", "id": 16595, "pid": 435, "tid": 435, "ts": 1776649790542857, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790542866, "dur": 15, + "args": { + "External id": 16596, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16596, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16596, "pid": 0, "tid": 7, "ts": 1776649790542866, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542858, "dur": 4, + "args": { + "External id": 16596, + "cbid": 211, "correlation": 16596 + } + }, + { + "ph": "s", "id": 16596, "pid": 435, "tid": 435, "ts": 1776649790542858, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790542901, "dur": 1, + "args": { + "External id": 2012, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16614, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16614, "pid": 0, "tid": 7, "ts": 1776649790542901, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542892, "dur": 6, + "args": { + "External id": 2012, + "cbid": 211, "correlation": 16614 + } + }, + { + "ph": "s", "id": 16614, "pid": 435, "tid": 435, "ts": 1776649790542892, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542904, "dur": 0, + "args": { + "External id": 16620, + "cbid": 200, "correlation": 16620 + } + }, + { + "ph": "f", "id": 16620, "pid": 435, "tid": 435, "ts": 1776649790542904, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542905, "dur": 0, + "args": { + "External id": 16621, + "cbid": 273, "correlation": 16621 + } + }, + { + "ph": "f", "id": 16621, "pid": 435, "tid": 435, "ts": 1776649790542905, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790542914, "dur": 15, + "args": { + "External id": 16622, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16622, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16622, "pid": 0, "tid": 7, "ts": 1776649790542914, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542906, "dur": 5, + "args": { + "External id": 16622, + "cbid": 211, "correlation": 16622 + } + }, + { + "ph": "s", "id": 16622, "pid": 435, "tid": 435, "ts": 1776649790542906, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790542949, "dur": 1, + "args": { + "External id": 2017, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16640, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16640, "pid": 0, "tid": 7, "ts": 1776649790542949, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542940, "dur": 6, + "args": { + "External id": 2017, + "cbid": 211, "correlation": 16640 + } + }, + { + "ph": "s", "id": 16640, "pid": 435, "tid": 435, "ts": 1776649790542940, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542952, "dur": 0, + "args": { + "External id": 16646, + "cbid": 200, "correlation": 16646 + } + }, + { + "ph": "f", "id": 16646, "pid": 435, "tid": 435, "ts": 1776649790542952, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790542953, "dur": 0, + "args": { + "External id": 16647, + "cbid": 273, "correlation": 16647 + } + }, + { + "ph": "f", "id": 16647, "pid": 435, "tid": 435, "ts": 1776649790542953, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790542962, "dur": 15, + "args": { + "External id": 16648, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16648, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16648, "pid": 0, "tid": 7, "ts": 1776649790542962, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542954, "dur": 4, + "args": { + "External id": 16648, + "cbid": 211, "correlation": 16648 + } + }, + { + "ph": "s", "id": 16648, "pid": 435, "tid": 435, "ts": 1776649790542954, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790542997, "dur": 1, + "args": { + "External id": 2022, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16666, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16666, "pid": 0, "tid": 7, "ts": 1776649790542997, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790542988, "dur": 6, + "args": { + "External id": 2022, + "cbid": 211, "correlation": 16666 + } + }, + { + "ph": "s", "id": 16666, "pid": 435, "tid": 435, "ts": 1776649790542988, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543000, "dur": 0, + "args": { + "External id": 16672, + "cbid": 200, "correlation": 16672 + } + }, + { + "ph": "f", "id": 16672, "pid": 435, "tid": 435, "ts": 1776649790543000, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543001, "dur": 0, + "args": { + "External id": 16673, + "cbid": 273, "correlation": 16673 + } + }, + { + "ph": "f", "id": 16673, "pid": 435, "tid": 435, "ts": 1776649790543001, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790543010, "dur": 15, + "args": { + "External id": 16674, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16674, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16674, "pid": 0, "tid": 7, "ts": 1776649790543010, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543002, "dur": 5, + "args": { + "External id": 16674, + "cbid": 211, "correlation": 16674 + } + }, + { + "ph": "s", "id": 16674, "pid": 435, "tid": 435, "ts": 1776649790543002, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790543045, "dur": 1, + "args": { + "External id": 2027, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16692, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16692, "pid": 0, "tid": 7, "ts": 1776649790543045, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543036, "dur": 6, + "args": { + "External id": 2027, + "cbid": 211, "correlation": 16692 + } + }, + { + "ph": "s", "id": 16692, "pid": 435, "tid": 435, "ts": 1776649790543036, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543048, "dur": 0, + "args": { + "External id": 16698, + "cbid": 200, "correlation": 16698 + } + }, + { + "ph": "f", "id": 16698, "pid": 435, "tid": 435, "ts": 1776649790543048, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543049, "dur": 0, + "args": { + "External id": 16699, + "cbid": 273, "correlation": 16699 + } + }, + { + "ph": "f", "id": 16699, "pid": 435, "tid": 435, "ts": 1776649790543049, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790543058, "dur": 15, + "args": { + "External id": 16700, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16700, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16700, "pid": 0, "tid": 7, "ts": 1776649790543058, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543050, "dur": 4, + "args": { + "External id": 16700, + "cbid": 211, "correlation": 16700 + } + }, + { + "ph": "s", "id": 16700, "pid": 435, "tid": 435, "ts": 1776649790543050, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790543093, "dur": 1, + "args": { + "External id": 2032, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16718, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16718, "pid": 0, "tid": 7, "ts": 1776649790543093, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543084, "dur": 6, + "args": { + "External id": 2032, + "cbid": 211, "correlation": 16718 + } + }, + { + "ph": "s", "id": 16718, "pid": 435, "tid": 435, "ts": 1776649790543084, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543096, "dur": 0, + "args": { + "External id": 16724, + "cbid": 200, "correlation": 16724 + } + }, + { + "ph": "f", "id": 16724, "pid": 435, "tid": 435, "ts": 1776649790543096, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543097, "dur": 0, + "args": { + "External id": 16725, + "cbid": 273, "correlation": 16725 + } + }, + { + "ph": "f", "id": 16725, "pid": 435, "tid": 435, "ts": 1776649790543097, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790543106, "dur": 15, + "args": { + "External id": 16726, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16726, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16726, "pid": 0, "tid": 7, "ts": 1776649790543106, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543098, "dur": 4, + "args": { + "External id": 16726, + "cbid": 211, "correlation": 16726 + } + }, + { + "ph": "s", "id": 16726, "pid": 435, "tid": 435, "ts": 1776649790543098, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790543141, "dur": 1, + "args": { + "External id": 2037, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16744, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16744, "pid": 0, "tid": 7, "ts": 1776649790543141, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543132, "dur": 6, + "args": { + "External id": 2037, + "cbid": 211, "correlation": 16744 + } + }, + { + "ph": "s", "id": 16744, "pid": 435, "tid": 435, "ts": 1776649790543132, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543144, "dur": 0, + "args": { + "External id": 16750, + "cbid": 200, "correlation": 16750 + } + }, + { + "ph": "f", "id": 16750, "pid": 435, "tid": 435, "ts": 1776649790543144, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543145, "dur": 0, + "args": { + "External id": 16751, + "cbid": 273, "correlation": 16751 + } + }, + { + "ph": "f", "id": 16751, "pid": 435, "tid": 435, "ts": 1776649790543145, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790543154, "dur": 17, + "args": { + "External id": 16752, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16752, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16752, "pid": 0, "tid": 7, "ts": 1776649790543154, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543146, "dur": 4, + "args": { + "External id": 16752, + "cbid": 211, "correlation": 16752 + } + }, + { + "ph": "s", "id": 16752, "pid": 435, "tid": 435, "ts": 1776649790543146, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790543190, "dur": 1, + "args": { + "External id": 2042, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16770, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16770, "pid": 0, "tid": 7, "ts": 1776649790543190, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543181, "dur": 6, + "args": { + "External id": 2042, + "cbid": 211, "correlation": 16770 + } + }, + { + "ph": "s", "id": 16770, "pid": 435, "tid": 435, "ts": 1776649790543181, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543193, "dur": 0, + "args": { + "External id": 16776, + "cbid": 200, "correlation": 16776 + } + }, + { + "ph": "f", "id": 16776, "pid": 435, "tid": 435, "ts": 1776649790543193, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543193, "dur": 0, + "args": { + "External id": 16777, + "cbid": 273, "correlation": 16777 + } + }, + { + "ph": "f", "id": 16777, "pid": 435, "tid": 435, "ts": 1776649790543193, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790543202, "dur": 15, + "args": { + "External id": 16778, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16778, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16778, "pid": 0, "tid": 7, "ts": 1776649790543202, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543194, "dur": 4, + "args": { + "External id": 16778, + "cbid": 211, "correlation": 16778 + } + }, + { + "ph": "s", "id": 16778, "pid": 435, "tid": 435, "ts": 1776649790543194, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790543237, "dur": 1, + "args": { + "External id": 2047, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16796, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16796, "pid": 0, "tid": 7, "ts": 1776649790543237, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543228, "dur": 6, + "args": { + "External id": 2047, + "cbid": 211, "correlation": 16796 + } + }, + { + "ph": "s", "id": 16796, "pid": 435, "tid": 435, "ts": 1776649790543228, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543240, "dur": 0, + "args": { + "External id": 16802, + "cbid": 200, "correlation": 16802 + } + }, + { + "ph": "f", "id": 16802, "pid": 435, "tid": 435, "ts": 1776649790543240, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543241, "dur": 0, + "args": { + "External id": 16803, + "cbid": 273, "correlation": 16803 + } + }, + { + "ph": "f", "id": 16803, "pid": 435, "tid": 435, "ts": 1776649790543241, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790543250, "dur": 15, + "args": { + "External id": 16804, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16804, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16804, "pid": 0, "tid": 7, "ts": 1776649790543250, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543242, "dur": 4, + "args": { + "External id": 16804, + "cbid": 211, "correlation": 16804 + } + }, + { + "ph": "s", "id": 16804, "pid": 435, "tid": 435, "ts": 1776649790543242, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790543294, "dur": 1, + "args": { + "External id": 2052, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16822, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16822, "pid": 0, "tid": 7, "ts": 1776649790543294, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543285, "dur": 6, + "args": { + "External id": 2052, + "cbid": 211, "correlation": 16822 + } + }, + { + "ph": "s", "id": 16822, "pid": 435, "tid": 435, "ts": 1776649790543285, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543297, "dur": 0, + "args": { + "External id": 16828, + "cbid": 200, "correlation": 16828 + } + }, + { + "ph": "f", "id": 16828, "pid": 435, "tid": 435, "ts": 1776649790543297, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543298, "dur": 0, + "args": { + "External id": 16829, + "cbid": 273, "correlation": 16829 + } + }, + { + "ph": "f", "id": 16829, "pid": 435, "tid": 435, "ts": 1776649790543298, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790543307, "dur": 15, + "args": { + "External id": 16830, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16830, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16830, "pid": 0, "tid": 7, "ts": 1776649790543307, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543299, "dur": 4, + "args": { + "External id": 16830, + "cbid": 211, "correlation": 16830 + } + }, + { + "ph": "s", "id": 16830, "pid": 435, "tid": 435, "ts": 1776649790543299, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790543342, "dur": 1, + "args": { + "External id": 2057, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16848, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16848, "pid": 0, "tid": 7, "ts": 1776649790543342, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543333, "dur": 6, + "args": { + "External id": 2057, + "cbid": 211, "correlation": 16848 + } + }, + { + "ph": "s", "id": 16848, "pid": 435, "tid": 435, "ts": 1776649790543333, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543345, "dur": 0, + "args": { + "External id": 16854, + "cbid": 200, "correlation": 16854 + } + }, + { + "ph": "f", "id": 16854, "pid": 435, "tid": 435, "ts": 1776649790543345, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543346, "dur": 0, + "args": { + "External id": 16855, + "cbid": 273, "correlation": 16855 + } + }, + { + "ph": "f", "id": 16855, "pid": 435, "tid": 435, "ts": 1776649790543346, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790543355, "dur": 15, + "args": { + "External id": 16856, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16856, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16856, "pid": 0, "tid": 7, "ts": 1776649790543355, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543347, "dur": 4, + "args": { + "External id": 16856, + "cbid": 211, "correlation": 16856 + } + }, + { + "ph": "s", "id": 16856, "pid": 435, "tid": 435, "ts": 1776649790543347, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790543389, "dur": 1, + "args": { + "External id": 2062, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16874, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16874, "pid": 0, "tid": 7, "ts": 1776649790543389, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543380, "dur": 6, + "args": { + "External id": 2062, + "cbid": 211, "correlation": 16874 + } + }, + { + "ph": "s", "id": 16874, "pid": 435, "tid": 435, "ts": 1776649790543380, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543392, "dur": 0, + "args": { + "External id": 16880, + "cbid": 200, "correlation": 16880 + } + }, + { + "ph": "f", "id": 16880, "pid": 435, "tid": 435, "ts": 1776649790543392, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543393, "dur": 0, + "args": { + "External id": 16881, + "cbid": 273, "correlation": 16881 + } + }, + { + "ph": "f", "id": 16881, "pid": 435, "tid": 435, "ts": 1776649790543393, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790543402, "dur": 15, + "args": { + "External id": 16882, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16882, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16882, "pid": 0, "tid": 7, "ts": 1776649790543402, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543394, "dur": 5, + "args": { + "External id": 16882, + "cbid": 211, "correlation": 16882 + } + }, + { + "ph": "s", "id": 16882, "pid": 435, "tid": 435, "ts": 1776649790543394, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790543437, "dur": 1, + "args": { + "External id": 2067, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16900, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16900, "pid": 0, "tid": 7, "ts": 1776649790543437, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543428, "dur": 6, + "args": { + "External id": 2067, + "cbid": 211, "correlation": 16900 + } + }, + { + "ph": "s", "id": 16900, "pid": 435, "tid": 435, "ts": 1776649790543428, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543440, "dur": 0, + "args": { + "External id": 16906, + "cbid": 200, "correlation": 16906 + } + }, + { + "ph": "f", "id": 16906, "pid": 435, "tid": 435, "ts": 1776649790543440, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543441, "dur": 0, + "args": { + "External id": 16907, + "cbid": 273, "correlation": 16907 + } + }, + { + "ph": "f", "id": 16907, "pid": 435, "tid": 435, "ts": 1776649790543441, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790543450, "dur": 15, + "args": { + "External id": 16908, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16908, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16908, "pid": 0, "tid": 7, "ts": 1776649790543450, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543442, "dur": 4, + "args": { + "External id": 16908, + "cbid": 211, "correlation": 16908 + } + }, + { + "ph": "s", "id": 16908, "pid": 435, "tid": 435, "ts": 1776649790543442, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790543485, "dur": 1, + "args": { + "External id": 2072, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16926, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16926, "pid": 0, "tid": 7, "ts": 1776649790543485, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543476, "dur": 6, + "args": { + "External id": 2072, + "cbid": 211, "correlation": 16926 + } + }, + { + "ph": "s", "id": 16926, "pid": 435, "tid": 435, "ts": 1776649790543476, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543488, "dur": 0, + "args": { + "External id": 16932, + "cbid": 200, "correlation": 16932 + } + }, + { + "ph": "f", "id": 16932, "pid": 435, "tid": 435, "ts": 1776649790543488, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543489, "dur": 0, + "args": { + "External id": 16933, + "cbid": 273, "correlation": 16933 + } + }, + { + "ph": "f", "id": 16933, "pid": 435, "tid": 435, "ts": 1776649790543489, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790543498, "dur": 15, + "args": { + "External id": 16934, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16934, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16934, "pid": 0, "tid": 7, "ts": 1776649790543498, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543490, "dur": 5, + "args": { + "External id": 16934, + "cbid": 211, "correlation": 16934 + } + }, + { + "ph": "s", "id": 16934, "pid": 435, "tid": 435, "ts": 1776649790543490, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790543533, "dur": 1, + "args": { + "External id": 2077, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16952, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16952, "pid": 0, "tid": 7, "ts": 1776649790543533, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543524, "dur": 6, + "args": { + "External id": 2077, + "cbid": 211, "correlation": 16952 + } + }, + { + "ph": "s", "id": 16952, "pid": 435, "tid": 435, "ts": 1776649790543524, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543536, "dur": 0, + "args": { + "External id": 16958, + "cbid": 200, "correlation": 16958 + } + }, + { + "ph": "f", "id": 16958, "pid": 435, "tid": 435, "ts": 1776649790543536, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543536, "dur": 0, + "args": { + "External id": 16959, + "cbid": 273, "correlation": 16959 + } + }, + { + "ph": "f", "id": 16959, "pid": 435, "tid": 435, "ts": 1776649790543536, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790543546, "dur": 15, + "args": { + "External id": 16960, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 16960, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 16960, "pid": 0, "tid": 7, "ts": 1776649790543546, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543537, "dur": 5, + "args": { + "External id": 16960, + "cbid": 211, "correlation": 16960 + } + }, + { + "ph": "s", "id": 16960, "pid": 435, "tid": 435, "ts": 1776649790543537, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790543567, "dur": 3, + "args": { + "External id": 16968, + "cbid": 135, "correlation": 16968 + } + }, + { + "ph": "f", "id": 16968, "pid": 435, "tid": 435, "ts": 1776649790543567, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790543582, "dur": 5, + "args": { + "External id": 16973, + "cbid": 165, "correlation": 16973 + } + }, + { + "ph": "s", "id": 16973, "pid": 435, "tid": 435, "ts": 1776649790543582, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649790543590, "dur": 4, + "args": { + "External id": 16976, + "cbid": 139, "correlation": 16976 + } + }, + { + "ph": "f", "id": 16976, "pid": 435, "tid": 435, "ts": 1776649790543590, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790543647, "dur": 1, + "args": { + "External id": 2082, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17004, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17004, "pid": 0, "tid": 7, "ts": 1776649790543647, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543635, "dur": 8, + "args": { + "External id": 2082, + "cbid": 211, "correlation": 17004 + } + }, + { + "ph": "s", "id": 17004, "pid": 435, "tid": 435, "ts": 1776649790543635, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543651, "dur": 0, + "args": { + "External id": 17010, + "cbid": 200, "correlation": 17010 + } + }, + { + "ph": "f", "id": 17010, "pid": 435, "tid": 435, "ts": 1776649790543651, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543652, "dur": 0, + "args": { + "External id": 17011, + "cbid": 273, "correlation": 17011 + } + }, + { + "ph": "f", "id": 17011, "pid": 435, "tid": 435, "ts": 1776649790543652, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790543661, "dur": 17, + "args": { + "External id": 17012, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17012, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17012, "pid": 0, "tid": 7, "ts": 1776649790543661, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543653, "dur": 5, + "args": { + "External id": 17012, + "cbid": 211, "correlation": 17012 + } + }, + { + "ph": "s", "id": 17012, "pid": 435, "tid": 435, "ts": 1776649790543653, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790543700, "dur": 1, + "args": { + "External id": 2087, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17030, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17030, "pid": 0, "tid": 7, "ts": 1776649790543700, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543690, "dur": 7, + "args": { + "External id": 2087, + "cbid": 211, "correlation": 17030 + } + }, + { + "ph": "s", "id": 17030, "pid": 435, "tid": 435, "ts": 1776649790543690, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543703, "dur": 0, + "args": { + "External id": 17036, + "cbid": 200, "correlation": 17036 + } + }, + { + "ph": "f", "id": 17036, "pid": 435, "tid": 435, "ts": 1776649790543703, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543703, "dur": 0, + "args": { + "External id": 17037, + "cbid": 273, "correlation": 17037 + } + }, + { + "ph": "f", "id": 17037, "pid": 435, "tid": 435, "ts": 1776649790543703, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790543712, "dur": 15, + "args": { + "External id": 17038, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17038, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17038, "pid": 0, "tid": 7, "ts": 1776649790543712, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543705, "dur": 4, + "args": { + "External id": 17038, + "cbid": 211, "correlation": 17038 + } + }, + { + "ph": "s", "id": 17038, "pid": 435, "tid": 435, "ts": 1776649790543705, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790543749, "dur": 1, + "args": { + "External id": 2092, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17056, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17056, "pid": 0, "tid": 7, "ts": 1776649790543749, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543740, "dur": 6, + "args": { + "External id": 2092, + "cbid": 211, "correlation": 17056 + } + }, + { + "ph": "s", "id": 17056, "pid": 435, "tid": 435, "ts": 1776649790543740, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543752, "dur": 0, + "args": { + "External id": 17062, + "cbid": 200, "correlation": 17062 + } + }, + { + "ph": "f", "id": 17062, "pid": 435, "tid": 435, "ts": 1776649790543752, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543753, "dur": 0, + "args": { + "External id": 17063, + "cbid": 273, "correlation": 17063 + } + }, + { + "ph": "f", "id": 17063, "pid": 435, "tid": 435, "ts": 1776649790543753, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790543762, "dur": 15, + "args": { + "External id": 17064, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17064, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17064, "pid": 0, "tid": 7, "ts": 1776649790543762, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543754, "dur": 4, + "args": { + "External id": 17064, + "cbid": 211, "correlation": 17064 + } + }, + { + "ph": "s", "id": 17064, "pid": 435, "tid": 435, "ts": 1776649790543754, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790543798, "dur": 1, + "args": { + "External id": 2097, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17082, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17082, "pid": 0, "tid": 7, "ts": 1776649790543798, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543789, "dur": 6, + "args": { + "External id": 2097, + "cbid": 211, "correlation": 17082 + } + }, + { + "ph": "s", "id": 17082, "pid": 435, "tid": 435, "ts": 1776649790543789, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543801, "dur": 0, + "args": { + "External id": 17088, + "cbid": 200, "correlation": 17088 + } + }, + { + "ph": "f", "id": 17088, "pid": 435, "tid": 435, "ts": 1776649790543801, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543802, "dur": 0, + "args": { + "External id": 17089, + "cbid": 273, "correlation": 17089 + } + }, + { + "ph": "f", "id": 17089, "pid": 435, "tid": 435, "ts": 1776649790543802, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790543811, "dur": 15, + "args": { + "External id": 17090, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17090, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17090, "pid": 0, "tid": 7, "ts": 1776649790543811, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543803, "dur": 4, + "args": { + "External id": 17090, + "cbid": 211, "correlation": 17090 + } + }, + { + "ph": "s", "id": 17090, "pid": 435, "tid": 435, "ts": 1776649790543803, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790543846, "dur": 1, + "args": { + "External id": 2102, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17108, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17108, "pid": 0, "tid": 7, "ts": 1776649790543846, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543837, "dur": 7, + "args": { + "External id": 2102, + "cbid": 211, "correlation": 17108 + } + }, + { + "ph": "s", "id": 17108, "pid": 435, "tid": 435, "ts": 1776649790543837, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543850, "dur": 0, + "args": { + "External id": 17114, + "cbid": 200, "correlation": 17114 + } + }, + { + "ph": "f", "id": 17114, "pid": 435, "tid": 435, "ts": 1776649790543850, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543850, "dur": 0, + "args": { + "External id": 17115, + "cbid": 273, "correlation": 17115 + } + }, + { + "ph": "f", "id": 17115, "pid": 435, "tid": 435, "ts": 1776649790543850, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790543859, "dur": 15, + "args": { + "External id": 17116, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17116, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17116, "pid": 0, "tid": 7, "ts": 1776649790543859, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543851, "dur": 4, + "args": { + "External id": 17116, + "cbid": 211, "correlation": 17116 + } + }, + { + "ph": "s", "id": 17116, "pid": 435, "tid": 435, "ts": 1776649790543851, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790543894, "dur": 1, + "args": { + "External id": 2107, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17134, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17134, "pid": 0, "tid": 7, "ts": 1776649790543894, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543885, "dur": 6, + "args": { + "External id": 2107, + "cbid": 211, "correlation": 17134 + } + }, + { + "ph": "s", "id": 17134, "pid": 435, "tid": 435, "ts": 1776649790543885, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543897, "dur": 0, + "args": { + "External id": 17140, + "cbid": 200, "correlation": 17140 + } + }, + { + "ph": "f", "id": 17140, "pid": 435, "tid": 435, "ts": 1776649790543897, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543898, "dur": 0, + "args": { + "External id": 17141, + "cbid": 273, "correlation": 17141 + } + }, + { + "ph": "f", "id": 17141, "pid": 435, "tid": 435, "ts": 1776649790543898, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790543907, "dur": 15, + "args": { + "External id": 17142, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17142, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17142, "pid": 0, "tid": 7, "ts": 1776649790543907, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543899, "dur": 4, + "args": { + "External id": 17142, + "cbid": 211, "correlation": 17142 + } + }, + { + "ph": "s", "id": 17142, "pid": 435, "tid": 435, "ts": 1776649790543899, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790543943, "dur": 1, + "args": { + "External id": 2112, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17160, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17160, "pid": 0, "tid": 7, "ts": 1776649790543943, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543933, "dur": 6, + "args": { + "External id": 2112, + "cbid": 211, "correlation": 17160 + } + }, + { + "ph": "s", "id": 17160, "pid": 435, "tid": 435, "ts": 1776649790543933, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543946, "dur": 0, + "args": { + "External id": 17166, + "cbid": 200, "correlation": 17166 + } + }, + { + "ph": "f", "id": 17166, "pid": 435, "tid": 435, "ts": 1776649790543946, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543946, "dur": 0, + "args": { + "External id": 17167, + "cbid": 273, "correlation": 17167 + } + }, + { + "ph": "f", "id": 17167, "pid": 435, "tid": 435, "ts": 1776649790543946, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790543955, "dur": 15, + "args": { + "External id": 17168, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17168, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17168, "pid": 0, "tid": 7, "ts": 1776649790543955, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543947, "dur": 5, + "args": { + "External id": 17168, + "cbid": 211, "correlation": 17168 + } + }, + { + "ph": "s", "id": 17168, "pid": 435, "tid": 435, "ts": 1776649790543947, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790543991, "dur": 1, + "args": { + "External id": 2117, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17186, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17186, "pid": 0, "tid": 7, "ts": 1776649790543991, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543981, "dur": 6, + "args": { + "External id": 2117, + "cbid": 211, "correlation": 17186 + } + }, + { + "ph": "s", "id": 17186, "pid": 435, "tid": 435, "ts": 1776649790543981, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543994, "dur": 0, + "args": { + "External id": 17192, + "cbid": 200, "correlation": 17192 + } + }, + { + "ph": "f", "id": 17192, "pid": 435, "tid": 435, "ts": 1776649790543994, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790543994, "dur": 0, + "args": { + "External id": 17193, + "cbid": 273, "correlation": 17193 + } + }, + { + "ph": "f", "id": 17193, "pid": 435, "tid": 435, "ts": 1776649790543994, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790544003, "dur": 15, + "args": { + "External id": 17194, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17194, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17194, "pid": 0, "tid": 7, "ts": 1776649790544003, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790543995, "dur": 5, + "args": { + "External id": 17194, + "cbid": 211, "correlation": 17194 + } + }, + { + "ph": "s", "id": 17194, "pid": 435, "tid": 435, "ts": 1776649790543995, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790544039, "dur": 1, + "args": { + "External id": 2122, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17212, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17212, "pid": 0, "tid": 7, "ts": 1776649790544039, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544030, "dur": 6, + "args": { + "External id": 2122, + "cbid": 211, "correlation": 17212 + } + }, + { + "ph": "s", "id": 17212, "pid": 435, "tid": 435, "ts": 1776649790544030, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544042, "dur": 0, + "args": { + "External id": 17218, + "cbid": 200, "correlation": 17218 + } + }, + { + "ph": "f", "id": 17218, "pid": 435, "tid": 435, "ts": 1776649790544042, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544043, "dur": 0, + "args": { + "External id": 17219, + "cbid": 273, "correlation": 17219 + } + }, + { + "ph": "f", "id": 17219, "pid": 435, "tid": 435, "ts": 1776649790544043, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790544052, "dur": 16, + "args": { + "External id": 17220, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17220, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17220, "pid": 0, "tid": 7, "ts": 1776649790544052, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544044, "dur": 4, + "args": { + "External id": 17220, + "cbid": 211, "correlation": 17220 + } + }, + { + "ph": "s", "id": 17220, "pid": 435, "tid": 435, "ts": 1776649790544044, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790544087, "dur": 1, + "args": { + "External id": 2127, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17238, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17238, "pid": 0, "tid": 7, "ts": 1776649790544087, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544078, "dur": 6, + "args": { + "External id": 2127, + "cbid": 211, "correlation": 17238 + } + }, + { + "ph": "s", "id": 17238, "pid": 435, "tid": 435, "ts": 1776649790544078, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544090, "dur": 0, + "args": { + "External id": 17244, + "cbid": 200, "correlation": 17244 + } + }, + { + "ph": "f", "id": 17244, "pid": 435, "tid": 435, "ts": 1776649790544090, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544091, "dur": 0, + "args": { + "External id": 17245, + "cbid": 273, "correlation": 17245 + } + }, + { + "ph": "f", "id": 17245, "pid": 435, "tid": 435, "ts": 1776649790544091, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790544099, "dur": 16, + "args": { + "External id": 17246, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17246, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17246, "pid": 0, "tid": 7, "ts": 1776649790544099, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544092, "dur": 4, + "args": { + "External id": 17246, + "cbid": 211, "correlation": 17246 + } + }, + { + "ph": "s", "id": 17246, "pid": 435, "tid": 435, "ts": 1776649790544092, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790544114, "dur": 6, + "args": { + "External id": 17249, + "cbid": 165, "correlation": 17249 + } + }, + { + "ph": "s", "id": 17249, "pid": 435, "tid": 435, "ts": 1776649790544114, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790544141, "dur": 3, + "args": { + "External id": 17258, + "cbid": 135, "correlation": 17258 + } + }, + { + "ph": "f", "id": 17258, "pid": 435, "tid": 435, "ts": 1776649790544141, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790544190, "dur": 1, + "args": { + "External id": 2132, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17278, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17278, "pid": 0, "tid": 7, "ts": 1776649790544190, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544179, "dur": 8, + "args": { + "External id": 2132, + "cbid": 211, "correlation": 17278 + } + }, + { + "ph": "s", "id": 17278, "pid": 435, "tid": 435, "ts": 1776649790544179, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544194, "dur": 0, + "args": { + "External id": 17284, + "cbid": 200, "correlation": 17284 + } + }, + { + "ph": "f", "id": 17284, "pid": 435, "tid": 435, "ts": 1776649790544194, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544195, "dur": 0, + "args": { + "External id": 17285, + "cbid": 273, "correlation": 17285 + } + }, + { + "ph": "f", "id": 17285, "pid": 435, "tid": 435, "ts": 1776649790544195, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790544204, "dur": 15, + "args": { + "External id": 17286, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17286, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17286, "pid": 0, "tid": 7, "ts": 1776649790544204, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544196, "dur": 5, + "args": { + "External id": 17286, + "cbid": 211, "correlation": 17286 + } + }, + { + "ph": "s", "id": 17286, "pid": 435, "tid": 435, "ts": 1776649790544196, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790544265, "dur": 1, + "args": { + "External id": 2137, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17304, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17304, "pid": 0, "tid": 7, "ts": 1776649790544265, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544254, "dur": 8, + "args": { + "External id": 2137, + "cbid": 211, "correlation": 17304 + } + }, + { + "ph": "s", "id": 17304, "pid": 435, "tid": 435, "ts": 1776649790544254, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544269, "dur": 0, + "args": { + "External id": 17310, + "cbid": 200, "correlation": 17310 + } + }, + { + "ph": "f", "id": 17310, "pid": 435, "tid": 435, "ts": 1776649790544269, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544269, "dur": 0, + "args": { + "External id": 17311, + "cbid": 273, "correlation": 17311 + } + }, + { + "ph": "f", "id": 17311, "pid": 435, "tid": 435, "ts": 1776649790544269, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790544279, "dur": 15, + "args": { + "External id": 17312, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17312, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17312, "pid": 0, "tid": 7, "ts": 1776649790544279, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544271, "dur": 5, + "args": { + "External id": 17312, + "cbid": 211, "correlation": 17312 + } + }, + { + "ph": "s", "id": 17312, "pid": 435, "tid": 435, "ts": 1776649790544271, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790544318, "dur": 1, + "args": { + "External id": 2142, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17330, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17330, "pid": 0, "tid": 7, "ts": 1776649790544318, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544308, "dur": 7, + "args": { + "External id": 2142, + "cbid": 211, "correlation": 17330 + } + }, + { + "ph": "s", "id": 17330, "pid": 435, "tid": 435, "ts": 1776649790544308, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544321, "dur": 0, + "args": { + "External id": 17336, + "cbid": 200, "correlation": 17336 + } + }, + { + "ph": "f", "id": 17336, "pid": 435, "tid": 435, "ts": 1776649790544321, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544322, "dur": 0, + "args": { + "External id": 17337, + "cbid": 273, "correlation": 17337 + } + }, + { + "ph": "f", "id": 17337, "pid": 435, "tid": 435, "ts": 1776649790544322, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790544331, "dur": 15, + "args": { + "External id": 17338, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17338, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17338, "pid": 0, "tid": 7, "ts": 1776649790544331, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544323, "dur": 4, + "args": { + "External id": 17338, + "cbid": 211, "correlation": 17338 + } + }, + { + "ph": "s", "id": 17338, "pid": 435, "tid": 435, "ts": 1776649790544323, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790544367, "dur": 1, + "args": { + "External id": 2147, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17356, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17356, "pid": 0, "tid": 7, "ts": 1776649790544367, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544358, "dur": 6, + "args": { + "External id": 2147, + "cbid": 211, "correlation": 17356 + } + }, + { + "ph": "s", "id": 17356, "pid": 435, "tid": 435, "ts": 1776649790544358, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544370, "dur": 0, + "args": { + "External id": 17362, + "cbid": 200, "correlation": 17362 + } + }, + { + "ph": "f", "id": 17362, "pid": 435, "tid": 435, "ts": 1776649790544370, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544371, "dur": 0, + "args": { + "External id": 17363, + "cbid": 273, "correlation": 17363 + } + }, + { + "ph": "f", "id": 17363, "pid": 435, "tid": 435, "ts": 1776649790544371, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790544380, "dur": 15, + "args": { + "External id": 17364, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17364, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17364, "pid": 0, "tid": 7, "ts": 1776649790544380, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544372, "dur": 5, + "args": { + "External id": 17364, + "cbid": 211, "correlation": 17364 + } + }, + { + "ph": "s", "id": 17364, "pid": 435, "tid": 435, "ts": 1776649790544372, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790544421, "dur": 1, + "args": { + "External id": 2152, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17382, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17382, "pid": 0, "tid": 7, "ts": 1776649790544421, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544411, "dur": 7, + "args": { + "External id": 2152, + "cbid": 211, "correlation": 17382 + } + }, + { + "ph": "s", "id": 17382, "pid": 435, "tid": 435, "ts": 1776649790544411, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544424, "dur": 0, + "args": { + "External id": 17388, + "cbid": 200, "correlation": 17388 + } + }, + { + "ph": "f", "id": 17388, "pid": 435, "tid": 435, "ts": 1776649790544424, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544424, "dur": 0, + "args": { + "External id": 17389, + "cbid": 273, "correlation": 17389 + } + }, + { + "ph": "f", "id": 17389, "pid": 435, "tid": 435, "ts": 1776649790544424, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790544433, "dur": 15, + "args": { + "External id": 17390, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17390, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17390, "pid": 0, "tid": 7, "ts": 1776649790544433, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544425, "dur": 4, + "args": { + "External id": 17390, + "cbid": 211, "correlation": 17390 + } + }, + { + "ph": "s", "id": 17390, "pid": 435, "tid": 435, "ts": 1776649790544425, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790544468, "dur": 1, + "args": { + "External id": 2157, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17408, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17408, "pid": 0, "tid": 7, "ts": 1776649790544468, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544459, "dur": 6, + "args": { + "External id": 2157, + "cbid": 211, "correlation": 17408 + } + }, + { + "ph": "s", "id": 17408, "pid": 435, "tid": 435, "ts": 1776649790544459, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544472, "dur": 0, + "args": { + "External id": 17414, + "cbid": 200, "correlation": 17414 + } + }, + { + "ph": "f", "id": 17414, "pid": 435, "tid": 435, "ts": 1776649790544472, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544472, "dur": 0, + "args": { + "External id": 17415, + "cbid": 273, "correlation": 17415 + } + }, + { + "ph": "f", "id": 17415, "pid": 435, "tid": 435, "ts": 1776649790544472, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790544481, "dur": 15, + "args": { + "External id": 17416, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17416, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17416, "pid": 0, "tid": 7, "ts": 1776649790544481, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544473, "dur": 4, + "args": { + "External id": 17416, + "cbid": 211, "correlation": 17416 + } + }, + { + "ph": "s", "id": 17416, "pid": 435, "tid": 435, "ts": 1776649790544473, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790544516, "dur": 1, + "args": { + "External id": 2162, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17434, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17434, "pid": 0, "tid": 7, "ts": 1776649790544516, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544507, "dur": 6, + "args": { + "External id": 2162, + "cbid": 211, "correlation": 17434 + } + }, + { + "ph": "s", "id": 17434, "pid": 435, "tid": 435, "ts": 1776649790544507, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544520, "dur": 0, + "args": { + "External id": 17440, + "cbid": 200, "correlation": 17440 + } + }, + { + "ph": "f", "id": 17440, "pid": 435, "tid": 435, "ts": 1776649790544520, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544520, "dur": 0, + "args": { + "External id": 17441, + "cbid": 273, "correlation": 17441 + } + }, + { + "ph": "f", "id": 17441, "pid": 435, "tid": 435, "ts": 1776649790544520, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790544530, "dur": 15, + "args": { + "External id": 17442, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17442, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17442, "pid": 0, "tid": 7, "ts": 1776649790544530, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544521, "dur": 4, + "args": { + "External id": 17442, + "cbid": 211, "correlation": 17442 + } + }, + { + "ph": "s", "id": 17442, "pid": 435, "tid": 435, "ts": 1776649790544521, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790544565, "dur": 1, + "args": { + "External id": 2167, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17460, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17460, "pid": 0, "tid": 7, "ts": 1776649790544565, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544555, "dur": 6, + "args": { + "External id": 2167, + "cbid": 211, "correlation": 17460 + } + }, + { + "ph": "s", "id": 17460, "pid": 435, "tid": 435, "ts": 1776649790544555, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544568, "dur": 0, + "args": { + "External id": 17466, + "cbid": 200, "correlation": 17466 + } + }, + { + "ph": "f", "id": 17466, "pid": 435, "tid": 435, "ts": 1776649790544568, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544568, "dur": 0, + "args": { + "External id": 17467, + "cbid": 273, "correlation": 17467 + } + }, + { + "ph": "f", "id": 17467, "pid": 435, "tid": 435, "ts": 1776649790544568, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790544577, "dur": 16, + "args": { + "External id": 17468, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17468, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17468, "pid": 0, "tid": 7, "ts": 1776649790544577, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544569, "dur": 5, + "args": { + "External id": 17468, + "cbid": 211, "correlation": 17468 + } + }, + { + "ph": "s", "id": 17468, "pid": 435, "tid": 435, "ts": 1776649790544569, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790544613, "dur": 1, + "args": { + "External id": 2172, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17486, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17486, "pid": 0, "tid": 7, "ts": 1776649790544613, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544604, "dur": 6, + "args": { + "External id": 2172, + "cbid": 211, "correlation": 17486 + } + }, + { + "ph": "s", "id": 17486, "pid": 435, "tid": 435, "ts": 1776649790544604, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544616, "dur": 0, + "args": { + "External id": 17492, + "cbid": 200, "correlation": 17492 + } + }, + { + "ph": "f", "id": 17492, "pid": 435, "tid": 435, "ts": 1776649790544616, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544617, "dur": 0, + "args": { + "External id": 17493, + "cbid": 273, "correlation": 17493 + } + }, + { + "ph": "f", "id": 17493, "pid": 435, "tid": 435, "ts": 1776649790544617, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790544626, "dur": 15, + "args": { + "External id": 17494, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17494, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17494, "pid": 0, "tid": 7, "ts": 1776649790544626, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544618, "dur": 4, + "args": { + "External id": 17494, + "cbid": 211, "correlation": 17494 + } + }, + { + "ph": "s", "id": 17494, "pid": 435, "tid": 435, "ts": 1776649790544618, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790544661, "dur": 1, + "args": { + "External id": 2177, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17512, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17512, "pid": 0, "tid": 7, "ts": 1776649790544661, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544652, "dur": 6, + "args": { + "External id": 2177, + "cbid": 211, "correlation": 17512 + } + }, + { + "ph": "s", "id": 17512, "pid": 435, "tid": 435, "ts": 1776649790544652, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544664, "dur": 0, + "args": { + "External id": 17518, + "cbid": 200, "correlation": 17518 + } + }, + { + "ph": "f", "id": 17518, "pid": 435, "tid": 435, "ts": 1776649790544664, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544664, "dur": 0, + "args": { + "External id": 17519, + "cbid": 273, "correlation": 17519 + } + }, + { + "ph": "f", "id": 17519, "pid": 435, "tid": 435, "ts": 1776649790544664, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790544673, "dur": 15, + "args": { + "External id": 17520, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17520, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17520, "pid": 0, "tid": 7, "ts": 1776649790544673, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544666, "dur": 4, + "args": { + "External id": 17520, + "cbid": 211, "correlation": 17520 + } + }, + { + "ph": "s", "id": 17520, "pid": 435, "tid": 435, "ts": 1776649790544666, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790544709, "dur": 1, + "args": { + "External id": 2182, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17538, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17538, "pid": 0, "tid": 7, "ts": 1776649790544709, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544700, "dur": 6, + "args": { + "External id": 2182, + "cbid": 211, "correlation": 17538 + } + }, + { + "ph": "s", "id": 17538, "pid": 435, "tid": 435, "ts": 1776649790544700, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544712, "dur": 0, + "args": { + "External id": 17544, + "cbid": 200, "correlation": 17544 + } + }, + { + "ph": "f", "id": 17544, "pid": 435, "tid": 435, "ts": 1776649790544712, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544713, "dur": 0, + "args": { + "External id": 17545, + "cbid": 273, "correlation": 17545 + } + }, + { + "ph": "f", "id": 17545, "pid": 435, "tid": 435, "ts": 1776649790544713, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790544722, "dur": 15, + "args": { + "External id": 17546, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17546, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17546, "pid": 0, "tid": 7, "ts": 1776649790544722, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544714, "dur": 5, + "args": { + "External id": 17546, + "cbid": 211, "correlation": 17546 + } + }, + { + "ph": "s", "id": 17546, "pid": 435, "tid": 435, "ts": 1776649790544714, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790544758, "dur": 1, + "args": { + "External id": 2187, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17564, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17564, "pid": 0, "tid": 7, "ts": 1776649790544758, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544748, "dur": 6, + "args": { + "External id": 2187, + "cbid": 211, "correlation": 17564 + } + }, + { + "ph": "s", "id": 17564, "pid": 435, "tid": 435, "ts": 1776649790544748, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544761, "dur": 0, + "args": { + "External id": 17570, + "cbid": 200, "correlation": 17570 + } + }, + { + "ph": "f", "id": 17570, "pid": 435, "tid": 435, "ts": 1776649790544761, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544761, "dur": 0, + "args": { + "External id": 17571, + "cbid": 273, "correlation": 17571 + } + }, + { + "ph": "f", "id": 17571, "pid": 435, "tid": 435, "ts": 1776649790544761, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790544770, "dur": 15, + "args": { + "External id": 17572, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17572, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17572, "pid": 0, "tid": 7, "ts": 1776649790544770, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544762, "dur": 4, + "args": { + "External id": 17572, + "cbid": 211, "correlation": 17572 + } + }, + { + "ph": "s", "id": 17572, "pid": 435, "tid": 435, "ts": 1776649790544762, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790544805, "dur": 1, + "args": { + "External id": 2192, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17590, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17590, "pid": 0, "tid": 7, "ts": 1776649790544805, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544796, "dur": 6, + "args": { + "External id": 2192, + "cbid": 211, "correlation": 17590 + } + }, + { + "ph": "s", "id": 17590, "pid": 435, "tid": 435, "ts": 1776649790544796, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544808, "dur": 0, + "args": { + "External id": 17596, + "cbid": 200, "correlation": 17596 + } + }, + { + "ph": "f", "id": 17596, "pid": 435, "tid": 435, "ts": 1776649790544808, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544809, "dur": 0, + "args": { + "External id": 17597, + "cbid": 273, "correlation": 17597 + } + }, + { + "ph": "f", "id": 17597, "pid": 435, "tid": 435, "ts": 1776649790544809, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790544818, "dur": 15, + "args": { + "External id": 17598, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17598, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17598, "pid": 0, "tid": 7, "ts": 1776649790544818, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544810, "dur": 5, + "args": { + "External id": 17598, + "cbid": 211, "correlation": 17598 + } + }, + { + "ph": "s", "id": 17598, "pid": 435, "tid": 435, "ts": 1776649790544810, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790544854, "dur": 1, + "args": { + "External id": 2197, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17616, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17616, "pid": 0, "tid": 7, "ts": 1776649790544854, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544844, "dur": 6, + "args": { + "External id": 2197, + "cbid": 211, "correlation": 17616 + } + }, + { + "ph": "s", "id": 17616, "pid": 435, "tid": 435, "ts": 1776649790544844, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544857, "dur": 0, + "args": { + "External id": 17622, + "cbid": 200, "correlation": 17622 + } + }, + { + "ph": "f", "id": 17622, "pid": 435, "tid": 435, "ts": 1776649790544857, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544857, "dur": 0, + "args": { + "External id": 17623, + "cbid": 273, "correlation": 17623 + } + }, + { + "ph": "f", "id": 17623, "pid": 435, "tid": 435, "ts": 1776649790544857, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790544866, "dur": 15, + "args": { + "External id": 17624, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17624, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17624, "pid": 0, "tid": 7, "ts": 1776649790544866, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544858, "dur": 4, + "args": { + "External id": 17624, + "cbid": 211, "correlation": 17624 + } + }, + { + "ph": "s", "id": 17624, "pid": 435, "tid": 435, "ts": 1776649790544858, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790544901, "dur": 1, + "args": { + "External id": 2202, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17642, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17642, "pid": 0, "tid": 7, "ts": 1776649790544901, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544892, "dur": 6, + "args": { + "External id": 2202, + "cbid": 211, "correlation": 17642 + } + }, + { + "ph": "s", "id": 17642, "pid": 435, "tid": 435, "ts": 1776649790544892, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544904, "dur": 0, + "args": { + "External id": 17648, + "cbid": 200, "correlation": 17648 + } + }, + { + "ph": "f", "id": 17648, "pid": 435, "tid": 435, "ts": 1776649790544904, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544905, "dur": 0, + "args": { + "External id": 17649, + "cbid": 273, "correlation": 17649 + } + }, + { + "ph": "f", "id": 17649, "pid": 435, "tid": 435, "ts": 1776649790544905, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790544914, "dur": 15, + "args": { + "External id": 17650, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17650, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17650, "pid": 0, "tid": 7, "ts": 1776649790544914, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544906, "dur": 4, + "args": { + "External id": 17650, + "cbid": 211, "correlation": 17650 + } + }, + { + "ph": "s", "id": 17650, "pid": 435, "tid": 435, "ts": 1776649790544906, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790544949, "dur": 1, + "args": { + "External id": 2207, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17668, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17668, "pid": 0, "tid": 7, "ts": 1776649790544949, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544940, "dur": 6, + "args": { + "External id": 2207, + "cbid": 211, "correlation": 17668 + } + }, + { + "ph": "s", "id": 17668, "pid": 435, "tid": 435, "ts": 1776649790544940, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544952, "dur": 0, + "args": { + "External id": 17674, + "cbid": 200, "correlation": 17674 + } + }, + { + "ph": "f", "id": 17674, "pid": 435, "tid": 435, "ts": 1776649790544952, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790544953, "dur": 0, + "args": { + "External id": 17675, + "cbid": 273, "correlation": 17675 + } + }, + { + "ph": "f", "id": 17675, "pid": 435, "tid": 435, "ts": 1776649790544953, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790544962, "dur": 17, + "args": { + "External id": 17676, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17676, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17676, "pid": 0, "tid": 7, "ts": 1776649790544962, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544954, "dur": 4, + "args": { + "External id": 17676, + "cbid": 211, "correlation": 17676 + } + }, + { + "ph": "s", "id": 17676, "pid": 435, "tid": 435, "ts": 1776649790544954, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790544997, "dur": 1, + "args": { + "External id": 2212, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17694, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17694, "pid": 0, "tid": 7, "ts": 1776649790544997, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790544988, "dur": 6, + "args": { + "External id": 2212, + "cbid": 211, "correlation": 17694 + } + }, + { + "ph": "s", "id": 17694, "pid": 435, "tid": 435, "ts": 1776649790544988, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545000, "dur": 0, + "args": { + "External id": 17700, + "cbid": 200, "correlation": 17700 + } + }, + { + "ph": "f", "id": 17700, "pid": 435, "tid": 435, "ts": 1776649790545000, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545001, "dur": 0, + "args": { + "External id": 17701, + "cbid": 273, "correlation": 17701 + } + }, + { + "ph": "f", "id": 17701, "pid": 435, "tid": 435, "ts": 1776649790545001, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790545010, "dur": 16, + "args": { + "External id": 17702, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17702, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17702, "pid": 0, "tid": 7, "ts": 1776649790545010, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545002, "dur": 5, + "args": { + "External id": 17702, + "cbid": 211, "correlation": 17702 + } + }, + { + "ph": "s", "id": 17702, "pid": 435, "tid": 435, "ts": 1776649790545002, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790545045, "dur": 1, + "args": { + "External id": 2217, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17720, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17720, "pid": 0, "tid": 7, "ts": 1776649790545045, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545036, "dur": 6, + "args": { + "External id": 2217, + "cbid": 211, "correlation": 17720 + } + }, + { + "ph": "s", "id": 17720, "pid": 435, "tid": 435, "ts": 1776649790545036, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545048, "dur": 0, + "args": { + "External id": 17726, + "cbid": 200, "correlation": 17726 + } + }, + { + "ph": "f", "id": 17726, "pid": 435, "tid": 435, "ts": 1776649790545048, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545048, "dur": 0, + "args": { + "External id": 17727, + "cbid": 273, "correlation": 17727 + } + }, + { + "ph": "f", "id": 17727, "pid": 435, "tid": 435, "ts": 1776649790545048, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790545057, "dur": 15, + "args": { + "External id": 17728, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17728, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17728, "pid": 0, "tid": 7, "ts": 1776649790545057, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545049, "dur": 4, + "args": { + "External id": 17728, + "cbid": 211, "correlation": 17728 + } + }, + { + "ph": "s", "id": 17728, "pid": 435, "tid": 435, "ts": 1776649790545049, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790545093, "dur": 1, + "args": { + "External id": 2222, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17746, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17746, "pid": 0, "tid": 7, "ts": 1776649790545093, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545084, "dur": 6, + "args": { + "External id": 2222, + "cbid": 211, "correlation": 17746 + } + }, + { + "ph": "s", "id": 17746, "pid": 435, "tid": 435, "ts": 1776649790545084, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545096, "dur": 0, + "args": { + "External id": 17752, + "cbid": 200, "correlation": 17752 + } + }, + { + "ph": "f", "id": 17752, "pid": 435, "tid": 435, "ts": 1776649790545096, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545097, "dur": 0, + "args": { + "External id": 17753, + "cbid": 273, "correlation": 17753 + } + }, + { + "ph": "f", "id": 17753, "pid": 435, "tid": 435, "ts": 1776649790545097, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790545106, "dur": 15, + "args": { + "External id": 17754, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17754, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17754, "pid": 0, "tid": 7, "ts": 1776649790545106, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545098, "dur": 4, + "args": { + "External id": 17754, + "cbid": 211, "correlation": 17754 + } + }, + { + "ph": "s", "id": 17754, "pid": 435, "tid": 435, "ts": 1776649790545098, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790545141, "dur": 1, + "args": { + "External id": 2227, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17772, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17772, "pid": 0, "tid": 7, "ts": 1776649790545141, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545132, "dur": 6, + "args": { + "External id": 2227, + "cbid": 211, "correlation": 17772 + } + }, + { + "ph": "s", "id": 17772, "pid": 435, "tid": 435, "ts": 1776649790545132, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545144, "dur": 0, + "args": { + "External id": 17778, + "cbid": 200, "correlation": 17778 + } + }, + { + "ph": "f", "id": 17778, "pid": 435, "tid": 435, "ts": 1776649790545144, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545144, "dur": 0, + "args": { + "External id": 17779, + "cbid": 273, "correlation": 17779 + } + }, + { + "ph": "f", "id": 17779, "pid": 435, "tid": 435, "ts": 1776649790545144, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790545153, "dur": 15, + "args": { + "External id": 17780, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17780, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17780, "pid": 0, "tid": 7, "ts": 1776649790545153, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545145, "dur": 5, + "args": { + "External id": 17780, + "cbid": 211, "correlation": 17780 + } + }, + { + "ph": "s", "id": 17780, "pid": 435, "tid": 435, "ts": 1776649790545145, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790545189, "dur": 1, + "args": { + "External id": 2232, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17798, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17798, "pid": 0, "tid": 7, "ts": 1776649790545189, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545180, "dur": 6, + "args": { + "External id": 2232, + "cbid": 211, "correlation": 17798 + } + }, + { + "ph": "s", "id": 17798, "pid": 435, "tid": 435, "ts": 1776649790545180, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545192, "dur": 0, + "args": { + "External id": 17804, + "cbid": 200, "correlation": 17804 + } + }, + { + "ph": "f", "id": 17804, "pid": 435, "tid": 435, "ts": 1776649790545192, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545193, "dur": 0, + "args": { + "External id": 17805, + "cbid": 273, "correlation": 17805 + } + }, + { + "ph": "f", "id": 17805, "pid": 435, "tid": 435, "ts": 1776649790545193, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790545202, "dur": 15, + "args": { + "External id": 17806, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17806, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17806, "pid": 0, "tid": 7, "ts": 1776649790545202, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545194, "dur": 4, + "args": { + "External id": 17806, + "cbid": 211, "correlation": 17806 + } + }, + { + "ph": "s", "id": 17806, "pid": 435, "tid": 435, "ts": 1776649790545194, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790545237, "dur": 1, + "args": { + "External id": 2237, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17824, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17824, "pid": 0, "tid": 7, "ts": 1776649790545237, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545228, "dur": 6, + "args": { + "External id": 2237, + "cbid": 211, "correlation": 17824 + } + }, + { + "ph": "s", "id": 17824, "pid": 435, "tid": 435, "ts": 1776649790545228, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545240, "dur": 0, + "args": { + "External id": 17830, + "cbid": 200, "correlation": 17830 + } + }, + { + "ph": "f", "id": 17830, "pid": 435, "tid": 435, "ts": 1776649790545240, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545241, "dur": 0, + "args": { + "External id": 17831, + "cbid": 273, "correlation": 17831 + } + }, + { + "ph": "f", "id": 17831, "pid": 435, "tid": 435, "ts": 1776649790545241, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790545250, "dur": 15, + "args": { + "External id": 17832, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17832, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17832, "pid": 0, "tid": 7, "ts": 1776649790545250, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545242, "dur": 4, + "args": { + "External id": 17832, + "cbid": 211, "correlation": 17832 + } + }, + { + "ph": "s", "id": 17832, "pid": 435, "tid": 435, "ts": 1776649790545242, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790545285, "dur": 1, + "args": { + "External id": 2242, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17850, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17850, "pid": 0, "tid": 7, "ts": 1776649790545285, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545276, "dur": 6, + "args": { + "External id": 2242, + "cbid": 211, "correlation": 17850 + } + }, + { + "ph": "s", "id": 17850, "pid": 435, "tid": 435, "ts": 1776649790545276, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545288, "dur": 0, + "args": { + "External id": 17856, + "cbid": 200, "correlation": 17856 + } + }, + { + "ph": "f", "id": 17856, "pid": 435, "tid": 435, "ts": 1776649790545288, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545289, "dur": 0, + "args": { + "External id": 17857, + "cbid": 273, "correlation": 17857 + } + }, + { + "ph": "f", "id": 17857, "pid": 435, "tid": 435, "ts": 1776649790545289, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790545298, "dur": 15, + "args": { + "External id": 17858, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17858, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17858, "pid": 0, "tid": 7, "ts": 1776649790545298, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545290, "dur": 4, + "args": { + "External id": 17858, + "cbid": 211, "correlation": 17858 + } + }, + { + "ph": "s", "id": 17858, "pid": 435, "tid": 435, "ts": 1776649790545290, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790545333, "dur": 1, + "args": { + "External id": 2247, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17876, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17876, "pid": 0, "tid": 7, "ts": 1776649790545333, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545324, "dur": 6, + "args": { + "External id": 2247, + "cbid": 211, "correlation": 17876 + } + }, + { + "ph": "s", "id": 17876, "pid": 435, "tid": 435, "ts": 1776649790545324, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545336, "dur": 0, + "args": { + "External id": 17882, + "cbid": 200, "correlation": 17882 + } + }, + { + "ph": "f", "id": 17882, "pid": 435, "tid": 435, "ts": 1776649790545336, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545337, "dur": 0, + "args": { + "External id": 17883, + "cbid": 273, "correlation": 17883 + } + }, + { + "ph": "f", "id": 17883, "pid": 435, "tid": 435, "ts": 1776649790545337, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790545345, "dur": 15, + "args": { + "External id": 17884, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17884, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17884, "pid": 0, "tid": 7, "ts": 1776649790545345, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545338, "dur": 5, + "args": { + "External id": 17884, + "cbid": 211, "correlation": 17884 + } + }, + { + "ph": "s", "id": 17884, "pid": 435, "tid": 435, "ts": 1776649790545338, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790545381, "dur": 1, + "args": { + "External id": 2252, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17902, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17902, "pid": 0, "tid": 7, "ts": 1776649790545381, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545372, "dur": 6, + "args": { + "External id": 2252, + "cbid": 211, "correlation": 17902 + } + }, + { + "ph": "s", "id": 17902, "pid": 435, "tid": 435, "ts": 1776649790545372, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545384, "dur": 0, + "args": { + "External id": 17908, + "cbid": 200, "correlation": 17908 + } + }, + { + "ph": "f", "id": 17908, "pid": 435, "tid": 435, "ts": 1776649790545384, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545385, "dur": 0, + "args": { + "External id": 17909, + "cbid": 273, "correlation": 17909 + } + }, + { + "ph": "f", "id": 17909, "pid": 435, "tid": 435, "ts": 1776649790545385, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790545394, "dur": 16, + "args": { + "External id": 17910, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17910, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17910, "pid": 0, "tid": 7, "ts": 1776649790545394, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545386, "dur": 4, + "args": { + "External id": 17910, + "cbid": 211, "correlation": 17910 + } + }, + { + "ph": "s", "id": 17910, "pid": 435, "tid": 435, "ts": 1776649790545386, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790545428, "dur": 1, + "args": { + "External id": 2257, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17928, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17928, "pid": 0, "tid": 7, "ts": 1776649790545428, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545419, "dur": 6, + "args": { + "External id": 2257, + "cbid": 211, "correlation": 17928 + } + }, + { + "ph": "s", "id": 17928, "pid": 435, "tid": 435, "ts": 1776649790545419, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545432, "dur": 0, + "args": { + "External id": 17934, + "cbid": 200, "correlation": 17934 + } + }, + { + "ph": "f", "id": 17934, "pid": 435, "tid": 435, "ts": 1776649790545432, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545432, "dur": 0, + "args": { + "External id": 17935, + "cbid": 273, "correlation": 17935 + } + }, + { + "ph": "f", "id": 17935, "pid": 435, "tid": 435, "ts": 1776649790545432, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790545441, "dur": 16, + "args": { + "External id": 17936, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17936, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17936, "pid": 0, "tid": 7, "ts": 1776649790545441, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545433, "dur": 4, + "args": { + "External id": 17936, + "cbid": 211, "correlation": 17936 + } + }, + { + "ph": "s", "id": 17936, "pid": 435, "tid": 435, "ts": 1776649790545433, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790545477, "dur": 1, + "args": { + "External id": 2262, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17954, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17954, "pid": 0, "tid": 7, "ts": 1776649790545477, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545467, "dur": 6, + "args": { + "External id": 2262, + "cbid": 211, "correlation": 17954 + } + }, + { + "ph": "s", "id": 17954, "pid": 435, "tid": 435, "ts": 1776649790545467, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545480, "dur": 0, + "args": { + "External id": 17960, + "cbid": 200, "correlation": 17960 + } + }, + { + "ph": "f", "id": 17960, "pid": 435, "tid": 435, "ts": 1776649790545480, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545480, "dur": 0, + "args": { + "External id": 17961, + "cbid": 273, "correlation": 17961 + } + }, + { + "ph": "f", "id": 17961, "pid": 435, "tid": 435, "ts": 1776649790545480, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790545490, "dur": 15, + "args": { + "External id": 17962, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17962, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17962, "pid": 0, "tid": 7, "ts": 1776649790545490, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545481, "dur": 4, + "args": { + "External id": 17962, + "cbid": 211, "correlation": 17962 + } + }, + { + "ph": "s", "id": 17962, "pid": 435, "tid": 435, "ts": 1776649790545481, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790545525, "dur": 1, + "args": { + "External id": 2267, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17980, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17980, "pid": 0, "tid": 7, "ts": 1776649790545525, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545515, "dur": 6, + "args": { + "External id": 2267, + "cbid": 211, "correlation": 17980 + } + }, + { + "ph": "s", "id": 17980, "pid": 435, "tid": 435, "ts": 1776649790545515, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545527, "dur": 0, + "args": { + "External id": 17986, + "cbid": 200, "correlation": 17986 + } + }, + { + "ph": "f", "id": 17986, "pid": 435, "tid": 435, "ts": 1776649790545527, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545528, "dur": 0, + "args": { + "External id": 17987, + "cbid": 273, "correlation": 17987 + } + }, + { + "ph": "f", "id": 17987, "pid": 435, "tid": 435, "ts": 1776649790545528, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790545538, "dur": 15, + "args": { + "External id": 17988, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 17988, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 17988, "pid": 0, "tid": 7, "ts": 1776649790545538, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545529, "dur": 4, + "args": { + "External id": 17988, + "cbid": 211, "correlation": 17988 + } + }, + { + "ph": "s", "id": 17988, "pid": 435, "tid": 435, "ts": 1776649790545529, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790545572, "dur": 1, + "args": { + "External id": 2272, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18006, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18006, "pid": 0, "tid": 7, "ts": 1776649790545572, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545563, "dur": 6, + "args": { + "External id": 2272, + "cbid": 211, "correlation": 18006 + } + }, + { + "ph": "s", "id": 18006, "pid": 435, "tid": 435, "ts": 1776649790545563, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545576, "dur": 0, + "args": { + "External id": 18012, + "cbid": 200, "correlation": 18012 + } + }, + { + "ph": "f", "id": 18012, "pid": 435, "tid": 435, "ts": 1776649790545576, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545576, "dur": 0, + "args": { + "External id": 18013, + "cbid": 273, "correlation": 18013 + } + }, + { + "ph": "f", "id": 18013, "pid": 435, "tid": 435, "ts": 1776649790545576, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790545585, "dur": 15, + "args": { + "External id": 18014, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18014, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18014, "pid": 0, "tid": 7, "ts": 1776649790545585, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545577, "dur": 4, + "args": { + "External id": 18014, + "cbid": 211, "correlation": 18014 + } + }, + { + "ph": "s", "id": 18014, "pid": 435, "tid": 435, "ts": 1776649790545577, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790545623, "dur": 1, + "args": { + "External id": 2277, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18032, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18032, "pid": 0, "tid": 7, "ts": 1776649790545623, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545614, "dur": 6, + "args": { + "External id": 2277, + "cbid": 211, "correlation": 18032 + } + }, + { + "ph": "s", "id": 18032, "pid": 435, "tid": 435, "ts": 1776649790545614, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545627, "dur": 0, + "args": { + "External id": 18038, + "cbid": 200, "correlation": 18038 + } + }, + { + "ph": "f", "id": 18038, "pid": 435, "tid": 435, "ts": 1776649790545627, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545627, "dur": 0, + "args": { + "External id": 18039, + "cbid": 273, "correlation": 18039 + } + }, + { + "ph": "f", "id": 18039, "pid": 435, "tid": 435, "ts": 1776649790545627, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790545636, "dur": 15, + "args": { + "External id": 18040, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18040, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18040, "pid": 0, "tid": 7, "ts": 1776649790545636, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545629, "dur": 5, + "args": { + "External id": 18040, + "cbid": 211, "correlation": 18040 + } + }, + { + "ph": "s", "id": 18040, "pid": 435, "tid": 435, "ts": 1776649790545629, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790545672, "dur": 1, + "args": { + "External id": 2282, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18058, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18058, "pid": 0, "tid": 7, "ts": 1776649790545672, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545663, "dur": 6, + "args": { + "External id": 2282, + "cbid": 211, "correlation": 18058 + } + }, + { + "ph": "s", "id": 18058, "pid": 435, "tid": 435, "ts": 1776649790545663, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545675, "dur": 0, + "args": { + "External id": 18064, + "cbid": 200, "correlation": 18064 + } + }, + { + "ph": "f", "id": 18064, "pid": 435, "tid": 435, "ts": 1776649790545675, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545676, "dur": 0, + "args": { + "External id": 18065, + "cbid": 273, "correlation": 18065 + } + }, + { + "ph": "f", "id": 18065, "pid": 435, "tid": 435, "ts": 1776649790545676, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790545685, "dur": 15, + "args": { + "External id": 18066, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18066, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18066, "pid": 0, "tid": 7, "ts": 1776649790545685, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545677, "dur": 5, + "args": { + "External id": 18066, + "cbid": 211, "correlation": 18066 + } + }, + { + "ph": "s", "id": 18066, "pid": 435, "tid": 435, "ts": 1776649790545677, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790545720, "dur": 1, + "args": { + "External id": 2287, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18084, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18084, "pid": 0, "tid": 7, "ts": 1776649790545720, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545711, "dur": 6, + "args": { + "External id": 2287, + "cbid": 211, "correlation": 18084 + } + }, + { + "ph": "s", "id": 18084, "pid": 435, "tid": 435, "ts": 1776649790545711, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545723, "dur": 0, + "args": { + "External id": 18090, + "cbid": 200, "correlation": 18090 + } + }, + { + "ph": "f", "id": 18090, "pid": 435, "tid": 435, "ts": 1776649790545723, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545724, "dur": 0, + "args": { + "External id": 18091, + "cbid": 273, "correlation": 18091 + } + }, + { + "ph": "f", "id": 18091, "pid": 435, "tid": 435, "ts": 1776649790545724, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790545733, "dur": 15, + "args": { + "External id": 18092, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18092, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18092, "pid": 0, "tid": 7, "ts": 1776649790545733, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545725, "dur": 5, + "args": { + "External id": 18092, + "cbid": 211, "correlation": 18092 + } + }, + { + "ph": "s", "id": 18092, "pid": 435, "tid": 435, "ts": 1776649790545725, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790545768, "dur": 1, + "args": { + "External id": 2292, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18110, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18110, "pid": 0, "tid": 7, "ts": 1776649790545768, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545759, "dur": 6, + "args": { + "External id": 2292, + "cbid": 211, "correlation": 18110 + } + }, + { + "ph": "s", "id": 18110, "pid": 435, "tid": 435, "ts": 1776649790545759, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545771, "dur": 0, + "args": { + "External id": 18116, + "cbid": 200, "correlation": 18116 + } + }, + { + "ph": "f", "id": 18116, "pid": 435, "tid": 435, "ts": 1776649790545771, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545772, "dur": 0, + "args": { + "External id": 18117, + "cbid": 273, "correlation": 18117 + } + }, + { + "ph": "f", "id": 18117, "pid": 435, "tid": 435, "ts": 1776649790545772, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790545781, "dur": 15, + "args": { + "External id": 18118, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18118, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18118, "pid": 0, "tid": 7, "ts": 1776649790545781, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545773, "dur": 4, + "args": { + "External id": 18118, + "cbid": 211, "correlation": 18118 + } + }, + { + "ph": "s", "id": 18118, "pid": 435, "tid": 435, "ts": 1776649790545773, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790545817, "dur": 1, + "args": { + "External id": 2297, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18136, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18136, "pid": 0, "tid": 7, "ts": 1776649790545817, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545807, "dur": 6, + "args": { + "External id": 2297, + "cbid": 211, "correlation": 18136 + } + }, + { + "ph": "s", "id": 18136, "pid": 435, "tid": 435, "ts": 1776649790545807, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545820, "dur": 0, + "args": { + "External id": 18142, + "cbid": 200, "correlation": 18142 + } + }, + { + "ph": "f", "id": 18142, "pid": 435, "tid": 435, "ts": 1776649790545820, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545820, "dur": 0, + "args": { + "External id": 18143, + "cbid": 273, "correlation": 18143 + } + }, + { + "ph": "f", "id": 18143, "pid": 435, "tid": 435, "ts": 1776649790545820, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790545829, "dur": 16, + "args": { + "External id": 18144, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18144, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18144, "pid": 0, "tid": 7, "ts": 1776649790545829, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545821, "dur": 4, + "args": { + "External id": 18144, + "cbid": 211, "correlation": 18144 + } + }, + { + "ph": "s", "id": 18144, "pid": 435, "tid": 435, "ts": 1776649790545821, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790545865, "dur": 1, + "args": { + "External id": 2302, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18162, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18162, "pid": 0, "tid": 7, "ts": 1776649790545865, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545856, "dur": 6, + "args": { + "External id": 2302, + "cbid": 211, "correlation": 18162 + } + }, + { + "ph": "s", "id": 18162, "pid": 435, "tid": 435, "ts": 1776649790545856, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545868, "dur": 0, + "args": { + "External id": 18168, + "cbid": 200, "correlation": 18168 + } + }, + { + "ph": "f", "id": 18168, "pid": 435, "tid": 435, "ts": 1776649790545868, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545869, "dur": 0, + "args": { + "External id": 18169, + "cbid": 273, "correlation": 18169 + } + }, + { + "ph": "f", "id": 18169, "pid": 435, "tid": 435, "ts": 1776649790545869, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790545878, "dur": 15, + "args": { + "External id": 18170, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18170, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18170, "pid": 0, "tid": 7, "ts": 1776649790545878, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545870, "dur": 4, + "args": { + "External id": 18170, + "cbid": 211, "correlation": 18170 + } + }, + { + "ph": "s", "id": 18170, "pid": 435, "tid": 435, "ts": 1776649790545870, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790545914, "dur": 1, + "args": { + "External id": 2307, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18188, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18188, "pid": 0, "tid": 7, "ts": 1776649790545914, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545904, "dur": 6, + "args": { + "External id": 2307, + "cbid": 211, "correlation": 18188 + } + }, + { + "ph": "s", "id": 18188, "pid": 435, "tid": 435, "ts": 1776649790545904, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545917, "dur": 0, + "args": { + "External id": 18194, + "cbid": 200, "correlation": 18194 + } + }, + { + "ph": "f", "id": 18194, "pid": 435, "tid": 435, "ts": 1776649790545917, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545917, "dur": 0, + "args": { + "External id": 18195, + "cbid": 273, "correlation": 18195 + } + }, + { + "ph": "f", "id": 18195, "pid": 435, "tid": 435, "ts": 1776649790545917, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790545926, "dur": 15, + "args": { + "External id": 18196, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18196, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18196, "pid": 0, "tid": 7, "ts": 1776649790545926, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545918, "dur": 4, + "args": { + "External id": 18196, + "cbid": 211, "correlation": 18196 + } + }, + { + "ph": "s", "id": 18196, "pid": 435, "tid": 435, "ts": 1776649790545918, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790545962, "dur": 1, + "args": { + "External id": 2312, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18214, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18214, "pid": 0, "tid": 7, "ts": 1776649790545962, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545952, "dur": 6, + "args": { + "External id": 2312, + "cbid": 211, "correlation": 18214 + } + }, + { + "ph": "s", "id": 18214, "pid": 435, "tid": 435, "ts": 1776649790545952, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545964, "dur": 0, + "args": { + "External id": 18220, + "cbid": 200, "correlation": 18220 + } + }, + { + "ph": "f", "id": 18220, "pid": 435, "tid": 435, "ts": 1776649790545964, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790545965, "dur": 0, + "args": { + "External id": 18221, + "cbid": 273, "correlation": 18221 + } + }, + { + "ph": "f", "id": 18221, "pid": 435, "tid": 435, "ts": 1776649790545965, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790545974, "dur": 15, + "args": { + "External id": 18222, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18222, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18222, "pid": 0, "tid": 7, "ts": 1776649790545974, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790545966, "dur": 5, + "args": { + "External id": 18222, + "cbid": 211, "correlation": 18222 + } + }, + { + "ph": "s", "id": 18222, "pid": 435, "tid": 435, "ts": 1776649790545966, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790546010, "dur": 1, + "args": { + "External id": 2317, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18240, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18240, "pid": 0, "tid": 7, "ts": 1776649790546010, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546000, "dur": 6, + "args": { + "External id": 2317, + "cbid": 211, "correlation": 18240 + } + }, + { + "ph": "s", "id": 18240, "pid": 435, "tid": 435, "ts": 1776649790546000, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790546013, "dur": 0, + "args": { + "External id": 18246, + "cbid": 200, "correlation": 18246 + } + }, + { + "ph": "f", "id": 18246, "pid": 435, "tid": 435, "ts": 1776649790546013, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790546013, "dur": 0, + "args": { + "External id": 18247, + "cbid": 273, "correlation": 18247 + } + }, + { + "ph": "f", "id": 18247, "pid": 435, "tid": 435, "ts": 1776649790546013, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790546022, "dur": 15, + "args": { + "External id": 18248, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18248, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18248, "pid": 0, "tid": 7, "ts": 1776649790546022, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546014, "dur": 4, + "args": { + "External id": 18248, + "cbid": 211, "correlation": 18248 + } + }, + { + "ph": "s", "id": 18248, "pid": 435, "tid": 435, "ts": 1776649790546014, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790546057, "dur": 1, + "args": { + "External id": 2322, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18266, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18266, "pid": 0, "tid": 7, "ts": 1776649790546057, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546048, "dur": 6, + "args": { + "External id": 2322, + "cbid": 211, "correlation": 18266 + } + }, + { + "ph": "s", "id": 18266, "pid": 435, "tid": 435, "ts": 1776649790546048, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790546061, "dur": 0, + "args": { + "External id": 18272, + "cbid": 200, "correlation": 18272 + } + }, + { + "ph": "f", "id": 18272, "pid": 435, "tid": 435, "ts": 1776649790546061, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790546062, "dur": 0, + "args": { + "External id": 18273, + "cbid": 273, "correlation": 18273 + } + }, + { + "ph": "f", "id": 18273, "pid": 435, "tid": 435, "ts": 1776649790546062, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790546071, "dur": 15, + "args": { + "External id": 18274, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18274, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18274, "pid": 0, "tid": 7, "ts": 1776649790546071, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546063, "dur": 4, + "args": { + "External id": 18274, + "cbid": 211, "correlation": 18274 + } + }, + { + "ph": "s", "id": 18274, "pid": 435, "tid": 435, "ts": 1776649790546063, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790546105, "dur": 1, + "args": { + "External id": 2327, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18292, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18292, "pid": 0, "tid": 7, "ts": 1776649790546105, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546096, "dur": 6, + "args": { + "External id": 2327, + "cbid": 211, "correlation": 18292 + } + }, + { + "ph": "s", "id": 18292, "pid": 435, "tid": 435, "ts": 1776649790546096, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790546109, "dur": 0, + "args": { + "External id": 18298, + "cbid": 200, "correlation": 18298 + } + }, + { + "ph": "f", "id": 18298, "pid": 435, "tid": 435, "ts": 1776649790546109, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790546110, "dur": 0, + "args": { + "External id": 18299, + "cbid": 273, "correlation": 18299 + } + }, + { + "ph": "f", "id": 18299, "pid": 435, "tid": 435, "ts": 1776649790546110, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790546119, "dur": 15, + "args": { + "External id": 18300, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18300, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18300, "pid": 0, "tid": 7, "ts": 1776649790546119, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546111, "dur": 4, + "args": { + "External id": 18300, + "cbid": 211, "correlation": 18300 + } + }, + { + "ph": "s", "id": 18300, "pid": 435, "tid": 435, "ts": 1776649790546111, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790546155, "dur": 1, + "args": { + "External id": 2332, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18318, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18318, "pid": 0, "tid": 7, "ts": 1776649790546155, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546145, "dur": 6, + "args": { + "External id": 2332, + "cbid": 211, "correlation": 18318 + } + }, + { + "ph": "s", "id": 18318, "pid": 435, "tid": 435, "ts": 1776649790546145, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790546157, "dur": 0, + "args": { + "External id": 18324, + "cbid": 200, "correlation": 18324 + } + }, + { + "ph": "f", "id": 18324, "pid": 435, "tid": 435, "ts": 1776649790546157, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790546158, "dur": 0, + "args": { + "External id": 18325, + "cbid": 273, "correlation": 18325 + } + }, + { + "ph": "f", "id": 18325, "pid": 435, "tid": 435, "ts": 1776649790546158, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790546166, "dur": 15, + "args": { + "External id": 18326, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18326, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18326, "pid": 0, "tid": 7, "ts": 1776649790546166, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546159, "dur": 4, + "args": { + "External id": 18326, + "cbid": 211, "correlation": 18326 + } + }, + { + "ph": "s", "id": 18326, "pid": 435, "tid": 435, "ts": 1776649790546159, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790546202, "dur": 1, + "args": { + "External id": 2337, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18344, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18344, "pid": 0, "tid": 7, "ts": 1776649790546202, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546193, "dur": 6, + "args": { + "External id": 2337, + "cbid": 211, "correlation": 18344 + } + }, + { + "ph": "s", "id": 18344, "pid": 435, "tid": 435, "ts": 1776649790546193, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790546205, "dur": 0, + "args": { + "External id": 18350, + "cbid": 200, "correlation": 18350 + } + }, + { + "ph": "f", "id": 18350, "pid": 435, "tid": 435, "ts": 1776649790546205, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790546205, "dur": 0, + "args": { + "External id": 18351, + "cbid": 273, "correlation": 18351 + } + }, + { + "ph": "f", "id": 18351, "pid": 435, "tid": 435, "ts": 1776649790546205, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790546214, "dur": 15, + "args": { + "External id": 18352, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18352, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18352, "pid": 0, "tid": 7, "ts": 1776649790546214, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546207, "dur": 4, + "args": { + "External id": 18352, + "cbid": 211, "correlation": 18352 + } + }, + { + "ph": "s", "id": 18352, "pid": 435, "tid": 435, "ts": 1776649790546207, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790546249, "dur": 1, + "args": { + "External id": 2342, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18370, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18370, "pid": 0, "tid": 7, "ts": 1776649790546249, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546240, "dur": 6, + "args": { + "External id": 2342, + "cbid": 211, "correlation": 18370 + } + }, + { + "ph": "s", "id": 18370, "pid": 435, "tid": 435, "ts": 1776649790546240, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790546253, "dur": 0, + "args": { + "External id": 18376, + "cbid": 200, "correlation": 18376 + } + }, + { + "ph": "f", "id": 18376, "pid": 435, "tid": 435, "ts": 1776649790546253, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790546253, "dur": 0, + "args": { + "External id": 18377, + "cbid": 273, "correlation": 18377 + } + }, + { + "ph": "f", "id": 18377, "pid": 435, "tid": 435, "ts": 1776649790546253, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790546263, "dur": 16, + "args": { + "External id": 18378, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18378, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18378, "pid": 0, "tid": 7, "ts": 1776649790546263, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546254, "dur": 4, + "args": { + "External id": 18378, + "cbid": 211, "correlation": 18378 + } + }, + { + "ph": "s", "id": 18378, "pid": 435, "tid": 435, "ts": 1776649790546254, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790546297, "dur": 1, + "args": { + "External id": 2347, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18396, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18396, "pid": 0, "tid": 7, "ts": 1776649790546297, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546288, "dur": 6, + "args": { + "External id": 2347, + "cbid": 211, "correlation": 18396 + } + }, + { + "ph": "s", "id": 18396, "pid": 435, "tid": 435, "ts": 1776649790546288, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790546300, "dur": 0, + "args": { + "External id": 18402, + "cbid": 200, "correlation": 18402 + } + }, + { + "ph": "f", "id": 18402, "pid": 435, "tid": 435, "ts": 1776649790546300, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790546301, "dur": 0, + "args": { + "External id": 18403, + "cbid": 273, "correlation": 18403 + } + }, + { + "ph": "f", "id": 18403, "pid": 435, "tid": 435, "ts": 1776649790546301, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790546310, "dur": 15, + "args": { + "External id": 18404, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18404, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18404, "pid": 0, "tid": 7, "ts": 1776649790546310, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546302, "dur": 5, + "args": { + "External id": 18404, + "cbid": 211, "correlation": 18404 + } + }, + { + "ph": "s", "id": 18404, "pid": 435, "tid": 435, "ts": 1776649790546302, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790546345, "dur": 1, + "args": { + "External id": 2352, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18422, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18422, "pid": 0, "tid": 7, "ts": 1776649790546345, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546336, "dur": 6, + "args": { + "External id": 2352, + "cbid": 211, "correlation": 18422 + } + }, + { + "ph": "s", "id": 18422, "pid": 435, "tid": 435, "ts": 1776649790546336, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790546349, "dur": 0, + "args": { + "External id": 18428, + "cbid": 200, "correlation": 18428 + } + }, + { + "ph": "f", "id": 18428, "pid": 435, "tid": 435, "ts": 1776649790546349, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790546349, "dur": 0, + "args": { + "External id": 18429, + "cbid": 273, "correlation": 18429 + } + }, + { + "ph": "f", "id": 18429, "pid": 435, "tid": 435, "ts": 1776649790546349, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790546359, "dur": 15, + "args": { + "External id": 18430, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18430, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18430, "pid": 0, "tid": 7, "ts": 1776649790546359, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546350, "dur": 4, + "args": { + "External id": 18430, + "cbid": 211, "correlation": 18430 + } + }, + { + "ph": "s", "id": 18430, "pid": 435, "tid": 435, "ts": 1776649790546350, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790546394, "dur": 1, + "args": { + "External id": 2357, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18448, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18448, "pid": 0, "tid": 7, "ts": 1776649790546394, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546385, "dur": 6, + "args": { + "External id": 2357, + "cbid": 211, "correlation": 18448 + } + }, + { + "ph": "s", "id": 18448, "pid": 435, "tid": 435, "ts": 1776649790546385, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790546397, "dur": 0, + "args": { + "External id": 18454, + "cbid": 200, "correlation": 18454 + } + }, + { + "ph": "f", "id": 18454, "pid": 435, "tid": 435, "ts": 1776649790546397, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790546398, "dur": 0, + "args": { + "External id": 18455, + "cbid": 273, "correlation": 18455 + } + }, + { + "ph": "f", "id": 18455, "pid": 435, "tid": 435, "ts": 1776649790546398, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790546407, "dur": 15, + "args": { + "External id": 18456, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18456, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18456, "pid": 0, "tid": 7, "ts": 1776649790546407, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546399, "dur": 4, + "args": { + "External id": 18456, + "cbid": 211, "correlation": 18456 + } + }, + { + "ph": "s", "id": 18456, "pid": 435, "tid": 435, "ts": 1776649790546399, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790546442, "dur": 1, + "args": { + "External id": 2362, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18474, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18474, "pid": 0, "tid": 7, "ts": 1776649790546442, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546433, "dur": 6, + "args": { + "External id": 2362, + "cbid": 211, "correlation": 18474 + } + }, + { + "ph": "s", "id": 18474, "pid": 435, "tid": 435, "ts": 1776649790546433, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790546445, "dur": 0, + "args": { + "External id": 18480, + "cbid": 200, "correlation": 18480 + } + }, + { + "ph": "f", "id": 18480, "pid": 435, "tid": 435, "ts": 1776649790546445, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790546446, "dur": 0, + "args": { + "External id": 18481, + "cbid": 273, "correlation": 18481 + } + }, + { + "ph": "f", "id": 18481, "pid": 435, "tid": 435, "ts": 1776649790546446, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790546455, "dur": 15, + "args": { + "External id": 18482, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18482, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18482, "pid": 0, "tid": 7, "ts": 1776649790546455, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546447, "dur": 4, + "args": { + "External id": 18482, + "cbid": 211, "correlation": 18482 + } + }, + { + "ph": "s", "id": 18482, "pid": 435, "tid": 435, "ts": 1776649790546447, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790546490, "dur": 1, + "args": { + "External id": 2367, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18500, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18500, "pid": 0, "tid": 7, "ts": 1776649790546490, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546480, "dur": 6, + "args": { + "External id": 2367, + "cbid": 211, "correlation": 18500 + } + }, + { + "ph": "s", "id": 18500, "pid": 435, "tid": 435, "ts": 1776649790546480, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790546493, "dur": 0, + "args": { + "External id": 18506, + "cbid": 200, "correlation": 18506 + } + }, + { + "ph": "f", "id": 18506, "pid": 435, "tid": 435, "ts": 1776649790546493, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790546494, "dur": 0, + "args": { + "External id": 18507, + "cbid": 273, "correlation": 18507 + } + }, + { + "ph": "f", "id": 18507, "pid": 435, "tid": 435, "ts": 1776649790546494, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790546503, "dur": 15, + "args": { + "External id": 18508, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18508, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18508, "pid": 0, "tid": 7, "ts": 1776649790546503, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546495, "dur": 5, + "args": { + "External id": 18508, + "cbid": 211, "correlation": 18508 + } + }, + { + "ph": "s", "id": 18508, "pid": 435, "tid": 435, "ts": 1776649790546495, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790546538, "dur": 1, + "args": { + "External id": 2372, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18526, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18526, "pid": 0, "tid": 7, "ts": 1776649790546538, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546529, "dur": 6, + "args": { + "External id": 2372, + "cbid": 211, "correlation": 18526 + } + }, + { + "ph": "s", "id": 18526, "pid": 435, "tid": 435, "ts": 1776649790546529, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790546541, "dur": 0, + "args": { + "External id": 18532, + "cbid": 200, "correlation": 18532 + } + }, + { + "ph": "f", "id": 18532, "pid": 435, "tid": 435, "ts": 1776649790546541, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790546542, "dur": 0, + "args": { + "External id": 18533, + "cbid": 273, "correlation": 18533 + } + }, + { + "ph": "f", "id": 18533, "pid": 435, "tid": 435, "ts": 1776649790546542, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790546551, "dur": 15, + "args": { + "External id": 18534, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18534, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18534, "pid": 0, "tid": 7, "ts": 1776649790546551, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546543, "dur": 4, + "args": { + "External id": 18534, + "cbid": 211, "correlation": 18534 + } + }, + { + "ph": "s", "id": 18534, "pid": 435, "tid": 435, "ts": 1776649790546543, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790546586, "dur": 1, + "args": { + "External id": 2377, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18552, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18552, "pid": 0, "tid": 7, "ts": 1776649790546586, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546576, "dur": 6, + "args": { + "External id": 2377, + "cbid": 211, "correlation": 18552 + } + }, + { + "ph": "s", "id": 18552, "pid": 435, "tid": 435, "ts": 1776649790546576, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790546589, "dur": 0, + "args": { + "External id": 18558, + "cbid": 200, "correlation": 18558 + } + }, + { + "ph": "f", "id": 18558, "pid": 435, "tid": 435, "ts": 1776649790546589, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790546589, "dur": 0, + "args": { + "External id": 18559, + "cbid": 273, "correlation": 18559 + } + }, + { + "ph": "f", "id": 18559, "pid": 435, "tid": 435, "ts": 1776649790546589, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790546598, "dur": 15, + "args": { + "External id": 18560, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18560, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18560, "pid": 0, "tid": 7, "ts": 1776649790546598, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546590, "dur": 4, + "args": { + "External id": 18560, + "cbid": 211, "correlation": 18560 + } + }, + { + "ph": "s", "id": 18560, "pid": 435, "tid": 435, "ts": 1776649790546590, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790546620, "dur": 3, + "args": { + "External id": 18568, + "cbid": 135, "correlation": 18568 + } + }, + { + "ph": "f", "id": 18568, "pid": 435, "tid": 435, "ts": 1776649790546620, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790546635, "dur": 5, + "args": { + "External id": 18573, + "cbid": 165, "correlation": 18573 + } + }, + { + "ph": "s", "id": 18573, "pid": 435, "tid": 435, "ts": 1776649790546635, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649790546644, "dur": 4, + "args": { + "External id": 18576, + "cbid": 139, "correlation": 18576 + } + }, + { + "ph": "f", "id": 18576, "pid": 435, "tid": 435, "ts": 1776649790546644, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649790546791, "dur": 2, + "args": { + "External id": 2380, + "cbid": 317, "correlation": 18601 + } + }, + { + "ph": "f", "id": 18601, "pid": 435, "tid": 435, "ts": 1776649790546791, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649790546795, "dur": 0, + "args": { + "External id": 2380, + "cbid": 317, "correlation": 18605 + } + }, + { + "ph": "f", "id": 18605, "pid": 435, "tid": 435, "ts": 1776649790546795, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::distribution_elementwise_grid_stride_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::(anonymous namespace)::distribution_nullary_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2} const&, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(int, float)#1}>(int, at::PhiloxCudaState, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::(anonymous namespace)::distribution_nullary_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2} const&, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(int, float)#1})", "pid": 0, "tid": 7, + "ts": 1776649790546812, "dur": 3, + "args": { + "External id": 2380, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18607, + "registers per thread": 38, + "shared memory": 0, + "blocks per SM": 0.592593, + "warps per SM": 4.740741, + "grid": [64, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 7 + } + }, + { + "ph": "f", "id": 18607, "pid": 0, "tid": 7, "ts": 1776649790546812, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546799, "dur": 12, + "args": { + "External id": 2380, + "cbid": 211, "correlation": 18607 + } + }, + { + "ph": "s", "id": 18607, "pid": 435, "tid": 435, "ts": 1776649790546799, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649790546824, "dur": 0, + "args": { + "External id": 18614, + "cbid": 317, "correlation": 18614 + } + }, + { + "ph": "f", "id": 18614, "pid": 435, "tid": 435, "ts": 1776649790546824, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649790546831, "dur": 0, + "args": { + "External id": 18618, + "cbid": 317, "correlation": 18618 + } + }, + { + "ph": "f", "id": 18618, "pid": 435, "tid": 435, "ts": 1776649790546831, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649790546854, "dur": 0, + "args": { + "External id": 2383, + "cbid": 317, "correlation": 18633 + } + }, + { + "ph": "f", "id": 18633, "pid": 435, "tid": 435, "ts": 1776649790546854, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649790546856, "dur": 0, + "args": { + "External id": 2383, + "cbid": 317, "correlation": 18637 + } + }, + { + "ph": "f", "id": 18637, "pid": 435, "tid": 435, "ts": 1776649790546856, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::distribution_elementwise_grid_stride_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::(anonymous namespace)::distribution_nullary_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2} const&, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(int, float)#1}>(int, at::PhiloxCudaState, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::(anonymous namespace)::distribution_nullary_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2} const&, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(int, float)#1})", "pid": 0, "tid": 7, + "ts": 1776649790546867, "dur": 85, + "args": { + "External id": 2383, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18639, + "registers per thread": 38, + "shared memory": 0, + "blocks per SM": 8.000000, + "warps per SM": 64.000000, + "grid": [864, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 75 + } + }, + { + "ph": "f", "id": 18639, "pid": 0, "tid": 7, "ts": 1776649790546867, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546857, "dur": 7, + "args": { + "External id": 2383, + "cbid": 211, "correlation": 18639 + } + }, + { + "ph": "s", "id": 18639, "pid": 435, "tid": 435, "ts": 1776649790546857, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor >, at::detail::Array >(int, at::native::AUnaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790546954, "dur": 38, + "args": { + "External id": 2384, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18649, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 18649, "pid": 0, "tid": 7, "ts": 1776649790546954, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790546926, "dur": 10, + "args": { + "External id": 2384, + "cbid": 211, "correlation": 18649 + } + }, + { + "ph": "s", "id": 18649, "pid": 435, "tid": 435, "ts": 1776649790546926, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AbsFunctor, at::detail::Array >(int, at::native::AbsFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790547023, "dur": 40, + "args": { + "External id": 2389, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18669, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 18669, "pid": 0, "tid": 7, "ts": 1776649790547023, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790547011, "dur": 9, + "args": { + "External id": 2389, + "cbid": 211, "correlation": 18669 + } + }, + { + "ph": "s", "id": 18669, "pid": 435, "tid": 435, "ts": 1776649790547011, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<128, 4, at::native::ReduceOp >, unsigned int, c10::Half, 4> >(at::native::ReduceOp >, unsigned int, c10::Half, 4>)", "pid": 0, "tid": 7, + "ts": 1776649790547077, "dur": 46, + "args": { + "External id": 2391, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18682, + "registers per thread": 49, + "shared memory": 1040, + "blocks per SM": 9.481482, + "warps per SM": 37.925926, + "grid": [1024, 1, 1], + "block": [32, 4, 1], + "est. achieved occupancy %": 56 + } + }, + { + "ph": "f", "id": 18682, "pid": 0, "tid": 7, "ts": 1776649790547077, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790547065, "dur": 10, + "args": { + "External id": 2391, + "cbid": 211, "correlation": 18682 + } + }, + { + "ph": "s", "id": 18682, "pid": 435, "tid": 435, "ts": 1776649790547065, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BUnaryFunctor >, at::detail::Array >(int, at::native::BUnaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790547125, "dur": 2, + "args": { + "External id": 2393, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18692, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 2.370370, + "warps per SM": 9.481482, + "grid": [256, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 15 + } + }, + { + "ph": "f", "id": 18692, "pid": 0, "tid": 7, "ts": 1776649790547125, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790547105, "dur": 8, + "args": { + "External id": 2393, + "cbid": 211, "correlation": 18692 + } + }, + { + "ph": "s", "id": 18692, "pid": 435, "tid": 435, "ts": 1776649790547105, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#8}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#8}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790547156, "dur": 3, + "args": { + "External id": 2394, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18702, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 2.370370, + "warps per SM": 9.481482, + "grid": [256, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 15 + } + }, + { + "ph": "f", "id": 18702, "pid": 0, "tid": 7, "ts": 1776649790547156, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790547146, "dur": 8, + "args": { + "External id": 2394, + "cbid": 211, "correlation": 18702 + } + }, + { + "ph": "s", "id": 18702, "pid": 435, "tid": 435, "ts": 1776649790547146, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649790547205, "dur": 99, + "args": { + "External id": 2399, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18712, + "registers per thread": 20, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 18712, "pid": 0, "tid": 7, "ts": 1776649790547205, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790547194, "dur": 8, + "args": { + "External id": 2399, + "cbid": 211, "correlation": 18712 + } + }, + { + "ph": "s", "id": 18712, "pid": 435, "tid": 435, "ts": 1776649790547194, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790547305, "dur": 39, + "args": { + "External id": 2400, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18722, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 18722, "pid": 0, "tid": 7, "ts": 1776649790547305, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790547224, "dur": 7, + "args": { + "External id": 2400, + "cbid": 211, "correlation": 18722 + } + }, + { + "ph": "s", "id": 18722, "pid": 435, "tid": 435, "ts": 1776649790547224, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#8}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#8}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790547346, "dur": 41, + "args": { + "External id": 2401, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18732, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 18732, "pid": 0, "tid": 7, "ts": 1776649790547346, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790547247, "dur": 6, + "args": { + "External id": 2401, + "cbid": 211, "correlation": 18732 + } + }, + { + "ph": "s", "id": 18732, "pid": 435, "tid": 435, "ts": 1776649790547247, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649790547389, "dur": 95, + "args": { + "External id": 2406, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18745, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 18745, "pid": 0, "tid": 7, "ts": 1776649790547389, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790547291, "dur": 9, + "args": { + "External id": 2406, + "cbid": 211, "correlation": 18745 + } + }, + { + "ph": "s", "id": 18745, "pid": 435, "tid": 435, "ts": 1776649790547291, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, at::detail::Array >(int, at::native::CUDAFunctorOnSelf_add, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790547486, "dur": 33, + "args": { + "External id": 2407, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18756, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 18756, "pid": 0, "tid": 7, "ts": 1776649790547486, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790547334, "dur": 8, + "args": { + "External id": 2407, + "cbid": 211, "correlation": 18756 + } + }, + { + "ph": "s", "id": 18756, "pid": 435, "tid": 435, "ts": 1776649790547334, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#1}::operator()() const::{lambda(unsigned char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649790547521, "dur": 57, + "args": { + "External id": 2411, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18769, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 18769, "pid": 0, "tid": 7, "ts": 1776649790547521, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790547365, "dur": 7, + "args": { + "External id": 2411, + "cbid": 211, "correlation": 18769 + } + }, + { + "ph": "s", "id": 18769, "pid": 435, "tid": 435, "ts": 1776649790547365, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast >(at::TensorIteratorBase&, at::native::BUnaryFunctor const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast >(at::TensorIteratorBase&, at::native::BUnaryFunctor const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649790547580, "dur": 36, + "args": { + "External id": 2422, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18789, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 18789, "pid": 0, "tid": 7, "ts": 1776649790547580, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790547469, "dur": 11, + "args": { + "External id": 2422, + "cbid": 211, "correlation": 18789 + } + }, + { + "ph": "s", "id": 18789, "pid": 435, "tid": 435, "ts": 1776649790547469, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649790547618, "dur": 39, + "args": { + "External id": 2425, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18799, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 18799, "pid": 0, "tid": 7, "ts": 1776649790547618, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790547508, "dur": 8, + "args": { + "External id": 2425, + "cbid": 211, "correlation": 18799 + } + }, + { + "ph": "s", "id": 18799, "pid": 435, "tid": 435, "ts": 1776649790547508, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor >, at::detail::Array >(int, at::native::AUnaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790547659, "dur": 17, + "args": { + "External id": 2429, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18809, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 18809, "pid": 0, "tid": 7, "ts": 1776649790547659, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790547563, "dur": 8, + "args": { + "External id": 2429, + "cbid": 211, "correlation": 18809 + } + }, + { + "ph": "s", "id": 18809, "pid": 435, "tid": 435, "ts": 1776649790547563, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649790547678, "dur": 30, + "args": { + "External id": 2433, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18822, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 18822, "pid": 0, "tid": 7, "ts": 1776649790547678, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790547595, "dur": 7, + "args": { + "External id": 2433, + "cbid": 211, "correlation": 18822 + } + }, + { + "ph": "s", "id": 18822, "pid": 435, "tid": 435, "ts": 1776649790547595, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BUnaryFunctor, at::detail::Array >(int, at::native::BUnaryFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790547709, "dur": 17, + "args": { + "External id": 2434, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18836, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 18836, "pid": 0, "tid": 7, "ts": 1776649790547709, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790547629, "dur": 8, + "args": { + "External id": 2434, + "cbid": 211, "correlation": 18836 + } + }, + { + "ph": "s", "id": 18836, "pid": 435, "tid": 435, "ts": 1776649790547629, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649790547728, "dur": 29, + "args": { + "External id": 2439, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18849, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 18849, "pid": 0, "tid": 7, "ts": 1776649790547728, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790547664, "dur": 7, + "args": { + "External id": 2439, + "cbid": 211, "correlation": 18849 + } + }, + { + "ph": "s", "id": 18849, "pid": 435, "tid": 435, "ts": 1776649790547664, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy_contig, unsigned int, 3, 128, 1>(at::native::(anonymous namespace)::OpaqueType<1u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata, unsigned int, 128, 1>, at::native::(anonymous namespace)::TensorSizeStride, int, unsigned int)", "pid": 0, "tid": 7, + "ts": 1776649790547759, "dur": 64, + "args": { + "External id": 2445, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18861, + "registers per thread": 21, + "shared memory": 0, + "blocks per SM": 4.000000, + "warps per SM": 64.000000, + "grid": [216, 2, 1], + "block": [512, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 18861, "pid": 0, "tid": 7, "ts": 1776649790547759, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790547729, "dur": 11, + "args": { + "External id": 2445, + "cbid": 211, "correlation": 18861 + } + }, + { + "ph": "s", "id": 18861, "pid": 435, "tid": 435, "ts": 1776649790547729, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649790547825, "dur": 71, + "args": { + "External id": 2456, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18874, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 18874, "pid": 0, "tid": 7, "ts": 1776649790547825, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790547808, "dur": 10, + "args": { + "External id": 2456, + "cbid": 211, "correlation": 18874 + } + }, + { + "ph": "s", "id": 18874, "pid": 435, "tid": 435, "ts": 1776649790547808, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, at::detail::Array >(int, at::native::CUDAFunctorOnSelf_add, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790547897, "dur": 33, + "args": { + "External id": 2459, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18885, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 18885, "pid": 0, "tid": 7, "ts": 1776649790547897, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790547851, "dur": 7, + "args": { + "External id": 2459, + "cbid": 211, "correlation": 18885 + } + }, + { + "ph": "s", "id": 18885, "pid": 435, "tid": 435, "ts": 1776649790547851, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649790547932, "dur": 91, + "args": { + "External id": 2463, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18898, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 18898, "pid": 0, "tid": 7, "ts": 1776649790547932, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790547880, "dur": 7, + "args": { + "External id": 2463, + "cbid": 211, "correlation": 18898 + } + }, + { + "ph": "s", "id": 18898, "pid": 435, "tid": 435, "ts": 1776649790547880, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor >, at::detail::Array >(int, at::native::BinaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790548025, "dur": 60, + "args": { + "External id": 2464, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18909, + "registers per thread": 22, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 18909, "pid": 0, "tid": 7, "ts": 1776649790548025, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790547908, "dur": 6, + "args": { + "External id": 2464, + "cbid": 211, "correlation": 18909 + } + }, + { + "ph": "s", "id": 18909, "pid": 435, "tid": 435, "ts": 1776649790547908, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790548127, "dur": 7, + "args": { + "External id": 2466, + "cbid": 210, "correlation": 18936 + } + }, + { + "ph": "f", "id": 18936, "pid": 435, "tid": 435, "ts": 1776649790548127, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790548145, "dur": 0, + "args": { + "External id": 2466, + "cbid": 210, "correlation": 18937 + } + }, + { + "ph": "f", "id": 18937, "pid": 435, "tid": 435, "ts": 1776649790548145, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790548164, "dur": 31, + "args": { + "External id": 2466, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18938, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18938, "pid": 0, "tid": 7, "ts": 1776649790548164, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790548150, "dur": 12, + "args": { + "External id": 2466, + "cbid": 211, "correlation": 18938 + } + }, + { + "ph": "s", "id": 18938, "pid": 435, "tid": 435, "ts": 1776649790548150, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790548197, "dur": 0, + "args": { + "External id": 2468, + "cbid": 210, "correlation": 18965 + } + }, + { + "ph": "f", "id": 18965, "pid": 435, "tid": 435, "ts": 1776649790548197, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790548208, "dur": 23, + "args": { + "External id": 2468, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18966, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18966, "pid": 0, "tid": 7, "ts": 1776649790548208, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790548199, "dur": 6, + "args": { + "External id": 2468, + "cbid": 211, "correlation": 18966 + } + }, + { + "ph": "s", "id": 18966, "pid": 435, "tid": 435, "ts": 1776649790548199, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790548250, "dur": 1, + "args": { + "External id": 2470, + "cbid": 210, "correlation": 18993 + } + }, + { + "ph": "f", "id": 18993, "pid": 435, "tid": 435, "ts": 1776649790548250, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790548261, "dur": 23, + "args": { + "External id": 2470, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 18994, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 18994, "pid": 0, "tid": 7, "ts": 1776649790548261, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790548252, "dur": 7, + "args": { + "External id": 2470, + "cbid": 211, "correlation": 18994 + } + }, + { + "ph": "s", "id": 18994, "pid": 435, "tid": 435, "ts": 1776649790548252, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790548284, "dur": 0, + "args": { + "External id": 2472, + "cbid": 210, "correlation": 19021 + } + }, + { + "ph": "f", "id": 19021, "pid": 435, "tid": 435, "ts": 1776649790548284, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790548295, "dur": 23, + "args": { + "External id": 2472, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19022, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19022, "pid": 0, "tid": 7, "ts": 1776649790548295, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790548286, "dur": 6, + "args": { + "External id": 2472, + "cbid": 211, "correlation": 19022 + } + }, + { + "ph": "s", "id": 19022, "pid": 435, "tid": 435, "ts": 1776649790548286, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790548315, "dur": 0, + "args": { + "External id": 2474, + "cbid": 210, "correlation": 19049 + } + }, + { + "ph": "f", "id": 19049, "pid": 435, "tid": 435, "ts": 1776649790548315, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790548325, "dur": 23, + "args": { + "External id": 2474, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19050, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19050, "pid": 0, "tid": 7, "ts": 1776649790548325, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790548316, "dur": 6, + "args": { + "External id": 2474, + "cbid": 211, "correlation": 19050 + } + }, + { + "ph": "s", "id": 19050, "pid": 435, "tid": 435, "ts": 1776649790548316, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790548345, "dur": 0, + "args": { + "External id": 2476, + "cbid": 210, "correlation": 19077 + } + }, + { + "ph": "f", "id": 19077, "pid": 435, "tid": 435, "ts": 1776649790548345, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790548356, "dur": 23, + "args": { + "External id": 2476, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19078, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19078, "pid": 0, "tid": 7, "ts": 1776649790548356, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790548347, "dur": 6, + "args": { + "External id": 2476, + "cbid": 211, "correlation": 19078 + } + }, + { + "ph": "s", "id": 19078, "pid": 435, "tid": 435, "ts": 1776649790548347, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790548376, "dur": 0, + "args": { + "External id": 2478, + "cbid": 210, "correlation": 19105 + } + }, + { + "ph": "f", "id": 19105, "pid": 435, "tid": 435, "ts": 1776649790548376, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790548386, "dur": 23, + "args": { + "External id": 2478, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19106, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19106, "pid": 0, "tid": 7, "ts": 1776649790548386, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790548377, "dur": 6, + "args": { + "External id": 2478, + "cbid": 211, "correlation": 19106 + } + }, + { + "ph": "s", "id": 19106, "pid": 435, "tid": 435, "ts": 1776649790548377, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790548405, "dur": 0, + "args": { + "External id": 2480, + "cbid": 210, "correlation": 19133 + } + }, + { + "ph": "f", "id": 19133, "pid": 435, "tid": 435, "ts": 1776649790548405, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790548415, "dur": 23, + "args": { + "External id": 2480, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19134, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19134, "pid": 0, "tid": 7, "ts": 1776649790548415, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790548407, "dur": 6, + "args": { + "External id": 2480, + "cbid": 211, "correlation": 19134 + } + }, + { + "ph": "s", "id": 19134, "pid": 435, "tid": 435, "ts": 1776649790548407, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790548435, "dur": 0, + "args": { + "External id": 2482, + "cbid": 210, "correlation": 19161 + } + }, + { + "ph": "f", "id": 19161, "pid": 435, "tid": 435, "ts": 1776649790548435, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790548447, "dur": 23, + "args": { + "External id": 2482, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19162, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19162, "pid": 0, "tid": 7, "ts": 1776649790548447, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790548436, "dur": 6, + "args": { + "External id": 2482, + "cbid": 211, "correlation": 19162 + } + }, + { + "ph": "s", "id": 19162, "pid": 435, "tid": 435, "ts": 1776649790548436, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790548465, "dur": 0, + "args": { + "External id": 2484, + "cbid": 210, "correlation": 19189 + } + }, + { + "ph": "f", "id": 19189, "pid": 435, "tid": 435, "ts": 1776649790548465, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790548476, "dur": 23, + "args": { + "External id": 2484, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19190, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19190, "pid": 0, "tid": 7, "ts": 1776649790548476, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790548466, "dur": 5, + "args": { + "External id": 2484, + "cbid": 211, "correlation": 19190 + } + }, + { + "ph": "s", "id": 19190, "pid": 435, "tid": 435, "ts": 1776649790548466, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790548496, "dur": 8, + "args": { + "External id": 19196, + "cbid": 165, "correlation": 19196 + } + }, + { + "ph": "s", "id": 19196, "pid": 435, "tid": 435, "ts": 1776649790548496, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790548530, "dur": 4, + "args": { + "External id": 19205, + "cbid": 135, "correlation": 19205 + } + }, + { + "ph": "f", "id": 19205, "pid": 435, "tid": 435, "ts": 1776649790548530, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790548563, "dur": 1, + "args": { + "External id": 2486, + "cbid": 210, "correlation": 19231 + } + }, + { + "ph": "f", "id": 19231, "pid": 435, "tid": 435, "ts": 1776649790548563, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790548575, "dur": 23, + "args": { + "External id": 2486, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19232, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19232, "pid": 0, "tid": 7, "ts": 1776649790548575, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790548565, "dur": 8, + "args": { + "External id": 2486, + "cbid": 211, "correlation": 19232 + } + }, + { + "ph": "s", "id": 19232, "pid": 435, "tid": 435, "ts": 1776649790548565, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790548599, "dur": 0, + "args": { + "External id": 2488, + "cbid": 210, "correlation": 19259 + } + }, + { + "ph": "f", "id": 19259, "pid": 435, "tid": 435, "ts": 1776649790548599, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790548610, "dur": 23, + "args": { + "External id": 2488, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19260, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19260, "pid": 0, "tid": 7, "ts": 1776649790548610, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790548601, "dur": 6, + "args": { + "External id": 2488, + "cbid": 211, "correlation": 19260 + } + }, + { + "ph": "s", "id": 19260, "pid": 435, "tid": 435, "ts": 1776649790548601, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790548630, "dur": 0, + "args": { + "External id": 2490, + "cbid": 210, "correlation": 19287 + } + }, + { + "ph": "f", "id": 19287, "pid": 435, "tid": 435, "ts": 1776649790548630, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790548640, "dur": 23, + "args": { + "External id": 2490, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19288, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19288, "pid": 0, "tid": 7, "ts": 1776649790548640, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790548631, "dur": 6, + "args": { + "External id": 2490, + "cbid": 211, "correlation": 19288 + } + }, + { + "ph": "s", "id": 19288, "pid": 435, "tid": 435, "ts": 1776649790548631, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790548660, "dur": 0, + "args": { + "External id": 2492, + "cbid": 210, "correlation": 19315 + } + }, + { + "ph": "f", "id": 19315, "pid": 435, "tid": 435, "ts": 1776649790548660, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790548670, "dur": 23, + "args": { + "External id": 2492, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19316, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19316, "pid": 0, "tid": 7, "ts": 1776649790548670, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790548662, "dur": 6, + "args": { + "External id": 2492, + "cbid": 211, "correlation": 19316 + } + }, + { + "ph": "s", "id": 19316, "pid": 435, "tid": 435, "ts": 1776649790548662, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790548690, "dur": 0, + "args": { + "External id": 2494, + "cbid": 210, "correlation": 19343 + } + }, + { + "ph": "f", "id": 19343, "pid": 435, "tid": 435, "ts": 1776649790548690, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790548700, "dur": 23, + "args": { + "External id": 2494, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19344, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19344, "pid": 0, "tid": 7, "ts": 1776649790548700, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790548691, "dur": 6, + "args": { + "External id": 2494, + "cbid": 211, "correlation": 19344 + } + }, + { + "ph": "s", "id": 19344, "pid": 435, "tid": 435, "ts": 1776649790548691, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790548720, "dur": 0, + "args": { + "External id": 2496, + "cbid": 210, "correlation": 19371 + } + }, + { + "ph": "f", "id": 19371, "pid": 435, "tid": 435, "ts": 1776649790548720, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790548731, "dur": 23, + "args": { + "External id": 2496, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19372, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19372, "pid": 0, "tid": 7, "ts": 1776649790548731, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790548721, "dur": 6, + "args": { + "External id": 2496, + "cbid": 211, "correlation": 19372 + } + }, + { + "ph": "s", "id": 19372, "pid": 435, "tid": 435, "ts": 1776649790548721, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790548750, "dur": 0, + "args": { + "External id": 2498, + "cbid": 210, "correlation": 19399 + } + }, + { + "ph": "f", "id": 19399, "pid": 435, "tid": 435, "ts": 1776649790548750, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790548761, "dur": 23, + "args": { + "External id": 2498, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19400, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19400, "pid": 0, "tid": 7, "ts": 1776649790548761, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790548751, "dur": 6, + "args": { + "External id": 2498, + "cbid": 211, "correlation": 19400 + } + }, + { + "ph": "s", "id": 19400, "pid": 435, "tid": 435, "ts": 1776649790548751, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790548780, "dur": 0, + "args": { + "External id": 2500, + "cbid": 210, "correlation": 19427 + } + }, + { + "ph": "f", "id": 19427, "pid": 435, "tid": 435, "ts": 1776649790548780, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790548791, "dur": 23, + "args": { + "External id": 2500, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19428, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19428, "pid": 0, "tid": 7, "ts": 1776649790548791, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790548781, "dur": 5, + "args": { + "External id": 2500, + "cbid": 211, "correlation": 19428 + } + }, + { + "ph": "s", "id": 19428, "pid": 435, "tid": 435, "ts": 1776649790548781, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790548809, "dur": 0, + "args": { + "External id": 2502, + "cbid": 210, "correlation": 19455 + } + }, + { + "ph": "f", "id": 19455, "pid": 435, "tid": 435, "ts": 1776649790548809, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790548820, "dur": 23, + "args": { + "External id": 2502, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19456, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19456, "pid": 0, "tid": 7, "ts": 1776649790548820, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790548810, "dur": 6, + "args": { + "External id": 2502, + "cbid": 211, "correlation": 19456 + } + }, + { + "ph": "s", "id": 19456, "pid": 435, "tid": 435, "ts": 1776649790548810, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790548839, "dur": 0, + "args": { + "External id": 2504, + "cbid": 210, "correlation": 19483 + } + }, + { + "ph": "f", "id": 19483, "pid": 435, "tid": 435, "ts": 1776649790548839, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790548850, "dur": 23, + "args": { + "External id": 2504, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19484, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19484, "pid": 0, "tid": 7, "ts": 1776649790548850, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790548840, "dur": 5, + "args": { + "External id": 2504, + "cbid": 211, "correlation": 19484 + } + }, + { + "ph": "s", "id": 19484, "pid": 435, "tid": 435, "ts": 1776649790548840, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790548868, "dur": 0, + "args": { + "External id": 2506, + "cbid": 210, "correlation": 19511 + } + }, + { + "ph": "f", "id": 19511, "pid": 435, "tid": 435, "ts": 1776649790548868, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790548880, "dur": 23, + "args": { + "External id": 2506, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19512, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19512, "pid": 0, "tid": 7, "ts": 1776649790548880, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790548869, "dur": 6, + "args": { + "External id": 2506, + "cbid": 211, "correlation": 19512 + } + }, + { + "ph": "s", "id": 19512, "pid": 435, "tid": 435, "ts": 1776649790548869, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790548898, "dur": 0, + "args": { + "External id": 2508, + "cbid": 210, "correlation": 19539 + } + }, + { + "ph": "f", "id": 19539, "pid": 435, "tid": 435, "ts": 1776649790548898, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790548909, "dur": 23, + "args": { + "External id": 2508, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19540, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19540, "pid": 0, "tid": 7, "ts": 1776649790548909, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790548899, "dur": 5, + "args": { + "External id": 2508, + "cbid": 211, "correlation": 19540 + } + }, + { + "ph": "s", "id": 19540, "pid": 435, "tid": 435, "ts": 1776649790548899, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790548927, "dur": 0, + "args": { + "External id": 2510, + "cbid": 210, "correlation": 19567 + } + }, + { + "ph": "f", "id": 19567, "pid": 435, "tid": 435, "ts": 1776649790548927, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790548939, "dur": 23, + "args": { + "External id": 2510, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19568, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19568, "pid": 0, "tid": 7, "ts": 1776649790548939, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790548928, "dur": 6, + "args": { + "External id": 2510, + "cbid": 211, "correlation": 19568 + } + }, + { + "ph": "s", "id": 19568, "pid": 435, "tid": 435, "ts": 1776649790548928, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790548957, "dur": 0, + "args": { + "External id": 2512, + "cbid": 210, "correlation": 19595 + } + }, + { + "ph": "f", "id": 19595, "pid": 435, "tid": 435, "ts": 1776649790548957, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790548968, "dur": 22, + "args": { + "External id": 2512, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19596, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19596, "pid": 0, "tid": 7, "ts": 1776649790548968, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790548958, "dur": 5, + "args": { + "External id": 2512, + "cbid": 211, "correlation": 19596 + } + }, + { + "ph": "s", "id": 19596, "pid": 435, "tid": 435, "ts": 1776649790548958, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790548986, "dur": 0, + "args": { + "External id": 2514, + "cbid": 210, "correlation": 19623 + } + }, + { + "ph": "f", "id": 19623, "pid": 435, "tid": 435, "ts": 1776649790548986, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790548997, "dur": 23, + "args": { + "External id": 2514, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19624, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19624, "pid": 0, "tid": 7, "ts": 1776649790548997, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790548987, "dur": 5, + "args": { + "External id": 2514, + "cbid": 211, "correlation": 19624 + } + }, + { + "ph": "s", "id": 19624, "pid": 435, "tid": 435, "ts": 1776649790548987, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549015, "dur": 0, + "args": { + "External id": 2516, + "cbid": 210, "correlation": 19651 + } + }, + { + "ph": "f", "id": 19651, "pid": 435, "tid": 435, "ts": 1776649790549015, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549027, "dur": 23, + "args": { + "External id": 2516, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19652, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19652, "pid": 0, "tid": 7, "ts": 1776649790549027, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549016, "dur": 5, + "args": { + "External id": 2516, + "cbid": 211, "correlation": 19652 + } + }, + { + "ph": "s", "id": 19652, "pid": 435, "tid": 435, "ts": 1776649790549016, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549044, "dur": 0, + "args": { + "External id": 2518, + "cbid": 210, "correlation": 19679 + } + }, + { + "ph": "f", "id": 19679, "pid": 435, "tid": 435, "ts": 1776649790549044, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549056, "dur": 23, + "args": { + "External id": 2518, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19680, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19680, "pid": 0, "tid": 7, "ts": 1776649790549056, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549046, "dur": 6, + "args": { + "External id": 2518, + "cbid": 211, "correlation": 19680 + } + }, + { + "ph": "s", "id": 19680, "pid": 435, "tid": 435, "ts": 1776649790549046, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549074, "dur": 0, + "args": { + "External id": 2520, + "cbid": 210, "correlation": 19707 + } + }, + { + "ph": "f", "id": 19707, "pid": 435, "tid": 435, "ts": 1776649790549074, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549085, "dur": 23, + "args": { + "External id": 2520, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19708, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19708, "pid": 0, "tid": 7, "ts": 1776649790549085, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549075, "dur": 6, + "args": { + "External id": 2520, + "cbid": 211, "correlation": 19708 + } + }, + { + "ph": "s", "id": 19708, "pid": 435, "tid": 435, "ts": 1776649790549075, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549104, "dur": 0, + "args": { + "External id": 2522, + "cbid": 210, "correlation": 19735 + } + }, + { + "ph": "f", "id": 19735, "pid": 435, "tid": 435, "ts": 1776649790549104, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549115, "dur": 23, + "args": { + "External id": 2522, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19736, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19736, "pid": 0, "tid": 7, "ts": 1776649790549115, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549105, "dur": 6, + "args": { + "External id": 2522, + "cbid": 211, "correlation": 19736 + } + }, + { + "ph": "s", "id": 19736, "pid": 435, "tid": 435, "ts": 1776649790549105, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549133, "dur": 0, + "args": { + "External id": 2524, + "cbid": 210, "correlation": 19763 + } + }, + { + "ph": "f", "id": 19763, "pid": 435, "tid": 435, "ts": 1776649790549133, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549144, "dur": 23, + "args": { + "External id": 2524, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19764, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19764, "pid": 0, "tid": 7, "ts": 1776649790549144, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549134, "dur": 5, + "args": { + "External id": 2524, + "cbid": 211, "correlation": 19764 + } + }, + { + "ph": "s", "id": 19764, "pid": 435, "tid": 435, "ts": 1776649790549134, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549162, "dur": 0, + "args": { + "External id": 2526, + "cbid": 210, "correlation": 19791 + } + }, + { + "ph": "f", "id": 19791, "pid": 435, "tid": 435, "ts": 1776649790549162, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549174, "dur": 23, + "args": { + "External id": 2526, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19792, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19792, "pid": 0, "tid": 7, "ts": 1776649790549174, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549163, "dur": 6, + "args": { + "External id": 2526, + "cbid": 211, "correlation": 19792 + } + }, + { + "ph": "s", "id": 19792, "pid": 435, "tid": 435, "ts": 1776649790549163, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549193, "dur": 0, + "args": { + "External id": 2528, + "cbid": 210, "correlation": 19819 + } + }, + { + "ph": "f", "id": 19819, "pid": 435, "tid": 435, "ts": 1776649790549193, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549205, "dur": 23, + "args": { + "External id": 2528, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19820, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19820, "pid": 0, "tid": 7, "ts": 1776649790549205, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549194, "dur": 5, + "args": { + "External id": 2528, + "cbid": 211, "correlation": 19820 + } + }, + { + "ph": "s", "id": 19820, "pid": 435, "tid": 435, "ts": 1776649790549194, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549223, "dur": 0, + "args": { + "External id": 2530, + "cbid": 210, "correlation": 19847 + } + }, + { + "ph": "f", "id": 19847, "pid": 435, "tid": 435, "ts": 1776649790549223, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549234, "dur": 23, + "args": { + "External id": 2530, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19848, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19848, "pid": 0, "tid": 7, "ts": 1776649790549234, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549224, "dur": 6, + "args": { + "External id": 2530, + "cbid": 211, "correlation": 19848 + } + }, + { + "ph": "s", "id": 19848, "pid": 435, "tid": 435, "ts": 1776649790549224, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549252, "dur": 0, + "args": { + "External id": 2532, + "cbid": 210, "correlation": 19875 + } + }, + { + "ph": "f", "id": 19875, "pid": 435, "tid": 435, "ts": 1776649790549252, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549264, "dur": 22, + "args": { + "External id": 2532, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19876, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19876, "pid": 0, "tid": 7, "ts": 1776649790549264, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549254, "dur": 5, + "args": { + "External id": 2532, + "cbid": 211, "correlation": 19876 + } + }, + { + "ph": "s", "id": 19876, "pid": 435, "tid": 435, "ts": 1776649790549254, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549283, "dur": 0, + "args": { + "External id": 2534, + "cbid": 210, "correlation": 19903 + } + }, + { + "ph": "f", "id": 19903, "pid": 435, "tid": 435, "ts": 1776649790549283, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549295, "dur": 23, + "args": { + "External id": 2534, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19904, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19904, "pid": 0, "tid": 7, "ts": 1776649790549295, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549284, "dur": 5, + "args": { + "External id": 2534, + "cbid": 211, "correlation": 19904 + } + }, + { + "ph": "s", "id": 19904, "pid": 435, "tid": 435, "ts": 1776649790549284, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549313, "dur": 0, + "args": { + "External id": 2536, + "cbid": 210, "correlation": 19931 + } + }, + { + "ph": "f", "id": 19931, "pid": 435, "tid": 435, "ts": 1776649790549313, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549324, "dur": 23, + "args": { + "External id": 2536, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19932, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19932, "pid": 0, "tid": 7, "ts": 1776649790549324, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549314, "dur": 5, + "args": { + "External id": 2536, + "cbid": 211, "correlation": 19932 + } + }, + { + "ph": "s", "id": 19932, "pid": 435, "tid": 435, "ts": 1776649790549314, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549342, "dur": 0, + "args": { + "External id": 2538, + "cbid": 210, "correlation": 19959 + } + }, + { + "ph": "f", "id": 19959, "pid": 435, "tid": 435, "ts": 1776649790549342, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549354, "dur": 23, + "args": { + "External id": 2538, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19960, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19960, "pid": 0, "tid": 7, "ts": 1776649790549354, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549344, "dur": 5, + "args": { + "External id": 2538, + "cbid": 211, "correlation": 19960 + } + }, + { + "ph": "s", "id": 19960, "pid": 435, "tid": 435, "ts": 1776649790549344, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549372, "dur": 0, + "args": { + "External id": 2540, + "cbid": 210, "correlation": 19987 + } + }, + { + "ph": "f", "id": 19987, "pid": 435, "tid": 435, "ts": 1776649790549372, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549384, "dur": 22, + "args": { + "External id": 2540, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 19988, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 19988, "pid": 0, "tid": 7, "ts": 1776649790549384, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549373, "dur": 5, + "args": { + "External id": 2540, + "cbid": 211, "correlation": 19988 + } + }, + { + "ph": "s", "id": 19988, "pid": 435, "tid": 435, "ts": 1776649790549373, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549401, "dur": 0, + "args": { + "External id": 2542, + "cbid": 210, "correlation": 20015 + } + }, + { + "ph": "f", "id": 20015, "pid": 435, "tid": 435, "ts": 1776649790549401, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549412, "dur": 23, + "args": { + "External id": 2542, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20016, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20016, "pid": 0, "tid": 7, "ts": 1776649790549412, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549403, "dur": 6, + "args": { + "External id": 2542, + "cbid": 211, "correlation": 20016 + } + }, + { + "ph": "s", "id": 20016, "pid": 435, "tid": 435, "ts": 1776649790549403, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549431, "dur": 0, + "args": { + "External id": 2544, + "cbid": 210, "correlation": 20043 + } + }, + { + "ph": "f", "id": 20043, "pid": 435, "tid": 435, "ts": 1776649790549431, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549442, "dur": 22, + "args": { + "External id": 2544, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20044, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20044, "pid": 0, "tid": 7, "ts": 1776649790549442, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549432, "dur": 5, + "args": { + "External id": 2544, + "cbid": 211, "correlation": 20044 + } + }, + { + "ph": "s", "id": 20044, "pid": 435, "tid": 435, "ts": 1776649790549432, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549460, "dur": 0, + "args": { + "External id": 2546, + "cbid": 210, "correlation": 20071 + } + }, + { + "ph": "f", "id": 20071, "pid": 435, "tid": 435, "ts": 1776649790549460, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549472, "dur": 23, + "args": { + "External id": 2546, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20072, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20072, "pid": 0, "tid": 7, "ts": 1776649790549472, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549461, "dur": 5, + "args": { + "External id": 2546, + "cbid": 211, "correlation": 20072 + } + }, + { + "ph": "s", "id": 20072, "pid": 435, "tid": 435, "ts": 1776649790549461, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549490, "dur": 0, + "args": { + "External id": 2548, + "cbid": 210, "correlation": 20099 + } + }, + { + "ph": "f", "id": 20099, "pid": 435, "tid": 435, "ts": 1776649790549490, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549501, "dur": 22, + "args": { + "External id": 2548, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20100, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20100, "pid": 0, "tid": 7, "ts": 1776649790549501, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549491, "dur": 5, + "args": { + "External id": 2548, + "cbid": 211, "correlation": 20100 + } + }, + { + "ph": "s", "id": 20100, "pid": 435, "tid": 435, "ts": 1776649790549491, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549519, "dur": 0, + "args": { + "External id": 2550, + "cbid": 210, "correlation": 20127 + } + }, + { + "ph": "f", "id": 20127, "pid": 435, "tid": 435, "ts": 1776649790549519, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549530, "dur": 23, + "args": { + "External id": 2550, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20128, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20128, "pid": 0, "tid": 7, "ts": 1776649790549530, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549520, "dur": 5, + "args": { + "External id": 2550, + "cbid": 211, "correlation": 20128 + } + }, + { + "ph": "s", "id": 20128, "pid": 435, "tid": 435, "ts": 1776649790549520, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549548, "dur": 0, + "args": { + "External id": 2552, + "cbid": 210, "correlation": 20155 + } + }, + { + "ph": "f", "id": 20155, "pid": 435, "tid": 435, "ts": 1776649790549548, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549559, "dur": 23, + "args": { + "External id": 2552, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20156, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20156, "pid": 0, "tid": 7, "ts": 1776649790549559, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549549, "dur": 5, + "args": { + "External id": 2552, + "cbid": 211, "correlation": 20156 + } + }, + { + "ph": "s", "id": 20156, "pid": 435, "tid": 435, "ts": 1776649790549549, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549577, "dur": 0, + "args": { + "External id": 2554, + "cbid": 210, "correlation": 20183 + } + }, + { + "ph": "f", "id": 20183, "pid": 435, "tid": 435, "ts": 1776649790549577, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549588, "dur": 23, + "args": { + "External id": 2554, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20184, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20184, "pid": 0, "tid": 7, "ts": 1776649790549588, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549578, "dur": 6, + "args": { + "External id": 2554, + "cbid": 211, "correlation": 20184 + } + }, + { + "ph": "s", "id": 20184, "pid": 435, "tid": 435, "ts": 1776649790549578, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549607, "dur": 0, + "args": { + "External id": 2556, + "cbid": 210, "correlation": 20211 + } + }, + { + "ph": "f", "id": 20211, "pid": 435, "tid": 435, "ts": 1776649790549607, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549618, "dur": 23, + "args": { + "External id": 2556, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20212, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20212, "pid": 0, "tid": 7, "ts": 1776649790549618, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549608, "dur": 5, + "args": { + "External id": 2556, + "cbid": 211, "correlation": 20212 + } + }, + { + "ph": "s", "id": 20212, "pid": 435, "tid": 435, "ts": 1776649790549608, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549636, "dur": 0, + "args": { + "External id": 2558, + "cbid": 210, "correlation": 20239 + } + }, + { + "ph": "f", "id": 20239, "pid": 435, "tid": 435, "ts": 1776649790549636, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549647, "dur": 23, + "args": { + "External id": 2558, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20240, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20240, "pid": 0, "tid": 7, "ts": 1776649790549647, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549638, "dur": 5, + "args": { + "External id": 2558, + "cbid": 211, "correlation": 20240 + } + }, + { + "ph": "s", "id": 20240, "pid": 435, "tid": 435, "ts": 1776649790549638, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549666, "dur": 0, + "args": { + "External id": 2560, + "cbid": 210, "correlation": 20267 + } + }, + { + "ph": "f", "id": 20267, "pid": 435, "tid": 435, "ts": 1776649790549666, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549677, "dur": 22, + "args": { + "External id": 2560, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20268, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20268, "pid": 0, "tid": 7, "ts": 1776649790549677, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549667, "dur": 5, + "args": { + "External id": 2560, + "cbid": 211, "correlation": 20268 + } + }, + { + "ph": "s", "id": 20268, "pid": 435, "tid": 435, "ts": 1776649790549667, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549705, "dur": 0, + "args": { + "External id": 2562, + "cbid": 210, "correlation": 20295 + } + }, + { + "ph": "f", "id": 20295, "pid": 435, "tid": 435, "ts": 1776649790549705, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549714, "dur": 23, + "args": { + "External id": 2562, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20296, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20296, "pid": 0, "tid": 7, "ts": 1776649790549714, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549706, "dur": 6, + "args": { + "External id": 2562, + "cbid": 211, "correlation": 20296 + } + }, + { + "ph": "s", "id": 20296, "pid": 435, "tid": 435, "ts": 1776649790549706, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549735, "dur": 0, + "args": { + "External id": 2564, + "cbid": 210, "correlation": 20323 + } + }, + { + "ph": "f", "id": 20323, "pid": 435, "tid": 435, "ts": 1776649790549735, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549745, "dur": 22, + "args": { + "External id": 2564, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20324, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20324, "pid": 0, "tid": 7, "ts": 1776649790549745, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549736, "dur": 6, + "args": { + "External id": 2564, + "cbid": 211, "correlation": 20324 + } + }, + { + "ph": "s", "id": 20324, "pid": 435, "tid": 435, "ts": 1776649790549736, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549765, "dur": 0, + "args": { + "External id": 2566, + "cbid": 210, "correlation": 20351 + } + }, + { + "ph": "f", "id": 20351, "pid": 435, "tid": 435, "ts": 1776649790549765, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549774, "dur": 23, + "args": { + "External id": 2566, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20352, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20352, "pid": 0, "tid": 7, "ts": 1776649790549774, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549766, "dur": 5, + "args": { + "External id": 2566, + "cbid": 211, "correlation": 20352 + } + }, + { + "ph": "s", "id": 20352, "pid": 435, "tid": 435, "ts": 1776649790549766, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549794, "dur": 0, + "args": { + "External id": 2568, + "cbid": 210, "correlation": 20379 + } + }, + { + "ph": "f", "id": 20379, "pid": 435, "tid": 435, "ts": 1776649790549794, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549806, "dur": 23, + "args": { + "External id": 2568, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20380, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20380, "pid": 0, "tid": 7, "ts": 1776649790549806, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549795, "dur": 6, + "args": { + "External id": 2568, + "cbid": 211, "correlation": 20380 + } + }, + { + "ph": "s", "id": 20380, "pid": 435, "tid": 435, "ts": 1776649790549795, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549824, "dur": 0, + "args": { + "External id": 2570, + "cbid": 210, "correlation": 20407 + } + }, + { + "ph": "f", "id": 20407, "pid": 435, "tid": 435, "ts": 1776649790549824, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549835, "dur": 23, + "args": { + "External id": 2570, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20408, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20408, "pid": 0, "tid": 7, "ts": 1776649790549835, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549825, "dur": 6, + "args": { + "External id": 2570, + "cbid": 211, "correlation": 20408 + } + }, + { + "ph": "s", "id": 20408, "pid": 435, "tid": 435, "ts": 1776649790549825, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549853, "dur": 0, + "args": { + "External id": 2572, + "cbid": 210, "correlation": 20435 + } + }, + { + "ph": "f", "id": 20435, "pid": 435, "tid": 435, "ts": 1776649790549853, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549865, "dur": 23, + "args": { + "External id": 2572, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20436, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20436, "pid": 0, "tid": 7, "ts": 1776649790549865, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549855, "dur": 6, + "args": { + "External id": 2572, + "cbid": 211, "correlation": 20436 + } + }, + { + "ph": "s", "id": 20436, "pid": 435, "tid": 435, "ts": 1776649790549855, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549883, "dur": 0, + "args": { + "External id": 2574, + "cbid": 210, "correlation": 20463 + } + }, + { + "ph": "f", "id": 20463, "pid": 435, "tid": 435, "ts": 1776649790549883, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549894, "dur": 23, + "args": { + "External id": 2574, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20464, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20464, "pid": 0, "tid": 7, "ts": 1776649790549894, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549884, "dur": 5, + "args": { + "External id": 2574, + "cbid": 211, "correlation": 20464 + } + }, + { + "ph": "s", "id": 20464, "pid": 435, "tid": 435, "ts": 1776649790549884, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549913, "dur": 0, + "args": { + "External id": 2576, + "cbid": 210, "correlation": 20491 + } + }, + { + "ph": "f", "id": 20491, "pid": 435, "tid": 435, "ts": 1776649790549913, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549924, "dur": 23, + "args": { + "External id": 2576, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20492, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20492, "pid": 0, "tid": 7, "ts": 1776649790549924, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549914, "dur": 6, + "args": { + "External id": 2576, + "cbid": 211, "correlation": 20492 + } + }, + { + "ph": "s", "id": 20492, "pid": 435, "tid": 435, "ts": 1776649790549914, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549943, "dur": 0, + "args": { + "External id": 2578, + "cbid": 210, "correlation": 20519 + } + }, + { + "ph": "f", "id": 20519, "pid": 435, "tid": 435, "ts": 1776649790549943, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549954, "dur": 23, + "args": { + "External id": 2578, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20520, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20520, "pid": 0, "tid": 7, "ts": 1776649790549954, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549944, "dur": 5, + "args": { + "External id": 2578, + "cbid": 211, "correlation": 20520 + } + }, + { + "ph": "s", "id": 20520, "pid": 435, "tid": 435, "ts": 1776649790549944, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790549972, "dur": 0, + "args": { + "External id": 2580, + "cbid": 210, "correlation": 20547 + } + }, + { + "ph": "f", "id": 20547, "pid": 435, "tid": 435, "ts": 1776649790549972, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790549984, "dur": 23, + "args": { + "External id": 2580, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20548, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20548, "pid": 0, "tid": 7, "ts": 1776649790549984, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790549974, "dur": 5, + "args": { + "External id": 2580, + "cbid": 211, "correlation": 20548 + } + }, + { + "ph": "s", "id": 20548, "pid": 435, "tid": 435, "ts": 1776649790549974, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550002, "dur": 0, + "args": { + "External id": 2582, + "cbid": 210, "correlation": 20575 + } + }, + { + "ph": "f", "id": 20575, "pid": 435, "tid": 435, "ts": 1776649790550002, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550014, "dur": 23, + "args": { + "External id": 2582, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20576, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20576, "pid": 0, "tid": 7, "ts": 1776649790550014, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550003, "dur": 5, + "args": { + "External id": 2582, + "cbid": 211, "correlation": 20576 + } + }, + { + "ph": "s", "id": 20576, "pid": 435, "tid": 435, "ts": 1776649790550003, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550032, "dur": 0, + "args": { + "External id": 2584, + "cbid": 210, "correlation": 20603 + } + }, + { + "ph": "f", "id": 20603, "pid": 435, "tid": 435, "ts": 1776649790550032, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550044, "dur": 23, + "args": { + "External id": 2584, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20604, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20604, "pid": 0, "tid": 7, "ts": 1776649790550044, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550033, "dur": 5, + "args": { + "External id": 2584, + "cbid": 211, "correlation": 20604 + } + }, + { + "ph": "s", "id": 20604, "pid": 435, "tid": 435, "ts": 1776649790550033, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790550067, "dur": 3, + "args": { + "External id": 20615, + "cbid": 135, "correlation": 20615 + } + }, + { + "ph": "f", "id": 20615, "pid": 435, "tid": 435, "ts": 1776649790550067, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790550082, "dur": 5, + "args": { + "External id": 20620, + "cbid": 165, "correlation": 20620 + } + }, + { + "ph": "s", "id": 20620, "pid": 435, "tid": 435, "ts": 1776649790550082, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649790550090, "dur": 5, + "args": { + "External id": 20623, + "cbid": 139, "correlation": 20623 + } + }, + { + "ph": "f", "id": 20623, "pid": 435, "tid": 435, "ts": 1776649790550090, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550128, "dur": 1, + "args": { + "External id": 2586, + "cbid": 210, "correlation": 20657 + } + }, + { + "ph": "f", "id": 20657, "pid": 435, "tid": 435, "ts": 1776649790550128, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550141, "dur": 23, + "args": { + "External id": 2586, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20658, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20658, "pid": 0, "tid": 7, "ts": 1776649790550141, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550130, "dur": 7, + "args": { + "External id": 2586, + "cbid": 211, "correlation": 20658 + } + }, + { + "ph": "s", "id": 20658, "pid": 435, "tid": 435, "ts": 1776649790550130, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550164, "dur": 0, + "args": { + "External id": 2588, + "cbid": 210, "correlation": 20685 + } + }, + { + "ph": "f", "id": 20685, "pid": 435, "tid": 435, "ts": 1776649790550164, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550175, "dur": 23, + "args": { + "External id": 2588, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20686, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20686, "pid": 0, "tid": 7, "ts": 1776649790550175, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550165, "dur": 6, + "args": { + "External id": 2588, + "cbid": 211, "correlation": 20686 + } + }, + { + "ph": "s", "id": 20686, "pid": 435, "tid": 435, "ts": 1776649790550165, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550195, "dur": 0, + "args": { + "External id": 2590, + "cbid": 210, "correlation": 20713 + } + }, + { + "ph": "f", "id": 20713, "pid": 435, "tid": 435, "ts": 1776649790550195, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550205, "dur": 23, + "args": { + "External id": 2590, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20714, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20714, "pid": 0, "tid": 7, "ts": 1776649790550205, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550196, "dur": 6, + "args": { + "External id": 2590, + "cbid": 211, "correlation": 20714 + } + }, + { + "ph": "s", "id": 20714, "pid": 435, "tid": 435, "ts": 1776649790550196, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550225, "dur": 0, + "args": { + "External id": 2592, + "cbid": 210, "correlation": 20741 + } + }, + { + "ph": "f", "id": 20741, "pid": 435, "tid": 435, "ts": 1776649790550225, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550236, "dur": 23, + "args": { + "External id": 2592, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20742, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20742, "pid": 0, "tid": 7, "ts": 1776649790550236, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550227, "dur": 6, + "args": { + "External id": 2592, + "cbid": 211, "correlation": 20742 + } + }, + { + "ph": "s", "id": 20742, "pid": 435, "tid": 435, "ts": 1776649790550227, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550255, "dur": 0, + "args": { + "External id": 2594, + "cbid": 210, "correlation": 20769 + } + }, + { + "ph": "f", "id": 20769, "pid": 435, "tid": 435, "ts": 1776649790550255, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550266, "dur": 23, + "args": { + "External id": 2594, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20770, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20770, "pid": 0, "tid": 7, "ts": 1776649790550266, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550256, "dur": 6, + "args": { + "External id": 2594, + "cbid": 211, "correlation": 20770 + } + }, + { + "ph": "s", "id": 20770, "pid": 435, "tid": 435, "ts": 1776649790550256, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550285, "dur": 0, + "args": { + "External id": 2596, + "cbid": 210, "correlation": 20797 + } + }, + { + "ph": "f", "id": 20797, "pid": 435, "tid": 435, "ts": 1776649790550285, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550295, "dur": 23, + "args": { + "External id": 2596, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20798, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20798, "pid": 0, "tid": 7, "ts": 1776649790550295, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550286, "dur": 6, + "args": { + "External id": 2596, + "cbid": 211, "correlation": 20798 + } + }, + { + "ph": "s", "id": 20798, "pid": 435, "tid": 435, "ts": 1776649790550286, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550315, "dur": 0, + "args": { + "External id": 2598, + "cbid": 210, "correlation": 20825 + } + }, + { + "ph": "f", "id": 20825, "pid": 435, "tid": 435, "ts": 1776649790550315, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550326, "dur": 23, + "args": { + "External id": 2598, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20826, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20826, "pid": 0, "tid": 7, "ts": 1776649790550326, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550317, "dur": 6, + "args": { + "External id": 2598, + "cbid": 211, "correlation": 20826 + } + }, + { + "ph": "s", "id": 20826, "pid": 435, "tid": 435, "ts": 1776649790550317, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550345, "dur": 0, + "args": { + "External id": 2600, + "cbid": 210, "correlation": 20853 + } + }, + { + "ph": "f", "id": 20853, "pid": 435, "tid": 435, "ts": 1776649790550345, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550355, "dur": 23, + "args": { + "External id": 2600, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20854, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20854, "pid": 0, "tid": 7, "ts": 1776649790550355, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550347, "dur": 6, + "args": { + "External id": 2600, + "cbid": 211, "correlation": 20854 + } + }, + { + "ph": "s", "id": 20854, "pid": 435, "tid": 435, "ts": 1776649790550347, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550375, "dur": 0, + "args": { + "External id": 2602, + "cbid": 210, "correlation": 20881 + } + }, + { + "ph": "f", "id": 20881, "pid": 435, "tid": 435, "ts": 1776649790550375, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550385, "dur": 23, + "args": { + "External id": 2602, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20882, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20882, "pid": 0, "tid": 7, "ts": 1776649790550385, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550376, "dur": 6, + "args": { + "External id": 2602, + "cbid": 211, "correlation": 20882 + } + }, + { + "ph": "s", "id": 20882, "pid": 435, "tid": 435, "ts": 1776649790550376, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550405, "dur": 0, + "args": { + "External id": 2604, + "cbid": 210, "correlation": 20909 + } + }, + { + "ph": "f", "id": 20909, "pid": 435, "tid": 435, "ts": 1776649790550405, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550416, "dur": 23, + "args": { + "External id": 2604, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20910, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20910, "pid": 0, "tid": 7, "ts": 1776649790550416, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550406, "dur": 6, + "args": { + "External id": 2604, + "cbid": 211, "correlation": 20910 + } + }, + { + "ph": "s", "id": 20910, "pid": 435, "tid": 435, "ts": 1776649790550406, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790550431, "dur": 12, + "args": { + "External id": 20916, + "cbid": 165, "correlation": 20916 + } + }, + { + "ph": "s", "id": 20916, "pid": 435, "tid": 435, "ts": 1776649790550431, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790550465, "dur": 3, + "args": { + "External id": 20925, + "cbid": 135, "correlation": 20925 + } + }, + { + "ph": "f", "id": 20925, "pid": 435, "tid": 435, "ts": 1776649790550465, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550495, "dur": 1, + "args": { + "External id": 2606, + "cbid": 210, "correlation": 20951 + } + }, + { + "ph": "f", "id": 20951, "pid": 435, "tid": 435, "ts": 1776649790550495, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550507, "dur": 23, + "args": { + "External id": 2606, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20952, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20952, "pid": 0, "tid": 7, "ts": 1776649790550507, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550497, "dur": 7, + "args": { + "External id": 2606, + "cbid": 211, "correlation": 20952 + } + }, + { + "ph": "s", "id": 20952, "pid": 435, "tid": 435, "ts": 1776649790550497, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550529, "dur": 0, + "args": { + "External id": 2608, + "cbid": 210, "correlation": 20979 + } + }, + { + "ph": "f", "id": 20979, "pid": 435, "tid": 435, "ts": 1776649790550529, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550540, "dur": 23, + "args": { + "External id": 2608, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 20980, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 20980, "pid": 0, "tid": 7, "ts": 1776649790550540, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550531, "dur": 6, + "args": { + "External id": 2608, + "cbid": 211, "correlation": 20980 + } + }, + { + "ph": "s", "id": 20980, "pid": 435, "tid": 435, "ts": 1776649790550531, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550560, "dur": 0, + "args": { + "External id": 2610, + "cbid": 210, "correlation": 21007 + } + }, + { + "ph": "f", "id": 21007, "pid": 435, "tid": 435, "ts": 1776649790550560, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550570, "dur": 23, + "args": { + "External id": 2610, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21008, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21008, "pid": 0, "tid": 7, "ts": 1776649790550570, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550561, "dur": 6, + "args": { + "External id": 2610, + "cbid": 211, "correlation": 21008 + } + }, + { + "ph": "s", "id": 21008, "pid": 435, "tid": 435, "ts": 1776649790550561, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550591, "dur": 0, + "args": { + "External id": 2612, + "cbid": 210, "correlation": 21035 + } + }, + { + "ph": "f", "id": 21035, "pid": 435, "tid": 435, "ts": 1776649790550591, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550601, "dur": 23, + "args": { + "External id": 2612, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21036, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21036, "pid": 0, "tid": 7, "ts": 1776649790550601, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550592, "dur": 6, + "args": { + "External id": 2612, + "cbid": 211, "correlation": 21036 + } + }, + { + "ph": "s", "id": 21036, "pid": 435, "tid": 435, "ts": 1776649790550592, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550621, "dur": 0, + "args": { + "External id": 2614, + "cbid": 210, "correlation": 21063 + } + }, + { + "ph": "f", "id": 21063, "pid": 435, "tid": 435, "ts": 1776649790550621, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550631, "dur": 22, + "args": { + "External id": 2614, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21064, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21064, "pid": 0, "tid": 7, "ts": 1776649790550631, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550622, "dur": 6, + "args": { + "External id": 2614, + "cbid": 211, "correlation": 21064 + } + }, + { + "ph": "s", "id": 21064, "pid": 435, "tid": 435, "ts": 1776649790550622, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550650, "dur": 0, + "args": { + "External id": 2616, + "cbid": 210, "correlation": 21091 + } + }, + { + "ph": "f", "id": 21091, "pid": 435, "tid": 435, "ts": 1776649790550650, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550661, "dur": 22, + "args": { + "External id": 2616, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21092, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21092, "pid": 0, "tid": 7, "ts": 1776649790550661, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550652, "dur": 6, + "args": { + "External id": 2616, + "cbid": 211, "correlation": 21092 + } + }, + { + "ph": "s", "id": 21092, "pid": 435, "tid": 435, "ts": 1776649790550652, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550680, "dur": 0, + "args": { + "External id": 2618, + "cbid": 210, "correlation": 21119 + } + }, + { + "ph": "f", "id": 21119, "pid": 435, "tid": 435, "ts": 1776649790550680, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550690, "dur": 23, + "args": { + "External id": 2618, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21120, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21120, "pid": 0, "tid": 7, "ts": 1776649790550690, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550681, "dur": 6, + "args": { + "External id": 2618, + "cbid": 211, "correlation": 21120 + } + }, + { + "ph": "s", "id": 21120, "pid": 435, "tid": 435, "ts": 1776649790550681, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550710, "dur": 0, + "args": { + "External id": 2620, + "cbid": 210, "correlation": 21147 + } + }, + { + "ph": "f", "id": 21147, "pid": 435, "tid": 435, "ts": 1776649790550710, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550722, "dur": 23, + "args": { + "External id": 2620, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21148, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21148, "pid": 0, "tid": 7, "ts": 1776649790550722, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550711, "dur": 6, + "args": { + "External id": 2620, + "cbid": 211, "correlation": 21148 + } + }, + { + "ph": "s", "id": 21148, "pid": 435, "tid": 435, "ts": 1776649790550711, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550740, "dur": 0, + "args": { + "External id": 2622, + "cbid": 210, "correlation": 21175 + } + }, + { + "ph": "f", "id": 21175, "pid": 435, "tid": 435, "ts": 1776649790550740, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550751, "dur": 22, + "args": { + "External id": 2622, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21176, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21176, "pid": 0, "tid": 7, "ts": 1776649790550751, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550741, "dur": 5, + "args": { + "External id": 2622, + "cbid": 211, "correlation": 21176 + } + }, + { + "ph": "s", "id": 21176, "pid": 435, "tid": 435, "ts": 1776649790550741, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550769, "dur": 0, + "args": { + "External id": 2624, + "cbid": 210, "correlation": 21203 + } + }, + { + "ph": "f", "id": 21203, "pid": 435, "tid": 435, "ts": 1776649790550769, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550781, "dur": 23, + "args": { + "External id": 2624, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21204, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21204, "pid": 0, "tid": 7, "ts": 1776649790550781, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550770, "dur": 5, + "args": { + "External id": 2624, + "cbid": 211, "correlation": 21204 + } + }, + { + "ph": "s", "id": 21204, "pid": 435, "tid": 435, "ts": 1776649790550770, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550798, "dur": 0, + "args": { + "External id": 2626, + "cbid": 210, "correlation": 21231 + } + }, + { + "ph": "f", "id": 21231, "pid": 435, "tid": 435, "ts": 1776649790550798, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550810, "dur": 23, + "args": { + "External id": 2626, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21232, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21232, "pid": 0, "tid": 7, "ts": 1776649790550810, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550800, "dur": 6, + "args": { + "External id": 2626, + "cbid": 211, "correlation": 21232 + } + }, + { + "ph": "s", "id": 21232, "pid": 435, "tid": 435, "ts": 1776649790550800, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550828, "dur": 0, + "args": { + "External id": 2628, + "cbid": 210, "correlation": 21259 + } + }, + { + "ph": "f", "id": 21259, "pid": 435, "tid": 435, "ts": 1776649790550828, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550840, "dur": 22, + "args": { + "External id": 2628, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21260, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21260, "pid": 0, "tid": 7, "ts": 1776649790550840, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550829, "dur": 5, + "args": { + "External id": 2628, + "cbid": 211, "correlation": 21260 + } + }, + { + "ph": "s", "id": 21260, "pid": 435, "tid": 435, "ts": 1776649790550829, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550858, "dur": 0, + "args": { + "External id": 2630, + "cbid": 210, "correlation": 21287 + } + }, + { + "ph": "f", "id": 21287, "pid": 435, "tid": 435, "ts": 1776649790550858, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550869, "dur": 23, + "args": { + "External id": 2630, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21288, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21288, "pid": 0, "tid": 7, "ts": 1776649790550869, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550859, "dur": 6, + "args": { + "External id": 2630, + "cbid": 211, "correlation": 21288 + } + }, + { + "ph": "s", "id": 21288, "pid": 435, "tid": 435, "ts": 1776649790550859, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550888, "dur": 0, + "args": { + "External id": 2632, + "cbid": 210, "correlation": 21315 + } + }, + { + "ph": "f", "id": 21315, "pid": 435, "tid": 435, "ts": 1776649790550888, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550899, "dur": 23, + "args": { + "External id": 2632, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21316, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21316, "pid": 0, "tid": 7, "ts": 1776649790550899, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550889, "dur": 6, + "args": { + "External id": 2632, + "cbid": 211, "correlation": 21316 + } + }, + { + "ph": "s", "id": 21316, "pid": 435, "tid": 435, "ts": 1776649790550889, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550918, "dur": 0, + "args": { + "External id": 2634, + "cbid": 210, "correlation": 21343 + } + }, + { + "ph": "f", "id": 21343, "pid": 435, "tid": 435, "ts": 1776649790550918, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550930, "dur": 23, + "args": { + "External id": 2634, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21344, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21344, "pid": 0, "tid": 7, "ts": 1776649790550930, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550919, "dur": 5, + "args": { + "External id": 2634, + "cbid": 211, "correlation": 21344 + } + }, + { + "ph": "s", "id": 21344, "pid": 435, "tid": 435, "ts": 1776649790550919, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550948, "dur": 0, + "args": { + "External id": 2636, + "cbid": 210, "correlation": 21371 + } + }, + { + "ph": "f", "id": 21371, "pid": 435, "tid": 435, "ts": 1776649790550948, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550959, "dur": 22, + "args": { + "External id": 2636, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21372, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21372, "pid": 0, "tid": 7, "ts": 1776649790550959, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550949, "dur": 5, + "args": { + "External id": 2636, + "cbid": 211, "correlation": 21372 + } + }, + { + "ph": "s", "id": 21372, "pid": 435, "tid": 435, "ts": 1776649790550949, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790550977, "dur": 0, + "args": { + "External id": 2638, + "cbid": 210, "correlation": 21399 + } + }, + { + "ph": "f", "id": 21399, "pid": 435, "tid": 435, "ts": 1776649790550977, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790550988, "dur": 23, + "args": { + "External id": 2638, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21400, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21400, "pid": 0, "tid": 7, "ts": 1776649790550988, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790550978, "dur": 5, + "args": { + "External id": 2638, + "cbid": 211, "correlation": 21400 + } + }, + { + "ph": "s", "id": 21400, "pid": 435, "tid": 435, "ts": 1776649790550978, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551007, "dur": 0, + "args": { + "External id": 2640, + "cbid": 210, "correlation": 21427 + } + }, + { + "ph": "f", "id": 21427, "pid": 435, "tid": 435, "ts": 1776649790551007, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551018, "dur": 22, + "args": { + "External id": 2640, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21428, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21428, "pid": 0, "tid": 7, "ts": 1776649790551018, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551008, "dur": 6, + "args": { + "External id": 2640, + "cbid": 211, "correlation": 21428 + } + }, + { + "ph": "s", "id": 21428, "pid": 435, "tid": 435, "ts": 1776649790551008, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551036, "dur": 0, + "args": { + "External id": 2642, + "cbid": 210, "correlation": 21455 + } + }, + { + "ph": "f", "id": 21455, "pid": 435, "tid": 435, "ts": 1776649790551036, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551047, "dur": 23, + "args": { + "External id": 2642, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21456, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21456, "pid": 0, "tid": 7, "ts": 1776649790551047, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551037, "dur": 5, + "args": { + "External id": 2642, + "cbid": 211, "correlation": 21456 + } + }, + { + "ph": "s", "id": 21456, "pid": 435, "tid": 435, "ts": 1776649790551037, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551066, "dur": 0, + "args": { + "External id": 2644, + "cbid": 210, "correlation": 21483 + } + }, + { + "ph": "f", "id": 21483, "pid": 435, "tid": 435, "ts": 1776649790551066, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551077, "dur": 23, + "args": { + "External id": 2644, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21484, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21484, "pid": 0, "tid": 7, "ts": 1776649790551077, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551067, "dur": 5, + "args": { + "External id": 2644, + "cbid": 211, "correlation": 21484 + } + }, + { + "ph": "s", "id": 21484, "pid": 435, "tid": 435, "ts": 1776649790551067, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551095, "dur": 0, + "args": { + "External id": 2646, + "cbid": 210, "correlation": 21511 + } + }, + { + "ph": "f", "id": 21511, "pid": 435, "tid": 435, "ts": 1776649790551095, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551106, "dur": 22, + "args": { + "External id": 2646, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21512, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21512, "pid": 0, "tid": 7, "ts": 1776649790551106, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551096, "dur": 5, + "args": { + "External id": 2646, + "cbid": 211, "correlation": 21512 + } + }, + { + "ph": "s", "id": 21512, "pid": 435, "tid": 435, "ts": 1776649790551096, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551125, "dur": 0, + "args": { + "External id": 2648, + "cbid": 210, "correlation": 21539 + } + }, + { + "ph": "f", "id": 21539, "pid": 435, "tid": 435, "ts": 1776649790551125, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551136, "dur": 22, + "args": { + "External id": 2648, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21540, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21540, "pid": 0, "tid": 7, "ts": 1776649790551136, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551126, "dur": 5, + "args": { + "External id": 2648, + "cbid": 211, "correlation": 21540 + } + }, + { + "ph": "s", "id": 21540, "pid": 435, "tid": 435, "ts": 1776649790551126, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551155, "dur": 0, + "args": { + "External id": 2650, + "cbid": 210, "correlation": 21567 + } + }, + { + "ph": "f", "id": 21567, "pid": 435, "tid": 435, "ts": 1776649790551155, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551166, "dur": 22, + "args": { + "External id": 2650, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21568, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21568, "pid": 0, "tid": 7, "ts": 1776649790551166, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551156, "dur": 6, + "args": { + "External id": 2650, + "cbid": 211, "correlation": 21568 + } + }, + { + "ph": "s", "id": 21568, "pid": 435, "tid": 435, "ts": 1776649790551156, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551184, "dur": 0, + "args": { + "External id": 2652, + "cbid": 210, "correlation": 21595 + } + }, + { + "ph": "f", "id": 21595, "pid": 435, "tid": 435, "ts": 1776649790551184, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551196, "dur": 22, + "args": { + "External id": 2652, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21596, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21596, "pid": 0, "tid": 7, "ts": 1776649790551196, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551186, "dur": 5, + "args": { + "External id": 2652, + "cbid": 211, "correlation": 21596 + } + }, + { + "ph": "s", "id": 21596, "pid": 435, "tid": 435, "ts": 1776649790551186, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551214, "dur": 0, + "args": { + "External id": 2654, + "cbid": 210, "correlation": 21623 + } + }, + { + "ph": "f", "id": 21623, "pid": 435, "tid": 435, "ts": 1776649790551214, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551225, "dur": 23, + "args": { + "External id": 2654, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21624, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21624, "pid": 0, "tid": 7, "ts": 1776649790551225, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551215, "dur": 5, + "args": { + "External id": 2654, + "cbid": 211, "correlation": 21624 + } + }, + { + "ph": "s", "id": 21624, "pid": 435, "tid": 435, "ts": 1776649790551215, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551243, "dur": 0, + "args": { + "External id": 2656, + "cbid": 210, "correlation": 21651 + } + }, + { + "ph": "f", "id": 21651, "pid": 435, "tid": 435, "ts": 1776649790551243, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551255, "dur": 23, + "args": { + "External id": 2656, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21652, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21652, "pid": 0, "tid": 7, "ts": 1776649790551255, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551244, "dur": 5, + "args": { + "External id": 2656, + "cbid": 211, "correlation": 21652 + } + }, + { + "ph": "s", "id": 21652, "pid": 435, "tid": 435, "ts": 1776649790551244, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551272, "dur": 0, + "args": { + "External id": 2658, + "cbid": 210, "correlation": 21679 + } + }, + { + "ph": "f", "id": 21679, "pid": 435, "tid": 435, "ts": 1776649790551272, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551284, "dur": 23, + "args": { + "External id": 2658, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21680, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21680, "pid": 0, "tid": 7, "ts": 1776649790551284, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551274, "dur": 6, + "args": { + "External id": 2658, + "cbid": 211, "correlation": 21680 + } + }, + { + "ph": "s", "id": 21680, "pid": 435, "tid": 435, "ts": 1776649790551274, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551302, "dur": 0, + "args": { + "External id": 2660, + "cbid": 210, "correlation": 21707 + } + }, + { + "ph": "f", "id": 21707, "pid": 435, "tid": 435, "ts": 1776649790551302, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551313, "dur": 23, + "args": { + "External id": 2660, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21708, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21708, "pid": 0, "tid": 7, "ts": 1776649790551313, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551303, "dur": 5, + "args": { + "External id": 2660, + "cbid": 211, "correlation": 21708 + } + }, + { + "ph": "s", "id": 21708, "pid": 435, "tid": 435, "ts": 1776649790551303, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551332, "dur": 0, + "args": { + "External id": 2662, + "cbid": 210, "correlation": 21735 + } + }, + { + "ph": "f", "id": 21735, "pid": 435, "tid": 435, "ts": 1776649790551332, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551344, "dur": 23, + "args": { + "External id": 2662, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21736, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21736, "pid": 0, "tid": 7, "ts": 1776649790551344, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551333, "dur": 6, + "args": { + "External id": 2662, + "cbid": 211, "correlation": 21736 + } + }, + { + "ph": "s", "id": 21736, "pid": 435, "tid": 435, "ts": 1776649790551333, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551362, "dur": 0, + "args": { + "External id": 2664, + "cbid": 210, "correlation": 21763 + } + }, + { + "ph": "f", "id": 21763, "pid": 435, "tid": 435, "ts": 1776649790551362, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551373, "dur": 23, + "args": { + "External id": 2664, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21764, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21764, "pid": 0, "tid": 7, "ts": 1776649790551373, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551363, "dur": 5, + "args": { + "External id": 2664, + "cbid": 211, "correlation": 21764 + } + }, + { + "ph": "s", "id": 21764, "pid": 435, "tid": 435, "ts": 1776649790551363, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551391, "dur": 0, + "args": { + "External id": 2666, + "cbid": 210, "correlation": 21791 + } + }, + { + "ph": "f", "id": 21791, "pid": 435, "tid": 435, "ts": 1776649790551391, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551402, "dur": 22, + "args": { + "External id": 2666, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21792, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21792, "pid": 0, "tid": 7, "ts": 1776649790551402, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551392, "dur": 5, + "args": { + "External id": 2666, + "cbid": 211, "correlation": 21792 + } + }, + { + "ph": "s", "id": 21792, "pid": 435, "tid": 435, "ts": 1776649790551392, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551420, "dur": 0, + "args": { + "External id": 2668, + "cbid": 210, "correlation": 21819 + } + }, + { + "ph": "f", "id": 21819, "pid": 435, "tid": 435, "ts": 1776649790551420, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551432, "dur": 23, + "args": { + "External id": 2668, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21820, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21820, "pid": 0, "tid": 7, "ts": 1776649790551432, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551422, "dur": 5, + "args": { + "External id": 2668, + "cbid": 211, "correlation": 21820 + } + }, + { + "ph": "s", "id": 21820, "pid": 435, "tid": 435, "ts": 1776649790551422, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551450, "dur": 0, + "args": { + "External id": 2670, + "cbid": 210, "correlation": 21847 + } + }, + { + "ph": "f", "id": 21847, "pid": 435, "tid": 435, "ts": 1776649790551450, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551461, "dur": 22, + "args": { + "External id": 2670, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21848, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21848, "pid": 0, "tid": 7, "ts": 1776649790551461, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551451, "dur": 6, + "args": { + "External id": 2670, + "cbid": 211, "correlation": 21848 + } + }, + { + "ph": "s", "id": 21848, "pid": 435, "tid": 435, "ts": 1776649790551451, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551479, "dur": 0, + "args": { + "External id": 2672, + "cbid": 210, "correlation": 21875 + } + }, + { + "ph": "f", "id": 21875, "pid": 435, "tid": 435, "ts": 1776649790551479, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551490, "dur": 23, + "args": { + "External id": 2672, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21876, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21876, "pid": 0, "tid": 7, "ts": 1776649790551490, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551480, "dur": 5, + "args": { + "External id": 2672, + "cbid": 211, "correlation": 21876 + } + }, + { + "ph": "s", "id": 21876, "pid": 435, "tid": 435, "ts": 1776649790551480, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551508, "dur": 0, + "args": { + "External id": 2674, + "cbid": 210, "correlation": 21903 + } + }, + { + "ph": "f", "id": 21903, "pid": 435, "tid": 435, "ts": 1776649790551508, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551520, "dur": 23, + "args": { + "External id": 2674, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21904, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21904, "pid": 0, "tid": 7, "ts": 1776649790551520, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551510, "dur": 5, + "args": { + "External id": 2674, + "cbid": 211, "correlation": 21904 + } + }, + { + "ph": "s", "id": 21904, "pid": 435, "tid": 435, "ts": 1776649790551510, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551538, "dur": 0, + "args": { + "External id": 2676, + "cbid": 210, "correlation": 21931 + } + }, + { + "ph": "f", "id": 21931, "pid": 435, "tid": 435, "ts": 1776649790551538, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551549, "dur": 23, + "args": { + "External id": 2676, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21932, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21932, "pid": 0, "tid": 7, "ts": 1776649790551549, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551539, "dur": 5, + "args": { + "External id": 2676, + "cbid": 211, "correlation": 21932 + } + }, + { + "ph": "s", "id": 21932, "pid": 435, "tid": 435, "ts": 1776649790551539, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551567, "dur": 0, + "args": { + "External id": 2678, + "cbid": 210, "correlation": 21959 + } + }, + { + "ph": "f", "id": 21959, "pid": 435, "tid": 435, "ts": 1776649790551567, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551579, "dur": 23, + "args": { + "External id": 2678, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21960, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21960, "pid": 0, "tid": 7, "ts": 1776649790551579, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551568, "dur": 6, + "args": { + "External id": 2678, + "cbid": 211, "correlation": 21960 + } + }, + { + "ph": "s", "id": 21960, "pid": 435, "tid": 435, "ts": 1776649790551568, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551597, "dur": 0, + "args": { + "External id": 2680, + "cbid": 210, "correlation": 21987 + } + }, + { + "ph": "f", "id": 21987, "pid": 435, "tid": 435, "ts": 1776649790551597, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551608, "dur": 23, + "args": { + "External id": 2680, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 21988, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 21988, "pid": 0, "tid": 7, "ts": 1776649790551608, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551598, "dur": 5, + "args": { + "External id": 2680, + "cbid": 211, "correlation": 21988 + } + }, + { + "ph": "s", "id": 21988, "pid": 435, "tid": 435, "ts": 1776649790551598, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551626, "dur": 0, + "args": { + "External id": 2682, + "cbid": 210, "correlation": 22015 + } + }, + { + "ph": "f", "id": 22015, "pid": 435, "tid": 435, "ts": 1776649790551626, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551637, "dur": 23, + "args": { + "External id": 2682, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22016, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22016, "pid": 0, "tid": 7, "ts": 1776649790551637, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551627, "dur": 6, + "args": { + "External id": 2682, + "cbid": 211, "correlation": 22016 + } + }, + { + "ph": "s", "id": 22016, "pid": 435, "tid": 435, "ts": 1776649790551627, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551655, "dur": 0, + "args": { + "External id": 2684, + "cbid": 210, "correlation": 22043 + } + }, + { + "ph": "f", "id": 22043, "pid": 435, "tid": 435, "ts": 1776649790551655, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551667, "dur": 23, + "args": { + "External id": 2684, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22044, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22044, "pid": 0, "tid": 7, "ts": 1776649790551667, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551656, "dur": 5, + "args": { + "External id": 2684, + "cbid": 211, "correlation": 22044 + } + }, + { + "ph": "s", "id": 22044, "pid": 435, "tid": 435, "ts": 1776649790551656, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551684, "dur": 0, + "args": { + "External id": 2686, + "cbid": 210, "correlation": 22071 + } + }, + { + "ph": "f", "id": 22071, "pid": 435, "tid": 435, "ts": 1776649790551684, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551696, "dur": 22, + "args": { + "External id": 2686, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22072, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22072, "pid": 0, "tid": 7, "ts": 1776649790551696, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551686, "dur": 6, + "args": { + "External id": 2686, + "cbid": 211, "correlation": 22072 + } + }, + { + "ph": "s", "id": 22072, "pid": 435, "tid": 435, "ts": 1776649790551686, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551714, "dur": 0, + "args": { + "External id": 2688, + "cbid": 210, "correlation": 22099 + } + }, + { + "ph": "f", "id": 22099, "pid": 435, "tid": 435, "ts": 1776649790551714, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551725, "dur": 22, + "args": { + "External id": 2688, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22100, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22100, "pid": 0, "tid": 7, "ts": 1776649790551725, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551715, "dur": 5, + "args": { + "External id": 2688, + "cbid": 211, "correlation": 22100 + } + }, + { + "ph": "s", "id": 22100, "pid": 435, "tid": 435, "ts": 1776649790551715, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551743, "dur": 0, + "args": { + "External id": 2690, + "cbid": 210, "correlation": 22127 + } + }, + { + "ph": "f", "id": 22127, "pid": 435, "tid": 435, "ts": 1776649790551743, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551755, "dur": 23, + "args": { + "External id": 2690, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22128, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22128, "pid": 0, "tid": 7, "ts": 1776649790551755, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551745, "dur": 5, + "args": { + "External id": 2690, + "cbid": 211, "correlation": 22128 + } + }, + { + "ph": "s", "id": 22128, "pid": 435, "tid": 435, "ts": 1776649790551745, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551773, "dur": 0, + "args": { + "External id": 2692, + "cbid": 210, "correlation": 22155 + } + }, + { + "ph": "f", "id": 22155, "pid": 435, "tid": 435, "ts": 1776649790551773, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551784, "dur": 23, + "args": { + "External id": 2692, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22156, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22156, "pid": 0, "tid": 7, "ts": 1776649790551784, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551774, "dur": 5, + "args": { + "External id": 2692, + "cbid": 211, "correlation": 22156 + } + }, + { + "ph": "s", "id": 22156, "pid": 435, "tid": 435, "ts": 1776649790551774, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551802, "dur": 0, + "args": { + "External id": 2694, + "cbid": 210, "correlation": 22183 + } + }, + { + "ph": "f", "id": 22183, "pid": 435, "tid": 435, "ts": 1776649790551802, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551813, "dur": 23, + "args": { + "External id": 2694, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22184, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22184, "pid": 0, "tid": 7, "ts": 1776649790551813, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551803, "dur": 5, + "args": { + "External id": 2694, + "cbid": 211, "correlation": 22184 + } + }, + { + "ph": "s", "id": 22184, "pid": 435, "tid": 435, "ts": 1776649790551803, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551831, "dur": 0, + "args": { + "External id": 2696, + "cbid": 210, "correlation": 22211 + } + }, + { + "ph": "f", "id": 22211, "pid": 435, "tid": 435, "ts": 1776649790551831, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551843, "dur": 23, + "args": { + "External id": 2696, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22212, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22212, "pid": 0, "tid": 7, "ts": 1776649790551843, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551833, "dur": 6, + "args": { + "External id": 2696, + "cbid": 211, "correlation": 22212 + } + }, + { + "ph": "s", "id": 22212, "pid": 435, "tid": 435, "ts": 1776649790551833, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551861, "dur": 0, + "args": { + "External id": 2698, + "cbid": 210, "correlation": 22239 + } + }, + { + "ph": "f", "id": 22239, "pid": 435, "tid": 435, "ts": 1776649790551861, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551872, "dur": 23, + "args": { + "External id": 2698, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22240, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22240, "pid": 0, "tid": 7, "ts": 1776649790551872, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551862, "dur": 5, + "args": { + "External id": 2698, + "cbid": 211, "correlation": 22240 + } + }, + { + "ph": "s", "id": 22240, "pid": 435, "tid": 435, "ts": 1776649790551862, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551891, "dur": 0, + "args": { + "External id": 2700, + "cbid": 210, "correlation": 22267 + } + }, + { + "ph": "f", "id": 22267, "pid": 435, "tid": 435, "ts": 1776649790551891, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551902, "dur": 23, + "args": { + "External id": 2700, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22268, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22268, "pid": 0, "tid": 7, "ts": 1776649790551902, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551892, "dur": 5, + "args": { + "External id": 2700, + "cbid": 211, "correlation": 22268 + } + }, + { + "ph": "s", "id": 22268, "pid": 435, "tid": 435, "ts": 1776649790551892, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551920, "dur": 0, + "args": { + "External id": 2702, + "cbid": 210, "correlation": 22295 + } + }, + { + "ph": "f", "id": 22295, "pid": 435, "tid": 435, "ts": 1776649790551920, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551932, "dur": 23, + "args": { + "External id": 2702, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22296, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22296, "pid": 0, "tid": 7, "ts": 1776649790551932, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551922, "dur": 5, + "args": { + "External id": 2702, + "cbid": 211, "correlation": 22296 + } + }, + { + "ph": "s", "id": 22296, "pid": 435, "tid": 435, "ts": 1776649790551922, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790551950, "dur": 0, + "args": { + "External id": 2704, + "cbid": 210, "correlation": 22323 + } + }, + { + "ph": "f", "id": 22323, "pid": 435, "tid": 435, "ts": 1776649790551950, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790551961, "dur": 23, + "args": { + "External id": 2704, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22324, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22324, "pid": 0, "tid": 7, "ts": 1776649790551961, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790551951, "dur": 6, + "args": { + "External id": 2704, + "cbid": 211, "correlation": 22324 + } + }, + { + "ph": "s", "id": 22324, "pid": 435, "tid": 435, "ts": 1776649790551951, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790551983, "dur": 4, + "args": { + "External id": 22335, + "cbid": 135, "correlation": 22335 + } + }, + { + "ph": "f", "id": 22335, "pid": 435, "tid": 435, "ts": 1776649790551983, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790551999, "dur": 5, + "args": { + "External id": 22340, + "cbid": 165, "correlation": 22340 + } + }, + { + "ph": "s", "id": 22340, "pid": 435, "tid": 435, "ts": 1776649790551999, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649790552007, "dur": 4, + "args": { + "External id": 22343, + "cbid": 139, "correlation": 22343 + } + }, + { + "ph": "f", "id": 22343, "pid": 435, "tid": 435, "ts": 1776649790552007, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552044, "dur": 1, + "args": { + "External id": 2706, + "cbid": 210, "correlation": 22377 + } + }, + { + "ph": "f", "id": 22377, "pid": 435, "tid": 435, "ts": 1776649790552044, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552057, "dur": 23, + "args": { + "External id": 2706, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22378, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22378, "pid": 0, "tid": 7, "ts": 1776649790552057, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552046, "dur": 8, + "args": { + "External id": 2706, + "cbid": 211, "correlation": 22378 + } + }, + { + "ph": "s", "id": 22378, "pid": 435, "tid": 435, "ts": 1776649790552046, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552080, "dur": 0, + "args": { + "External id": 2708, + "cbid": 210, "correlation": 22405 + } + }, + { + "ph": "f", "id": 22405, "pid": 435, "tid": 435, "ts": 1776649790552080, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552090, "dur": 23, + "args": { + "External id": 2708, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22406, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22406, "pid": 0, "tid": 7, "ts": 1776649790552090, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552081, "dur": 6, + "args": { + "External id": 2708, + "cbid": 211, "correlation": 22406 + } + }, + { + "ph": "s", "id": 22406, "pid": 435, "tid": 435, "ts": 1776649790552081, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552111, "dur": 0, + "args": { + "External id": 2710, + "cbid": 210, "correlation": 22433 + } + }, + { + "ph": "f", "id": 22433, "pid": 435, "tid": 435, "ts": 1776649790552111, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552121, "dur": 23, + "args": { + "External id": 2710, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22434, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22434, "pid": 0, "tid": 7, "ts": 1776649790552121, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552112, "dur": 6, + "args": { + "External id": 2710, + "cbid": 211, "correlation": 22434 + } + }, + { + "ph": "s", "id": 22434, "pid": 435, "tid": 435, "ts": 1776649790552112, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552141, "dur": 0, + "args": { + "External id": 2712, + "cbid": 210, "correlation": 22461 + } + }, + { + "ph": "f", "id": 22461, "pid": 435, "tid": 435, "ts": 1776649790552141, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552151, "dur": 23, + "args": { + "External id": 2712, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22462, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22462, "pid": 0, "tid": 7, "ts": 1776649790552151, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552142, "dur": 6, + "args": { + "External id": 2712, + "cbid": 211, "correlation": 22462 + } + }, + { + "ph": "s", "id": 22462, "pid": 435, "tid": 435, "ts": 1776649790552142, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552172, "dur": 0, + "args": { + "External id": 2714, + "cbid": 210, "correlation": 22489 + } + }, + { + "ph": "f", "id": 22489, "pid": 435, "tid": 435, "ts": 1776649790552172, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552182, "dur": 22, + "args": { + "External id": 2714, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22490, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22490, "pid": 0, "tid": 7, "ts": 1776649790552182, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552173, "dur": 6, + "args": { + "External id": 2714, + "cbid": 211, "correlation": 22490 + } + }, + { + "ph": "s", "id": 22490, "pid": 435, "tid": 435, "ts": 1776649790552173, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552202, "dur": 0, + "args": { + "External id": 2716, + "cbid": 210, "correlation": 22517 + } + }, + { + "ph": "f", "id": 22517, "pid": 435, "tid": 435, "ts": 1776649790552202, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552212, "dur": 23, + "args": { + "External id": 2716, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22518, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22518, "pid": 0, "tid": 7, "ts": 1776649790552212, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552203, "dur": 5, + "args": { + "External id": 2716, + "cbid": 211, "correlation": 22518 + } + }, + { + "ph": "s", "id": 22518, "pid": 435, "tid": 435, "ts": 1776649790552203, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552254, "dur": 0, + "args": { + "External id": 2718, + "cbid": 210, "correlation": 22545 + } + }, + { + "ph": "f", "id": 22545, "pid": 435, "tid": 435, "ts": 1776649790552254, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552264, "dur": 23, + "args": { + "External id": 2718, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22546, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22546, "pid": 0, "tid": 7, "ts": 1776649790552264, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552255, "dur": 6, + "args": { + "External id": 2718, + "cbid": 211, "correlation": 22546 + } + }, + { + "ph": "s", "id": 22546, "pid": 435, "tid": 435, "ts": 1776649790552255, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552285, "dur": 0, + "args": { + "External id": 2720, + "cbid": 210, "correlation": 22573 + } + }, + { + "ph": "f", "id": 22573, "pid": 435, "tid": 435, "ts": 1776649790552285, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552295, "dur": 22, + "args": { + "External id": 2720, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22574, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22574, "pid": 0, "tid": 7, "ts": 1776649790552295, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552287, "dur": 6, + "args": { + "External id": 2720, + "cbid": 211, "correlation": 22574 + } + }, + { + "ph": "s", "id": 22574, "pid": 435, "tid": 435, "ts": 1776649790552287, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552316, "dur": 0, + "args": { + "External id": 2722, + "cbid": 210, "correlation": 22601 + } + }, + { + "ph": "f", "id": 22601, "pid": 435, "tid": 435, "ts": 1776649790552316, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552326, "dur": 23, + "args": { + "External id": 2722, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22602, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22602, "pid": 0, "tid": 7, "ts": 1776649790552326, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552317, "dur": 5, + "args": { + "External id": 2722, + "cbid": 211, "correlation": 22602 + } + }, + { + "ph": "s", "id": 22602, "pid": 435, "tid": 435, "ts": 1776649790552317, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552346, "dur": 0, + "args": { + "External id": 2724, + "cbid": 210, "correlation": 22629 + } + }, + { + "ph": "f", "id": 22629, "pid": 435, "tid": 435, "ts": 1776649790552346, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552356, "dur": 23, + "args": { + "External id": 2724, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22630, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22630, "pid": 0, "tid": 7, "ts": 1776649790552356, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552347, "dur": 6, + "args": { + "External id": 2724, + "cbid": 211, "correlation": 22630 + } + }, + { + "ph": "s", "id": 22630, "pid": 435, "tid": 435, "ts": 1776649790552347, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790552374, "dur": 9, + "args": { + "External id": 22636, + "cbid": 165, "correlation": 22636 + } + }, + { + "ph": "s", "id": 22636, "pid": 435, "tid": 435, "ts": 1776649790552374, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790552404, "dur": 3, + "args": { + "External id": 22645, + "cbid": 135, "correlation": 22645 + } + }, + { + "ph": "f", "id": 22645, "pid": 435, "tid": 435, "ts": 1776649790552404, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552435, "dur": 1, + "args": { + "External id": 2726, + "cbid": 210, "correlation": 22671 + } + }, + { + "ph": "f", "id": 22671, "pid": 435, "tid": 435, "ts": 1776649790552435, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552447, "dur": 23, + "args": { + "External id": 2726, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22672, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22672, "pid": 0, "tid": 7, "ts": 1776649790552447, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552436, "dur": 7, + "args": { + "External id": 2726, + "cbid": 211, "correlation": 22672 + } + }, + { + "ph": "s", "id": 22672, "pid": 435, "tid": 435, "ts": 1776649790552436, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552469, "dur": 0, + "args": { + "External id": 2728, + "cbid": 210, "correlation": 22699 + } + }, + { + "ph": "f", "id": 22699, "pid": 435, "tid": 435, "ts": 1776649790552469, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552480, "dur": 23, + "args": { + "External id": 2728, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22700, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22700, "pid": 0, "tid": 7, "ts": 1776649790552480, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552471, "dur": 6, + "args": { + "External id": 2728, + "cbid": 211, "correlation": 22700 + } + }, + { + "ph": "s", "id": 22700, "pid": 435, "tid": 435, "ts": 1776649790552471, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552500, "dur": 0, + "args": { + "External id": 2730, + "cbid": 210, "correlation": 22727 + } + }, + { + "ph": "f", "id": 22727, "pid": 435, "tid": 435, "ts": 1776649790552500, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552510, "dur": 23, + "args": { + "External id": 2730, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22728, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22728, "pid": 0, "tid": 7, "ts": 1776649790552510, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552502, "dur": 5, + "args": { + "External id": 2730, + "cbid": 211, "correlation": 22728 + } + }, + { + "ph": "s", "id": 22728, "pid": 435, "tid": 435, "ts": 1776649790552502, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552531, "dur": 0, + "args": { + "External id": 2732, + "cbid": 210, "correlation": 22755 + } + }, + { + "ph": "f", "id": 22755, "pid": 435, "tid": 435, "ts": 1776649790552531, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552541, "dur": 22, + "args": { + "External id": 2732, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22756, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22756, "pid": 0, "tid": 7, "ts": 1776649790552541, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552532, "dur": 5, + "args": { + "External id": 2732, + "cbid": 211, "correlation": 22756 + } + }, + { + "ph": "s", "id": 22756, "pid": 435, "tid": 435, "ts": 1776649790552532, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552561, "dur": 0, + "args": { + "External id": 2734, + "cbid": 210, "correlation": 22783 + } + }, + { + "ph": "f", "id": 22783, "pid": 435, "tid": 435, "ts": 1776649790552561, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552571, "dur": 23, + "args": { + "External id": 2734, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22784, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22784, "pid": 0, "tid": 7, "ts": 1776649790552571, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552562, "dur": 6, + "args": { + "External id": 2734, + "cbid": 211, "correlation": 22784 + } + }, + { + "ph": "s", "id": 22784, "pid": 435, "tid": 435, "ts": 1776649790552562, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552591, "dur": 0, + "args": { + "External id": 2736, + "cbid": 210, "correlation": 22811 + } + }, + { + "ph": "f", "id": 22811, "pid": 435, "tid": 435, "ts": 1776649790552591, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552601, "dur": 22, + "args": { + "External id": 2736, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22812, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22812, "pid": 0, "tid": 7, "ts": 1776649790552601, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552592, "dur": 6, + "args": { + "External id": 2736, + "cbid": 211, "correlation": 22812 + } + }, + { + "ph": "s", "id": 22812, "pid": 435, "tid": 435, "ts": 1776649790552592, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552621, "dur": 0, + "args": { + "External id": 2738, + "cbid": 210, "correlation": 22839 + } + }, + { + "ph": "f", "id": 22839, "pid": 435, "tid": 435, "ts": 1776649790552621, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552631, "dur": 23, + "args": { + "External id": 2738, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22840, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22840, "pid": 0, "tid": 7, "ts": 1776649790552631, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552622, "dur": 6, + "args": { + "External id": 2738, + "cbid": 211, "correlation": 22840 + } + }, + { + "ph": "s", "id": 22840, "pid": 435, "tid": 435, "ts": 1776649790552622, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552651, "dur": 0, + "args": { + "External id": 2740, + "cbid": 210, "correlation": 22867 + } + }, + { + "ph": "f", "id": 22867, "pid": 435, "tid": 435, "ts": 1776649790552651, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552662, "dur": 22, + "args": { + "External id": 2740, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22868, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22868, "pid": 0, "tid": 7, "ts": 1776649790552662, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552653, "dur": 6, + "args": { + "External id": 2740, + "cbid": 211, "correlation": 22868 + } + }, + { + "ph": "s", "id": 22868, "pid": 435, "tid": 435, "ts": 1776649790552653, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552682, "dur": 0, + "args": { + "External id": 2742, + "cbid": 210, "correlation": 22895 + } + }, + { + "ph": "f", "id": 22895, "pid": 435, "tid": 435, "ts": 1776649790552682, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552692, "dur": 23, + "args": { + "External id": 2742, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22896, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22896, "pid": 0, "tid": 7, "ts": 1776649790552692, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552683, "dur": 6, + "args": { + "External id": 2742, + "cbid": 211, "correlation": 22896 + } + }, + { + "ph": "s", "id": 22896, "pid": 435, "tid": 435, "ts": 1776649790552683, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552712, "dur": 0, + "args": { + "External id": 2744, + "cbid": 210, "correlation": 22923 + } + }, + { + "ph": "f", "id": 22923, "pid": 435, "tid": 435, "ts": 1776649790552712, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552722, "dur": 23, + "args": { + "External id": 2744, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22924, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22924, "pid": 0, "tid": 7, "ts": 1776649790552722, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552713, "dur": 6, + "args": { + "External id": 2744, + "cbid": 211, "correlation": 22924 + } + }, + { + "ph": "s", "id": 22924, "pid": 435, "tid": 435, "ts": 1776649790552713, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552742, "dur": 0, + "args": { + "External id": 2746, + "cbid": 210, "correlation": 22951 + } + }, + { + "ph": "f", "id": 22951, "pid": 435, "tid": 435, "ts": 1776649790552742, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552752, "dur": 23, + "args": { + "External id": 2746, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22952, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22952, "pid": 0, "tid": 7, "ts": 1776649790552752, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552743, "dur": 6, + "args": { + "External id": 2746, + "cbid": 211, "correlation": 22952 + } + }, + { + "ph": "s", "id": 22952, "pid": 435, "tid": 435, "ts": 1776649790552743, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552772, "dur": 0, + "args": { + "External id": 2748, + "cbid": 210, "correlation": 22979 + } + }, + { + "ph": "f", "id": 22979, "pid": 435, "tid": 435, "ts": 1776649790552772, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552783, "dur": 23, + "args": { + "External id": 2748, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 22980, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 22980, "pid": 0, "tid": 7, "ts": 1776649790552783, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552774, "dur": 6, + "args": { + "External id": 2748, + "cbid": 211, "correlation": 22980 + } + }, + { + "ph": "s", "id": 22980, "pid": 435, "tid": 435, "ts": 1776649790552774, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552803, "dur": 0, + "args": { + "External id": 2750, + "cbid": 210, "correlation": 23007 + } + }, + { + "ph": "f", "id": 23007, "pid": 435, "tid": 435, "ts": 1776649790552803, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552813, "dur": 23, + "args": { + "External id": 2750, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23008, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23008, "pid": 0, "tid": 7, "ts": 1776649790552813, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552804, "dur": 6, + "args": { + "External id": 2750, + "cbid": 211, "correlation": 23008 + } + }, + { + "ph": "s", "id": 23008, "pid": 435, "tid": 435, "ts": 1776649790552804, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552833, "dur": 0, + "args": { + "External id": 2752, + "cbid": 210, "correlation": 23035 + } + }, + { + "ph": "f", "id": 23035, "pid": 435, "tid": 435, "ts": 1776649790552833, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552843, "dur": 23, + "args": { + "External id": 2752, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23036, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23036, "pid": 0, "tid": 7, "ts": 1776649790552843, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552834, "dur": 6, + "args": { + "External id": 2752, + "cbid": 211, "correlation": 23036 + } + }, + { + "ph": "s", "id": 23036, "pid": 435, "tid": 435, "ts": 1776649790552834, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552863, "dur": 0, + "args": { + "External id": 2754, + "cbid": 210, "correlation": 23063 + } + }, + { + "ph": "f", "id": 23063, "pid": 435, "tid": 435, "ts": 1776649790552863, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552874, "dur": 23, + "args": { + "External id": 2754, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23064, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23064, "pid": 0, "tid": 7, "ts": 1776649790552874, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552865, "dur": 6, + "args": { + "External id": 2754, + "cbid": 211, "correlation": 23064 + } + }, + { + "ph": "s", "id": 23064, "pid": 435, "tid": 435, "ts": 1776649790552865, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552894, "dur": 0, + "args": { + "External id": 2756, + "cbid": 210, "correlation": 23091 + } + }, + { + "ph": "f", "id": 23091, "pid": 435, "tid": 435, "ts": 1776649790552894, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552904, "dur": 23, + "args": { + "External id": 2756, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23092, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23092, "pid": 0, "tid": 7, "ts": 1776649790552904, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552895, "dur": 6, + "args": { + "External id": 2756, + "cbid": 211, "correlation": 23092 + } + }, + { + "ph": "s", "id": 23092, "pid": 435, "tid": 435, "ts": 1776649790552895, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552924, "dur": 0, + "args": { + "External id": 2758, + "cbid": 210, "correlation": 23119 + } + }, + { + "ph": "f", "id": 23119, "pid": 435, "tid": 435, "ts": 1776649790552924, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552934, "dur": 22, + "args": { + "External id": 2758, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23120, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23120, "pid": 0, "tid": 7, "ts": 1776649790552934, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552925, "dur": 6, + "args": { + "External id": 2758, + "cbid": 211, "correlation": 23120 + } + }, + { + "ph": "s", "id": 23120, "pid": 435, "tid": 435, "ts": 1776649790552925, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552954, "dur": 0, + "args": { + "External id": 2760, + "cbid": 210, "correlation": 23147 + } + }, + { + "ph": "f", "id": 23147, "pid": 435, "tid": 435, "ts": 1776649790552954, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552964, "dur": 23, + "args": { + "External id": 2760, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23148, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23148, "pid": 0, "tid": 7, "ts": 1776649790552964, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552955, "dur": 6, + "args": { + "External id": 2760, + "cbid": 211, "correlation": 23148 + } + }, + { + "ph": "s", "id": 23148, "pid": 435, "tid": 435, "ts": 1776649790552955, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790552985, "dur": 0, + "args": { + "External id": 2762, + "cbid": 210, "correlation": 23175 + } + }, + { + "ph": "f", "id": 23175, "pid": 435, "tid": 435, "ts": 1776649790552985, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790552995, "dur": 22, + "args": { + "External id": 2762, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23176, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23176, "pid": 0, "tid": 7, "ts": 1776649790552995, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790552986, "dur": 6, + "args": { + "External id": 2762, + "cbid": 211, "correlation": 23176 + } + }, + { + "ph": "s", "id": 23176, "pid": 435, "tid": 435, "ts": 1776649790552986, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553014, "dur": 0, + "args": { + "External id": 2764, + "cbid": 210, "correlation": 23203 + } + }, + { + "ph": "f", "id": 23203, "pid": 435, "tid": 435, "ts": 1776649790553014, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553025, "dur": 23, + "args": { + "External id": 2764, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23204, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23204, "pid": 0, "tid": 7, "ts": 1776649790553025, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553016, "dur": 6, + "args": { + "External id": 2764, + "cbid": 211, "correlation": 23204 + } + }, + { + "ph": "s", "id": 23204, "pid": 435, "tid": 435, "ts": 1776649790553016, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553045, "dur": 0, + "args": { + "External id": 2766, + "cbid": 210, "correlation": 23231 + } + }, + { + "ph": "f", "id": 23231, "pid": 435, "tid": 435, "ts": 1776649790553045, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553055, "dur": 23, + "args": { + "External id": 2766, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23232, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23232, "pid": 0, "tid": 7, "ts": 1776649790553055, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553046, "dur": 6, + "args": { + "External id": 2766, + "cbid": 211, "correlation": 23232 + } + }, + { + "ph": "s", "id": 23232, "pid": 435, "tid": 435, "ts": 1776649790553046, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553075, "dur": 0, + "args": { + "External id": 2768, + "cbid": 210, "correlation": 23259 + } + }, + { + "ph": "f", "id": 23259, "pid": 435, "tid": 435, "ts": 1776649790553075, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553085, "dur": 23, + "args": { + "External id": 2768, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23260, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23260, "pid": 0, "tid": 7, "ts": 1776649790553085, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553076, "dur": 6, + "args": { + "External id": 2768, + "cbid": 211, "correlation": 23260 + } + }, + { + "ph": "s", "id": 23260, "pid": 435, "tid": 435, "ts": 1776649790553076, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553105, "dur": 0, + "args": { + "External id": 2770, + "cbid": 210, "correlation": 23287 + } + }, + { + "ph": "f", "id": 23287, "pid": 435, "tid": 435, "ts": 1776649790553105, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553115, "dur": 23, + "args": { + "External id": 2770, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23288, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23288, "pid": 0, "tid": 7, "ts": 1776649790553115, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553106, "dur": 6, + "args": { + "External id": 2770, + "cbid": 211, "correlation": 23288 + } + }, + { + "ph": "s", "id": 23288, "pid": 435, "tid": 435, "ts": 1776649790553106, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553135, "dur": 0, + "args": { + "External id": 2772, + "cbid": 210, "correlation": 23315 + } + }, + { + "ph": "f", "id": 23315, "pid": 435, "tid": 435, "ts": 1776649790553135, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553145, "dur": 23, + "args": { + "External id": 2772, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23316, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23316, "pid": 0, "tid": 7, "ts": 1776649790553145, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553136, "dur": 6, + "args": { + "External id": 2772, + "cbid": 211, "correlation": 23316 + } + }, + { + "ph": "s", "id": 23316, "pid": 435, "tid": 435, "ts": 1776649790553136, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553165, "dur": 0, + "args": { + "External id": 2774, + "cbid": 210, "correlation": 23343 + } + }, + { + "ph": "f", "id": 23343, "pid": 435, "tid": 435, "ts": 1776649790553165, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553175, "dur": 23, + "args": { + "External id": 2774, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23344, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23344, "pid": 0, "tid": 7, "ts": 1776649790553175, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553166, "dur": 6, + "args": { + "External id": 2774, + "cbid": 211, "correlation": 23344 + } + }, + { + "ph": "s", "id": 23344, "pid": 435, "tid": 435, "ts": 1776649790553166, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553195, "dur": 0, + "args": { + "External id": 2776, + "cbid": 210, "correlation": 23371 + } + }, + { + "ph": "f", "id": 23371, "pid": 435, "tid": 435, "ts": 1776649790553195, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553205, "dur": 23, + "args": { + "External id": 2776, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23372, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23372, "pid": 0, "tid": 7, "ts": 1776649790553205, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553196, "dur": 6, + "args": { + "External id": 2776, + "cbid": 211, "correlation": 23372 + } + }, + { + "ph": "s", "id": 23372, "pid": 435, "tid": 435, "ts": 1776649790553196, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553225, "dur": 0, + "args": { + "External id": 2778, + "cbid": 210, "correlation": 23399 + } + }, + { + "ph": "f", "id": 23399, "pid": 435, "tid": 435, "ts": 1776649790553225, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553235, "dur": 23, + "args": { + "External id": 2778, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23400, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23400, "pid": 0, "tid": 7, "ts": 1776649790553235, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553226, "dur": 5, + "args": { + "External id": 2778, + "cbid": 211, "correlation": 23400 + } + }, + { + "ph": "s", "id": 23400, "pid": 435, "tid": 435, "ts": 1776649790553226, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553254, "dur": 0, + "args": { + "External id": 2780, + "cbid": 210, "correlation": 23427 + } + }, + { + "ph": "f", "id": 23427, "pid": 435, "tid": 435, "ts": 1776649790553254, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553265, "dur": 23, + "args": { + "External id": 2780, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23428, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23428, "pid": 0, "tid": 7, "ts": 1776649790553265, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553256, "dur": 6, + "args": { + "External id": 2780, + "cbid": 211, "correlation": 23428 + } + }, + { + "ph": "s", "id": 23428, "pid": 435, "tid": 435, "ts": 1776649790553256, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553285, "dur": 0, + "args": { + "External id": 2782, + "cbid": 210, "correlation": 23455 + } + }, + { + "ph": "f", "id": 23455, "pid": 435, "tid": 435, "ts": 1776649790553285, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553295, "dur": 23, + "args": { + "External id": 2782, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23456, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23456, "pid": 0, "tid": 7, "ts": 1776649790553295, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553286, "dur": 6, + "args": { + "External id": 2782, + "cbid": 211, "correlation": 23456 + } + }, + { + "ph": "s", "id": 23456, "pid": 435, "tid": 435, "ts": 1776649790553286, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553315, "dur": 0, + "args": { + "External id": 2784, + "cbid": 210, "correlation": 23483 + } + }, + { + "ph": "f", "id": 23483, "pid": 435, "tid": 435, "ts": 1776649790553315, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553325, "dur": 23, + "args": { + "External id": 2784, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23484, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23484, "pid": 0, "tid": 7, "ts": 1776649790553325, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553316, "dur": 6, + "args": { + "External id": 2784, + "cbid": 211, "correlation": 23484 + } + }, + { + "ph": "s", "id": 23484, "pid": 435, "tid": 435, "ts": 1776649790553316, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553345, "dur": 0, + "args": { + "External id": 2786, + "cbid": 210, "correlation": 23511 + } + }, + { + "ph": "f", "id": 23511, "pid": 435, "tid": 435, "ts": 1776649790553345, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553355, "dur": 22, + "args": { + "External id": 2786, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23512, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23512, "pid": 0, "tid": 7, "ts": 1776649790553355, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553346, "dur": 6, + "args": { + "External id": 2786, + "cbid": 211, "correlation": 23512 + } + }, + { + "ph": "s", "id": 23512, "pid": 435, "tid": 435, "ts": 1776649790553346, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553375, "dur": 0, + "args": { + "External id": 2788, + "cbid": 210, "correlation": 23539 + } + }, + { + "ph": "f", "id": 23539, "pid": 435, "tid": 435, "ts": 1776649790553375, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553385, "dur": 23, + "args": { + "External id": 2788, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23540, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23540, "pid": 0, "tid": 7, "ts": 1776649790553385, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553376, "dur": 5, + "args": { + "External id": 2788, + "cbid": 211, "correlation": 23540 + } + }, + { + "ph": "s", "id": 23540, "pid": 435, "tid": 435, "ts": 1776649790553376, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553406, "dur": 0, + "args": { + "External id": 2790, + "cbid": 210, "correlation": 23567 + } + }, + { + "ph": "f", "id": 23567, "pid": 435, "tid": 435, "ts": 1776649790553406, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553416, "dur": 23, + "args": { + "External id": 2790, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23568, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23568, "pid": 0, "tid": 7, "ts": 1776649790553416, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553407, "dur": 6, + "args": { + "External id": 2790, + "cbid": 211, "correlation": 23568 + } + }, + { + "ph": "s", "id": 23568, "pid": 435, "tid": 435, "ts": 1776649790553407, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553436, "dur": 0, + "args": { + "External id": 2792, + "cbid": 210, "correlation": 23595 + } + }, + { + "ph": "f", "id": 23595, "pid": 435, "tid": 435, "ts": 1776649790553436, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553446, "dur": 23, + "args": { + "External id": 2792, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23596, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23596, "pid": 0, "tid": 7, "ts": 1776649790553446, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553437, "dur": 6, + "args": { + "External id": 2792, + "cbid": 211, "correlation": 23596 + } + }, + { + "ph": "s", "id": 23596, "pid": 435, "tid": 435, "ts": 1776649790553437, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553466, "dur": 0, + "args": { + "External id": 2794, + "cbid": 210, "correlation": 23623 + } + }, + { + "ph": "f", "id": 23623, "pid": 435, "tid": 435, "ts": 1776649790553466, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553476, "dur": 23, + "args": { + "External id": 2794, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23624, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23624, "pid": 0, "tid": 7, "ts": 1776649790553476, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553467, "dur": 6, + "args": { + "External id": 2794, + "cbid": 211, "correlation": 23624 + } + }, + { + "ph": "s", "id": 23624, "pid": 435, "tid": 435, "ts": 1776649790553467, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553496, "dur": 0, + "args": { + "External id": 2796, + "cbid": 210, "correlation": 23651 + } + }, + { + "ph": "f", "id": 23651, "pid": 435, "tid": 435, "ts": 1776649790553496, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553507, "dur": 23, + "args": { + "External id": 2796, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23652, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23652, "pid": 0, "tid": 7, "ts": 1776649790553507, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553498, "dur": 6, + "args": { + "External id": 2796, + "cbid": 211, "correlation": 23652 + } + }, + { + "ph": "s", "id": 23652, "pid": 435, "tid": 435, "ts": 1776649790553498, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553527, "dur": 0, + "args": { + "External id": 2798, + "cbid": 210, "correlation": 23679 + } + }, + { + "ph": "f", "id": 23679, "pid": 435, "tid": 435, "ts": 1776649790553527, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553537, "dur": 22, + "args": { + "External id": 2798, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23680, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23680, "pid": 0, "tid": 7, "ts": 1776649790553537, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553528, "dur": 6, + "args": { + "External id": 2798, + "cbid": 211, "correlation": 23680 + } + }, + { + "ph": "s", "id": 23680, "pid": 435, "tid": 435, "ts": 1776649790553528, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553557, "dur": 0, + "args": { + "External id": 2800, + "cbid": 210, "correlation": 23707 + } + }, + { + "ph": "f", "id": 23707, "pid": 435, "tid": 435, "ts": 1776649790553557, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553567, "dur": 23, + "args": { + "External id": 2800, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23708, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23708, "pid": 0, "tid": 7, "ts": 1776649790553567, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553558, "dur": 6, + "args": { + "External id": 2800, + "cbid": 211, "correlation": 23708 + } + }, + { + "ph": "s", "id": 23708, "pid": 435, "tid": 435, "ts": 1776649790553558, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553586, "dur": 0, + "args": { + "External id": 2802, + "cbid": 210, "correlation": 23735 + } + }, + { + "ph": "f", "id": 23735, "pid": 435, "tid": 435, "ts": 1776649790553586, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553596, "dur": 23, + "args": { + "External id": 2802, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23736, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23736, "pid": 0, "tid": 7, "ts": 1776649790553596, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553588, "dur": 6, + "args": { + "External id": 2802, + "cbid": 211, "correlation": 23736 + } + }, + { + "ph": "s", "id": 23736, "pid": 435, "tid": 435, "ts": 1776649790553588, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553616, "dur": 0, + "args": { + "External id": 2804, + "cbid": 210, "correlation": 23763 + } + }, + { + "ph": "f", "id": 23763, "pid": 435, "tid": 435, "ts": 1776649790553616, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553626, "dur": 22, + "args": { + "External id": 2804, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23764, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23764, "pid": 0, "tid": 7, "ts": 1776649790553626, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553618, "dur": 6, + "args": { + "External id": 2804, + "cbid": 211, "correlation": 23764 + } + }, + { + "ph": "s", "id": 23764, "pid": 435, "tid": 435, "ts": 1776649790553618, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553646, "dur": 0, + "args": { + "External id": 2806, + "cbid": 210, "correlation": 23791 + } + }, + { + "ph": "f", "id": 23791, "pid": 435, "tid": 435, "ts": 1776649790553646, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553657, "dur": 23, + "args": { + "External id": 2806, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23792, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23792, "pid": 0, "tid": 7, "ts": 1776649790553657, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553648, "dur": 6, + "args": { + "External id": 2806, + "cbid": 211, "correlation": 23792 + } + }, + { + "ph": "s", "id": 23792, "pid": 435, "tid": 435, "ts": 1776649790553648, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553676, "dur": 0, + "args": { + "External id": 2808, + "cbid": 210, "correlation": 23819 + } + }, + { + "ph": "f", "id": 23819, "pid": 435, "tid": 435, "ts": 1776649790553676, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553687, "dur": 23, + "args": { + "External id": 2808, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23820, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23820, "pid": 0, "tid": 7, "ts": 1776649790553687, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553678, "dur": 6, + "args": { + "External id": 2808, + "cbid": 211, "correlation": 23820 + } + }, + { + "ph": "s", "id": 23820, "pid": 435, "tid": 435, "ts": 1776649790553678, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553707, "dur": 0, + "args": { + "External id": 2810, + "cbid": 210, "correlation": 23847 + } + }, + { + "ph": "f", "id": 23847, "pid": 435, "tid": 435, "ts": 1776649790553707, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553717, "dur": 23, + "args": { + "External id": 2810, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23848, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23848, "pid": 0, "tid": 7, "ts": 1776649790553717, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553708, "dur": 6, + "args": { + "External id": 2810, + "cbid": 211, "correlation": 23848 + } + }, + { + "ph": "s", "id": 23848, "pid": 435, "tid": 435, "ts": 1776649790553708, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553737, "dur": 0, + "args": { + "External id": 2812, + "cbid": 210, "correlation": 23875 + } + }, + { + "ph": "f", "id": 23875, "pid": 435, "tid": 435, "ts": 1776649790553737, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553747, "dur": 23, + "args": { + "External id": 2812, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23876, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23876, "pid": 0, "tid": 7, "ts": 1776649790553747, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553738, "dur": 6, + "args": { + "External id": 2812, + "cbid": 211, "correlation": 23876 + } + }, + { + "ph": "s", "id": 23876, "pid": 435, "tid": 435, "ts": 1776649790553738, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553767, "dur": 0, + "args": { + "External id": 2814, + "cbid": 210, "correlation": 23903 + } + }, + { + "ph": "f", "id": 23903, "pid": 435, "tid": 435, "ts": 1776649790553767, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553777, "dur": 23, + "args": { + "External id": 2814, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23904, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23904, "pid": 0, "tid": 7, "ts": 1776649790553777, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553768, "dur": 6, + "args": { + "External id": 2814, + "cbid": 211, "correlation": 23904 + } + }, + { + "ph": "s", "id": 23904, "pid": 435, "tid": 435, "ts": 1776649790553768, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553797, "dur": 0, + "args": { + "External id": 2816, + "cbid": 210, "correlation": 23931 + } + }, + { + "ph": "f", "id": 23931, "pid": 435, "tid": 435, "ts": 1776649790553797, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553807, "dur": 22, + "args": { + "External id": 2816, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23932, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23932, "pid": 0, "tid": 7, "ts": 1776649790553807, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553798, "dur": 6, + "args": { + "External id": 2816, + "cbid": 211, "correlation": 23932 + } + }, + { + "ph": "s", "id": 23932, "pid": 435, "tid": 435, "ts": 1776649790553798, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553827, "dur": 0, + "args": { + "External id": 2818, + "cbid": 210, "correlation": 23959 + } + }, + { + "ph": "f", "id": 23959, "pid": 435, "tid": 435, "ts": 1776649790553827, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553838, "dur": 23, + "args": { + "External id": 2818, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23960, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23960, "pid": 0, "tid": 7, "ts": 1776649790553838, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553828, "dur": 6, + "args": { + "External id": 2818, + "cbid": 211, "correlation": 23960 + } + }, + { + "ph": "s", "id": 23960, "pid": 435, "tid": 435, "ts": 1776649790553828, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553857, "dur": 0, + "args": { + "External id": 2820, + "cbid": 210, "correlation": 23987 + } + }, + { + "ph": "f", "id": 23987, "pid": 435, "tid": 435, "ts": 1776649790553857, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553869, "dur": 23, + "args": { + "External id": 2820, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 23988, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 23988, "pid": 0, "tid": 7, "ts": 1776649790553869, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553858, "dur": 6, + "args": { + "External id": 2820, + "cbid": 211, "correlation": 23988 + } + }, + { + "ph": "s", "id": 23988, "pid": 435, "tid": 435, "ts": 1776649790553858, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553887, "dur": 0, + "args": { + "External id": 2822, + "cbid": 210, "correlation": 24015 + } + }, + { + "ph": "f", "id": 24015, "pid": 435, "tid": 435, "ts": 1776649790553887, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553899, "dur": 23, + "args": { + "External id": 2822, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24016, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24016, "pid": 0, "tid": 7, "ts": 1776649790553899, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553888, "dur": 6, + "args": { + "External id": 2822, + "cbid": 211, "correlation": 24016 + } + }, + { + "ph": "s", "id": 24016, "pid": 435, "tid": 435, "ts": 1776649790553888, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790553917, "dur": 0, + "args": { + "External id": 2824, + "cbid": 210, "correlation": 24043 + } + }, + { + "ph": "f", "id": 24043, "pid": 435, "tid": 435, "ts": 1776649790553917, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790553929, "dur": 23, + "args": { + "External id": 2824, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24044, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24044, "pid": 0, "tid": 7, "ts": 1776649790553929, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790553918, "dur": 6, + "args": { + "External id": 2824, + "cbid": 211, "correlation": 24044 + } + }, + { + "ph": "s", "id": 24044, "pid": 435, "tid": 435, "ts": 1776649790553918, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790553951, "dur": 4, + "args": { + "External id": 24055, + "cbid": 135, "correlation": 24055 + } + }, + { + "ph": "f", "id": 24055, "pid": 435, "tid": 435, "ts": 1776649790553951, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790553966, "dur": 5, + "args": { + "External id": 24060, + "cbid": 165, "correlation": 24060 + } + }, + { + "ph": "s", "id": 24060, "pid": 435, "tid": 435, "ts": 1776649790553966, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649790553974, "dur": 5, + "args": { + "External id": 24063, + "cbid": 139, "correlation": 24063 + } + }, + { + "ph": "f", "id": 24063, "pid": 435, "tid": 435, "ts": 1776649790553974, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554011, "dur": 1, + "args": { + "External id": 2826, + "cbid": 210, "correlation": 24097 + } + }, + { + "ph": "f", "id": 24097, "pid": 435, "tid": 435, "ts": 1776649790554011, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554024, "dur": 23, + "args": { + "External id": 2826, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24098, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24098, "pid": 0, "tid": 7, "ts": 1776649790554024, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554013, "dur": 7, + "args": { + "External id": 2826, + "cbid": 211, "correlation": 24098 + } + }, + { + "ph": "s", "id": 24098, "pid": 435, "tid": 435, "ts": 1776649790554013, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554047, "dur": 0, + "args": { + "External id": 2828, + "cbid": 210, "correlation": 24125 + } + }, + { + "ph": "f", "id": 24125, "pid": 435, "tid": 435, "ts": 1776649790554047, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554058, "dur": 23, + "args": { + "External id": 2828, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24126, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24126, "pid": 0, "tid": 7, "ts": 1776649790554058, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554049, "dur": 6, + "args": { + "External id": 2828, + "cbid": 211, "correlation": 24126 + } + }, + { + "ph": "s", "id": 24126, "pid": 435, "tid": 435, "ts": 1776649790554049, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554078, "dur": 0, + "args": { + "External id": 2830, + "cbid": 210, "correlation": 24153 + } + }, + { + "ph": "f", "id": 24153, "pid": 435, "tid": 435, "ts": 1776649790554078, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554089, "dur": 23, + "args": { + "External id": 2830, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24154, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24154, "pid": 0, "tid": 7, "ts": 1776649790554089, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554080, "dur": 6, + "args": { + "External id": 2830, + "cbid": 211, "correlation": 24154 + } + }, + { + "ph": "s", "id": 24154, "pid": 435, "tid": 435, "ts": 1776649790554080, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554109, "dur": 0, + "args": { + "External id": 2832, + "cbid": 210, "correlation": 24181 + } + }, + { + "ph": "f", "id": 24181, "pid": 435, "tid": 435, "ts": 1776649790554109, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554119, "dur": 23, + "args": { + "External id": 2832, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24182, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24182, "pid": 0, "tid": 7, "ts": 1776649790554119, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554110, "dur": 6, + "args": { + "External id": 2832, + "cbid": 211, "correlation": 24182 + } + }, + { + "ph": "s", "id": 24182, "pid": 435, "tid": 435, "ts": 1776649790554110, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554139, "dur": 0, + "args": { + "External id": 2834, + "cbid": 210, "correlation": 24209 + } + }, + { + "ph": "f", "id": 24209, "pid": 435, "tid": 435, "ts": 1776649790554139, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554150, "dur": 23, + "args": { + "External id": 2834, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24210, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24210, "pid": 0, "tid": 7, "ts": 1776649790554150, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554141, "dur": 6, + "args": { + "External id": 2834, + "cbid": 211, "correlation": 24210 + } + }, + { + "ph": "s", "id": 24210, "pid": 435, "tid": 435, "ts": 1776649790554141, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554170, "dur": 0, + "args": { + "External id": 2836, + "cbid": 210, "correlation": 24237 + } + }, + { + "ph": "f", "id": 24237, "pid": 435, "tid": 435, "ts": 1776649790554170, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554180, "dur": 23, + "args": { + "External id": 2836, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24238, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24238, "pid": 0, "tid": 7, "ts": 1776649790554180, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554171, "dur": 6, + "args": { + "External id": 2836, + "cbid": 211, "correlation": 24238 + } + }, + { + "ph": "s", "id": 24238, "pid": 435, "tid": 435, "ts": 1776649790554171, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554200, "dur": 0, + "args": { + "External id": 2838, + "cbid": 210, "correlation": 24265 + } + }, + { + "ph": "f", "id": 24265, "pid": 435, "tid": 435, "ts": 1776649790554200, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554211, "dur": 23, + "args": { + "External id": 2838, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24266, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24266, "pid": 0, "tid": 7, "ts": 1776649790554211, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554201, "dur": 6, + "args": { + "External id": 2838, + "cbid": 211, "correlation": 24266 + } + }, + { + "ph": "s", "id": 24266, "pid": 435, "tid": 435, "ts": 1776649790554201, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554230, "dur": 0, + "args": { + "External id": 2840, + "cbid": 210, "correlation": 24293 + } + }, + { + "ph": "f", "id": 24293, "pid": 435, "tid": 435, "ts": 1776649790554230, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554242, "dur": 23, + "args": { + "External id": 2840, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24294, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24294, "pid": 0, "tid": 7, "ts": 1776649790554242, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554231, "dur": 6, + "args": { + "External id": 2840, + "cbid": 211, "correlation": 24294 + } + }, + { + "ph": "s", "id": 24294, "pid": 435, "tid": 435, "ts": 1776649790554231, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554260, "dur": 0, + "args": { + "External id": 2842, + "cbid": 210, "correlation": 24321 + } + }, + { + "ph": "f", "id": 24321, "pid": 435, "tid": 435, "ts": 1776649790554260, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554272, "dur": 22, + "args": { + "External id": 2842, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24322, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24322, "pid": 0, "tid": 7, "ts": 1776649790554272, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554261, "dur": 6, + "args": { + "External id": 2842, + "cbid": 211, "correlation": 24322 + } + }, + { + "ph": "s", "id": 24322, "pid": 435, "tid": 435, "ts": 1776649790554261, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554290, "dur": 0, + "args": { + "External id": 2844, + "cbid": 210, "correlation": 24349 + } + }, + { + "ph": "f", "id": 24349, "pid": 435, "tid": 435, "ts": 1776649790554290, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554302, "dur": 22, + "args": { + "External id": 2844, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24350, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24350, "pid": 0, "tid": 7, "ts": 1776649790554302, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554291, "dur": 6, + "args": { + "External id": 2844, + "cbid": 211, "correlation": 24350 + } + }, + { + "ph": "s", "id": 24350, "pid": 435, "tid": 435, "ts": 1776649790554291, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790554317, "dur": 12, + "args": { + "External id": 24356, + "cbid": 165, "correlation": 24356 + } + }, + { + "ph": "s", "id": 24356, "pid": 435, "tid": 435, "ts": 1776649790554317, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790554349, "dur": 3, + "args": { + "External id": 24365, + "cbid": 135, "correlation": 24365 + } + }, + { + "ph": "f", "id": 24365, "pid": 435, "tid": 435, "ts": 1776649790554349, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554379, "dur": 1, + "args": { + "External id": 2846, + "cbid": 210, "correlation": 24391 + } + }, + { + "ph": "f", "id": 24391, "pid": 435, "tid": 435, "ts": 1776649790554379, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554391, "dur": 23, + "args": { + "External id": 2846, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24392, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24392, "pid": 0, "tid": 7, "ts": 1776649790554391, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554381, "dur": 7, + "args": { + "External id": 2846, + "cbid": 211, "correlation": 24392 + } + }, + { + "ph": "s", "id": 24392, "pid": 435, "tid": 435, "ts": 1776649790554381, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554413, "dur": 0, + "args": { + "External id": 2848, + "cbid": 210, "correlation": 24419 + } + }, + { + "ph": "f", "id": 24419, "pid": 435, "tid": 435, "ts": 1776649790554413, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554423, "dur": 23, + "args": { + "External id": 2848, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24420, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24420, "pid": 0, "tid": 7, "ts": 1776649790554423, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554414, "dur": 6, + "args": { + "External id": 2848, + "cbid": 211, "correlation": 24420 + } + }, + { + "ph": "s", "id": 24420, "pid": 435, "tid": 435, "ts": 1776649790554414, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554444, "dur": 0, + "args": { + "External id": 2850, + "cbid": 210, "correlation": 24447 + } + }, + { + "ph": "f", "id": 24447, "pid": 435, "tid": 435, "ts": 1776649790554444, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554454, "dur": 23, + "args": { + "External id": 2850, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24448, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24448, "pid": 0, "tid": 7, "ts": 1776649790554454, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554445, "dur": 6, + "args": { + "External id": 2850, + "cbid": 211, "correlation": 24448 + } + }, + { + "ph": "s", "id": 24448, "pid": 435, "tid": 435, "ts": 1776649790554445, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554474, "dur": 0, + "args": { + "External id": 2852, + "cbid": 210, "correlation": 24475 + } + }, + { + "ph": "f", "id": 24475, "pid": 435, "tid": 435, "ts": 1776649790554474, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554484, "dur": 23, + "args": { + "External id": 2852, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24476, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24476, "pid": 0, "tid": 7, "ts": 1776649790554484, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554475, "dur": 6, + "args": { + "External id": 2852, + "cbid": 211, "correlation": 24476 + } + }, + { + "ph": "s", "id": 24476, "pid": 435, "tid": 435, "ts": 1776649790554475, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554504, "dur": 0, + "args": { + "External id": 2854, + "cbid": 210, "correlation": 24503 + } + }, + { + "ph": "f", "id": 24503, "pid": 435, "tid": 435, "ts": 1776649790554504, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554514, "dur": 23, + "args": { + "External id": 2854, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24504, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24504, "pid": 0, "tid": 7, "ts": 1776649790554514, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554505, "dur": 6, + "args": { + "External id": 2854, + "cbid": 211, "correlation": 24504 + } + }, + { + "ph": "s", "id": 24504, "pid": 435, "tid": 435, "ts": 1776649790554505, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554534, "dur": 0, + "args": { + "External id": 2856, + "cbid": 210, "correlation": 24531 + } + }, + { + "ph": "f", "id": 24531, "pid": 435, "tid": 435, "ts": 1776649790554534, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554544, "dur": 23, + "args": { + "External id": 2856, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24532, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24532, "pid": 0, "tid": 7, "ts": 1776649790554544, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554535, "dur": 6, + "args": { + "External id": 2856, + "cbid": 211, "correlation": 24532 + } + }, + { + "ph": "s", "id": 24532, "pid": 435, "tid": 435, "ts": 1776649790554535, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554563, "dur": 0, + "args": { + "External id": 2858, + "cbid": 210, "correlation": 24559 + } + }, + { + "ph": "f", "id": 24559, "pid": 435, "tid": 435, "ts": 1776649790554563, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554575, "dur": 23, + "args": { + "External id": 2858, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24560, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24560, "pid": 0, "tid": 7, "ts": 1776649790554575, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554564, "dur": 6, + "args": { + "External id": 2858, + "cbid": 211, "correlation": 24560 + } + }, + { + "ph": "s", "id": 24560, "pid": 435, "tid": 435, "ts": 1776649790554564, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554593, "dur": 0, + "args": { + "External id": 2860, + "cbid": 210, "correlation": 24587 + } + }, + { + "ph": "f", "id": 24587, "pid": 435, "tid": 435, "ts": 1776649790554593, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554605, "dur": 23, + "args": { + "External id": 2860, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24588, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24588, "pid": 0, "tid": 7, "ts": 1776649790554605, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554594, "dur": 5, + "args": { + "External id": 2860, + "cbid": 211, "correlation": 24588 + } + }, + { + "ph": "s", "id": 24588, "pid": 435, "tid": 435, "ts": 1776649790554594, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554623, "dur": 0, + "args": { + "External id": 2862, + "cbid": 210, "correlation": 24615 + } + }, + { + "ph": "f", "id": 24615, "pid": 435, "tid": 435, "ts": 1776649790554623, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554634, "dur": 23, + "args": { + "External id": 2862, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24616, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24616, "pid": 0, "tid": 7, "ts": 1776649790554634, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554624, "dur": 6, + "args": { + "External id": 2862, + "cbid": 211, "correlation": 24616 + } + }, + { + "ph": "s", "id": 24616, "pid": 435, "tid": 435, "ts": 1776649790554624, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554652, "dur": 0, + "args": { + "External id": 2864, + "cbid": 210, "correlation": 24643 + } + }, + { + "ph": "f", "id": 24643, "pid": 435, "tid": 435, "ts": 1776649790554652, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554664, "dur": 23, + "args": { + "External id": 2864, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24644, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24644, "pid": 0, "tid": 7, "ts": 1776649790554664, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554654, "dur": 5, + "args": { + "External id": 2864, + "cbid": 211, "correlation": 24644 + } + }, + { + "ph": "s", "id": 24644, "pid": 435, "tid": 435, "ts": 1776649790554654, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554682, "dur": 0, + "args": { + "External id": 2866, + "cbid": 210, "correlation": 24671 + } + }, + { + "ph": "f", "id": 24671, "pid": 435, "tid": 435, "ts": 1776649790554682, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554694, "dur": 22, + "args": { + "External id": 2866, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24672, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24672, "pid": 0, "tid": 7, "ts": 1776649790554694, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554683, "dur": 6, + "args": { + "External id": 2866, + "cbid": 211, "correlation": 24672 + } + }, + { + "ph": "s", "id": 24672, "pid": 435, "tid": 435, "ts": 1776649790554683, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554712, "dur": 0, + "args": { + "External id": 2868, + "cbid": 210, "correlation": 24699 + } + }, + { + "ph": "f", "id": 24699, "pid": 435, "tid": 435, "ts": 1776649790554712, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554724, "dur": 23, + "args": { + "External id": 2868, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24700, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24700, "pid": 0, "tid": 7, "ts": 1776649790554724, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554713, "dur": 6, + "args": { + "External id": 2868, + "cbid": 211, "correlation": 24700 + } + }, + { + "ph": "s", "id": 24700, "pid": 435, "tid": 435, "ts": 1776649790554713, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554742, "dur": 0, + "args": { + "External id": 2870, + "cbid": 210, "correlation": 24727 + } + }, + { + "ph": "f", "id": 24727, "pid": 435, "tid": 435, "ts": 1776649790554742, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554754, "dur": 22, + "args": { + "External id": 2870, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24728, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24728, "pid": 0, "tid": 7, "ts": 1776649790554754, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554744, "dur": 6, + "args": { + "External id": 2870, + "cbid": 211, "correlation": 24728 + } + }, + { + "ph": "s", "id": 24728, "pid": 435, "tid": 435, "ts": 1776649790554744, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554772, "dur": 0, + "args": { + "External id": 2872, + "cbid": 210, "correlation": 24755 + } + }, + { + "ph": "f", "id": 24755, "pid": 435, "tid": 435, "ts": 1776649790554772, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554784, "dur": 23, + "args": { + "External id": 2872, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24756, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24756, "pid": 0, "tid": 7, "ts": 1776649790554784, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554773, "dur": 5, + "args": { + "External id": 2872, + "cbid": 211, "correlation": 24756 + } + }, + { + "ph": "s", "id": 24756, "pid": 435, "tid": 435, "ts": 1776649790554773, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554802, "dur": 0, + "args": { + "External id": 2874, + "cbid": 210, "correlation": 24783 + } + }, + { + "ph": "f", "id": 24783, "pid": 435, "tid": 435, "ts": 1776649790554802, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554814, "dur": 23, + "args": { + "External id": 2874, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24784, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24784, "pid": 0, "tid": 7, "ts": 1776649790554814, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554803, "dur": 6, + "args": { + "External id": 2874, + "cbid": 211, "correlation": 24784 + } + }, + { + "ph": "s", "id": 24784, "pid": 435, "tid": 435, "ts": 1776649790554803, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554832, "dur": 0, + "args": { + "External id": 2876, + "cbid": 210, "correlation": 24811 + } + }, + { + "ph": "f", "id": 24811, "pid": 435, "tid": 435, "ts": 1776649790554832, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554843, "dur": 22, + "args": { + "External id": 2876, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24812, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24812, "pid": 0, "tid": 7, "ts": 1776649790554843, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554833, "dur": 6, + "args": { + "External id": 2876, + "cbid": 211, "correlation": 24812 + } + }, + { + "ph": "s", "id": 24812, "pid": 435, "tid": 435, "ts": 1776649790554833, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554861, "dur": 0, + "args": { + "External id": 2878, + "cbid": 210, "correlation": 24839 + } + }, + { + "ph": "f", "id": 24839, "pid": 435, "tid": 435, "ts": 1776649790554861, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554873, "dur": 23, + "args": { + "External id": 2878, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24840, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24840, "pid": 0, "tid": 7, "ts": 1776649790554873, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554863, "dur": 6, + "args": { + "External id": 2878, + "cbid": 211, "correlation": 24840 + } + }, + { + "ph": "s", "id": 24840, "pid": 435, "tid": 435, "ts": 1776649790554863, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554891, "dur": 0, + "args": { + "External id": 2880, + "cbid": 210, "correlation": 24867 + } + }, + { + "ph": "f", "id": 24867, "pid": 435, "tid": 435, "ts": 1776649790554891, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554903, "dur": 22, + "args": { + "External id": 2880, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24868, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24868, "pid": 0, "tid": 7, "ts": 1776649790554903, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554892, "dur": 5, + "args": { + "External id": 2880, + "cbid": 211, "correlation": 24868 + } + }, + { + "ph": "s", "id": 24868, "pid": 435, "tid": 435, "ts": 1776649790554892, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554921, "dur": 0, + "args": { + "External id": 2882, + "cbid": 210, "correlation": 24895 + } + }, + { + "ph": "f", "id": 24895, "pid": 435, "tid": 435, "ts": 1776649790554921, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554932, "dur": 23, + "args": { + "External id": 2882, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24896, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24896, "pid": 0, "tid": 7, "ts": 1776649790554932, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554922, "dur": 5, + "args": { + "External id": 2882, + "cbid": 211, "correlation": 24896 + } + }, + { + "ph": "s", "id": 24896, "pid": 435, "tid": 435, "ts": 1776649790554922, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554950, "dur": 0, + "args": { + "External id": 2884, + "cbid": 210, "correlation": 24923 + } + }, + { + "ph": "f", "id": 24923, "pid": 435, "tid": 435, "ts": 1776649790554950, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554962, "dur": 23, + "args": { + "External id": 2884, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24924, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24924, "pid": 0, "tid": 7, "ts": 1776649790554962, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554951, "dur": 5, + "args": { + "External id": 2884, + "cbid": 211, "correlation": 24924 + } + }, + { + "ph": "s", "id": 24924, "pid": 435, "tid": 435, "ts": 1776649790554951, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790554980, "dur": 0, + "args": { + "External id": 2886, + "cbid": 210, "correlation": 24951 + } + }, + { + "ph": "f", "id": 24951, "pid": 435, "tid": 435, "ts": 1776649790554980, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790554992, "dur": 23, + "args": { + "External id": 2886, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24952, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24952, "pid": 0, "tid": 7, "ts": 1776649790554992, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790554982, "dur": 5, + "args": { + "External id": 2886, + "cbid": 211, "correlation": 24952 + } + }, + { + "ph": "s", "id": 24952, "pid": 435, "tid": 435, "ts": 1776649790554982, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555010, "dur": 0, + "args": { + "External id": 2888, + "cbid": 210, "correlation": 24979 + } + }, + { + "ph": "f", "id": 24979, "pid": 435, "tid": 435, "ts": 1776649790555010, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555022, "dur": 23, + "args": { + "External id": 2888, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 24980, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 24980, "pid": 0, "tid": 7, "ts": 1776649790555022, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555012, "dur": 5, + "args": { + "External id": 2888, + "cbid": 211, "correlation": 24980 + } + }, + { + "ph": "s", "id": 24980, "pid": 435, "tid": 435, "ts": 1776649790555012, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555040, "dur": 0, + "args": { + "External id": 2890, + "cbid": 210, "correlation": 25007 + } + }, + { + "ph": "f", "id": 25007, "pid": 435, "tid": 435, "ts": 1776649790555040, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555051, "dur": 23, + "args": { + "External id": 2890, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25008, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25008, "pid": 0, "tid": 7, "ts": 1776649790555051, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555041, "dur": 5, + "args": { + "External id": 2890, + "cbid": 211, "correlation": 25008 + } + }, + { + "ph": "s", "id": 25008, "pid": 435, "tid": 435, "ts": 1776649790555041, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555069, "dur": 0, + "args": { + "External id": 2892, + "cbid": 210, "correlation": 25035 + } + }, + { + "ph": "f", "id": 25035, "pid": 435, "tid": 435, "ts": 1776649790555069, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555081, "dur": 23, + "args": { + "External id": 2892, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25036, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25036, "pid": 0, "tid": 7, "ts": 1776649790555081, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555071, "dur": 5, + "args": { + "External id": 2892, + "cbid": 211, "correlation": 25036 + } + }, + { + "ph": "s", "id": 25036, "pid": 435, "tid": 435, "ts": 1776649790555071, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555099, "dur": 0, + "args": { + "External id": 2894, + "cbid": 210, "correlation": 25063 + } + }, + { + "ph": "f", "id": 25063, "pid": 435, "tid": 435, "ts": 1776649790555099, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555110, "dur": 23, + "args": { + "External id": 2894, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25064, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25064, "pid": 0, "tid": 7, "ts": 1776649790555110, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555100, "dur": 5, + "args": { + "External id": 2894, + "cbid": 211, "correlation": 25064 + } + }, + { + "ph": "s", "id": 25064, "pid": 435, "tid": 435, "ts": 1776649790555100, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555128, "dur": 0, + "args": { + "External id": 2896, + "cbid": 210, "correlation": 25091 + } + }, + { + "ph": "f", "id": 25091, "pid": 435, "tid": 435, "ts": 1776649790555128, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555140, "dur": 22, + "args": { + "External id": 2896, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25092, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25092, "pid": 0, "tid": 7, "ts": 1776649790555140, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555129, "dur": 5, + "args": { + "External id": 2896, + "cbid": 211, "correlation": 25092 + } + }, + { + "ph": "s", "id": 25092, "pid": 435, "tid": 435, "ts": 1776649790555129, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555157, "dur": 0, + "args": { + "External id": 2898, + "cbid": 210, "correlation": 25119 + } + }, + { + "ph": "f", "id": 25119, "pid": 435, "tid": 435, "ts": 1776649790555157, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555169, "dur": 22, + "args": { + "External id": 2898, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25120, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25120, "pid": 0, "tid": 7, "ts": 1776649790555169, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555159, "dur": 5, + "args": { + "External id": 2898, + "cbid": 211, "correlation": 25120 + } + }, + { + "ph": "s", "id": 25120, "pid": 435, "tid": 435, "ts": 1776649790555159, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555187, "dur": 0, + "args": { + "External id": 2900, + "cbid": 210, "correlation": 25147 + } + }, + { + "ph": "f", "id": 25147, "pid": 435, "tid": 435, "ts": 1776649790555187, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555198, "dur": 23, + "args": { + "External id": 2900, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25148, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25148, "pid": 0, "tid": 7, "ts": 1776649790555198, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555188, "dur": 5, + "args": { + "External id": 2900, + "cbid": 211, "correlation": 25148 + } + }, + { + "ph": "s", "id": 25148, "pid": 435, "tid": 435, "ts": 1776649790555188, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555216, "dur": 0, + "args": { + "External id": 2902, + "cbid": 210, "correlation": 25175 + } + }, + { + "ph": "f", "id": 25175, "pid": 435, "tid": 435, "ts": 1776649790555216, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555227, "dur": 23, + "args": { + "External id": 2902, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25176, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25176, "pid": 0, "tid": 7, "ts": 1776649790555227, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555217, "dur": 6, + "args": { + "External id": 2902, + "cbid": 211, "correlation": 25176 + } + }, + { + "ph": "s", "id": 25176, "pid": 435, "tid": 435, "ts": 1776649790555217, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555245, "dur": 0, + "args": { + "External id": 2904, + "cbid": 210, "correlation": 25203 + } + }, + { + "ph": "f", "id": 25203, "pid": 435, "tid": 435, "ts": 1776649790555245, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555256, "dur": 22, + "args": { + "External id": 2904, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25204, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25204, "pid": 0, "tid": 7, "ts": 1776649790555256, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555246, "dur": 5, + "args": { + "External id": 2904, + "cbid": 211, "correlation": 25204 + } + }, + { + "ph": "s", "id": 25204, "pid": 435, "tid": 435, "ts": 1776649790555246, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555274, "dur": 0, + "args": { + "External id": 2906, + "cbid": 210, "correlation": 25231 + } + }, + { + "ph": "f", "id": 25231, "pid": 435, "tid": 435, "ts": 1776649790555274, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555285, "dur": 23, + "args": { + "External id": 2906, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25232, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25232, "pid": 0, "tid": 7, "ts": 1776649790555285, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555275, "dur": 5, + "args": { + "External id": 2906, + "cbid": 211, "correlation": 25232 + } + }, + { + "ph": "s", "id": 25232, "pid": 435, "tid": 435, "ts": 1776649790555275, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555303, "dur": 0, + "args": { + "External id": 2908, + "cbid": 210, "correlation": 25259 + } + }, + { + "ph": "f", "id": 25259, "pid": 435, "tid": 435, "ts": 1776649790555303, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555315, "dur": 23, + "args": { + "External id": 2908, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25260, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25260, "pid": 0, "tid": 7, "ts": 1776649790555315, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555305, "dur": 5, + "args": { + "External id": 2908, + "cbid": 211, "correlation": 25260 + } + }, + { + "ph": "s", "id": 25260, "pid": 435, "tid": 435, "ts": 1776649790555305, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555337, "dur": 0, + "args": { + "External id": 2910, + "cbid": 210, "correlation": 25287 + } + }, + { + "ph": "f", "id": 25287, "pid": 435, "tid": 435, "ts": 1776649790555337, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555347, "dur": 23, + "args": { + "External id": 2910, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25288, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25288, "pid": 0, "tid": 7, "ts": 1776649790555347, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555338, "dur": 6, + "args": { + "External id": 2910, + "cbid": 211, "correlation": 25288 + } + }, + { + "ph": "s", "id": 25288, "pid": 435, "tid": 435, "ts": 1776649790555338, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555366, "dur": 0, + "args": { + "External id": 2912, + "cbid": 210, "correlation": 25315 + } + }, + { + "ph": "f", "id": 25315, "pid": 435, "tid": 435, "ts": 1776649790555366, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555376, "dur": 23, + "args": { + "External id": 2912, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25316, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25316, "pid": 0, "tid": 7, "ts": 1776649790555376, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555368, "dur": 6, + "args": { + "External id": 2912, + "cbid": 211, "correlation": 25316 + } + }, + { + "ph": "s", "id": 25316, "pid": 435, "tid": 435, "ts": 1776649790555368, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555396, "dur": 0, + "args": { + "External id": 2914, + "cbid": 210, "correlation": 25343 + } + }, + { + "ph": "f", "id": 25343, "pid": 435, "tid": 435, "ts": 1776649790555396, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555407, "dur": 23, + "args": { + "External id": 2914, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25344, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25344, "pid": 0, "tid": 7, "ts": 1776649790555407, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555397, "dur": 5, + "args": { + "External id": 2914, + "cbid": 211, "correlation": 25344 + } + }, + { + "ph": "s", "id": 25344, "pid": 435, "tid": 435, "ts": 1776649790555397, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555425, "dur": 0, + "args": { + "External id": 2916, + "cbid": 210, "correlation": 25371 + } + }, + { + "ph": "f", "id": 25371, "pid": 435, "tid": 435, "ts": 1776649790555425, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555436, "dur": 23, + "args": { + "External id": 2916, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25372, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25372, "pid": 0, "tid": 7, "ts": 1776649790555436, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555426, "dur": 6, + "args": { + "External id": 2916, + "cbid": 211, "correlation": 25372 + } + }, + { + "ph": "s", "id": 25372, "pid": 435, "tid": 435, "ts": 1776649790555426, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555455, "dur": 0, + "args": { + "External id": 2918, + "cbid": 210, "correlation": 25399 + } + }, + { + "ph": "f", "id": 25399, "pid": 435, "tid": 435, "ts": 1776649790555455, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555466, "dur": 23, + "args": { + "External id": 2918, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25400, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25400, "pid": 0, "tid": 7, "ts": 1776649790555466, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555456, "dur": 5, + "args": { + "External id": 2918, + "cbid": 211, "correlation": 25400 + } + }, + { + "ph": "s", "id": 25400, "pid": 435, "tid": 435, "ts": 1776649790555456, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555484, "dur": 0, + "args": { + "External id": 2920, + "cbid": 210, "correlation": 25427 + } + }, + { + "ph": "f", "id": 25427, "pid": 435, "tid": 435, "ts": 1776649790555484, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555495, "dur": 23, + "args": { + "External id": 2920, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25428, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25428, "pid": 0, "tid": 7, "ts": 1776649790555495, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555485, "dur": 5, + "args": { + "External id": 2920, + "cbid": 211, "correlation": 25428 + } + }, + { + "ph": "s", "id": 25428, "pid": 435, "tid": 435, "ts": 1776649790555485, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555513, "dur": 0, + "args": { + "External id": 2922, + "cbid": 210, "correlation": 25455 + } + }, + { + "ph": "f", "id": 25455, "pid": 435, "tid": 435, "ts": 1776649790555513, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555525, "dur": 22, + "args": { + "External id": 2922, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25456, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25456, "pid": 0, "tid": 7, "ts": 1776649790555525, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555515, "dur": 5, + "args": { + "External id": 2922, + "cbid": 211, "correlation": 25456 + } + }, + { + "ph": "s", "id": 25456, "pid": 435, "tid": 435, "ts": 1776649790555515, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555543, "dur": 0, + "args": { + "External id": 2924, + "cbid": 210, "correlation": 25483 + } + }, + { + "ph": "f", "id": 25483, "pid": 435, "tid": 435, "ts": 1776649790555543, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555554, "dur": 23, + "args": { + "External id": 2924, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25484, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25484, "pid": 0, "tid": 7, "ts": 1776649790555554, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555544, "dur": 6, + "args": { + "External id": 2924, + "cbid": 211, "correlation": 25484 + } + }, + { + "ph": "s", "id": 25484, "pid": 435, "tid": 435, "ts": 1776649790555544, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555572, "dur": 0, + "args": { + "External id": 2926, + "cbid": 210, "correlation": 25511 + } + }, + { + "ph": "f", "id": 25511, "pid": 435, "tid": 435, "ts": 1776649790555572, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555584, "dur": 22, + "args": { + "External id": 2926, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25512, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25512, "pid": 0, "tid": 7, "ts": 1776649790555584, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555573, "dur": 5, + "args": { + "External id": 2926, + "cbid": 211, "correlation": 25512 + } + }, + { + "ph": "s", "id": 25512, "pid": 435, "tid": 435, "ts": 1776649790555573, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555601, "dur": 0, + "args": { + "External id": 2928, + "cbid": 210, "correlation": 25539 + } + }, + { + "ph": "f", "id": 25539, "pid": 435, "tid": 435, "ts": 1776649790555601, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555613, "dur": 23, + "args": { + "External id": 2928, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25540, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25540, "pid": 0, "tid": 7, "ts": 1776649790555613, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555602, "dur": 6, + "args": { + "External id": 2928, + "cbid": 211, "correlation": 25540 + } + }, + { + "ph": "s", "id": 25540, "pid": 435, "tid": 435, "ts": 1776649790555602, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555630, "dur": 0, + "args": { + "External id": 2930, + "cbid": 210, "correlation": 25567 + } + }, + { + "ph": "f", "id": 25567, "pid": 435, "tid": 435, "ts": 1776649790555630, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555641, "dur": 23, + "args": { + "External id": 2930, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25568, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25568, "pid": 0, "tid": 7, "ts": 1776649790555641, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555631, "dur": 5, + "args": { + "External id": 2930, + "cbid": 211, "correlation": 25568 + } + }, + { + "ph": "s", "id": 25568, "pid": 435, "tid": 435, "ts": 1776649790555631, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555659, "dur": 0, + "args": { + "External id": 2932, + "cbid": 210, "correlation": 25595 + } + }, + { + "ph": "f", "id": 25595, "pid": 435, "tid": 435, "ts": 1776649790555659, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555671, "dur": 23, + "args": { + "External id": 2932, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25596, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25596, "pid": 0, "tid": 7, "ts": 1776649790555671, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555660, "dur": 5, + "args": { + "External id": 2932, + "cbid": 211, "correlation": 25596 + } + }, + { + "ph": "s", "id": 25596, "pid": 435, "tid": 435, "ts": 1776649790555660, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555688, "dur": 0, + "args": { + "External id": 2934, + "cbid": 210, "correlation": 25623 + } + }, + { + "ph": "f", "id": 25623, "pid": 435, "tid": 435, "ts": 1776649790555688, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555699, "dur": 23, + "args": { + "External id": 2934, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25624, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25624, "pid": 0, "tid": 7, "ts": 1776649790555699, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555689, "dur": 5, + "args": { + "External id": 2934, + "cbid": 211, "correlation": 25624 + } + }, + { + "ph": "s", "id": 25624, "pid": 435, "tid": 435, "ts": 1776649790555689, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555717, "dur": 0, + "args": { + "External id": 2936, + "cbid": 210, "correlation": 25651 + } + }, + { + "ph": "f", "id": 25651, "pid": 435, "tid": 435, "ts": 1776649790555717, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555729, "dur": 22, + "args": { + "External id": 2936, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25652, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25652, "pid": 0, "tid": 7, "ts": 1776649790555729, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555718, "dur": 5, + "args": { + "External id": 2936, + "cbid": 211, "correlation": 25652 + } + }, + { + "ph": "s", "id": 25652, "pid": 435, "tid": 435, "ts": 1776649790555718, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555746, "dur": 0, + "args": { + "External id": 2938, + "cbid": 210, "correlation": 25679 + } + }, + { + "ph": "f", "id": 25679, "pid": 435, "tid": 435, "ts": 1776649790555746, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555758, "dur": 22, + "args": { + "External id": 2938, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25680, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25680, "pid": 0, "tid": 7, "ts": 1776649790555758, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555748, "dur": 6, + "args": { + "External id": 2938, + "cbid": 211, "correlation": 25680 + } + }, + { + "ph": "s", "id": 25680, "pid": 435, "tid": 435, "ts": 1776649790555748, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555776, "dur": 0, + "args": { + "External id": 2940, + "cbid": 210, "correlation": 25707 + } + }, + { + "ph": "f", "id": 25707, "pid": 435, "tid": 435, "ts": 1776649790555776, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555787, "dur": 22, + "args": { + "External id": 2940, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25708, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25708, "pid": 0, "tid": 7, "ts": 1776649790555787, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555777, "dur": 5, + "args": { + "External id": 2940, + "cbid": 211, "correlation": 25708 + } + }, + { + "ph": "s", "id": 25708, "pid": 435, "tid": 435, "ts": 1776649790555777, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555805, "dur": 0, + "args": { + "External id": 2942, + "cbid": 210, "correlation": 25735 + } + }, + { + "ph": "f", "id": 25735, "pid": 435, "tid": 435, "ts": 1776649790555805, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555816, "dur": 23, + "args": { + "External id": 2942, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25736, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25736, "pid": 0, "tid": 7, "ts": 1776649790555816, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555806, "dur": 5, + "args": { + "External id": 2942, + "cbid": 211, "correlation": 25736 + } + }, + { + "ph": "s", "id": 25736, "pid": 435, "tid": 435, "ts": 1776649790555806, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555834, "dur": 0, + "args": { + "External id": 2944, + "cbid": 210, "correlation": 25763 + } + }, + { + "ph": "f", "id": 25763, "pid": 435, "tid": 435, "ts": 1776649790555834, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555845, "dur": 23, + "args": { + "External id": 2944, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25764, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25764, "pid": 0, "tid": 7, "ts": 1776649790555845, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555835, "dur": 5, + "args": { + "External id": 2944, + "cbid": 211, "correlation": 25764 + } + }, + { + "ph": "s", "id": 25764, "pid": 435, "tid": 435, "ts": 1776649790555835, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790555867, "dur": 3, + "args": { + "External id": 25775, + "cbid": 135, "correlation": 25775 + } + }, + { + "ph": "f", "id": 25775, "pid": 435, "tid": 435, "ts": 1776649790555867, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790555881, "dur": 5, + "args": { + "External id": 25780, + "cbid": 165, "correlation": 25780 + } + }, + { + "ph": "s", "id": 25780, "pid": 435, "tid": 435, "ts": 1776649790555881, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649790555889, "dur": 4, + "args": { + "External id": 25783, + "cbid": 139, "correlation": 25783 + } + }, + { + "ph": "f", "id": 25783, "pid": 435, "tid": 435, "ts": 1776649790555889, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555926, "dur": 1, + "args": { + "External id": 2946, + "cbid": 210, "correlation": 25817 + } + }, + { + "ph": "f", "id": 25817, "pid": 435, "tid": 435, "ts": 1776649790555926, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555938, "dur": 23, + "args": { + "External id": 2946, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25818, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25818, "pid": 0, "tid": 7, "ts": 1776649790555938, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555928, "dur": 8, + "args": { + "External id": 2946, + "cbid": 211, "correlation": 25818 + } + }, + { + "ph": "s", "id": 25818, "pid": 435, "tid": 435, "ts": 1776649790555928, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555961, "dur": 0, + "args": { + "External id": 2948, + "cbid": 210, "correlation": 25845 + } + }, + { + "ph": "f", "id": 25845, "pid": 435, "tid": 435, "ts": 1776649790555961, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790555971, "dur": 23, + "args": { + "External id": 2948, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25846, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25846, "pid": 0, "tid": 7, "ts": 1776649790555971, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555963, "dur": 6, + "args": { + "External id": 2948, + "cbid": 211, "correlation": 25846 + } + }, + { + "ph": "s", "id": 25846, "pid": 435, "tid": 435, "ts": 1776649790555963, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790555991, "dur": 0, + "args": { + "External id": 2950, + "cbid": 210, "correlation": 25873 + } + }, + { + "ph": "f", "id": 25873, "pid": 435, "tid": 435, "ts": 1776649790555991, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556002, "dur": 23, + "args": { + "External id": 2950, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25874, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25874, "pid": 0, "tid": 7, "ts": 1776649790556002, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790555993, "dur": 6, + "args": { + "External id": 2950, + "cbid": 211, "correlation": 25874 + } + }, + { + "ph": "s", "id": 25874, "pid": 435, "tid": 435, "ts": 1776649790555993, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556022, "dur": 0, + "args": { + "External id": 2952, + "cbid": 210, "correlation": 25901 + } + }, + { + "ph": "f", "id": 25901, "pid": 435, "tid": 435, "ts": 1776649790556022, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556032, "dur": 22, + "args": { + "External id": 2952, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25902, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25902, "pid": 0, "tid": 7, "ts": 1776649790556032, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556023, "dur": 6, + "args": { + "External id": 2952, + "cbid": 211, "correlation": 25902 + } + }, + { + "ph": "s", "id": 25902, "pid": 435, "tid": 435, "ts": 1776649790556023, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556052, "dur": 0, + "args": { + "External id": 2954, + "cbid": 210, "correlation": 25929 + } + }, + { + "ph": "f", "id": 25929, "pid": 435, "tid": 435, "ts": 1776649790556052, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556062, "dur": 23, + "args": { + "External id": 2954, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25930, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25930, "pid": 0, "tid": 7, "ts": 1776649790556062, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556053, "dur": 6, + "args": { + "External id": 2954, + "cbid": 211, "correlation": 25930 + } + }, + { + "ph": "s", "id": 25930, "pid": 435, "tid": 435, "ts": 1776649790556053, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556081, "dur": 0, + "args": { + "External id": 2956, + "cbid": 210, "correlation": 25957 + } + }, + { + "ph": "f", "id": 25957, "pid": 435, "tid": 435, "ts": 1776649790556081, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556093, "dur": 22, + "args": { + "External id": 2956, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25958, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25958, "pid": 0, "tid": 7, "ts": 1776649790556093, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556082, "dur": 6, + "args": { + "External id": 2956, + "cbid": 211, "correlation": 25958 + } + }, + { + "ph": "s", "id": 25958, "pid": 435, "tid": 435, "ts": 1776649790556082, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556110, "dur": 0, + "args": { + "External id": 2958, + "cbid": 210, "correlation": 25985 + } + }, + { + "ph": "f", "id": 25985, "pid": 435, "tid": 435, "ts": 1776649790556110, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556121, "dur": 22, + "args": { + "External id": 2958, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 25986, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 25986, "pid": 0, "tid": 7, "ts": 1776649790556121, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556112, "dur": 5, + "args": { + "External id": 2958, + "cbid": 211, "correlation": 25986 + } + }, + { + "ph": "s", "id": 25986, "pid": 435, "tid": 435, "ts": 1776649790556112, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556140, "dur": 0, + "args": { + "External id": 2960, + "cbid": 210, "correlation": 26013 + } + }, + { + "ph": "f", "id": 26013, "pid": 435, "tid": 435, "ts": 1776649790556140, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556151, "dur": 22, + "args": { + "External id": 2960, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26014, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26014, "pid": 0, "tid": 7, "ts": 1776649790556151, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556141, "dur": 6, + "args": { + "External id": 2960, + "cbid": 211, "correlation": 26014 + } + }, + { + "ph": "s", "id": 26014, "pid": 435, "tid": 435, "ts": 1776649790556141, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556169, "dur": 0, + "args": { + "External id": 2962, + "cbid": 210, "correlation": 26041 + } + }, + { + "ph": "f", "id": 26041, "pid": 435, "tid": 435, "ts": 1776649790556169, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556180, "dur": 22, + "args": { + "External id": 2962, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26042, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26042, "pid": 0, "tid": 7, "ts": 1776649790556180, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556170, "dur": 5, + "args": { + "External id": 2962, + "cbid": 211, "correlation": 26042 + } + }, + { + "ph": "s", "id": 26042, "pid": 435, "tid": 435, "ts": 1776649790556170, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556198, "dur": 0, + "args": { + "External id": 2964, + "cbid": 210, "correlation": 26069 + } + }, + { + "ph": "f", "id": 26069, "pid": 435, "tid": 435, "ts": 1776649790556198, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556209, "dur": 22, + "args": { + "External id": 2964, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26070, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26070, "pid": 0, "tid": 7, "ts": 1776649790556209, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556199, "dur": 5, + "args": { + "External id": 2964, + "cbid": 211, "correlation": 26070 + } + }, + { + "ph": "s", "id": 26070, "pid": 435, "tid": 435, "ts": 1776649790556199, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790556244, "dur": 5, + "args": { + "External id": 26076, + "cbid": 165, "correlation": 26076 + } + }, + { + "ph": "s", "id": 26076, "pid": 435, "tid": 435, "ts": 1776649790556244, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790556271, "dur": 3, + "args": { + "External id": 26085, + "cbid": 135, "correlation": 26085 + } + }, + { + "ph": "f", "id": 26085, "pid": 435, "tid": 435, "ts": 1776649790556271, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556303, "dur": 1, + "args": { + "External id": 2966, + "cbid": 210, "correlation": 26111 + } + }, + { + "ph": "f", "id": 26111, "pid": 435, "tid": 435, "ts": 1776649790556303, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556315, "dur": 22, + "args": { + "External id": 2966, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26112, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26112, "pid": 0, "tid": 7, "ts": 1776649790556315, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556305, "dur": 8, + "args": { + "External id": 2966, + "cbid": 211, "correlation": 26112 + } + }, + { + "ph": "s", "id": 26112, "pid": 435, "tid": 435, "ts": 1776649790556305, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556338, "dur": 0, + "args": { + "External id": 2968, + "cbid": 210, "correlation": 26139 + } + }, + { + "ph": "f", "id": 26139, "pid": 435, "tid": 435, "ts": 1776649790556338, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556348, "dur": 23, + "args": { + "External id": 2968, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26140, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26140, "pid": 0, "tid": 7, "ts": 1776649790556348, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556339, "dur": 6, + "args": { + "External id": 2968, + "cbid": 211, "correlation": 26140 + } + }, + { + "ph": "s", "id": 26140, "pid": 435, "tid": 435, "ts": 1776649790556339, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556367, "dur": 0, + "args": { + "External id": 2970, + "cbid": 210, "correlation": 26167 + } + }, + { + "ph": "f", "id": 26167, "pid": 435, "tid": 435, "ts": 1776649790556367, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556379, "dur": 22, + "args": { + "External id": 2970, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26168, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26168, "pid": 0, "tid": 7, "ts": 1776649790556379, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556369, "dur": 5, + "args": { + "External id": 2970, + "cbid": 211, "correlation": 26168 + } + }, + { + "ph": "s", "id": 26168, "pid": 435, "tid": 435, "ts": 1776649790556369, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556396, "dur": 0, + "args": { + "External id": 2972, + "cbid": 210, "correlation": 26195 + } + }, + { + "ph": "f", "id": 26195, "pid": 435, "tid": 435, "ts": 1776649790556396, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556408, "dur": 23, + "args": { + "External id": 2972, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26196, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26196, "pid": 0, "tid": 7, "ts": 1776649790556408, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556398, "dur": 6, + "args": { + "External id": 2972, + "cbid": 211, "correlation": 26196 + } + }, + { + "ph": "s", "id": 26196, "pid": 435, "tid": 435, "ts": 1776649790556398, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556426, "dur": 0, + "args": { + "External id": 2974, + "cbid": 210, "correlation": 26223 + } + }, + { + "ph": "f", "id": 26223, "pid": 435, "tid": 435, "ts": 1776649790556426, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556437, "dur": 23, + "args": { + "External id": 2974, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26224, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26224, "pid": 0, "tid": 7, "ts": 1776649790556437, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556427, "dur": 6, + "args": { + "External id": 2974, + "cbid": 211, "correlation": 26224 + } + }, + { + "ph": "s", "id": 26224, "pid": 435, "tid": 435, "ts": 1776649790556427, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556455, "dur": 0, + "args": { + "External id": 2976, + "cbid": 210, "correlation": 26251 + } + }, + { + "ph": "f", "id": 26251, "pid": 435, "tid": 435, "ts": 1776649790556455, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556467, "dur": 22, + "args": { + "External id": 2976, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26252, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26252, "pid": 0, "tid": 7, "ts": 1776649790556467, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556456, "dur": 5, + "args": { + "External id": 2976, + "cbid": 211, "correlation": 26252 + } + }, + { + "ph": "s", "id": 26252, "pid": 435, "tid": 435, "ts": 1776649790556456, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556484, "dur": 0, + "args": { + "External id": 2978, + "cbid": 210, "correlation": 26279 + } + }, + { + "ph": "f", "id": 26279, "pid": 435, "tid": 435, "ts": 1776649790556484, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556496, "dur": 22, + "args": { + "External id": 2978, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26280, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26280, "pid": 0, "tid": 7, "ts": 1776649790556496, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556485, "dur": 5, + "args": { + "External id": 2978, + "cbid": 211, "correlation": 26280 + } + }, + { + "ph": "s", "id": 26280, "pid": 435, "tid": 435, "ts": 1776649790556485, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556513, "dur": 0, + "args": { + "External id": 2980, + "cbid": 210, "correlation": 26307 + } + }, + { + "ph": "f", "id": 26307, "pid": 435, "tid": 435, "ts": 1776649790556513, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556525, "dur": 23, + "args": { + "External id": 2980, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26308, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26308, "pid": 0, "tid": 7, "ts": 1776649790556525, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556515, "dur": 5, + "args": { + "External id": 2980, + "cbid": 211, "correlation": 26308 + } + }, + { + "ph": "s", "id": 26308, "pid": 435, "tid": 435, "ts": 1776649790556515, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556543, "dur": 0, + "args": { + "External id": 2982, + "cbid": 210, "correlation": 26335 + } + }, + { + "ph": "f", "id": 26335, "pid": 435, "tid": 435, "ts": 1776649790556543, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556554, "dur": 22, + "args": { + "External id": 2982, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26336, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26336, "pid": 0, "tid": 7, "ts": 1776649790556554, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556544, "dur": 5, + "args": { + "External id": 2982, + "cbid": 211, "correlation": 26336 + } + }, + { + "ph": "s", "id": 26336, "pid": 435, "tid": 435, "ts": 1776649790556544, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556572, "dur": 0, + "args": { + "External id": 2984, + "cbid": 210, "correlation": 26363 + } + }, + { + "ph": "f", "id": 26363, "pid": 435, "tid": 435, "ts": 1776649790556572, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556583, "dur": 23, + "args": { + "External id": 2984, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26364, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26364, "pid": 0, "tid": 7, "ts": 1776649790556583, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556573, "dur": 5, + "args": { + "External id": 2984, + "cbid": 211, "correlation": 26364 + } + }, + { + "ph": "s", "id": 26364, "pid": 435, "tid": 435, "ts": 1776649790556573, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556601, "dur": 0, + "args": { + "External id": 2986, + "cbid": 210, "correlation": 26391 + } + }, + { + "ph": "f", "id": 26391, "pid": 435, "tid": 435, "ts": 1776649790556601, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556612, "dur": 23, + "args": { + "External id": 2986, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26392, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26392, "pid": 0, "tid": 7, "ts": 1776649790556612, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556602, "dur": 5, + "args": { + "External id": 2986, + "cbid": 211, "correlation": 26392 + } + }, + { + "ph": "s", "id": 26392, "pid": 435, "tid": 435, "ts": 1776649790556602, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556630, "dur": 0, + "args": { + "External id": 2988, + "cbid": 210, "correlation": 26419 + } + }, + { + "ph": "f", "id": 26419, "pid": 435, "tid": 435, "ts": 1776649790556630, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556642, "dur": 22, + "args": { + "External id": 2988, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26420, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26420, "pid": 0, "tid": 7, "ts": 1776649790556642, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556632, "dur": 5, + "args": { + "External id": 2988, + "cbid": 211, "correlation": 26420 + } + }, + { + "ph": "s", "id": 26420, "pid": 435, "tid": 435, "ts": 1776649790556632, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556659, "dur": 0, + "args": { + "External id": 2990, + "cbid": 210, "correlation": 26447 + } + }, + { + "ph": "f", "id": 26447, "pid": 435, "tid": 435, "ts": 1776649790556659, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556671, "dur": 23, + "args": { + "External id": 2990, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26448, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26448, "pid": 0, "tid": 7, "ts": 1776649790556671, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556661, "dur": 6, + "args": { + "External id": 2990, + "cbid": 211, "correlation": 26448 + } + }, + { + "ph": "s", "id": 26448, "pid": 435, "tid": 435, "ts": 1776649790556661, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556688, "dur": 0, + "args": { + "External id": 2992, + "cbid": 210, "correlation": 26475 + } + }, + { + "ph": "f", "id": 26475, "pid": 435, "tid": 435, "ts": 1776649790556688, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556700, "dur": 23, + "args": { + "External id": 2992, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26476, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26476, "pid": 0, "tid": 7, "ts": 1776649790556700, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556690, "dur": 5, + "args": { + "External id": 2992, + "cbid": 211, "correlation": 26476 + } + }, + { + "ph": "s", "id": 26476, "pid": 435, "tid": 435, "ts": 1776649790556690, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556717, "dur": 0, + "args": { + "External id": 2994, + "cbid": 210, "correlation": 26503 + } + }, + { + "ph": "f", "id": 26503, "pid": 435, "tid": 435, "ts": 1776649790556717, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556729, "dur": 23, + "args": { + "External id": 2994, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26504, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26504, "pid": 0, "tid": 7, "ts": 1776649790556729, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556719, "dur": 5, + "args": { + "External id": 2994, + "cbid": 211, "correlation": 26504 + } + }, + { + "ph": "s", "id": 26504, "pid": 435, "tid": 435, "ts": 1776649790556719, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556746, "dur": 0, + "args": { + "External id": 2996, + "cbid": 210, "correlation": 26531 + } + }, + { + "ph": "f", "id": 26531, "pid": 435, "tid": 435, "ts": 1776649790556746, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556758, "dur": 22, + "args": { + "External id": 2996, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26532, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26532, "pid": 0, "tid": 7, "ts": 1776649790556758, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556748, "dur": 5, + "args": { + "External id": 2996, + "cbid": 211, "correlation": 26532 + } + }, + { + "ph": "s", "id": 26532, "pid": 435, "tid": 435, "ts": 1776649790556748, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556775, "dur": 0, + "args": { + "External id": 2998, + "cbid": 210, "correlation": 26559 + } + }, + { + "ph": "f", "id": 26559, "pid": 435, "tid": 435, "ts": 1776649790556775, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556787, "dur": 22, + "args": { + "External id": 2998, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26560, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26560, "pid": 0, "tid": 7, "ts": 1776649790556787, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556776, "dur": 5, + "args": { + "External id": 2998, + "cbid": 211, "correlation": 26560 + } + }, + { + "ph": "s", "id": 26560, "pid": 435, "tid": 435, "ts": 1776649790556776, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556804, "dur": 0, + "args": { + "External id": 3000, + "cbid": 210, "correlation": 26587 + } + }, + { + "ph": "f", "id": 26587, "pid": 435, "tid": 435, "ts": 1776649790556804, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556816, "dur": 22, + "args": { + "External id": 3000, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26588, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26588, "pid": 0, "tid": 7, "ts": 1776649790556816, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556806, "dur": 5, + "args": { + "External id": 3000, + "cbid": 211, "correlation": 26588 + } + }, + { + "ph": "s", "id": 26588, "pid": 435, "tid": 435, "ts": 1776649790556806, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556833, "dur": 0, + "args": { + "External id": 3002, + "cbid": 210, "correlation": 26615 + } + }, + { + "ph": "f", "id": 26615, "pid": 435, "tid": 435, "ts": 1776649790556833, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556845, "dur": 23, + "args": { + "External id": 3002, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26616, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26616, "pid": 0, "tid": 7, "ts": 1776649790556845, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556835, "dur": 6, + "args": { + "External id": 3002, + "cbid": 211, "correlation": 26616 + } + }, + { + "ph": "s", "id": 26616, "pid": 435, "tid": 435, "ts": 1776649790556835, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556863, "dur": 0, + "args": { + "External id": 3004, + "cbid": 210, "correlation": 26643 + } + }, + { + "ph": "f", "id": 26643, "pid": 435, "tid": 435, "ts": 1776649790556863, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556874, "dur": 23, + "args": { + "External id": 3004, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26644, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26644, "pid": 0, "tid": 7, "ts": 1776649790556874, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556864, "dur": 5, + "args": { + "External id": 3004, + "cbid": 211, "correlation": 26644 + } + }, + { + "ph": "s", "id": 26644, "pid": 435, "tid": 435, "ts": 1776649790556864, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556892, "dur": 0, + "args": { + "External id": 3006, + "cbid": 210, "correlation": 26671 + } + }, + { + "ph": "f", "id": 26671, "pid": 435, "tid": 435, "ts": 1776649790556892, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556903, "dur": 23, + "args": { + "External id": 3006, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26672, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26672, "pid": 0, "tid": 7, "ts": 1776649790556903, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556893, "dur": 5, + "args": { + "External id": 3006, + "cbid": 211, "correlation": 26672 + } + }, + { + "ph": "s", "id": 26672, "pid": 435, "tid": 435, "ts": 1776649790556893, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556921, "dur": 0, + "args": { + "External id": 3008, + "cbid": 210, "correlation": 26699 + } + }, + { + "ph": "f", "id": 26699, "pid": 435, "tid": 435, "ts": 1776649790556921, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556933, "dur": 22, + "args": { + "External id": 3008, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26700, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26700, "pid": 0, "tid": 7, "ts": 1776649790556933, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556922, "dur": 6, + "args": { + "External id": 3008, + "cbid": 211, "correlation": 26700 + } + }, + { + "ph": "s", "id": 26700, "pid": 435, "tid": 435, "ts": 1776649790556922, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556950, "dur": 0, + "args": { + "External id": 3010, + "cbid": 210, "correlation": 26727 + } + }, + { + "ph": "f", "id": 26727, "pid": 435, "tid": 435, "ts": 1776649790556950, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556961, "dur": 23, + "args": { + "External id": 3010, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26728, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26728, "pid": 0, "tid": 7, "ts": 1776649790556961, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556951, "dur": 6, + "args": { + "External id": 3010, + "cbid": 211, "correlation": 26728 + } + }, + { + "ph": "s", "id": 26728, "pid": 435, "tid": 435, "ts": 1776649790556951, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790556980, "dur": 0, + "args": { + "External id": 3012, + "cbid": 210, "correlation": 26755 + } + }, + { + "ph": "f", "id": 26755, "pid": 435, "tid": 435, "ts": 1776649790556980, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790556991, "dur": 22, + "args": { + "External id": 3012, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26756, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26756, "pid": 0, "tid": 7, "ts": 1776649790556991, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790556981, "dur": 5, + "args": { + "External id": 3012, + "cbid": 211, "correlation": 26756 + } + }, + { + "ph": "s", "id": 26756, "pid": 435, "tid": 435, "ts": 1776649790556981, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790557009, "dur": 0, + "args": { + "External id": 3014, + "cbid": 210, "correlation": 26783 + } + }, + { + "ph": "f", "id": 26783, "pid": 435, "tid": 435, "ts": 1776649790557009, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790557020, "dur": 22, + "args": { + "External id": 3014, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26784, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26784, "pid": 0, "tid": 7, "ts": 1776649790557020, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557010, "dur": 5, + "args": { + "External id": 3014, + "cbid": 211, "correlation": 26784 + } + }, + { + "ph": "s", "id": 26784, "pid": 435, "tid": 435, "ts": 1776649790557010, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790557038, "dur": 0, + "args": { + "External id": 3016, + "cbid": 210, "correlation": 26811 + } + }, + { + "ph": "f", "id": 26811, "pid": 435, "tid": 435, "ts": 1776649790557038, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790557049, "dur": 23, + "args": { + "External id": 3016, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26812, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26812, "pid": 0, "tid": 7, "ts": 1776649790557049, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557039, "dur": 5, + "args": { + "External id": 3016, + "cbid": 211, "correlation": 26812 + } + }, + { + "ph": "s", "id": 26812, "pid": 435, "tid": 435, "ts": 1776649790557039, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790557067, "dur": 0, + "args": { + "External id": 3018, + "cbid": 210, "correlation": 26839 + } + }, + { + "ph": "f", "id": 26839, "pid": 435, "tid": 435, "ts": 1776649790557067, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790557078, "dur": 23, + "args": { + "External id": 3018, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26840, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26840, "pid": 0, "tid": 7, "ts": 1776649790557078, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557068, "dur": 6, + "args": { + "External id": 3018, + "cbid": 211, "correlation": 26840 + } + }, + { + "ph": "s", "id": 26840, "pid": 435, "tid": 435, "ts": 1776649790557068, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790557096, "dur": 0, + "args": { + "External id": 3020, + "cbid": 210, "correlation": 26867 + } + }, + { + "ph": "f", "id": 26867, "pid": 435, "tid": 435, "ts": 1776649790557096, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790557108, "dur": 22, + "args": { + "External id": 3020, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26868, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26868, "pid": 0, "tid": 7, "ts": 1776649790557108, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557097, "dur": 6, + "args": { + "External id": 3020, + "cbid": 211, "correlation": 26868 + } + }, + { + "ph": "s", "id": 26868, "pid": 435, "tid": 435, "ts": 1776649790557097, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790557125, "dur": 0, + "args": { + "External id": 3022, + "cbid": 210, "correlation": 26895 + } + }, + { + "ph": "f", "id": 26895, "pid": 435, "tid": 435, "ts": 1776649790557125, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790557137, "dur": 23, + "args": { + "External id": 3022, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26896, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26896, "pid": 0, "tid": 7, "ts": 1776649790557137, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557126, "dur": 6, + "args": { + "External id": 3022, + "cbid": 211, "correlation": 26896 + } + }, + { + "ph": "s", "id": 26896, "pid": 435, "tid": 435, "ts": 1776649790557126, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790557154, "dur": 0, + "args": { + "External id": 3024, + "cbid": 210, "correlation": 26923 + } + }, + { + "ph": "f", "id": 26923, "pid": 435, "tid": 435, "ts": 1776649790557154, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790557166, "dur": 23, + "args": { + "External id": 3024, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26924, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26924, "pid": 0, "tid": 7, "ts": 1776649790557166, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557156, "dur": 6, + "args": { + "External id": 3024, + "cbid": 211, "correlation": 26924 + } + }, + { + "ph": "s", "id": 26924, "pid": 435, "tid": 435, "ts": 1776649790557156, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790557183, "dur": 0, + "args": { + "External id": 3026, + "cbid": 210, "correlation": 26951 + } + }, + { + "ph": "f", "id": 26951, "pid": 435, "tid": 435, "ts": 1776649790557183, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790557195, "dur": 23, + "args": { + "External id": 3026, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26952, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26952, "pid": 0, "tid": 7, "ts": 1776649790557195, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557185, "dur": 5, + "args": { + "External id": 3026, + "cbid": 211, "correlation": 26952 + } + }, + { + "ph": "s", "id": 26952, "pid": 435, "tid": 435, "ts": 1776649790557185, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790557212, "dur": 0, + "args": { + "External id": 3028, + "cbid": 210, "correlation": 26979 + } + }, + { + "ph": "f", "id": 26979, "pid": 435, "tid": 435, "ts": 1776649790557212, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790557224, "dur": 23, + "args": { + "External id": 3028, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 26980, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 26980, "pid": 0, "tid": 7, "ts": 1776649790557224, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557214, "dur": 5, + "args": { + "External id": 3028, + "cbid": 211, "correlation": 26980 + } + }, + { + "ph": "s", "id": 26980, "pid": 435, "tid": 435, "ts": 1776649790557214, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790557242, "dur": 0, + "args": { + "External id": 3030, + "cbid": 210, "correlation": 27007 + } + }, + { + "ph": "f", "id": 27007, "pid": 435, "tid": 435, "ts": 1776649790557242, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790557253, "dur": 23, + "args": { + "External id": 3030, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27008, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 27008, "pid": 0, "tid": 7, "ts": 1776649790557253, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557243, "dur": 6, + "args": { + "External id": 3030, + "cbid": 211, "correlation": 27008 + } + }, + { + "ph": "s", "id": 27008, "pid": 435, "tid": 435, "ts": 1776649790557243, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790557271, "dur": 0, + "args": { + "External id": 3032, + "cbid": 210, "correlation": 27035 + } + }, + { + "ph": "f", "id": 27035, "pid": 435, "tid": 435, "ts": 1776649790557271, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790557282, "dur": 23, + "args": { + "External id": 3032, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27036, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 27036, "pid": 0, "tid": 7, "ts": 1776649790557282, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557272, "dur": 5, + "args": { + "External id": 3032, + "cbid": 211, "correlation": 27036 + } + }, + { + "ph": "s", "id": 27036, "pid": 435, "tid": 435, "ts": 1776649790557272, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790557300, "dur": 0, + "args": { + "External id": 3034, + "cbid": 210, "correlation": 27063 + } + }, + { + "ph": "f", "id": 27063, "pid": 435, "tid": 435, "ts": 1776649790557300, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790557311, "dur": 23, + "args": { + "External id": 3034, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27064, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 27064, "pid": 0, "tid": 7, "ts": 1776649790557311, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557301, "dur": 6, + "args": { + "External id": 3034, + "cbid": 211, "correlation": 27064 + } + }, + { + "ph": "s", "id": 27064, "pid": 435, "tid": 435, "ts": 1776649790557301, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790557329, "dur": 0, + "args": { + "External id": 3036, + "cbid": 210, "correlation": 27091 + } + }, + { + "ph": "f", "id": 27091, "pid": 435, "tid": 435, "ts": 1776649790557329, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790557341, "dur": 23, + "args": { + "External id": 3036, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27092, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 27092, "pid": 0, "tid": 7, "ts": 1776649790557341, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557330, "dur": 5, + "args": { + "External id": 3036, + "cbid": 211, "correlation": 27092 + } + }, + { + "ph": "s", "id": 27092, "pid": 435, "tid": 435, "ts": 1776649790557330, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790557358, "dur": 0, + "args": { + "External id": 3038, + "cbid": 210, "correlation": 27119 + } + }, + { + "ph": "f", "id": 27119, "pid": 435, "tid": 435, "ts": 1776649790557358, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790557370, "dur": 23, + "args": { + "External id": 3038, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27120, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 27120, "pid": 0, "tid": 7, "ts": 1776649790557370, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557360, "dur": 6, + "args": { + "External id": 3038, + "cbid": 211, "correlation": 27120 + } + }, + { + "ph": "s", "id": 27120, "pid": 435, "tid": 435, "ts": 1776649790557360, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790557388, "dur": 0, + "args": { + "External id": 3040, + "cbid": 210, "correlation": 27147 + } + }, + { + "ph": "f", "id": 27147, "pid": 435, "tid": 435, "ts": 1776649790557388, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790557399, "dur": 22, + "args": { + "External id": 3040, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27148, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 27148, "pid": 0, "tid": 7, "ts": 1776649790557399, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557389, "dur": 5, + "args": { + "External id": 3040, + "cbid": 211, "correlation": 27148 + } + }, + { + "ph": "s", "id": 27148, "pid": 435, "tid": 435, "ts": 1776649790557389, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790557417, "dur": 0, + "args": { + "External id": 3042, + "cbid": 210, "correlation": 27175 + } + }, + { + "ph": "f", "id": 27175, "pid": 435, "tid": 435, "ts": 1776649790557417, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790557428, "dur": 22, + "args": { + "External id": 3042, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27176, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 27176, "pid": 0, "tid": 7, "ts": 1776649790557428, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557418, "dur": 6, + "args": { + "External id": 3042, + "cbid": 211, "correlation": 27176 + } + }, + { + "ph": "s", "id": 27176, "pid": 435, "tid": 435, "ts": 1776649790557418, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790557447, "dur": 0, + "args": { + "External id": 3044, + "cbid": 210, "correlation": 27203 + } + }, + { + "ph": "f", "id": 27203, "pid": 435, "tid": 435, "ts": 1776649790557447, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790557458, "dur": 22, + "args": { + "External id": 3044, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27204, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 27204, "pid": 0, "tid": 7, "ts": 1776649790557458, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557448, "dur": 5, + "args": { + "External id": 3044, + "cbid": 211, "correlation": 27204 + } + }, + { + "ph": "s", "id": 27204, "pid": 435, "tid": 435, "ts": 1776649790557448, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790557477, "dur": 0, + "args": { + "External id": 3046, + "cbid": 210, "correlation": 27231 + } + }, + { + "ph": "f", "id": 27231, "pid": 435, "tid": 435, "ts": 1776649790557477, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790557489, "dur": 23, + "args": { + "External id": 3046, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27232, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 27232, "pid": 0, "tid": 7, "ts": 1776649790557489, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557478, "dur": 6, + "args": { + "External id": 3046, + "cbid": 211, "correlation": 27232 + } + }, + { + "ph": "s", "id": 27232, "pid": 435, "tid": 435, "ts": 1776649790557478, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790557507, "dur": 0, + "args": { + "External id": 3048, + "cbid": 210, "correlation": 27259 + } + }, + { + "ph": "f", "id": 27259, "pid": 435, "tid": 435, "ts": 1776649790557507, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790557518, "dur": 23, + "args": { + "External id": 3048, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27260, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 27260, "pid": 0, "tid": 7, "ts": 1776649790557518, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557508, "dur": 5, + "args": { + "External id": 3048, + "cbid": 211, "correlation": 27260 + } + }, + { + "ph": "s", "id": 27260, "pid": 435, "tid": 435, "ts": 1776649790557508, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790557536, "dur": 0, + "args": { + "External id": 3050, + "cbid": 210, "correlation": 27287 + } + }, + { + "ph": "f", "id": 27287, "pid": 435, "tid": 435, "ts": 1776649790557536, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790557547, "dur": 23, + "args": { + "External id": 3050, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27288, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 27288, "pid": 0, "tid": 7, "ts": 1776649790557547, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557537, "dur": 5, + "args": { + "External id": 3050, + "cbid": 211, "correlation": 27288 + } + }, + { + "ph": "s", "id": 27288, "pid": 435, "tid": 435, "ts": 1776649790557537, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790557565, "dur": 0, + "args": { + "External id": 3052, + "cbid": 210, "correlation": 27315 + } + }, + { + "ph": "f", "id": 27315, "pid": 435, "tid": 435, "ts": 1776649790557565, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790557577, "dur": 22, + "args": { + "External id": 3052, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27316, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 27316, "pid": 0, "tid": 7, "ts": 1776649790557577, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557566, "dur": 5, + "args": { + "External id": 3052, + "cbid": 211, "correlation": 27316 + } + }, + { + "ph": "s", "id": 27316, "pid": 435, "tid": 435, "ts": 1776649790557566, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790557594, "dur": 0, + "args": { + "External id": 3054, + "cbid": 210, "correlation": 27343 + } + }, + { + "ph": "f", "id": 27343, "pid": 435, "tid": 435, "ts": 1776649790557594, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790557606, "dur": 23, + "args": { + "External id": 3054, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27344, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 27344, "pid": 0, "tid": 7, "ts": 1776649790557606, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557595, "dur": 5, + "args": { + "External id": 3054, + "cbid": 211, "correlation": 27344 + } + }, + { + "ph": "s", "id": 27344, "pid": 435, "tid": 435, "ts": 1776649790557595, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790557623, "dur": 0, + "args": { + "External id": 3056, + "cbid": 210, "correlation": 27371 + } + }, + { + "ph": "f", "id": 27371, "pid": 435, "tid": 435, "ts": 1776649790557623, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790557635, "dur": 23, + "args": { + "External id": 3056, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27372, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 27372, "pid": 0, "tid": 7, "ts": 1776649790557635, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557625, "dur": 5, + "args": { + "External id": 3056, + "cbid": 211, "correlation": 27372 + } + }, + { + "ph": "s", "id": 27372, "pid": 435, "tid": 435, "ts": 1776649790557625, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790557658, "dur": 1, + "args": { + "External id": 3058, + "cbid": 210, "correlation": 27399 + } + }, + { + "ph": "f", "id": 27399, "pid": 435, "tid": 435, "ts": 1776649790557658, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790557669, "dur": 23, + "args": { + "External id": 3058, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27400, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 27400, "pid": 0, "tid": 7, "ts": 1776649790557669, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557660, "dur": 6, + "args": { + "External id": 3058, + "cbid": 211, "correlation": 27400 + } + }, + { + "ph": "s", "id": 27400, "pid": 435, "tid": 435, "ts": 1776649790557660, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790557690, "dur": 0, + "args": { + "External id": 3060, + "cbid": 210, "correlation": 27427 + } + }, + { + "ph": "f", "id": 27427, "pid": 435, "tid": 435, "ts": 1776649790557690, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790557700, "dur": 22, + "args": { + "External id": 3060, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27428, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 27428, "pid": 0, "tid": 7, "ts": 1776649790557700, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557691, "dur": 5, + "args": { + "External id": 3060, + "cbid": 211, "correlation": 27428 + } + }, + { + "ph": "s", "id": 27428, "pid": 435, "tid": 435, "ts": 1776649790557691, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790557719, "dur": 0, + "args": { + "External id": 3062, + "cbid": 210, "correlation": 27455 + } + }, + { + "ph": "f", "id": 27455, "pid": 435, "tid": 435, "ts": 1776649790557719, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790557729, "dur": 23, + "args": { + "External id": 3062, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27456, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 27456, "pid": 0, "tid": 7, "ts": 1776649790557729, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557720, "dur": 6, + "args": { + "External id": 3062, + "cbid": 211, "correlation": 27456 + } + }, + { + "ph": "s", "id": 27456, "pid": 435, "tid": 435, "ts": 1776649790557720, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790557749, "dur": 0, + "args": { + "External id": 3064, + "cbid": 210, "correlation": 27483 + } + }, + { + "ph": "f", "id": 27483, "pid": 435, "tid": 435, "ts": 1776649790557749, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790557759, "dur": 23, + "args": { + "External id": 3064, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27484, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 27484, "pid": 0, "tid": 7, "ts": 1776649790557759, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557750, "dur": 5, + "args": { + "External id": 3064, + "cbid": 211, "correlation": 27484 + } + }, + { + "ph": "s", "id": 27484, "pid": 435, "tid": 435, "ts": 1776649790557750, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790557783, "dur": 3, + "args": { + "External id": 27495, + "cbid": 135, "correlation": 27495 + } + }, + { + "ph": "f", "id": 27495, "pid": 435, "tid": 435, "ts": 1776649790557783, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790557798, "dur": 5, + "args": { + "External id": 27500, + "cbid": 165, "correlation": 27500 + } + }, + { + "ph": "s", "id": 27500, "pid": 435, "tid": 435, "ts": 1776649790557798, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649790557806, "dur": 4, + "args": { + "External id": 27503, + "cbid": 139, "correlation": 27503 + } + }, + { + "ph": "f", "id": 27503, "pid": 435, "tid": 435, "ts": 1776649790557806, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor >, at::detail::Array >(int, at::native::AUnaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790557895, "dur": 18, + "args": { + "External id": 3067, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27520, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 27520, "pid": 0, "tid": 7, "ts": 1776649790557895, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557883, "dur": 10, + "args": { + "External id": 3067, + "cbid": 211, "correlation": 27520 + } + }, + { + "ph": "s", "id": 27520, "pid": 435, "tid": 435, "ts": 1776649790557883, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649790557934, "dur": 30, + "args": { + "External id": 3071, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27533, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 27533, "pid": 0, "tid": 7, "ts": 1776649790557934, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557924, "dur": 7, + "args": { + "External id": 3071, + "cbid": 211, "correlation": 27533 + } + }, + { + "ph": "s", "id": 27533, "pid": 435, "tid": 435, "ts": 1776649790557924, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BUnaryFunctor, at::detail::Array >(int, at::native::BUnaryFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790557977, "dur": 17, + "args": { + "External id": 3072, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27547, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 27547, "pid": 0, "tid": 7, "ts": 1776649790557977, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557967, "dur": 7, + "args": { + "External id": 3072, + "cbid": 211, "correlation": 27547 + } + }, + { + "ph": "s", "id": 27547, "pid": 435, "tid": 435, "ts": 1776649790557967, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649790558003, "dur": 29, + "args": { + "External id": 3077, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27560, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 27560, "pid": 0, "tid": 7, "ts": 1776649790558003, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790557994, "dur": 6, + "args": { + "External id": 3077, + "cbid": 211, "correlation": 27560 + } + }, + { + "ph": "s", "id": 27560, "pid": 435, "tid": 435, "ts": 1776649790557994, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy_contig, unsigned int, 3, 128, 1>(at::native::(anonymous namespace)::OpaqueType<1u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata, unsigned int, 128, 1>, at::native::(anonymous namespace)::TensorSizeStride, int, unsigned int)", "pid": 0, "tid": 7, + "ts": 1776649790558053, "dur": 64, + "args": { + "External id": 3083, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27572, + "registers per thread": 21, + "shared memory": 0, + "blocks per SM": 4.000000, + "warps per SM": 64.000000, + "grid": [216, 2, 1], + "block": [512, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 27572, "pid": 0, "tid": 7, "ts": 1776649790558053, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790558043, "dur": 8, + "args": { + "External id": 3083, + "cbid": 211, "correlation": 27572 + } + }, + { + "ph": "s", "id": 27572, "pid": 435, "tid": 435, "ts": 1776649790558043, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649790558119, "dur": 71, + "args": { + "External id": 3094, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27585, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 27585, "pid": 0, "tid": 7, "ts": 1776649790558119, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790558106, "dur": 9, + "args": { + "External id": 3094, + "cbid": 211, "correlation": 27585 + } + }, + { + "ph": "s", "id": 27585, "pid": 435, "tid": 435, "ts": 1776649790558106, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, at::detail::Array >(int, at::native::CUDAFunctorOnSelf_add, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790558192, "dur": 33, + "args": { + "External id": 3097, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27596, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 27596, "pid": 0, "tid": 7, "ts": 1776649790558192, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790558143, "dur": 8, + "args": { + "External id": 3097, + "cbid": 211, "correlation": 27596 + } + }, + { + "ph": "s", "id": 27596, "pid": 435, "tid": 435, "ts": 1776649790558143, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649790558226, "dur": 91, + "args": { + "External id": 3101, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27609, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 27609, "pid": 0, "tid": 7, "ts": 1776649790558226, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790558173, "dur": 7, + "args": { + "External id": 3101, + "cbid": 211, "correlation": 27609 + } + }, + { + "ph": "s", "id": 27609, "pid": 435, "tid": 435, "ts": 1776649790558173, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor >, at::detail::Array >(int, at::native::BinaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790558319, "dur": 60, + "args": { + "External id": 3102, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27620, + "registers per thread": 22, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 27620, "pid": 0, "tid": 7, "ts": 1776649790558319, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790558200, "dur": 6, + "args": { + "External id": 3102, + "cbid": 211, "correlation": 27620 + } + }, + { + "ph": "s", "id": 27620, "pid": 435, "tid": 435, "ts": 1776649790558200, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649790558381, "dur": 221, + "args": { + "External id": 3110, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27634, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 27634, "pid": 0, "tid": 7, "ts": 1776649790558381, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790558255, "dur": 8, + "args": { + "External id": 3110, + "cbid": 211, "correlation": 27634 + } + }, + { + "ph": "s", "id": 27634, "pid": 435, "tid": 435, "ts": 1776649790558255, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649790676665, "dur": 3484, + "args": { + "External id": 3126, + "device": 0, "context": 1, + "stream": 7, "correlation": 27650, + "bytes": 33554432, "memory bandwidth (GB/s)": 9.630079303488527 + } + }, + { + "ph": "f", "id": 27650, "pid": 0, "tid": 7, "ts": 1776649790676665, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649790676522, "dur": 3595, + "args": { + "External id": 3126, + "cbid": 41, "correlation": 27650 + } + }, + { + "ph": "s", "id": 27650, "pid": 435, "tid": 435, "ts": 1776649790676522, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790680125, "dur": 35, + "args": { + "External id": 3126, + "cbid": 131, "correlation": 27651 + } + }, + { + "ph": "s", "id": 27651, "pid": 435, "tid": 435, "ts": 1776649790680125, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7, + "ts": 1776649790780522, "dur": 39, + "args": { + "External id": 3129, + "device": 0, "context": 1, + "stream": 7, "correlation": 27656, + "bytes": 33554432, "memory bandwidth (GB/s)": 844.2640901771337 + } + }, + { + "ph": "f", "id": 27656, "pid": 0, "tid": 7, "ts": 1776649790780522, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649790780186, "dur": 336, + "args": { + "External id": 3129, + "cbid": 41, "correlation": 27656 + } + }, + { + "ph": "s", "id": 27656, "pid": 435, "tid": 435, "ts": 1776649790780186, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649790781125, "dur": 2121, + "args": { + "External id": 3138, + "device": 0, "context": 1, + "stream": 7, "correlation": 27671, + "bytes": 8388608, "memory bandwidth (GB/s)": 3.9547583543400893 + } + }, + { + "ph": "f", "id": 27671, "pid": 0, "tid": 7, "ts": 1776649790781125, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649790780799, "dur": 2390, + "args": { + "External id": 3138, + "cbid": 41, "correlation": 27671 + } + }, + { + "ph": "s", "id": 27671, "pid": 435, "tid": 435, "ts": 1776649790780799, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790783191, "dur": 66, + "args": { + "External id": 3138, + "cbid": 131, "correlation": 27672 + } + }, + { + "ph": "s", "id": 27672, "pid": 435, "tid": 435, "ts": 1776649790783191, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649790783318, "dur": 17, + "args": { + "External id": 3142, + "device": 0, "context": 1, + "stream": 7, "correlation": 27684, + "bytes": 262144, "memory bandwidth (GB/s)": 14.895391783624069 + } + }, + { + "ph": "f", "id": 27684, "pid": 0, "tid": 7, "ts": 1776649790783318, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649790783288, "dur": 27, + "args": { + "External id": 3142, + "cbid": 41, "correlation": 27684 + } + }, + { + "ph": "s", "id": 27684, "pid": 435, "tid": 435, "ts": 1776649790783288, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790783316, "dur": 26, + "args": { + "External id": 3142, + "cbid": 131, "correlation": 27685 + } + }, + { + "ph": "s", "id": 27685, "pid": 435, "tid": 435, "ts": 1776649790783316, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649790783375, "dur": 1, + "args": { + "External id": 3146, + "device": 0, "context": 1, + "stream": 7, "correlation": 27697, + "bytes": 2048, "memory bandwidth (GB/s)": 1.4883720930232558 + } + }, + { + "ph": "f", "id": 27697, "pid": 0, "tid": 7, "ts": 1776649790783375, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649790783364, "dur": 6, + "args": { + "External id": 3146, + "cbid": 41, "correlation": 27697 + } + }, + { + "ph": "s", "id": 27697, "pid": 435, "tid": 435, "ts": 1776649790783364, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790783371, "dur": 10, + "args": { + "External id": 3146, + "cbid": 131, "correlation": 27698 + } + }, + { + "ph": "s", "id": 27698, "pid": 435, "tid": 435, "ts": 1776649790783371, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649790783541, "dur": 229, + "args": { + "External id": 3161, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27710, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 27710, "pid": 0, "tid": 7, "ts": 1776649790783541, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790783520, "dur": 22, + "args": { + "External id": 3161, + "cbid": 211, "correlation": 27710 + } + }, + { + "ph": "s", "id": 27710, "pid": 435, "tid": 435, "ts": 1776649790783520, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649790783772, "dur": 5, + "args": { + "External id": 3167, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27724, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 2.370370, + "warps per SM": 9.481482, + "grid": [256, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 15 + } + }, + { + "ph": "f", "id": 27724, "pid": 0, "tid": 7, "ts": 1776649790783772, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790783581, "dur": 8, + "args": { + "External id": 3167, + "cbid": 211, "correlation": 27724 + } + }, + { + "ph": "s", "id": 27724, "pid": 435, "tid": 435, "ts": 1776649790783581, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649790783779, "dur": 126, + "args": { + "External id": 3169, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27735, + "registers per thread": 20, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 27735, "pid": 0, "tid": 7, "ts": 1776649790783779, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790783628, "dur": 8, + "args": { + "External id": 3169, + "cbid": 211, "correlation": 27735 + } + }, + { + "ph": "s", "id": 27735, "pid": 435, "tid": 435, "ts": 1776649790783628, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790783907, "dur": 39, + "args": { + "External id": 3170, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27745, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 27745, "pid": 0, "tid": 7, "ts": 1776649790783907, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790783661, "dur": 7, + "args": { + "External id": 3170, + "cbid": 211, "correlation": 27745 + } + }, + { + "ph": "s", "id": 27745, "pid": 435, "tid": 435, "ts": 1776649790783661, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649790783948, "dur": 99, + "args": { + "External id": 3174, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27758, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 27758, "pid": 0, "tid": 7, "ts": 1776649790783948, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790783695, "dur": 9, + "args": { + "External id": 3174, + "cbid": 211, "correlation": 27758 + } + }, + { + "ph": "s", "id": 27758, "pid": 435, "tid": 435, "ts": 1776649790783695, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, at::detail::Array >(int, at::native::CUDAFunctorOnSelf_add, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790784048, "dur": 75, + "args": { + "External id": 3175, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27765, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 27765, "pid": 0, "tid": 7, "ts": 1776649790784048, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790783740, "dur": 9, + "args": { + "External id": 3175, + "cbid": 211, "correlation": 27765 + } + }, + { + "ph": "s", "id": 27765, "pid": 435, "tid": 435, "ts": 1776649790783740, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1}, at::detail::Array >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790784125, "dur": 78, + "args": { + "External id": 3176, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27775, + "registers per thread": 22, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 27775, "pid": 0, "tid": 7, "ts": 1776649790784125, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790783783, "dur": 8, + "args": { + "External id": 3176, + "cbid": 211, "correlation": 27775 + } + }, + { + "ph": "s", "id": 27775, "pid": 435, "tid": 435, "ts": 1776649790783783, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649790784205, "dur": 102, + "args": { + "External id": 3186, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27788, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 27788, "pid": 0, "tid": 7, "ts": 1776649790784205, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790783841, "dur": 8, + "args": { + "External id": 3186, + "cbid": 211, "correlation": 27788 + } + }, + { + "ph": "s", "id": 27788, "pid": 435, "tid": 435, "ts": 1776649790783841, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649790784310, "dur": 1, + "args": { + "External id": 3196, + "device": 0, "context": 1, + "stream": 7, "correlation": 27803, + "bytes": 512, "memory bandwidth (GB/s)": 0.3902439024390244 + } + }, + { + "ph": "f", "id": 27803, "pid": 0, "tid": 7, "ts": 1776649790784310, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649790783932, "dur": 10, + "args": { + "External id": 3196, + "cbid": 41, "correlation": 27803 + } + }, + { + "ph": "s", "id": 27803, "pid": 435, "tid": 435, "ts": 1776649790783932, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790783943, "dur": 375, + "args": { + "External id": 3196, + "cbid": 131, "correlation": 27804 + } + }, + { + "ph": "s", "id": 27804, "pid": 435, "tid": 435, "ts": 1776649790783943, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649790784381, "dur": 9, + "args": { + "External id": 3198, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27813, + "registers per thread": 40, + "shared memory": 0, + "blocks per SM": 2.370370, + "warps per SM": 9.481482, + "grid": [256, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 15 + } + }, + { + "ph": "f", "id": 27813, "pid": 0, "tid": 7, "ts": 1776649790784381, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790784367, "dur": 13, + "args": { + "External id": 3198, + "cbid": 211, "correlation": 27813 + } + }, + { + "ph": "s", "id": 27813, "pid": 435, "tid": 435, "ts": 1776649790784367, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649790784448, "dur": 111, + "args": { + "External id": 3212, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27828, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 27828, "pid": 0, "tid": 7, "ts": 1776649790784448, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790784437, "dur": 8, + "args": { + "External id": 3212, + "cbid": 211, "correlation": 27828 + } + }, + { + "ph": "s", "id": 27828, "pid": 435, "tid": 435, "ts": 1776649790784437, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649790784562, "dur": 1, + "args": { + "External id": 3222, + "device": 0, "context": 1, + "stream": 7, "correlation": 27843, + "bytes": 8192, "memory bandwidth (GB/s)": 4.491228070175438 + } + }, + { + "ph": "f", "id": 27843, "pid": 0, "tid": 7, "ts": 1776649790784562, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649790784497, "dur": 10, + "args": { + "External id": 3222, + "cbid": 41, "correlation": 27843 + } + }, + { + "ph": "s", "id": 27843, "pid": 435, "tid": 435, "ts": 1776649790784497, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790784508, "dur": 61, + "args": { + "External id": 3222, + "cbid": 131, "correlation": 27844 + } + }, + { + "ph": "s", "id": 27844, "pid": 435, "tid": 435, "ts": 1776649790784508, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649790784600, "dur": 127, + "args": { + "External id": 3218, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27853, + "registers per thread": 40, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 75 + } + }, + { + "ph": "f", "id": 27853, "pid": 0, "tid": 7, "ts": 1776649790784600, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790784588, "dur": 10, + "args": { + "External id": 3218, + "cbid": 211, "correlation": 27853 + } + }, + { + "ph": "s", "id": 27853, "pid": 435, "tid": 435, "ts": 1776649790784588, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoH (Device -> Pageable)", "pid": 0, "tid": 7, + "ts": 1776649790798644, "dur": 43955, + "args": { + "External id": 3231, + "device": 0, "context": 1, + "stream": 7, "correlation": 27861, + "bytes": 67108864, "memory bandwidth (GB/s)": 1.5267504122364144 + } + }, + { + "ph": "f", "id": 27861, "pid": 0, "tid": 7, "ts": 1776649790798644, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649790798624, "dur": 45216, + "args": { + "External id": 3231, + "cbid": 41, "correlation": 27861 + } + }, + { + "ph": "s", "id": 27861, "pid": 435, "tid": 435, "ts": 1776649790798624, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790843845, "dur": 12, + "args": { + "External id": 3231, + "cbid": 131, "correlation": 27862 + } + }, + { + "ph": "s", "id": 27862, "pid": 435, "tid": 435, "ts": 1776649790843845, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649790904291, "dur": 498, + "args": { + "External id": 3241, + "device": 0, "context": 1, + "stream": 7, "correlation": 27874, + "bytes": 8388608, "memory bandwidth (GB/s)": 16.811713636381857 + } + }, + { + "ph": "f", "id": 27874, "pid": 0, "tid": 7, "ts": 1776649790904291, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649790904117, "dur": 617, + "args": { + "External id": 3241, + "cbid": 41, "correlation": 27874 + } + }, + { + "ph": "s", "id": 27874, "pid": 435, "tid": 435, "ts": 1776649790904117, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790904736, "dur": 64, + "args": { + "External id": 3241, + "cbid": 131, "correlation": 27875 + } + }, + { + "ph": "s", "id": 27875, "pid": 435, "tid": 435, "ts": 1776649790904736, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7, + "ts": 1776649790906150, "dur": 9, + "args": { + "External id": 3247, + "device": 0, "context": 1, + "stream": 7, "correlation": 27880, + "bytes": 8388608, "memory bandwidth (GB/s)": 873.8133333333334 + } + }, + { + "ph": "f", "id": 27880, "pid": 0, "tid": 7, "ts": 1776649790906150, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649790906106, "dur": 42, + "args": { + "External id": 3247, + "cbid": 41, "correlation": 27880 + } + }, + { + "ph": "s", "id": 27880, "pid": 435, "tid": 435, "ts": 1776649790906106, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7, + "ts": 1776649790906202, "dur": 3, + "args": { + "External id": 3253, + "device": 0, "context": 1, + "stream": 7, "correlation": 27888, + "bytes": 262144, "memory bandwidth (GB/s)": 77.28301886792453 + } + }, + { + "ph": "f", "id": 27888, "pid": 0, "tid": 7, "ts": 1776649790906202, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649790906185, "dur": 13, + "args": { + "External id": 3253, + "cbid": 41, "correlation": 27888 + } + }, + { + "ph": "s", "id": 27888, "pid": 435, "tid": 435, "ts": 1776649790906185, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790906942, "dur": 1, + "args": { + "External id": 3258, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27911, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 27911, "pid": 0, "tid": 7, "ts": 1776649790906942, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790906919, "dur": 23, + "args": { + "External id": 3258, + "cbid": 211, "correlation": 27911 + } + }, + { + "ph": "s", "id": 27911, "pid": 435, "tid": 435, "ts": 1776649790906919, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790906960, "dur": 2, + "args": { + "External id": 27917, + "cbid": 200, "correlation": 27917 + } + }, + { + "ph": "f", "id": 27917, "pid": 435, "tid": 435, "ts": 1776649790906960, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790906964, "dur": 3, + "args": { + "External id": 27918, + "cbid": 273, "correlation": 27918 + } + }, + { + "ph": "f", "id": 27918, "pid": 435, "tid": 435, "ts": 1776649790906964, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790906978, "dur": 16, + "args": { + "External id": 27919, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27919, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 27919, "pid": 0, "tid": 7, "ts": 1776649790906978, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790906969, "dur": 6, + "args": { + "External id": 27919, + "cbid": 211, "correlation": 27919 + } + }, + { + "ph": "s", "id": 27919, "pid": 435, "tid": 435, "ts": 1776649790906969, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790907023, "dur": 1, + "args": { + "External id": 3263, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27937, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 27937, "pid": 0, "tid": 7, "ts": 1776649790907023, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907013, "dur": 7, + "args": { + "External id": 3263, + "cbid": 211, "correlation": 27937 + } + }, + { + "ph": "s", "id": 27937, "pid": 435, "tid": 435, "ts": 1776649790907013, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907027, "dur": 0, + "args": { + "External id": 27943, + "cbid": 200, "correlation": 27943 + } + }, + { + "ph": "f", "id": 27943, "pid": 435, "tid": 435, "ts": 1776649790907027, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907027, "dur": 0, + "args": { + "External id": 27944, + "cbid": 273, "correlation": 27944 + } + }, + { + "ph": "f", "id": 27944, "pid": 435, "tid": 435, "ts": 1776649790907027, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790907037, "dur": 15, + "args": { + "External id": 27945, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27945, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 27945, "pid": 0, "tid": 7, "ts": 1776649790907037, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907029, "dur": 5, + "args": { + "External id": 27945, + "cbid": 211, "correlation": 27945 + } + }, + { + "ph": "s", "id": 27945, "pid": 435, "tid": 435, "ts": 1776649790907029, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790907075, "dur": 1, + "args": { + "External id": 3268, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27963, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 27963, "pid": 0, "tid": 7, "ts": 1776649790907075, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907065, "dur": 7, + "args": { + "External id": 3268, + "cbid": 211, "correlation": 27963 + } + }, + { + "ph": "s", "id": 27963, "pid": 435, "tid": 435, "ts": 1776649790907065, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907079, "dur": 0, + "args": { + "External id": 27969, + "cbid": 200, "correlation": 27969 + } + }, + { + "ph": "f", "id": 27969, "pid": 435, "tid": 435, "ts": 1776649790907079, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907080, "dur": 0, + "args": { + "External id": 27970, + "cbid": 273, "correlation": 27970 + } + }, + { + "ph": "f", "id": 27970, "pid": 435, "tid": 435, "ts": 1776649790907080, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790907089, "dur": 15, + "args": { + "External id": 27971, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27971, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 27971, "pid": 0, "tid": 7, "ts": 1776649790907089, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907081, "dur": 5, + "args": { + "External id": 27971, + "cbid": 211, "correlation": 27971 + } + }, + { + "ph": "s", "id": 27971, "pid": 435, "tid": 435, "ts": 1776649790907081, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790907126, "dur": 1, + "args": { + "External id": 3273, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27989, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 27989, "pid": 0, "tid": 7, "ts": 1776649790907126, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907117, "dur": 7, + "args": { + "External id": 3273, + "cbid": 211, "correlation": 27989 + } + }, + { + "ph": "s", "id": 27989, "pid": 435, "tid": 435, "ts": 1776649790907117, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907130, "dur": 0, + "args": { + "External id": 27995, + "cbid": 200, "correlation": 27995 + } + }, + { + "ph": "f", "id": 27995, "pid": 435, "tid": 435, "ts": 1776649790907130, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907130, "dur": 0, + "args": { + "External id": 27996, + "cbid": 273, "correlation": 27996 + } + }, + { + "ph": "f", "id": 27996, "pid": 435, "tid": 435, "ts": 1776649790907130, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790907139, "dur": 17, + "args": { + "External id": 27997, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 27997, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 27997, "pid": 0, "tid": 7, "ts": 1776649790907139, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907131, "dur": 4, + "args": { + "External id": 27997, + "cbid": 211, "correlation": 27997 + } + }, + { + "ph": "s", "id": 27997, "pid": 435, "tid": 435, "ts": 1776649790907131, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790907176, "dur": 1, + "args": { + "External id": 3278, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28015, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28015, "pid": 0, "tid": 7, "ts": 1776649790907176, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907166, "dur": 7, + "args": { + "External id": 3278, + "cbid": 211, "correlation": 28015 + } + }, + { + "ph": "s", "id": 28015, "pid": 435, "tid": 435, "ts": 1776649790907166, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907179, "dur": 0, + "args": { + "External id": 28021, + "cbid": 200, "correlation": 28021 + } + }, + { + "ph": "f", "id": 28021, "pid": 435, "tid": 435, "ts": 1776649790907179, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907180, "dur": 0, + "args": { + "External id": 28022, + "cbid": 273, "correlation": 28022 + } + }, + { + "ph": "f", "id": 28022, "pid": 435, "tid": 435, "ts": 1776649790907180, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790907189, "dur": 16, + "args": { + "External id": 28023, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28023, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28023, "pid": 0, "tid": 7, "ts": 1776649790907189, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907181, "dur": 5, + "args": { + "External id": 28023, + "cbid": 211, "correlation": 28023 + } + }, + { + "ph": "s", "id": 28023, "pid": 435, "tid": 435, "ts": 1776649790907181, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790907225, "dur": 1, + "args": { + "External id": 3283, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28041, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28041, "pid": 0, "tid": 7, "ts": 1776649790907225, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907216, "dur": 6, + "args": { + "External id": 3283, + "cbid": 211, "correlation": 28041 + } + }, + { + "ph": "s", "id": 28041, "pid": 435, "tid": 435, "ts": 1776649790907216, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907229, "dur": 0, + "args": { + "External id": 28047, + "cbid": 200, "correlation": 28047 + } + }, + { + "ph": "f", "id": 28047, "pid": 435, "tid": 435, "ts": 1776649790907229, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907229, "dur": 0, + "args": { + "External id": 28048, + "cbid": 273, "correlation": 28048 + } + }, + { + "ph": "f", "id": 28048, "pid": 435, "tid": 435, "ts": 1776649790907229, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790907238, "dur": 15, + "args": { + "External id": 28049, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28049, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28049, "pid": 0, "tid": 7, "ts": 1776649790907238, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907230, "dur": 4, + "args": { + "External id": 28049, + "cbid": 211, "correlation": 28049 + } + }, + { + "ph": "s", "id": 28049, "pid": 435, "tid": 435, "ts": 1776649790907230, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790907274, "dur": 1, + "args": { + "External id": 3288, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28067, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28067, "pid": 0, "tid": 7, "ts": 1776649790907274, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907265, "dur": 7, + "args": { + "External id": 3288, + "cbid": 211, "correlation": 28067 + } + }, + { + "ph": "s", "id": 28067, "pid": 435, "tid": 435, "ts": 1776649790907265, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907278, "dur": 0, + "args": { + "External id": 28073, + "cbid": 200, "correlation": 28073 + } + }, + { + "ph": "f", "id": 28073, "pid": 435, "tid": 435, "ts": 1776649790907278, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907278, "dur": 0, + "args": { + "External id": 28074, + "cbid": 273, "correlation": 28074 + } + }, + { + "ph": "f", "id": 28074, "pid": 435, "tid": 435, "ts": 1776649790907278, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790907287, "dur": 15, + "args": { + "External id": 28075, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28075, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28075, "pid": 0, "tid": 7, "ts": 1776649790907287, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907279, "dur": 5, + "args": { + "External id": 28075, + "cbid": 211, "correlation": 28075 + } + }, + { + "ph": "s", "id": 28075, "pid": 435, "tid": 435, "ts": 1776649790907279, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790907324, "dur": 1, + "args": { + "External id": 3293, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28093, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28093, "pid": 0, "tid": 7, "ts": 1776649790907324, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907315, "dur": 6, + "args": { + "External id": 3293, + "cbid": 211, "correlation": 28093 + } + }, + { + "ph": "s", "id": 28093, "pid": 435, "tid": 435, "ts": 1776649790907315, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907327, "dur": 0, + "args": { + "External id": 28099, + "cbid": 200, "correlation": 28099 + } + }, + { + "ph": "f", "id": 28099, "pid": 435, "tid": 435, "ts": 1776649790907327, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907328, "dur": 0, + "args": { + "External id": 28100, + "cbid": 273, "correlation": 28100 + } + }, + { + "ph": "f", "id": 28100, "pid": 435, "tid": 435, "ts": 1776649790907328, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790907337, "dur": 15, + "args": { + "External id": 28101, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28101, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28101, "pid": 0, "tid": 7, "ts": 1776649790907337, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907329, "dur": 4, + "args": { + "External id": 28101, + "cbid": 211, "correlation": 28101 + } + }, + { + "ph": "s", "id": 28101, "pid": 435, "tid": 435, "ts": 1776649790907329, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790907372, "dur": 1, + "args": { + "External id": 3298, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28119, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28119, "pid": 0, "tid": 7, "ts": 1776649790907372, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907363, "dur": 6, + "args": { + "External id": 3298, + "cbid": 211, "correlation": 28119 + } + }, + { + "ph": "s", "id": 28119, "pid": 435, "tid": 435, "ts": 1776649790907363, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907375, "dur": 0, + "args": { + "External id": 28125, + "cbid": 200, "correlation": 28125 + } + }, + { + "ph": "f", "id": 28125, "pid": 435, "tid": 435, "ts": 1776649790907375, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907376, "dur": 0, + "args": { + "External id": 28126, + "cbid": 273, "correlation": 28126 + } + }, + { + "ph": "f", "id": 28126, "pid": 435, "tid": 435, "ts": 1776649790907376, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790907385, "dur": 16, + "args": { + "External id": 28127, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28127, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28127, "pid": 0, "tid": 7, "ts": 1776649790907385, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907377, "dur": 5, + "args": { + "External id": 28127, + "cbid": 211, "correlation": 28127 + } + }, + { + "ph": "s", "id": 28127, "pid": 435, "tid": 435, "ts": 1776649790907377, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790907421, "dur": 1, + "args": { + "External id": 3303, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28145, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28145, "pid": 0, "tid": 7, "ts": 1776649790907421, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907412, "dur": 6, + "args": { + "External id": 3303, + "cbid": 211, "correlation": 28145 + } + }, + { + "ph": "s", "id": 28145, "pid": 435, "tid": 435, "ts": 1776649790907412, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907424, "dur": 0, + "args": { + "External id": 28151, + "cbid": 200, "correlation": 28151 + } + }, + { + "ph": "f", "id": 28151, "pid": 435, "tid": 435, "ts": 1776649790907424, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907425, "dur": 0, + "args": { + "External id": 28152, + "cbid": 273, "correlation": 28152 + } + }, + { + "ph": "f", "id": 28152, "pid": 435, "tid": 435, "ts": 1776649790907425, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790907434, "dur": 16, + "args": { + "External id": 28153, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28153, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28153, "pid": 0, "tid": 7, "ts": 1776649790907434, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907426, "dur": 5, + "args": { + "External id": 28153, + "cbid": 211, "correlation": 28153 + } + }, + { + "ph": "s", "id": 28153, "pid": 435, "tid": 435, "ts": 1776649790907426, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790907474, "dur": 14, + "args": { + "External id": 28156, + "cbid": 165, "correlation": 28156 + } + }, + { + "ph": "s", "id": 28156, "pid": 435, "tid": 435, "ts": 1776649790907474, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790907533, "dur": 8, + "args": { + "External id": 28165, + "cbid": 135, "correlation": 28165 + } + }, + { + "ph": "f", "id": 28165, "pid": 435, "tid": 435, "ts": 1776649790907533, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790907589, "dur": 1, + "args": { + "External id": 3308, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28185, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28185, "pid": 0, "tid": 7, "ts": 1776649790907589, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907577, "dur": 9, + "args": { + "External id": 3308, + "cbid": 211, "correlation": 28185 + } + }, + { + "ph": "s", "id": 28185, "pid": 435, "tid": 435, "ts": 1776649790907577, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907593, "dur": 0, + "args": { + "External id": 28191, + "cbid": 200, "correlation": 28191 + } + }, + { + "ph": "f", "id": 28191, "pid": 435, "tid": 435, "ts": 1776649790907593, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907594, "dur": 1, + "args": { + "External id": 28192, + "cbid": 273, "correlation": 28192 + } + }, + { + "ph": "f", "id": 28192, "pid": 435, "tid": 435, "ts": 1776649790907594, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790907604, "dur": 15, + "args": { + "External id": 28193, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28193, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28193, "pid": 0, "tid": 7, "ts": 1776649790907604, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907596, "dur": 5, + "args": { + "External id": 28193, + "cbid": 211, "correlation": 28193 + } + }, + { + "ph": "s", "id": 28193, "pid": 435, "tid": 435, "ts": 1776649790907596, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790907642, "dur": 1, + "args": { + "External id": 3313, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28211, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28211, "pid": 0, "tid": 7, "ts": 1776649790907642, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907633, "dur": 7, + "args": { + "External id": 3313, + "cbid": 211, "correlation": 28211 + } + }, + { + "ph": "s", "id": 28211, "pid": 435, "tid": 435, "ts": 1776649790907633, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907645, "dur": 0, + "args": { + "External id": 28217, + "cbid": 200, "correlation": 28217 + } + }, + { + "ph": "f", "id": 28217, "pid": 435, "tid": 435, "ts": 1776649790907645, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907646, "dur": 0, + "args": { + "External id": 28218, + "cbid": 273, "correlation": 28218 + } + }, + { + "ph": "f", "id": 28218, "pid": 435, "tid": 435, "ts": 1776649790907646, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790907655, "dur": 15, + "args": { + "External id": 28219, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28219, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28219, "pid": 0, "tid": 7, "ts": 1776649790907655, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907647, "dur": 5, + "args": { + "External id": 28219, + "cbid": 211, "correlation": 28219 + } + }, + { + "ph": "s", "id": 28219, "pid": 435, "tid": 435, "ts": 1776649790907647, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790907691, "dur": 1, + "args": { + "External id": 3318, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28237, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28237, "pid": 0, "tid": 7, "ts": 1776649790907691, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907682, "dur": 6, + "args": { + "External id": 3318, + "cbid": 211, "correlation": 28237 + } + }, + { + "ph": "s", "id": 28237, "pid": 435, "tid": 435, "ts": 1776649790907682, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907695, "dur": 0, + "args": { + "External id": 28243, + "cbid": 200, "correlation": 28243 + } + }, + { + "ph": "f", "id": 28243, "pid": 435, "tid": 435, "ts": 1776649790907695, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907695, "dur": 0, + "args": { + "External id": 28244, + "cbid": 273, "correlation": 28244 + } + }, + { + "ph": "f", "id": 28244, "pid": 435, "tid": 435, "ts": 1776649790907695, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790907704, "dur": 16, + "args": { + "External id": 28245, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28245, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28245, "pid": 0, "tid": 7, "ts": 1776649790907704, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907696, "dur": 5, + "args": { + "External id": 28245, + "cbid": 211, "correlation": 28245 + } + }, + { + "ph": "s", "id": 28245, "pid": 435, "tid": 435, "ts": 1776649790907696, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790907741, "dur": 1, + "args": { + "External id": 3323, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28263, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28263, "pid": 0, "tid": 7, "ts": 1776649790907741, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907731, "dur": 6, + "args": { + "External id": 3323, + "cbid": 211, "correlation": 28263 + } + }, + { + "ph": "s", "id": 28263, "pid": 435, "tid": 435, "ts": 1776649790907731, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907743, "dur": 0, + "args": { + "External id": 28269, + "cbid": 200, "correlation": 28269 + } + }, + { + "ph": "f", "id": 28269, "pid": 435, "tid": 435, "ts": 1776649790907743, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907744, "dur": 0, + "args": { + "External id": 28270, + "cbid": 273, "correlation": 28270 + } + }, + { + "ph": "f", "id": 28270, "pid": 435, "tid": 435, "ts": 1776649790907744, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790907753, "dur": 15, + "args": { + "External id": 28271, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28271, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28271, "pid": 0, "tid": 7, "ts": 1776649790907753, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907745, "dur": 4, + "args": { + "External id": 28271, + "cbid": 211, "correlation": 28271 + } + }, + { + "ph": "s", "id": 28271, "pid": 435, "tid": 435, "ts": 1776649790907745, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790907788, "dur": 1, + "args": { + "External id": 3328, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28289, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28289, "pid": 0, "tid": 7, "ts": 1776649790907788, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907779, "dur": 6, + "args": { + "External id": 3328, + "cbid": 211, "correlation": 28289 + } + }, + { + "ph": "s", "id": 28289, "pid": 435, "tid": 435, "ts": 1776649790907779, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907792, "dur": 0, + "args": { + "External id": 28295, + "cbid": 200, "correlation": 28295 + } + }, + { + "ph": "f", "id": 28295, "pid": 435, "tid": 435, "ts": 1776649790907792, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907793, "dur": 0, + "args": { + "External id": 28296, + "cbid": 273, "correlation": 28296 + } + }, + { + "ph": "f", "id": 28296, "pid": 435, "tid": 435, "ts": 1776649790907793, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790907802, "dur": 15, + "args": { + "External id": 28297, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28297, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28297, "pid": 0, "tid": 7, "ts": 1776649790907802, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907794, "dur": 5, + "args": { + "External id": 28297, + "cbid": 211, "correlation": 28297 + } + }, + { + "ph": "s", "id": 28297, "pid": 435, "tid": 435, "ts": 1776649790907794, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790907838, "dur": 1, + "args": { + "External id": 3333, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28315, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28315, "pid": 0, "tid": 7, "ts": 1776649790907838, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907829, "dur": 6, + "args": { + "External id": 3333, + "cbid": 211, "correlation": 28315 + } + }, + { + "ph": "s", "id": 28315, "pid": 435, "tid": 435, "ts": 1776649790907829, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907841, "dur": 0, + "args": { + "External id": 28321, + "cbid": 200, "correlation": 28321 + } + }, + { + "ph": "f", "id": 28321, "pid": 435, "tid": 435, "ts": 1776649790907841, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907842, "dur": 0, + "args": { + "External id": 28322, + "cbid": 273, "correlation": 28322 + } + }, + { + "ph": "f", "id": 28322, "pid": 435, "tid": 435, "ts": 1776649790907842, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790907851, "dur": 15, + "args": { + "External id": 28323, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28323, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28323, "pid": 0, "tid": 7, "ts": 1776649790907851, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907843, "dur": 4, + "args": { + "External id": 28323, + "cbid": 211, "correlation": 28323 + } + }, + { + "ph": "s", "id": 28323, "pid": 435, "tid": 435, "ts": 1776649790907843, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790907887, "dur": 1, + "args": { + "External id": 3338, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28341, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28341, "pid": 0, "tid": 7, "ts": 1776649790907887, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907877, "dur": 6, + "args": { + "External id": 3338, + "cbid": 211, "correlation": 28341 + } + }, + { + "ph": "s", "id": 28341, "pid": 435, "tid": 435, "ts": 1776649790907877, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907890, "dur": 0, + "args": { + "External id": 28347, + "cbid": 200, "correlation": 28347 + } + }, + { + "ph": "f", "id": 28347, "pid": 435, "tid": 435, "ts": 1776649790907890, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907891, "dur": 0, + "args": { + "External id": 28348, + "cbid": 273, "correlation": 28348 + } + }, + { + "ph": "f", "id": 28348, "pid": 435, "tid": 435, "ts": 1776649790907891, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790907900, "dur": 15, + "args": { + "External id": 28349, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28349, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28349, "pid": 0, "tid": 7, "ts": 1776649790907900, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907892, "dur": 5, + "args": { + "External id": 28349, + "cbid": 211, "correlation": 28349 + } + }, + { + "ph": "s", "id": 28349, "pid": 435, "tid": 435, "ts": 1776649790907892, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790907936, "dur": 1, + "args": { + "External id": 3343, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28367, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28367, "pid": 0, "tid": 7, "ts": 1776649790907936, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907926, "dur": 6, + "args": { + "External id": 3343, + "cbid": 211, "correlation": 28367 + } + }, + { + "ph": "s", "id": 28367, "pid": 435, "tid": 435, "ts": 1776649790907926, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907938, "dur": 0, + "args": { + "External id": 28373, + "cbid": 200, "correlation": 28373 + } + }, + { + "ph": "f", "id": 28373, "pid": 435, "tid": 435, "ts": 1776649790907938, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907939, "dur": 0, + "args": { + "External id": 28374, + "cbid": 273, "correlation": 28374 + } + }, + { + "ph": "f", "id": 28374, "pid": 435, "tid": 435, "ts": 1776649790907939, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790907948, "dur": 16, + "args": { + "External id": 28375, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28375, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28375, "pid": 0, "tid": 7, "ts": 1776649790907948, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907940, "dur": 4, + "args": { + "External id": 28375, + "cbid": 211, "correlation": 28375 + } + }, + { + "ph": "s", "id": 28375, "pid": 435, "tid": 435, "ts": 1776649790907940, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790907984, "dur": 1, + "args": { + "External id": 3348, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28393, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28393, "pid": 0, "tid": 7, "ts": 1776649790907984, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907975, "dur": 6, + "args": { + "External id": 3348, + "cbid": 211, "correlation": 28393 + } + }, + { + "ph": "s", "id": 28393, "pid": 435, "tid": 435, "ts": 1776649790907975, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907987, "dur": 0, + "args": { + "External id": 28399, + "cbid": 200, "correlation": 28399 + } + }, + { + "ph": "f", "id": 28399, "pid": 435, "tid": 435, "ts": 1776649790907987, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790907987, "dur": 0, + "args": { + "External id": 28400, + "cbid": 273, "correlation": 28400 + } + }, + { + "ph": "f", "id": 28400, "pid": 435, "tid": 435, "ts": 1776649790907987, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790907997, "dur": 16, + "args": { + "External id": 28401, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28401, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28401, "pid": 0, "tid": 7, "ts": 1776649790907997, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790907988, "dur": 4, + "args": { + "External id": 28401, + "cbid": 211, "correlation": 28401 + } + }, + { + "ph": "s", "id": 28401, "pid": 435, "tid": 435, "ts": 1776649790907988, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790908032, "dur": 1, + "args": { + "External id": 3353, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28419, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28419, "pid": 0, "tid": 7, "ts": 1776649790908032, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908023, "dur": 6, + "args": { + "External id": 3353, + "cbid": 211, "correlation": 28419 + } + }, + { + "ph": "s", "id": 28419, "pid": 435, "tid": 435, "ts": 1776649790908023, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908035, "dur": 0, + "args": { + "External id": 28425, + "cbid": 200, "correlation": 28425 + } + }, + { + "ph": "f", "id": 28425, "pid": 435, "tid": 435, "ts": 1776649790908035, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908036, "dur": 0, + "args": { + "External id": 28426, + "cbid": 273, "correlation": 28426 + } + }, + { + "ph": "f", "id": 28426, "pid": 435, "tid": 435, "ts": 1776649790908036, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790908044, "dur": 15, + "args": { + "External id": 28427, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28427, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28427, "pid": 0, "tid": 7, "ts": 1776649790908044, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908037, "dur": 4, + "args": { + "External id": 28427, + "cbid": 211, "correlation": 28427 + } + }, + { + "ph": "s", "id": 28427, "pid": 435, "tid": 435, "ts": 1776649790908037, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790908081, "dur": 1, + "args": { + "External id": 3358, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28445, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28445, "pid": 0, "tid": 7, "ts": 1776649790908081, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908071, "dur": 6, + "args": { + "External id": 3358, + "cbid": 211, "correlation": 28445 + } + }, + { + "ph": "s", "id": 28445, "pid": 435, "tid": 435, "ts": 1776649790908071, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908084, "dur": 0, + "args": { + "External id": 28451, + "cbid": 200, "correlation": 28451 + } + }, + { + "ph": "f", "id": 28451, "pid": 435, "tid": 435, "ts": 1776649790908084, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908084, "dur": 0, + "args": { + "External id": 28452, + "cbid": 273, "correlation": 28452 + } + }, + { + "ph": "f", "id": 28452, "pid": 435, "tid": 435, "ts": 1776649790908084, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790908093, "dur": 17, + "args": { + "External id": 28453, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28453, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28453, "pid": 0, "tid": 7, "ts": 1776649790908093, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908085, "dur": 5, + "args": { + "External id": 28453, + "cbid": 211, "correlation": 28453 + } + }, + { + "ph": "s", "id": 28453, "pid": 435, "tid": 435, "ts": 1776649790908085, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790908129, "dur": 1, + "args": { + "External id": 3363, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28471, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28471, "pid": 0, "tid": 7, "ts": 1776649790908129, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908120, "dur": 6, + "args": { + "External id": 3363, + "cbid": 211, "correlation": 28471 + } + }, + { + "ph": "s", "id": 28471, "pid": 435, "tid": 435, "ts": 1776649790908120, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908132, "dur": 0, + "args": { + "External id": 28477, + "cbid": 200, "correlation": 28477 + } + }, + { + "ph": "f", "id": 28477, "pid": 435, "tid": 435, "ts": 1776649790908132, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908133, "dur": 0, + "args": { + "External id": 28478, + "cbid": 273, "correlation": 28478 + } + }, + { + "ph": "f", "id": 28478, "pid": 435, "tid": 435, "ts": 1776649790908133, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790908142, "dur": 16, + "args": { + "External id": 28479, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28479, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28479, "pid": 0, "tid": 7, "ts": 1776649790908142, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908134, "dur": 4, + "args": { + "External id": 28479, + "cbid": 211, "correlation": 28479 + } + }, + { + "ph": "s", "id": 28479, "pid": 435, "tid": 435, "ts": 1776649790908134, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790908177, "dur": 1, + "args": { + "External id": 3368, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28497, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28497, "pid": 0, "tid": 7, "ts": 1776649790908177, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908168, "dur": 6, + "args": { + "External id": 3368, + "cbid": 211, "correlation": 28497 + } + }, + { + "ph": "s", "id": 28497, "pid": 435, "tid": 435, "ts": 1776649790908168, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908180, "dur": 0, + "args": { + "External id": 28503, + "cbid": 200, "correlation": 28503 + } + }, + { + "ph": "f", "id": 28503, "pid": 435, "tid": 435, "ts": 1776649790908180, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908181, "dur": 0, + "args": { + "External id": 28504, + "cbid": 273, "correlation": 28504 + } + }, + { + "ph": "f", "id": 28504, "pid": 435, "tid": 435, "ts": 1776649790908181, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790908190, "dur": 16, + "args": { + "External id": 28505, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28505, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28505, "pid": 0, "tid": 7, "ts": 1776649790908190, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908182, "dur": 4, + "args": { + "External id": 28505, + "cbid": 211, "correlation": 28505 + } + }, + { + "ph": "s", "id": 28505, "pid": 435, "tid": 435, "ts": 1776649790908182, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790908263, "dur": 1, + "args": { + "External id": 3373, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28523, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28523, "pid": 0, "tid": 7, "ts": 1776649790908263, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908251, "dur": 8, + "args": { + "External id": 3373, + "cbid": 211, "correlation": 28523 + } + }, + { + "ph": "s", "id": 28523, "pid": 435, "tid": 435, "ts": 1776649790908251, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908267, "dur": 0, + "args": { + "External id": 28529, + "cbid": 200, "correlation": 28529 + } + }, + { + "ph": "f", "id": 28529, "pid": 435, "tid": 435, "ts": 1776649790908267, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908268, "dur": 1, + "args": { + "External id": 28530, + "cbid": 273, "correlation": 28530 + } + }, + { + "ph": "f", "id": 28530, "pid": 435, "tid": 435, "ts": 1776649790908268, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790908278, "dur": 15, + "args": { + "External id": 28531, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28531, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28531, "pid": 0, "tid": 7, "ts": 1776649790908278, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908270, "dur": 5, + "args": { + "External id": 28531, + "cbid": 211, "correlation": 28531 + } + }, + { + "ph": "s", "id": 28531, "pid": 435, "tid": 435, "ts": 1776649790908270, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790908319, "dur": 1, + "args": { + "External id": 3378, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28549, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28549, "pid": 0, "tid": 7, "ts": 1776649790908319, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908310, "dur": 7, + "args": { + "External id": 3378, + "cbid": 211, "correlation": 28549 + } + }, + { + "ph": "s", "id": 28549, "pid": 435, "tid": 435, "ts": 1776649790908310, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908323, "dur": 0, + "args": { + "External id": 28555, + "cbid": 200, "correlation": 28555 + } + }, + { + "ph": "f", "id": 28555, "pid": 435, "tid": 435, "ts": 1776649790908323, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908323, "dur": 0, + "args": { + "External id": 28556, + "cbid": 273, "correlation": 28556 + } + }, + { + "ph": "f", "id": 28556, "pid": 435, "tid": 435, "ts": 1776649790908323, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790908332, "dur": 16, + "args": { + "External id": 28557, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28557, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28557, "pid": 0, "tid": 7, "ts": 1776649790908332, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908324, "dur": 5, + "args": { + "External id": 28557, + "cbid": 211, "correlation": 28557 + } + }, + { + "ph": "s", "id": 28557, "pid": 435, "tid": 435, "ts": 1776649790908324, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790908369, "dur": 1, + "args": { + "External id": 3383, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28575, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28575, "pid": 0, "tid": 7, "ts": 1776649790908369, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908359, "dur": 6, + "args": { + "External id": 3383, + "cbid": 211, "correlation": 28575 + } + }, + { + "ph": "s", "id": 28575, "pid": 435, "tid": 435, "ts": 1776649790908359, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908372, "dur": 0, + "args": { + "External id": 28581, + "cbid": 200, "correlation": 28581 + } + }, + { + "ph": "f", "id": 28581, "pid": 435, "tid": 435, "ts": 1776649790908372, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908372, "dur": 0, + "args": { + "External id": 28582, + "cbid": 273, "correlation": 28582 + } + }, + { + "ph": "f", "id": 28582, "pid": 435, "tid": 435, "ts": 1776649790908372, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790908382, "dur": 16, + "args": { + "External id": 28583, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28583, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28583, "pid": 0, "tid": 7, "ts": 1776649790908382, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908373, "dur": 5, + "args": { + "External id": 28583, + "cbid": 211, "correlation": 28583 + } + }, + { + "ph": "s", "id": 28583, "pid": 435, "tid": 435, "ts": 1776649790908373, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790908417, "dur": 1, + "args": { + "External id": 3388, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28601, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28601, "pid": 0, "tid": 7, "ts": 1776649790908417, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908408, "dur": 6, + "args": { + "External id": 3388, + "cbid": 211, "correlation": 28601 + } + }, + { + "ph": "s", "id": 28601, "pid": 435, "tid": 435, "ts": 1776649790908408, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908420, "dur": 0, + "args": { + "External id": 28607, + "cbid": 200, "correlation": 28607 + } + }, + { + "ph": "f", "id": 28607, "pid": 435, "tid": 435, "ts": 1776649790908420, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908421, "dur": 0, + "args": { + "External id": 28608, + "cbid": 273, "correlation": 28608 + } + }, + { + "ph": "f", "id": 28608, "pid": 435, "tid": 435, "ts": 1776649790908421, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790908430, "dur": 16, + "args": { + "External id": 28609, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28609, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28609, "pid": 0, "tid": 7, "ts": 1776649790908430, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908422, "dur": 4, + "args": { + "External id": 28609, + "cbid": 211, "correlation": 28609 + } + }, + { + "ph": "s", "id": 28609, "pid": 435, "tid": 435, "ts": 1776649790908422, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790908465, "dur": 1, + "args": { + "External id": 3393, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28627, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28627, "pid": 0, "tid": 7, "ts": 1776649790908465, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908456, "dur": 6, + "args": { + "External id": 3393, + "cbid": 211, "correlation": 28627 + } + }, + { + "ph": "s", "id": 28627, "pid": 435, "tid": 435, "ts": 1776649790908456, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908468, "dur": 0, + "args": { + "External id": 28633, + "cbid": 200, "correlation": 28633 + } + }, + { + "ph": "f", "id": 28633, "pid": 435, "tid": 435, "ts": 1776649790908468, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908469, "dur": 0, + "args": { + "External id": 28634, + "cbid": 273, "correlation": 28634 + } + }, + { + "ph": "f", "id": 28634, "pid": 435, "tid": 435, "ts": 1776649790908469, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790908478, "dur": 15, + "args": { + "External id": 28635, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28635, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28635, "pid": 0, "tid": 7, "ts": 1776649790908478, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908470, "dur": 5, + "args": { + "External id": 28635, + "cbid": 211, "correlation": 28635 + } + }, + { + "ph": "s", "id": 28635, "pid": 435, "tid": 435, "ts": 1776649790908470, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790908514, "dur": 1, + "args": { + "External id": 3398, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28653, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28653, "pid": 0, "tid": 7, "ts": 1776649790908514, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908505, "dur": 7, + "args": { + "External id": 3398, + "cbid": 211, "correlation": 28653 + } + }, + { + "ph": "s", "id": 28653, "pid": 435, "tid": 435, "ts": 1776649790908505, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908517, "dur": 0, + "args": { + "External id": 28659, + "cbid": 200, "correlation": 28659 + } + }, + { + "ph": "f", "id": 28659, "pid": 435, "tid": 435, "ts": 1776649790908517, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908518, "dur": 0, + "args": { + "External id": 28660, + "cbid": 273, "correlation": 28660 + } + }, + { + "ph": "f", "id": 28660, "pid": 435, "tid": 435, "ts": 1776649790908518, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790908527, "dur": 15, + "args": { + "External id": 28661, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28661, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28661, "pid": 0, "tid": 7, "ts": 1776649790908527, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908519, "dur": 4, + "args": { + "External id": 28661, + "cbid": 211, "correlation": 28661 + } + }, + { + "ph": "s", "id": 28661, "pid": 435, "tid": 435, "ts": 1776649790908519, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790908563, "dur": 1, + "args": { + "External id": 3403, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28679, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28679, "pid": 0, "tid": 7, "ts": 1776649790908563, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908554, "dur": 6, + "args": { + "External id": 3403, + "cbid": 211, "correlation": 28679 + } + }, + { + "ph": "s", "id": 28679, "pid": 435, "tid": 435, "ts": 1776649790908554, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908566, "dur": 0, + "args": { + "External id": 28685, + "cbid": 200, "correlation": 28685 + } + }, + { + "ph": "f", "id": 28685, "pid": 435, "tid": 435, "ts": 1776649790908566, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908566, "dur": 0, + "args": { + "External id": 28686, + "cbid": 273, "correlation": 28686 + } + }, + { + "ph": "f", "id": 28686, "pid": 435, "tid": 435, "ts": 1776649790908566, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790908575, "dur": 17, + "args": { + "External id": 28687, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28687, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28687, "pid": 0, "tid": 7, "ts": 1776649790908575, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908568, "dur": 4, + "args": { + "External id": 28687, + "cbid": 211, "correlation": 28687 + } + }, + { + "ph": "s", "id": 28687, "pid": 435, "tid": 435, "ts": 1776649790908568, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790908611, "dur": 1, + "args": { + "External id": 3408, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28705, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28705, "pid": 0, "tid": 7, "ts": 1776649790908611, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908601, "dur": 6, + "args": { + "External id": 3408, + "cbid": 211, "correlation": 28705 + } + }, + { + "ph": "s", "id": 28705, "pid": 435, "tid": 435, "ts": 1776649790908601, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908614, "dur": 0, + "args": { + "External id": 28711, + "cbid": 200, "correlation": 28711 + } + }, + { + "ph": "f", "id": 28711, "pid": 435, "tid": 435, "ts": 1776649790908614, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908614, "dur": 0, + "args": { + "External id": 28712, + "cbid": 273, "correlation": 28712 + } + }, + { + "ph": "f", "id": 28712, "pid": 435, "tid": 435, "ts": 1776649790908614, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790908624, "dur": 15, + "args": { + "External id": 28713, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28713, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28713, "pid": 0, "tid": 7, "ts": 1776649790908624, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908615, "dur": 5, + "args": { + "External id": 28713, + "cbid": 211, "correlation": 28713 + } + }, + { + "ph": "s", "id": 28713, "pid": 435, "tid": 435, "ts": 1776649790908615, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790908660, "dur": 1, + "args": { + "External id": 3413, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28731, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28731, "pid": 0, "tid": 7, "ts": 1776649790908660, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908651, "dur": 6, + "args": { + "External id": 3413, + "cbid": 211, "correlation": 28731 + } + }, + { + "ph": "s", "id": 28731, "pid": 435, "tid": 435, "ts": 1776649790908651, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908663, "dur": 0, + "args": { + "External id": 28737, + "cbid": 200, "correlation": 28737 + } + }, + { + "ph": "f", "id": 28737, "pid": 435, "tid": 435, "ts": 1776649790908663, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908663, "dur": 0, + "args": { + "External id": 28738, + "cbid": 273, "correlation": 28738 + } + }, + { + "ph": "f", "id": 28738, "pid": 435, "tid": 435, "ts": 1776649790908663, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790908673, "dur": 16, + "args": { + "External id": 28739, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28739, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28739, "pid": 0, "tid": 7, "ts": 1776649790908673, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908665, "dur": 5, + "args": { + "External id": 28739, + "cbid": 211, "correlation": 28739 + } + }, + { + "ph": "s", "id": 28739, "pid": 435, "tid": 435, "ts": 1776649790908665, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790908708, "dur": 1, + "args": { + "External id": 3418, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28757, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28757, "pid": 0, "tid": 7, "ts": 1776649790908708, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908699, "dur": 6, + "args": { + "External id": 3418, + "cbid": 211, "correlation": 28757 + } + }, + { + "ph": "s", "id": 28757, "pid": 435, "tid": 435, "ts": 1776649790908699, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908711, "dur": 0, + "args": { + "External id": 28763, + "cbid": 200, "correlation": 28763 + } + }, + { + "ph": "f", "id": 28763, "pid": 435, "tid": 435, "ts": 1776649790908711, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908712, "dur": 0, + "args": { + "External id": 28764, + "cbid": 273, "correlation": 28764 + } + }, + { + "ph": "f", "id": 28764, "pid": 435, "tid": 435, "ts": 1776649790908712, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790908721, "dur": 15, + "args": { + "External id": 28765, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28765, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28765, "pid": 0, "tid": 7, "ts": 1776649790908721, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908713, "dur": 4, + "args": { + "External id": 28765, + "cbid": 211, "correlation": 28765 + } + }, + { + "ph": "s", "id": 28765, "pid": 435, "tid": 435, "ts": 1776649790908713, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790908757, "dur": 1, + "args": { + "External id": 3423, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28783, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28783, "pid": 0, "tid": 7, "ts": 1776649790908757, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908747, "dur": 6, + "args": { + "External id": 3423, + "cbid": 211, "correlation": 28783 + } + }, + { + "ph": "s", "id": 28783, "pid": 435, "tid": 435, "ts": 1776649790908747, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908760, "dur": 0, + "args": { + "External id": 28789, + "cbid": 200, "correlation": 28789 + } + }, + { + "ph": "f", "id": 28789, "pid": 435, "tid": 435, "ts": 1776649790908760, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908761, "dur": 0, + "args": { + "External id": 28790, + "cbid": 273, "correlation": 28790 + } + }, + { + "ph": "f", "id": 28790, "pid": 435, "tid": 435, "ts": 1776649790908761, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790908770, "dur": 16, + "args": { + "External id": 28791, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28791, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28791, "pid": 0, "tid": 7, "ts": 1776649790908770, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908762, "dur": 5, + "args": { + "External id": 28791, + "cbid": 211, "correlation": 28791 + } + }, + { + "ph": "s", "id": 28791, "pid": 435, "tid": 435, "ts": 1776649790908762, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790908805, "dur": 1, + "args": { + "External id": 3428, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28809, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28809, "pid": 0, "tid": 7, "ts": 1776649790908805, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908796, "dur": 6, + "args": { + "External id": 3428, + "cbid": 211, "correlation": 28809 + } + }, + { + "ph": "s", "id": 28809, "pid": 435, "tid": 435, "ts": 1776649790908796, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908808, "dur": 0, + "args": { + "External id": 28815, + "cbid": 200, "correlation": 28815 + } + }, + { + "ph": "f", "id": 28815, "pid": 435, "tid": 435, "ts": 1776649790908808, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908809, "dur": 0, + "args": { + "External id": 28816, + "cbid": 273, "correlation": 28816 + } + }, + { + "ph": "f", "id": 28816, "pid": 435, "tid": 435, "ts": 1776649790908809, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790908817, "dur": 15, + "args": { + "External id": 28817, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28817, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28817, "pid": 0, "tid": 7, "ts": 1776649790908817, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908810, "dur": 4, + "args": { + "External id": 28817, + "cbid": 211, "correlation": 28817 + } + }, + { + "ph": "s", "id": 28817, "pid": 435, "tid": 435, "ts": 1776649790908810, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790908853, "dur": 1, + "args": { + "External id": 3433, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28835, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28835, "pid": 0, "tid": 7, "ts": 1776649790908853, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908844, "dur": 6, + "args": { + "External id": 3433, + "cbid": 211, "correlation": 28835 + } + }, + { + "ph": "s", "id": 28835, "pid": 435, "tid": 435, "ts": 1776649790908844, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908856, "dur": 0, + "args": { + "External id": 28841, + "cbid": 200, "correlation": 28841 + } + }, + { + "ph": "f", "id": 28841, "pid": 435, "tid": 435, "ts": 1776649790908856, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908856, "dur": 0, + "args": { + "External id": 28842, + "cbid": 273, "correlation": 28842 + } + }, + { + "ph": "f", "id": 28842, "pid": 435, "tid": 435, "ts": 1776649790908856, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790908865, "dur": 15, + "args": { + "External id": 28843, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28843, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28843, "pid": 0, "tid": 7, "ts": 1776649790908865, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908857, "dur": 4, + "args": { + "External id": 28843, + "cbid": 211, "correlation": 28843 + } + }, + { + "ph": "s", "id": 28843, "pid": 435, "tid": 435, "ts": 1776649790908857, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790908900, "dur": 1, + "args": { + "External id": 3438, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28861, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28861, "pid": 0, "tid": 7, "ts": 1776649790908900, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908891, "dur": 6, + "args": { + "External id": 3438, + "cbid": 211, "correlation": 28861 + } + }, + { + "ph": "s", "id": 28861, "pid": 435, "tid": 435, "ts": 1776649790908891, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908903, "dur": 0, + "args": { + "External id": 28867, + "cbid": 200, "correlation": 28867 + } + }, + { + "ph": "f", "id": 28867, "pid": 435, "tid": 435, "ts": 1776649790908903, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908904, "dur": 0, + "args": { + "External id": 28868, + "cbid": 273, "correlation": 28868 + } + }, + { + "ph": "f", "id": 28868, "pid": 435, "tid": 435, "ts": 1776649790908904, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790908913, "dur": 16, + "args": { + "External id": 28869, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28869, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28869, "pid": 0, "tid": 7, "ts": 1776649790908913, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908905, "dur": 4, + "args": { + "External id": 28869, + "cbid": 211, "correlation": 28869 + } + }, + { + "ph": "s", "id": 28869, "pid": 435, "tid": 435, "ts": 1776649790908905, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790908948, "dur": 1, + "args": { + "External id": 3443, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28887, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28887, "pid": 0, "tid": 7, "ts": 1776649790908948, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908939, "dur": 6, + "args": { + "External id": 3443, + "cbid": 211, "correlation": 28887 + } + }, + { + "ph": "s", "id": 28887, "pid": 435, "tid": 435, "ts": 1776649790908939, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908951, "dur": 0, + "args": { + "External id": 28893, + "cbid": 200, "correlation": 28893 + } + }, + { + "ph": "f", "id": 28893, "pid": 435, "tid": 435, "ts": 1776649790908951, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908952, "dur": 0, + "args": { + "External id": 28894, + "cbid": 273, "correlation": 28894 + } + }, + { + "ph": "f", "id": 28894, "pid": 435, "tid": 435, "ts": 1776649790908952, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790908961, "dur": 17, + "args": { + "External id": 28895, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28895, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28895, "pid": 0, "tid": 7, "ts": 1776649790908961, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908953, "dur": 5, + "args": { + "External id": 28895, + "cbid": 211, "correlation": 28895 + } + }, + { + "ph": "s", "id": 28895, "pid": 435, "tid": 435, "ts": 1776649790908953, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790908996, "dur": 1, + "args": { + "External id": 3448, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28913, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28913, "pid": 0, "tid": 7, "ts": 1776649790908996, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790908987, "dur": 6, + "args": { + "External id": 3448, + "cbid": 211, "correlation": 28913 + } + }, + { + "ph": "s", "id": 28913, "pid": 435, "tid": 435, "ts": 1776649790908987, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790908999, "dur": 0, + "args": { + "External id": 28919, + "cbid": 200, "correlation": 28919 + } + }, + { + "ph": "f", "id": 28919, "pid": 435, "tid": 435, "ts": 1776649790908999, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909000, "dur": 0, + "args": { + "External id": 28920, + "cbid": 273, "correlation": 28920 + } + }, + { + "ph": "f", "id": 28920, "pid": 435, "tid": 435, "ts": 1776649790909000, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790909009, "dur": 16, + "args": { + "External id": 28921, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28921, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28921, "pid": 0, "tid": 7, "ts": 1776649790909009, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909001, "dur": 4, + "args": { + "External id": 28921, + "cbid": 211, "correlation": 28921 + } + }, + { + "ph": "s", "id": 28921, "pid": 435, "tid": 435, "ts": 1776649790909001, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790909044, "dur": 1, + "args": { + "External id": 3453, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28939, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28939, "pid": 0, "tid": 7, "ts": 1776649790909044, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909035, "dur": 6, + "args": { + "External id": 3453, + "cbid": 211, "correlation": 28939 + } + }, + { + "ph": "s", "id": 28939, "pid": 435, "tid": 435, "ts": 1776649790909035, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909047, "dur": 0, + "args": { + "External id": 28945, + "cbid": 200, "correlation": 28945 + } + }, + { + "ph": "f", "id": 28945, "pid": 435, "tid": 435, "ts": 1776649790909047, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909048, "dur": 0, + "args": { + "External id": 28946, + "cbid": 273, "correlation": 28946 + } + }, + { + "ph": "f", "id": 28946, "pid": 435, "tid": 435, "ts": 1776649790909048, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790909056, "dur": 15, + "args": { + "External id": 28947, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28947, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28947, "pid": 0, "tid": 7, "ts": 1776649790909056, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909049, "dur": 4, + "args": { + "External id": 28947, + "cbid": 211, "correlation": 28947 + } + }, + { + "ph": "s", "id": 28947, "pid": 435, "tid": 435, "ts": 1776649790909049, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790909092, "dur": 1, + "args": { + "External id": 3458, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28965, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28965, "pid": 0, "tid": 7, "ts": 1776649790909092, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909082, "dur": 6, + "args": { + "External id": 3458, + "cbid": 211, "correlation": 28965 + } + }, + { + "ph": "s", "id": 28965, "pid": 435, "tid": 435, "ts": 1776649790909082, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909095, "dur": 0, + "args": { + "External id": 28971, + "cbid": 200, "correlation": 28971 + } + }, + { + "ph": "f", "id": 28971, "pid": 435, "tid": 435, "ts": 1776649790909095, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909095, "dur": 0, + "args": { + "External id": 28972, + "cbid": 273, "correlation": 28972 + } + }, + { + "ph": "f", "id": 28972, "pid": 435, "tid": 435, "ts": 1776649790909095, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790909104, "dur": 16, + "args": { + "External id": 28973, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28973, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28973, "pid": 0, "tid": 7, "ts": 1776649790909104, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909096, "dur": 4, + "args": { + "External id": 28973, + "cbid": 211, "correlation": 28973 + } + }, + { + "ph": "s", "id": 28973, "pid": 435, "tid": 435, "ts": 1776649790909096, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790909140, "dur": 1, + "args": { + "External id": 3463, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28991, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28991, "pid": 0, "tid": 7, "ts": 1776649790909140, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909131, "dur": 6, + "args": { + "External id": 3463, + "cbid": 211, "correlation": 28991 + } + }, + { + "ph": "s", "id": 28991, "pid": 435, "tid": 435, "ts": 1776649790909131, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909143, "dur": 0, + "args": { + "External id": 28997, + "cbid": 200, "correlation": 28997 + } + }, + { + "ph": "f", "id": 28997, "pid": 435, "tid": 435, "ts": 1776649790909143, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909143, "dur": 0, + "args": { + "External id": 28998, + "cbid": 273, "correlation": 28998 + } + }, + { + "ph": "f", "id": 28998, "pid": 435, "tid": 435, "ts": 1776649790909143, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790909152, "dur": 16, + "args": { + "External id": 28999, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 28999, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 28999, "pid": 0, "tid": 7, "ts": 1776649790909152, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909144, "dur": 5, + "args": { + "External id": 28999, + "cbid": 211, "correlation": 28999 + } + }, + { + "ph": "s", "id": 28999, "pid": 435, "tid": 435, "ts": 1776649790909144, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790909188, "dur": 1, + "args": { + "External id": 3468, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29017, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29017, "pid": 0, "tid": 7, "ts": 1776649790909188, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909178, "dur": 6, + "args": { + "External id": 3468, + "cbid": 211, "correlation": 29017 + } + }, + { + "ph": "s", "id": 29017, "pid": 435, "tid": 435, "ts": 1776649790909178, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909191, "dur": 0, + "args": { + "External id": 29023, + "cbid": 200, "correlation": 29023 + } + }, + { + "ph": "f", "id": 29023, "pid": 435, "tid": 435, "ts": 1776649790909191, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909191, "dur": 0, + "args": { + "External id": 29024, + "cbid": 273, "correlation": 29024 + } + }, + { + "ph": "f", "id": 29024, "pid": 435, "tid": 435, "ts": 1776649790909191, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790909200, "dur": 15, + "args": { + "External id": 29025, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29025, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29025, "pid": 0, "tid": 7, "ts": 1776649790909200, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909193, "dur": 5, + "args": { + "External id": 29025, + "cbid": 211, "correlation": 29025 + } + }, + { + "ph": "s", "id": 29025, "pid": 435, "tid": 435, "ts": 1776649790909193, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790909236, "dur": 1, + "args": { + "External id": 3473, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29043, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29043, "pid": 0, "tid": 7, "ts": 1776649790909236, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909227, "dur": 6, + "args": { + "External id": 3473, + "cbid": 211, "correlation": 29043 + } + }, + { + "ph": "s", "id": 29043, "pid": 435, "tid": 435, "ts": 1776649790909227, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909239, "dur": 0, + "args": { + "External id": 29049, + "cbid": 200, "correlation": 29049 + } + }, + { + "ph": "f", "id": 29049, "pid": 435, "tid": 435, "ts": 1776649790909239, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909240, "dur": 0, + "args": { + "External id": 29050, + "cbid": 273, "correlation": 29050 + } + }, + { + "ph": "f", "id": 29050, "pid": 435, "tid": 435, "ts": 1776649790909240, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790909249, "dur": 16, + "args": { + "External id": 29051, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29051, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29051, "pid": 0, "tid": 7, "ts": 1776649790909249, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909241, "dur": 5, + "args": { + "External id": 29051, + "cbid": 211, "correlation": 29051 + } + }, + { + "ph": "s", "id": 29051, "pid": 435, "tid": 435, "ts": 1776649790909241, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790909284, "dur": 1, + "args": { + "External id": 3478, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29069, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29069, "pid": 0, "tid": 7, "ts": 1776649790909284, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909275, "dur": 6, + "args": { + "External id": 3478, + "cbid": 211, "correlation": 29069 + } + }, + { + "ph": "s", "id": 29069, "pid": 435, "tid": 435, "ts": 1776649790909275, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909287, "dur": 0, + "args": { + "External id": 29075, + "cbid": 200, "correlation": 29075 + } + }, + { + "ph": "f", "id": 29075, "pid": 435, "tid": 435, "ts": 1776649790909287, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909287, "dur": 0, + "args": { + "External id": 29076, + "cbid": 273, "correlation": 29076 + } + }, + { + "ph": "f", "id": 29076, "pid": 435, "tid": 435, "ts": 1776649790909287, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790909296, "dur": 16, + "args": { + "External id": 29077, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29077, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29077, "pid": 0, "tid": 7, "ts": 1776649790909296, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909288, "dur": 10, + "args": { + "External id": 29077, + "cbid": 211, "correlation": 29077 + } + }, + { + "ph": "s", "id": 29077, "pid": 435, "tid": 435, "ts": 1776649790909288, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790909341, "dur": 1, + "args": { + "External id": 3483, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29095, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29095, "pid": 0, "tid": 7, "ts": 1776649790909341, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909331, "dur": 7, + "args": { + "External id": 3483, + "cbid": 211, "correlation": 29095 + } + }, + { + "ph": "s", "id": 29095, "pid": 435, "tid": 435, "ts": 1776649790909331, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909344, "dur": 0, + "args": { + "External id": 29101, + "cbid": 200, "correlation": 29101 + } + }, + { + "ph": "f", "id": 29101, "pid": 435, "tid": 435, "ts": 1776649790909344, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909345, "dur": 0, + "args": { + "External id": 29102, + "cbid": 273, "correlation": 29102 + } + }, + { + "ph": "f", "id": 29102, "pid": 435, "tid": 435, "ts": 1776649790909345, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790909354, "dur": 16, + "args": { + "External id": 29103, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29103, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29103, "pid": 0, "tid": 7, "ts": 1776649790909354, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909346, "dur": 4, + "args": { + "External id": 29103, + "cbid": 211, "correlation": 29103 + } + }, + { + "ph": "s", "id": 29103, "pid": 435, "tid": 435, "ts": 1776649790909346, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790909390, "dur": 1, + "args": { + "External id": 3488, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29121, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29121, "pid": 0, "tid": 7, "ts": 1776649790909390, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909380, "dur": 7, + "args": { + "External id": 3488, + "cbid": 211, "correlation": 29121 + } + }, + { + "ph": "s", "id": 29121, "pid": 435, "tid": 435, "ts": 1776649790909380, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909393, "dur": 0, + "args": { + "External id": 29127, + "cbid": 200, "correlation": 29127 + } + }, + { + "ph": "f", "id": 29127, "pid": 435, "tid": 435, "ts": 1776649790909393, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909393, "dur": 0, + "args": { + "External id": 29128, + "cbid": 273, "correlation": 29128 + } + }, + { + "ph": "f", "id": 29128, "pid": 435, "tid": 435, "ts": 1776649790909393, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790909402, "dur": 16, + "args": { + "External id": 29129, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29129, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29129, "pid": 0, "tid": 7, "ts": 1776649790909402, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909395, "dur": 4, + "args": { + "External id": 29129, + "cbid": 211, "correlation": 29129 + } + }, + { + "ph": "s", "id": 29129, "pid": 435, "tid": 435, "ts": 1776649790909395, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790909438, "dur": 1, + "args": { + "External id": 3493, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29147, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29147, "pid": 0, "tid": 7, "ts": 1776649790909438, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909429, "dur": 6, + "args": { + "External id": 3493, + "cbid": 211, "correlation": 29147 + } + }, + { + "ph": "s", "id": 29147, "pid": 435, "tid": 435, "ts": 1776649790909429, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909441, "dur": 0, + "args": { + "External id": 29153, + "cbid": 200, "correlation": 29153 + } + }, + { + "ph": "f", "id": 29153, "pid": 435, "tid": 435, "ts": 1776649790909441, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909442, "dur": 0, + "args": { + "External id": 29154, + "cbid": 273, "correlation": 29154 + } + }, + { + "ph": "f", "id": 29154, "pid": 435, "tid": 435, "ts": 1776649790909442, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790909451, "dur": 16, + "args": { + "External id": 29155, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29155, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29155, "pid": 0, "tid": 7, "ts": 1776649790909451, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909443, "dur": 4, + "args": { + "External id": 29155, + "cbid": 211, "correlation": 29155 + } + }, + { + "ph": "s", "id": 29155, "pid": 435, "tid": 435, "ts": 1776649790909443, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790909486, "dur": 1, + "args": { + "External id": 3498, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29173, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29173, "pid": 0, "tid": 7, "ts": 1776649790909486, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909477, "dur": 6, + "args": { + "External id": 3498, + "cbid": 211, "correlation": 29173 + } + }, + { + "ph": "s", "id": 29173, "pid": 435, "tid": 435, "ts": 1776649790909477, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909489, "dur": 0, + "args": { + "External id": 29179, + "cbid": 200, "correlation": 29179 + } + }, + { + "ph": "f", "id": 29179, "pid": 435, "tid": 435, "ts": 1776649790909489, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909490, "dur": 0, + "args": { + "External id": 29180, + "cbid": 273, "correlation": 29180 + } + }, + { + "ph": "f", "id": 29180, "pid": 435, "tid": 435, "ts": 1776649790909490, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790909499, "dur": 15, + "args": { + "External id": 29181, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29181, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29181, "pid": 0, "tid": 7, "ts": 1776649790909499, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909491, "dur": 4, + "args": { + "External id": 29181, + "cbid": 211, "correlation": 29181 + } + }, + { + "ph": "s", "id": 29181, "pid": 435, "tid": 435, "ts": 1776649790909491, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790909534, "dur": 1, + "args": { + "External id": 3503, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29199, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29199, "pid": 0, "tid": 7, "ts": 1776649790909534, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909525, "dur": 6, + "args": { + "External id": 3503, + "cbid": 211, "correlation": 29199 + } + }, + { + "ph": "s", "id": 29199, "pid": 435, "tid": 435, "ts": 1776649790909525, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909537, "dur": 0, + "args": { + "External id": 29205, + "cbid": 200, "correlation": 29205 + } + }, + { + "ph": "f", "id": 29205, "pid": 435, "tid": 435, "ts": 1776649790909537, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909538, "dur": 0, + "args": { + "External id": 29206, + "cbid": 273, "correlation": 29206 + } + }, + { + "ph": "f", "id": 29206, "pid": 435, "tid": 435, "ts": 1776649790909538, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790909547, "dur": 15, + "args": { + "External id": 29207, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29207, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29207, "pid": 0, "tid": 7, "ts": 1776649790909547, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909539, "dur": 5, + "args": { + "External id": 29207, + "cbid": 211, "correlation": 29207 + } + }, + { + "ph": "s", "id": 29207, "pid": 435, "tid": 435, "ts": 1776649790909539, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790909583, "dur": 1, + "args": { + "External id": 3508, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29225, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29225, "pid": 0, "tid": 7, "ts": 1776649790909583, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909573, "dur": 6, + "args": { + "External id": 3508, + "cbid": 211, "correlation": 29225 + } + }, + { + "ph": "s", "id": 29225, "pid": 435, "tid": 435, "ts": 1776649790909573, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909586, "dur": 0, + "args": { + "External id": 29231, + "cbid": 200, "correlation": 29231 + } + }, + { + "ph": "f", "id": 29231, "pid": 435, "tid": 435, "ts": 1776649790909586, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909586, "dur": 0, + "args": { + "External id": 29232, + "cbid": 273, "correlation": 29232 + } + }, + { + "ph": "f", "id": 29232, "pid": 435, "tid": 435, "ts": 1776649790909586, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790909595, "dur": 16, + "args": { + "External id": 29233, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29233, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29233, "pid": 0, "tid": 7, "ts": 1776649790909595, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909587, "dur": 4, + "args": { + "External id": 29233, + "cbid": 211, "correlation": 29233 + } + }, + { + "ph": "s", "id": 29233, "pid": 435, "tid": 435, "ts": 1776649790909587, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790909631, "dur": 1, + "args": { + "External id": 3513, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29251, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29251, "pid": 0, "tid": 7, "ts": 1776649790909631, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909621, "dur": 6, + "args": { + "External id": 3513, + "cbid": 211, "correlation": 29251 + } + }, + { + "ph": "s", "id": 29251, "pid": 435, "tid": 435, "ts": 1776649790909621, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909634, "dur": 0, + "args": { + "External id": 29257, + "cbid": 200, "correlation": 29257 + } + }, + { + "ph": "f", "id": 29257, "pid": 435, "tid": 435, "ts": 1776649790909634, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909634, "dur": 0, + "args": { + "External id": 29258, + "cbid": 273, "correlation": 29258 + } + }, + { + "ph": "f", "id": 29258, "pid": 435, "tid": 435, "ts": 1776649790909634, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790909643, "dur": 16, + "args": { + "External id": 29259, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29259, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29259, "pid": 0, "tid": 7, "ts": 1776649790909643, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909635, "dur": 4, + "args": { + "External id": 29259, + "cbid": 211, "correlation": 29259 + } + }, + { + "ph": "s", "id": 29259, "pid": 435, "tid": 435, "ts": 1776649790909635, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790909679, "dur": 1, + "args": { + "External id": 3518, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29277, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29277, "pid": 0, "tid": 7, "ts": 1776649790909679, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909670, "dur": 6, + "args": { + "External id": 3518, + "cbid": 211, "correlation": 29277 + } + }, + { + "ph": "s", "id": 29277, "pid": 435, "tid": 435, "ts": 1776649790909670, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909682, "dur": 0, + "args": { + "External id": 29283, + "cbid": 200, "correlation": 29283 + } + }, + { + "ph": "f", "id": 29283, "pid": 435, "tid": 435, "ts": 1776649790909682, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909682, "dur": 0, + "args": { + "External id": 29284, + "cbid": 273, "correlation": 29284 + } + }, + { + "ph": "f", "id": 29284, "pid": 435, "tid": 435, "ts": 1776649790909682, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790909691, "dur": 15, + "args": { + "External id": 29285, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29285, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29285, "pid": 0, "tid": 7, "ts": 1776649790909691, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909684, "dur": 4, + "args": { + "External id": 29285, + "cbid": 211, "correlation": 29285 + } + }, + { + "ph": "s", "id": 29285, "pid": 435, "tid": 435, "ts": 1776649790909684, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790909727, "dur": 1, + "args": { + "External id": 3523, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29303, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29303, "pid": 0, "tid": 7, "ts": 1776649790909727, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909718, "dur": 6, + "args": { + "External id": 3523, + "cbid": 211, "correlation": 29303 + } + }, + { + "ph": "s", "id": 29303, "pid": 435, "tid": 435, "ts": 1776649790909718, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909730, "dur": 0, + "args": { + "External id": 29309, + "cbid": 200, "correlation": 29309 + } + }, + { + "ph": "f", "id": 29309, "pid": 435, "tid": 435, "ts": 1776649790909730, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909731, "dur": 0, + "args": { + "External id": 29310, + "cbid": 273, "correlation": 29310 + } + }, + { + "ph": "f", "id": 29310, "pid": 435, "tid": 435, "ts": 1776649790909731, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790909740, "dur": 15, + "args": { + "External id": 29311, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29311, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29311, "pid": 0, "tid": 7, "ts": 1776649790909740, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909732, "dur": 5, + "args": { + "External id": 29311, + "cbid": 211, "correlation": 29311 + } + }, + { + "ph": "s", "id": 29311, "pid": 435, "tid": 435, "ts": 1776649790909732, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790909775, "dur": 1, + "args": { + "External id": 3528, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29329, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29329, "pid": 0, "tid": 7, "ts": 1776649790909775, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909766, "dur": 6, + "args": { + "External id": 3528, + "cbid": 211, "correlation": 29329 + } + }, + { + "ph": "s", "id": 29329, "pid": 435, "tid": 435, "ts": 1776649790909766, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909778, "dur": 0, + "args": { + "External id": 29335, + "cbid": 200, "correlation": 29335 + } + }, + { + "ph": "f", "id": 29335, "pid": 435, "tid": 435, "ts": 1776649790909778, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909779, "dur": 0, + "args": { + "External id": 29336, + "cbid": 273, "correlation": 29336 + } + }, + { + "ph": "f", "id": 29336, "pid": 435, "tid": 435, "ts": 1776649790909779, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790909788, "dur": 15, + "args": { + "External id": 29337, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29337, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29337, "pid": 0, "tid": 7, "ts": 1776649790909788, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909780, "dur": 4, + "args": { + "External id": 29337, + "cbid": 211, "correlation": 29337 + } + }, + { + "ph": "s", "id": 29337, "pid": 435, "tid": 435, "ts": 1776649790909780, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790909824, "dur": 1, + "args": { + "External id": 3533, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29355, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29355, "pid": 0, "tid": 7, "ts": 1776649790909824, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909814, "dur": 6, + "args": { + "External id": 3533, + "cbid": 211, "correlation": 29355 + } + }, + { + "ph": "s", "id": 29355, "pid": 435, "tid": 435, "ts": 1776649790909814, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909827, "dur": 0, + "args": { + "External id": 29361, + "cbid": 200, "correlation": 29361 + } + }, + { + "ph": "f", "id": 29361, "pid": 435, "tid": 435, "ts": 1776649790909827, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909827, "dur": 0, + "args": { + "External id": 29362, + "cbid": 273, "correlation": 29362 + } + }, + { + "ph": "f", "id": 29362, "pid": 435, "tid": 435, "ts": 1776649790909827, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790909836, "dur": 16, + "args": { + "External id": 29363, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29363, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29363, "pid": 0, "tid": 7, "ts": 1776649790909836, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909828, "dur": 4, + "args": { + "External id": 29363, + "cbid": 211, "correlation": 29363 + } + }, + { + "ph": "s", "id": 29363, "pid": 435, "tid": 435, "ts": 1776649790909828, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790909871, "dur": 1, + "args": { + "External id": 3538, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29381, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29381, "pid": 0, "tid": 7, "ts": 1776649790909871, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909862, "dur": 6, + "args": { + "External id": 3538, + "cbid": 211, "correlation": 29381 + } + }, + { + "ph": "s", "id": 29381, "pid": 435, "tid": 435, "ts": 1776649790909862, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909874, "dur": 0, + "args": { + "External id": 29387, + "cbid": 200, "correlation": 29387 + } + }, + { + "ph": "f", "id": 29387, "pid": 435, "tid": 435, "ts": 1776649790909874, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909875, "dur": 0, + "args": { + "External id": 29388, + "cbid": 273, "correlation": 29388 + } + }, + { + "ph": "f", "id": 29388, "pid": 435, "tid": 435, "ts": 1776649790909875, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790909884, "dur": 16, + "args": { + "External id": 29389, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29389, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29389, "pid": 0, "tid": 7, "ts": 1776649790909884, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909876, "dur": 4, + "args": { + "External id": 29389, + "cbid": 211, "correlation": 29389 + } + }, + { + "ph": "s", "id": 29389, "pid": 435, "tid": 435, "ts": 1776649790909876, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790909919, "dur": 1, + "args": { + "External id": 3543, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29407, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29407, "pid": 0, "tid": 7, "ts": 1776649790909919, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909910, "dur": 6, + "args": { + "External id": 3543, + "cbid": 211, "correlation": 29407 + } + }, + { + "ph": "s", "id": 29407, "pid": 435, "tid": 435, "ts": 1776649790909910, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909922, "dur": 0, + "args": { + "External id": 29413, + "cbid": 200, "correlation": 29413 + } + }, + { + "ph": "f", "id": 29413, "pid": 435, "tid": 435, "ts": 1776649790909922, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909923, "dur": 0, + "args": { + "External id": 29414, + "cbid": 273, "correlation": 29414 + } + }, + { + "ph": "f", "id": 29414, "pid": 435, "tid": 435, "ts": 1776649790909923, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790909932, "dur": 16, + "args": { + "External id": 29415, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29415, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29415, "pid": 0, "tid": 7, "ts": 1776649790909932, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909924, "dur": 5, + "args": { + "External id": 29415, + "cbid": 211, "correlation": 29415 + } + }, + { + "ph": "s", "id": 29415, "pid": 435, "tid": 435, "ts": 1776649790909924, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790909967, "dur": 1, + "args": { + "External id": 3548, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29433, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29433, "pid": 0, "tid": 7, "ts": 1776649790909967, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909958, "dur": 6, + "args": { + "External id": 3548, + "cbid": 211, "correlation": 29433 + } + }, + { + "ph": "s", "id": 29433, "pid": 435, "tid": 435, "ts": 1776649790909958, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909970, "dur": 0, + "args": { + "External id": 29439, + "cbid": 200, "correlation": 29439 + } + }, + { + "ph": "f", "id": 29439, "pid": 435, "tid": 435, "ts": 1776649790909970, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790909971, "dur": 0, + "args": { + "External id": 29440, + "cbid": 273, "correlation": 29440 + } + }, + { + "ph": "f", "id": 29440, "pid": 435, "tid": 435, "ts": 1776649790909971, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790909980, "dur": 15, + "args": { + "External id": 29441, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29441, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29441, "pid": 0, "tid": 7, "ts": 1776649790909980, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790909972, "dur": 4, + "args": { + "External id": 29441, + "cbid": 211, "correlation": 29441 + } + }, + { + "ph": "s", "id": 29441, "pid": 435, "tid": 435, "ts": 1776649790909972, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790910015, "dur": 1, + "args": { + "External id": 3553, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29459, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29459, "pid": 0, "tid": 7, "ts": 1776649790910015, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910006, "dur": 6, + "args": { + "External id": 3553, + "cbid": 211, "correlation": 29459 + } + }, + { + "ph": "s", "id": 29459, "pid": 435, "tid": 435, "ts": 1776649790910006, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910018, "dur": 0, + "args": { + "External id": 29465, + "cbid": 200, "correlation": 29465 + } + }, + { + "ph": "f", "id": 29465, "pid": 435, "tid": 435, "ts": 1776649790910018, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910019, "dur": 0, + "args": { + "External id": 29466, + "cbid": 273, "correlation": 29466 + } + }, + { + "ph": "f", "id": 29466, "pid": 435, "tid": 435, "ts": 1776649790910019, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790910028, "dur": 16, + "args": { + "External id": 29467, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29467, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29467, "pid": 0, "tid": 7, "ts": 1776649790910028, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910020, "dur": 4, + "args": { + "External id": 29467, + "cbid": 211, "correlation": 29467 + } + }, + { + "ph": "s", "id": 29467, "pid": 435, "tid": 435, "ts": 1776649790910020, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790910052, "dur": 4, + "args": { + "External id": 29475, + "cbid": 135, "correlation": 29475 + } + }, + { + "ph": "f", "id": 29475, "pid": 435, "tid": 435, "ts": 1776649790910052, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790910067, "dur": 5, + "args": { + "External id": 29480, + "cbid": 165, "correlation": 29480 + } + }, + { + "ph": "s", "id": 29480, "pid": 435, "tid": 435, "ts": 1776649790910067, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649790910077, "dur": 6, + "args": { + "External id": 29483, + "cbid": 139, "correlation": 29483 + } + }, + { + "ph": "f", "id": 29483, "pid": 435, "tid": 435, "ts": 1776649790910077, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790910140, "dur": 1, + "args": { + "External id": 3558, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29511, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29511, "pid": 0, "tid": 7, "ts": 1776649790910140, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910129, "dur": 8, + "args": { + "External id": 3558, + "cbid": 211, "correlation": 29511 + } + }, + { + "ph": "s", "id": 29511, "pid": 435, "tid": 435, "ts": 1776649790910129, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910144, "dur": 0, + "args": { + "External id": 29517, + "cbid": 200, "correlation": 29517 + } + }, + { + "ph": "f", "id": 29517, "pid": 435, "tid": 435, "ts": 1776649790910144, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910145, "dur": 1, + "args": { + "External id": 29518, + "cbid": 273, "correlation": 29518 + } + }, + { + "ph": "f", "id": 29518, "pid": 435, "tid": 435, "ts": 1776649790910145, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790910155, "dur": 15, + "args": { + "External id": 29519, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29519, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29519, "pid": 0, "tid": 7, "ts": 1776649790910155, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910146, "dur": 5, + "args": { + "External id": 29519, + "cbid": 211, "correlation": 29519 + } + }, + { + "ph": "s", "id": 29519, "pid": 435, "tid": 435, "ts": 1776649790910146, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790910192, "dur": 1, + "args": { + "External id": 3563, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29537, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29537, "pid": 0, "tid": 7, "ts": 1776649790910192, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910183, "dur": 6, + "args": { + "External id": 3563, + "cbid": 211, "correlation": 29537 + } + }, + { + "ph": "s", "id": 29537, "pid": 435, "tid": 435, "ts": 1776649790910183, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910196, "dur": 0, + "args": { + "External id": 29543, + "cbid": 200, "correlation": 29543 + } + }, + { + "ph": "f", "id": 29543, "pid": 435, "tid": 435, "ts": 1776649790910196, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910197, "dur": 0, + "args": { + "External id": 29544, + "cbid": 273, "correlation": 29544 + } + }, + { + "ph": "f", "id": 29544, "pid": 435, "tid": 435, "ts": 1776649790910197, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790910206, "dur": 16, + "args": { + "External id": 29545, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29545, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29545, "pid": 0, "tid": 7, "ts": 1776649790910206, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910198, "dur": 5, + "args": { + "External id": 29545, + "cbid": 211, "correlation": 29545 + } + }, + { + "ph": "s", "id": 29545, "pid": 435, "tid": 435, "ts": 1776649790910198, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790910242, "dur": 1, + "args": { + "External id": 3568, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29563, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29563, "pid": 0, "tid": 7, "ts": 1776649790910242, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910233, "dur": 6, + "args": { + "External id": 3568, + "cbid": 211, "correlation": 29563 + } + }, + { + "ph": "s", "id": 29563, "pid": 435, "tid": 435, "ts": 1776649790910233, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910246, "dur": 0, + "args": { + "External id": 29569, + "cbid": 200, "correlation": 29569 + } + }, + { + "ph": "f", "id": 29569, "pid": 435, "tid": 435, "ts": 1776649790910246, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910246, "dur": 0, + "args": { + "External id": 29570, + "cbid": 273, "correlation": 29570 + } + }, + { + "ph": "f", "id": 29570, "pid": 435, "tid": 435, "ts": 1776649790910246, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790910255, "dur": 15, + "args": { + "External id": 29571, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29571, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29571, "pid": 0, "tid": 7, "ts": 1776649790910255, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910247, "dur": 4, + "args": { + "External id": 29571, + "cbid": 211, "correlation": 29571 + } + }, + { + "ph": "s", "id": 29571, "pid": 435, "tid": 435, "ts": 1776649790910247, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790910291, "dur": 1, + "args": { + "External id": 3573, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29589, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29589, "pid": 0, "tid": 7, "ts": 1776649790910291, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910282, "dur": 6, + "args": { + "External id": 3573, + "cbid": 211, "correlation": 29589 + } + }, + { + "ph": "s", "id": 29589, "pid": 435, "tid": 435, "ts": 1776649790910282, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910295, "dur": 0, + "args": { + "External id": 29595, + "cbid": 200, "correlation": 29595 + } + }, + { + "ph": "f", "id": 29595, "pid": 435, "tid": 435, "ts": 1776649790910295, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910295, "dur": 0, + "args": { + "External id": 29596, + "cbid": 273, "correlation": 29596 + } + }, + { + "ph": "f", "id": 29596, "pid": 435, "tid": 435, "ts": 1776649790910295, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790910304, "dur": 17, + "args": { + "External id": 29597, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29597, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29597, "pid": 0, "tid": 7, "ts": 1776649790910304, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910296, "dur": 4, + "args": { + "External id": 29597, + "cbid": 211, "correlation": 29597 + } + }, + { + "ph": "s", "id": 29597, "pid": 435, "tid": 435, "ts": 1776649790910296, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790910340, "dur": 1, + "args": { + "External id": 3578, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29615, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29615, "pid": 0, "tid": 7, "ts": 1776649790910340, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910331, "dur": 6, + "args": { + "External id": 3578, + "cbid": 211, "correlation": 29615 + } + }, + { + "ph": "s", "id": 29615, "pid": 435, "tid": 435, "ts": 1776649790910331, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910343, "dur": 0, + "args": { + "External id": 29621, + "cbid": 200, "correlation": 29621 + } + }, + { + "ph": "f", "id": 29621, "pid": 435, "tid": 435, "ts": 1776649790910343, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910344, "dur": 0, + "args": { + "External id": 29622, + "cbid": 273, "correlation": 29622 + } + }, + { + "ph": "f", "id": 29622, "pid": 435, "tid": 435, "ts": 1776649790910344, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790910353, "dur": 16, + "args": { + "External id": 29623, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29623, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29623, "pid": 0, "tid": 7, "ts": 1776649790910353, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910345, "dur": 4, + "args": { + "External id": 29623, + "cbid": 211, "correlation": 29623 + } + }, + { + "ph": "s", "id": 29623, "pid": 435, "tid": 435, "ts": 1776649790910345, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790910388, "dur": 1, + "args": { + "External id": 3583, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29641, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29641, "pid": 0, "tid": 7, "ts": 1776649790910388, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910379, "dur": 6, + "args": { + "External id": 3583, + "cbid": 211, "correlation": 29641 + } + }, + { + "ph": "s", "id": 29641, "pid": 435, "tid": 435, "ts": 1776649790910379, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910391, "dur": 0, + "args": { + "External id": 29647, + "cbid": 200, "correlation": 29647 + } + }, + { + "ph": "f", "id": 29647, "pid": 435, "tid": 435, "ts": 1776649790910391, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910392, "dur": 0, + "args": { + "External id": 29648, + "cbid": 273, "correlation": 29648 + } + }, + { + "ph": "f", "id": 29648, "pid": 435, "tid": 435, "ts": 1776649790910392, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790910401, "dur": 16, + "args": { + "External id": 29649, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29649, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29649, "pid": 0, "tid": 7, "ts": 1776649790910401, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910393, "dur": 5, + "args": { + "External id": 29649, + "cbid": 211, "correlation": 29649 + } + }, + { + "ph": "s", "id": 29649, "pid": 435, "tid": 435, "ts": 1776649790910393, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790910446, "dur": 1, + "args": { + "External id": 3588, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29667, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29667, "pid": 0, "tid": 7, "ts": 1776649790910446, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910437, "dur": 6, + "args": { + "External id": 3588, + "cbid": 211, "correlation": 29667 + } + }, + { + "ph": "s", "id": 29667, "pid": 435, "tid": 435, "ts": 1776649790910437, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910450, "dur": 0, + "args": { + "External id": 29673, + "cbid": 200, "correlation": 29673 + } + }, + { + "ph": "f", "id": 29673, "pid": 435, "tid": 435, "ts": 1776649790910450, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910450, "dur": 0, + "args": { + "External id": 29674, + "cbid": 273, "correlation": 29674 + } + }, + { + "ph": "f", "id": 29674, "pid": 435, "tid": 435, "ts": 1776649790910450, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790910460, "dur": 15, + "args": { + "External id": 29675, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29675, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29675, "pid": 0, "tid": 7, "ts": 1776649790910460, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910452, "dur": 5, + "args": { + "External id": 29675, + "cbid": 211, "correlation": 29675 + } + }, + { + "ph": "s", "id": 29675, "pid": 435, "tid": 435, "ts": 1776649790910452, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790910496, "dur": 1, + "args": { + "External id": 3593, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29693, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29693, "pid": 0, "tid": 7, "ts": 1776649790910496, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910487, "dur": 6, + "args": { + "External id": 3593, + "cbid": 211, "correlation": 29693 + } + }, + { + "ph": "s", "id": 29693, "pid": 435, "tid": 435, "ts": 1776649790910487, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910499, "dur": 0, + "args": { + "External id": 29699, + "cbid": 200, "correlation": 29699 + } + }, + { + "ph": "f", "id": 29699, "pid": 435, "tid": 435, "ts": 1776649790910499, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910500, "dur": 0, + "args": { + "External id": 29700, + "cbid": 273, "correlation": 29700 + } + }, + { + "ph": "f", "id": 29700, "pid": 435, "tid": 435, "ts": 1776649790910500, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790910508, "dur": 15, + "args": { + "External id": 29701, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29701, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29701, "pid": 0, "tid": 7, "ts": 1776649790910508, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910501, "dur": 4, + "args": { + "External id": 29701, + "cbid": 211, "correlation": 29701 + } + }, + { + "ph": "s", "id": 29701, "pid": 435, "tid": 435, "ts": 1776649790910501, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790910544, "dur": 1, + "args": { + "External id": 3598, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29719, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29719, "pid": 0, "tid": 7, "ts": 1776649790910544, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910534, "dur": 6, + "args": { + "External id": 3598, + "cbid": 211, "correlation": 29719 + } + }, + { + "ph": "s", "id": 29719, "pid": 435, "tid": 435, "ts": 1776649790910534, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910547, "dur": 0, + "args": { + "External id": 29725, + "cbid": 200, "correlation": 29725 + } + }, + { + "ph": "f", "id": 29725, "pid": 435, "tid": 435, "ts": 1776649790910547, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910547, "dur": 0, + "args": { + "External id": 29726, + "cbid": 273, "correlation": 29726 + } + }, + { + "ph": "f", "id": 29726, "pid": 435, "tid": 435, "ts": 1776649790910547, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790910557, "dur": 16, + "args": { + "External id": 29727, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29727, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29727, "pid": 0, "tid": 7, "ts": 1776649790910557, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910549, "dur": 5, + "args": { + "External id": 29727, + "cbid": 211, "correlation": 29727 + } + }, + { + "ph": "s", "id": 29727, "pid": 435, "tid": 435, "ts": 1776649790910549, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790910592, "dur": 1, + "args": { + "External id": 3603, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29745, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29745, "pid": 0, "tid": 7, "ts": 1776649790910592, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910583, "dur": 6, + "args": { + "External id": 3603, + "cbid": 211, "correlation": 29745 + } + }, + { + "ph": "s", "id": 29745, "pid": 435, "tid": 435, "ts": 1776649790910583, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910595, "dur": 0, + "args": { + "External id": 29751, + "cbid": 200, "correlation": 29751 + } + }, + { + "ph": "f", "id": 29751, "pid": 435, "tid": 435, "ts": 1776649790910595, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910596, "dur": 0, + "args": { + "External id": 29752, + "cbid": 273, "correlation": 29752 + } + }, + { + "ph": "f", "id": 29752, "pid": 435, "tid": 435, "ts": 1776649790910596, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790910605, "dur": 16, + "args": { + "External id": 29753, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29753, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29753, "pid": 0, "tid": 7, "ts": 1776649790910605, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910597, "dur": 4, + "args": { + "External id": 29753, + "cbid": 211, "correlation": 29753 + } + }, + { + "ph": "s", "id": 29753, "pid": 435, "tid": 435, "ts": 1776649790910597, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790910619, "dur": 6, + "args": { + "External id": 29756, + "cbid": 165, "correlation": 29756 + } + }, + { + "ph": "s", "id": 29756, "pid": 435, "tid": 435, "ts": 1776649790910619, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790910647, "dur": 3, + "args": { + "External id": 29765, + "cbid": 135, "correlation": 29765 + } + }, + { + "ph": "f", "id": 29765, "pid": 435, "tid": 435, "ts": 1776649790910647, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790910696, "dur": 1, + "args": { + "External id": 3608, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29785, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29785, "pid": 0, "tid": 7, "ts": 1776649790910696, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910685, "dur": 8, + "args": { + "External id": 3608, + "cbid": 211, "correlation": 29785 + } + }, + { + "ph": "s", "id": 29785, "pid": 435, "tid": 435, "ts": 1776649790910685, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910700, "dur": 0, + "args": { + "External id": 29791, + "cbid": 200, "correlation": 29791 + } + }, + { + "ph": "f", "id": 29791, "pid": 435, "tid": 435, "ts": 1776649790910700, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910701, "dur": 1, + "args": { + "External id": 29792, + "cbid": 273, "correlation": 29792 + } + }, + { + "ph": "f", "id": 29792, "pid": 435, "tid": 435, "ts": 1776649790910701, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790910711, "dur": 15, + "args": { + "External id": 29793, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29793, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29793, "pid": 0, "tid": 7, "ts": 1776649790910711, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910703, "dur": 5, + "args": { + "External id": 29793, + "cbid": 211, "correlation": 29793 + } + }, + { + "ph": "s", "id": 29793, "pid": 435, "tid": 435, "ts": 1776649790910703, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790910749, "dur": 1, + "args": { + "External id": 3613, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29811, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29811, "pid": 0, "tid": 7, "ts": 1776649790910749, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910740, "dur": 7, + "args": { + "External id": 3613, + "cbid": 211, "correlation": 29811 + } + }, + { + "ph": "s", "id": 29811, "pid": 435, "tid": 435, "ts": 1776649790910740, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910753, "dur": 0, + "args": { + "External id": 29817, + "cbid": 200, "correlation": 29817 + } + }, + { + "ph": "f", "id": 29817, "pid": 435, "tid": 435, "ts": 1776649790910753, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910753, "dur": 0, + "args": { + "External id": 29818, + "cbid": 273, "correlation": 29818 + } + }, + { + "ph": "f", "id": 29818, "pid": 435, "tid": 435, "ts": 1776649790910753, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790910763, "dur": 17, + "args": { + "External id": 29819, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29819, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29819, "pid": 0, "tid": 7, "ts": 1776649790910763, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910755, "dur": 5, + "args": { + "External id": 29819, + "cbid": 211, "correlation": 29819 + } + }, + { + "ph": "s", "id": 29819, "pid": 435, "tid": 435, "ts": 1776649790910755, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790910799, "dur": 1, + "args": { + "External id": 3618, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29837, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29837, "pid": 0, "tid": 7, "ts": 1776649790910799, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910789, "dur": 6, + "args": { + "External id": 3618, + "cbid": 211, "correlation": 29837 + } + }, + { + "ph": "s", "id": 29837, "pid": 435, "tid": 435, "ts": 1776649790910789, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910802, "dur": 0, + "args": { + "External id": 29843, + "cbid": 200, "correlation": 29843 + } + }, + { + "ph": "f", "id": 29843, "pid": 435, "tid": 435, "ts": 1776649790910802, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910803, "dur": 0, + "args": { + "External id": 29844, + "cbid": 273, "correlation": 29844 + } + }, + { + "ph": "f", "id": 29844, "pid": 435, "tid": 435, "ts": 1776649790910803, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790910812, "dur": 16, + "args": { + "External id": 29845, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29845, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29845, "pid": 0, "tid": 7, "ts": 1776649790910812, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910804, "dur": 5, + "args": { + "External id": 29845, + "cbid": 211, "correlation": 29845 + } + }, + { + "ph": "s", "id": 29845, "pid": 435, "tid": 435, "ts": 1776649790910804, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790910847, "dur": 1, + "args": { + "External id": 3623, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29863, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29863, "pid": 0, "tid": 7, "ts": 1776649790910847, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910838, "dur": 6, + "args": { + "External id": 3623, + "cbid": 211, "correlation": 29863 + } + }, + { + "ph": "s", "id": 29863, "pid": 435, "tid": 435, "ts": 1776649790910838, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910850, "dur": 0, + "args": { + "External id": 29869, + "cbid": 200, "correlation": 29869 + } + }, + { + "ph": "f", "id": 29869, "pid": 435, "tid": 435, "ts": 1776649790910850, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910851, "dur": 0, + "args": { + "External id": 29870, + "cbid": 273, "correlation": 29870 + } + }, + { + "ph": "f", "id": 29870, "pid": 435, "tid": 435, "ts": 1776649790910851, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790910860, "dur": 15, + "args": { + "External id": 29871, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29871, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29871, "pid": 0, "tid": 7, "ts": 1776649790910860, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910852, "dur": 5, + "args": { + "External id": 29871, + "cbid": 211, "correlation": 29871 + } + }, + { + "ph": "s", "id": 29871, "pid": 435, "tid": 435, "ts": 1776649790910852, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790910896, "dur": 1, + "args": { + "External id": 3628, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29889, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29889, "pid": 0, "tid": 7, "ts": 1776649790910896, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910887, "dur": 6, + "args": { + "External id": 3628, + "cbid": 211, "correlation": 29889 + } + }, + { + "ph": "s", "id": 29889, "pid": 435, "tid": 435, "ts": 1776649790910887, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910899, "dur": 0, + "args": { + "External id": 29895, + "cbid": 200, "correlation": 29895 + } + }, + { + "ph": "f", "id": 29895, "pid": 435, "tid": 435, "ts": 1776649790910899, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910900, "dur": 0, + "args": { + "External id": 29896, + "cbid": 273, "correlation": 29896 + } + }, + { + "ph": "f", "id": 29896, "pid": 435, "tid": 435, "ts": 1776649790910900, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790910909, "dur": 16, + "args": { + "External id": 29897, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29897, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29897, "pid": 0, "tid": 7, "ts": 1776649790910909, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910901, "dur": 5, + "args": { + "External id": 29897, + "cbid": 211, "correlation": 29897 + } + }, + { + "ph": "s", "id": 29897, "pid": 435, "tid": 435, "ts": 1776649790910901, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790910944, "dur": 1, + "args": { + "External id": 3633, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29915, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29915, "pid": 0, "tid": 7, "ts": 1776649790910944, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910935, "dur": 6, + "args": { + "External id": 3633, + "cbid": 211, "correlation": 29915 + } + }, + { + "ph": "s", "id": 29915, "pid": 435, "tid": 435, "ts": 1776649790910935, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910948, "dur": 0, + "args": { + "External id": 29921, + "cbid": 200, "correlation": 29921 + } + }, + { + "ph": "f", "id": 29921, "pid": 435, "tid": 435, "ts": 1776649790910948, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910948, "dur": 0, + "args": { + "External id": 29922, + "cbid": 273, "correlation": 29922 + } + }, + { + "ph": "f", "id": 29922, "pid": 435, "tid": 435, "ts": 1776649790910948, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790910957, "dur": 16, + "args": { + "External id": 29923, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29923, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29923, "pid": 0, "tid": 7, "ts": 1776649790910957, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910950, "dur": 4, + "args": { + "External id": 29923, + "cbid": 211, "correlation": 29923 + } + }, + { + "ph": "s", "id": 29923, "pid": 435, "tid": 435, "ts": 1776649790910950, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790910993, "dur": 1, + "args": { + "External id": 3638, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29941, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29941, "pid": 0, "tid": 7, "ts": 1776649790910993, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910983, "dur": 6, + "args": { + "External id": 3638, + "cbid": 211, "correlation": 29941 + } + }, + { + "ph": "s", "id": 29941, "pid": 435, "tid": 435, "ts": 1776649790910983, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910996, "dur": 0, + "args": { + "External id": 29947, + "cbid": 200, "correlation": 29947 + } + }, + { + "ph": "f", "id": 29947, "pid": 435, "tid": 435, "ts": 1776649790910996, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790910997, "dur": 0, + "args": { + "External id": 29948, + "cbid": 273, "correlation": 29948 + } + }, + { + "ph": "f", "id": 29948, "pid": 435, "tid": 435, "ts": 1776649790910997, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790911006, "dur": 15, + "args": { + "External id": 29949, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29949, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29949, "pid": 0, "tid": 7, "ts": 1776649790911006, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790910998, "dur": 5, + "args": { + "External id": 29949, + "cbid": 211, "correlation": 29949 + } + }, + { + "ph": "s", "id": 29949, "pid": 435, "tid": 435, "ts": 1776649790910998, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790911042, "dur": 1, + "args": { + "External id": 3643, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29967, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29967, "pid": 0, "tid": 7, "ts": 1776649790911042, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911033, "dur": 6, + "args": { + "External id": 3643, + "cbid": 211, "correlation": 29967 + } + }, + { + "ph": "s", "id": 29967, "pid": 435, "tid": 435, "ts": 1776649790911033, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911045, "dur": 0, + "args": { + "External id": 29973, + "cbid": 200, "correlation": 29973 + } + }, + { + "ph": "f", "id": 29973, "pid": 435, "tid": 435, "ts": 1776649790911045, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911046, "dur": 0, + "args": { + "External id": 29974, + "cbid": 273, "correlation": 29974 + } + }, + { + "ph": "f", "id": 29974, "pid": 435, "tid": 435, "ts": 1776649790911046, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790911055, "dur": 16, + "args": { + "External id": 29975, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29975, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29975, "pid": 0, "tid": 7, "ts": 1776649790911055, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911047, "dur": 5, + "args": { + "External id": 29975, + "cbid": 211, "correlation": 29975 + } + }, + { + "ph": "s", "id": 29975, "pid": 435, "tid": 435, "ts": 1776649790911047, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790911091, "dur": 1, + "args": { + "External id": 3648, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 29993, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 29993, "pid": 0, "tid": 7, "ts": 1776649790911091, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911082, "dur": 6, + "args": { + "External id": 3648, + "cbid": 211, "correlation": 29993 + } + }, + { + "ph": "s", "id": 29993, "pid": 435, "tid": 435, "ts": 1776649790911082, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911094, "dur": 0, + "args": { + "External id": 29999, + "cbid": 200, "correlation": 29999 + } + }, + { + "ph": "f", "id": 29999, "pid": 435, "tid": 435, "ts": 1776649790911094, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911095, "dur": 0, + "args": { + "External id": 30000, + "cbid": 273, "correlation": 30000 + } + }, + { + "ph": "f", "id": 30000, "pid": 435, "tid": 435, "ts": 1776649790911095, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790911104, "dur": 15, + "args": { + "External id": 30001, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30001, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30001, "pid": 0, "tid": 7, "ts": 1776649790911104, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911096, "dur": 4, + "args": { + "External id": 30001, + "cbid": 211, "correlation": 30001 + } + }, + { + "ph": "s", "id": 30001, "pid": 435, "tid": 435, "ts": 1776649790911096, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790911139, "dur": 1, + "args": { + "External id": 3653, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30019, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30019, "pid": 0, "tid": 7, "ts": 1776649790911139, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911130, "dur": 6, + "args": { + "External id": 3653, + "cbid": 211, "correlation": 30019 + } + }, + { + "ph": "s", "id": 30019, "pid": 435, "tid": 435, "ts": 1776649790911130, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911142, "dur": 0, + "args": { + "External id": 30025, + "cbid": 200, "correlation": 30025 + } + }, + { + "ph": "f", "id": 30025, "pid": 435, "tid": 435, "ts": 1776649790911142, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911143, "dur": 0, + "args": { + "External id": 30026, + "cbid": 273, "correlation": 30026 + } + }, + { + "ph": "f", "id": 30026, "pid": 435, "tid": 435, "ts": 1776649790911143, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790911152, "dur": 16, + "args": { + "External id": 30027, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30027, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30027, "pid": 0, "tid": 7, "ts": 1776649790911152, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911144, "dur": 4, + "args": { + "External id": 30027, + "cbid": 211, "correlation": 30027 + } + }, + { + "ph": "s", "id": 30027, "pid": 435, "tid": 435, "ts": 1776649790911144, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790911187, "dur": 1, + "args": { + "External id": 3658, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30045, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30045, "pid": 0, "tid": 7, "ts": 1776649790911187, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911178, "dur": 6, + "args": { + "External id": 3658, + "cbid": 211, "correlation": 30045 + } + }, + { + "ph": "s", "id": 30045, "pid": 435, "tid": 435, "ts": 1776649790911178, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911190, "dur": 0, + "args": { + "External id": 30051, + "cbid": 200, "correlation": 30051 + } + }, + { + "ph": "f", "id": 30051, "pid": 435, "tid": 435, "ts": 1776649790911190, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911191, "dur": 0, + "args": { + "External id": 30052, + "cbid": 273, "correlation": 30052 + } + }, + { + "ph": "f", "id": 30052, "pid": 435, "tid": 435, "ts": 1776649790911191, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790911200, "dur": 16, + "args": { + "External id": 30053, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30053, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30053, "pid": 0, "tid": 7, "ts": 1776649790911200, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911192, "dur": 5, + "args": { + "External id": 30053, + "cbid": 211, "correlation": 30053 + } + }, + { + "ph": "s", "id": 30053, "pid": 435, "tid": 435, "ts": 1776649790911192, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790911236, "dur": 1, + "args": { + "External id": 3663, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30071, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30071, "pid": 0, "tid": 7, "ts": 1776649790911236, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911227, "dur": 6, + "args": { + "External id": 3663, + "cbid": 211, "correlation": 30071 + } + }, + { + "ph": "s", "id": 30071, "pid": 435, "tid": 435, "ts": 1776649790911227, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911239, "dur": 0, + "args": { + "External id": 30077, + "cbid": 200, "correlation": 30077 + } + }, + { + "ph": "f", "id": 30077, "pid": 435, "tid": 435, "ts": 1776649790911239, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911240, "dur": 0, + "args": { + "External id": 30078, + "cbid": 273, "correlation": 30078 + } + }, + { + "ph": "f", "id": 30078, "pid": 435, "tid": 435, "ts": 1776649790911240, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790911249, "dur": 16, + "args": { + "External id": 30079, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30079, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30079, "pid": 0, "tid": 7, "ts": 1776649790911249, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911241, "dur": 5, + "args": { + "External id": 30079, + "cbid": 211, "correlation": 30079 + } + }, + { + "ph": "s", "id": 30079, "pid": 435, "tid": 435, "ts": 1776649790911241, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790911285, "dur": 1, + "args": { + "External id": 3668, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30097, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30097, "pid": 0, "tid": 7, "ts": 1776649790911285, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911276, "dur": 6, + "args": { + "External id": 3668, + "cbid": 211, "correlation": 30097 + } + }, + { + "ph": "s", "id": 30097, "pid": 435, "tid": 435, "ts": 1776649790911276, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911288, "dur": 0, + "args": { + "External id": 30103, + "cbid": 200, "correlation": 30103 + } + }, + { + "ph": "f", "id": 30103, "pid": 435, "tid": 435, "ts": 1776649790911288, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911289, "dur": 0, + "args": { + "External id": 30104, + "cbid": 273, "correlation": 30104 + } + }, + { + "ph": "f", "id": 30104, "pid": 435, "tid": 435, "ts": 1776649790911289, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790911298, "dur": 15, + "args": { + "External id": 30105, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30105, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30105, "pid": 0, "tid": 7, "ts": 1776649790911298, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911290, "dur": 5, + "args": { + "External id": 30105, + "cbid": 211, "correlation": 30105 + } + }, + { + "ph": "s", "id": 30105, "pid": 435, "tid": 435, "ts": 1776649790911290, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790911334, "dur": 1, + "args": { + "External id": 3673, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30123, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30123, "pid": 0, "tid": 7, "ts": 1776649790911334, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911325, "dur": 6, + "args": { + "External id": 3673, + "cbid": 211, "correlation": 30123 + } + }, + { + "ph": "s", "id": 30123, "pid": 435, "tid": 435, "ts": 1776649790911325, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911337, "dur": 0, + "args": { + "External id": 30129, + "cbid": 200, "correlation": 30129 + } + }, + { + "ph": "f", "id": 30129, "pid": 435, "tid": 435, "ts": 1776649790911337, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911338, "dur": 0, + "args": { + "External id": 30130, + "cbid": 273, "correlation": 30130 + } + }, + { + "ph": "f", "id": 30130, "pid": 435, "tid": 435, "ts": 1776649790911338, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790911346, "dur": 16, + "args": { + "External id": 30131, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30131, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30131, "pid": 0, "tid": 7, "ts": 1776649790911346, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911339, "dur": 4, + "args": { + "External id": 30131, + "cbid": 211, "correlation": 30131 + } + }, + { + "ph": "s", "id": 30131, "pid": 435, "tid": 435, "ts": 1776649790911339, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790911382, "dur": 1, + "args": { + "External id": 3678, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30149, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30149, "pid": 0, "tid": 7, "ts": 1776649790911382, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911373, "dur": 7, + "args": { + "External id": 3678, + "cbid": 211, "correlation": 30149 + } + }, + { + "ph": "s", "id": 30149, "pid": 435, "tid": 435, "ts": 1776649790911373, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911385, "dur": 0, + "args": { + "External id": 30155, + "cbid": 200, "correlation": 30155 + } + }, + { + "ph": "f", "id": 30155, "pid": 435, "tid": 435, "ts": 1776649790911385, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911386, "dur": 0, + "args": { + "External id": 30156, + "cbid": 273, "correlation": 30156 + } + }, + { + "ph": "f", "id": 30156, "pid": 435, "tid": 435, "ts": 1776649790911386, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790911395, "dur": 16, + "args": { + "External id": 30157, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30157, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30157, "pid": 0, "tid": 7, "ts": 1776649790911395, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911387, "dur": 5, + "args": { + "External id": 30157, + "cbid": 211, "correlation": 30157 + } + }, + { + "ph": "s", "id": 30157, "pid": 435, "tid": 435, "ts": 1776649790911387, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790911431, "dur": 1, + "args": { + "External id": 3683, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30175, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30175, "pid": 0, "tid": 7, "ts": 1776649790911431, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911421, "dur": 6, + "args": { + "External id": 3683, + "cbid": 211, "correlation": 30175 + } + }, + { + "ph": "s", "id": 30175, "pid": 435, "tid": 435, "ts": 1776649790911421, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911434, "dur": 0, + "args": { + "External id": 30181, + "cbid": 200, "correlation": 30181 + } + }, + { + "ph": "f", "id": 30181, "pid": 435, "tid": 435, "ts": 1776649790911434, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911434, "dur": 0, + "args": { + "External id": 30182, + "cbid": 273, "correlation": 30182 + } + }, + { + "ph": "f", "id": 30182, "pid": 435, "tid": 435, "ts": 1776649790911434, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790911443, "dur": 16, + "args": { + "External id": 30183, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30183, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30183, "pid": 0, "tid": 7, "ts": 1776649790911443, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911435, "dur": 4, + "args": { + "External id": 30183, + "cbid": 211, "correlation": 30183 + } + }, + { + "ph": "s", "id": 30183, "pid": 435, "tid": 435, "ts": 1776649790911435, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790911479, "dur": 1, + "args": { + "External id": 3688, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30201, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30201, "pid": 0, "tid": 7, "ts": 1776649790911479, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911470, "dur": 6, + "args": { + "External id": 3688, + "cbid": 211, "correlation": 30201 + } + }, + { + "ph": "s", "id": 30201, "pid": 435, "tid": 435, "ts": 1776649790911470, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911482, "dur": 0, + "args": { + "External id": 30207, + "cbid": 200, "correlation": 30207 + } + }, + { + "ph": "f", "id": 30207, "pid": 435, "tid": 435, "ts": 1776649790911482, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911483, "dur": 0, + "args": { + "External id": 30208, + "cbid": 273, "correlation": 30208 + } + }, + { + "ph": "f", "id": 30208, "pid": 435, "tid": 435, "ts": 1776649790911483, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790911492, "dur": 16, + "args": { + "External id": 30209, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30209, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30209, "pid": 0, "tid": 7, "ts": 1776649790911492, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911484, "dur": 5, + "args": { + "External id": 30209, + "cbid": 211, "correlation": 30209 + } + }, + { + "ph": "s", "id": 30209, "pid": 435, "tid": 435, "ts": 1776649790911484, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790911527, "dur": 1, + "args": { + "External id": 3693, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30227, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30227, "pid": 0, "tid": 7, "ts": 1776649790911527, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911518, "dur": 6, + "args": { + "External id": 3693, + "cbid": 211, "correlation": 30227 + } + }, + { + "ph": "s", "id": 30227, "pid": 435, "tid": 435, "ts": 1776649790911518, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911531, "dur": 0, + "args": { + "External id": 30233, + "cbid": 200, "correlation": 30233 + } + }, + { + "ph": "f", "id": 30233, "pid": 435, "tid": 435, "ts": 1776649790911531, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911531, "dur": 0, + "args": { + "External id": 30234, + "cbid": 273, "correlation": 30234 + } + }, + { + "ph": "f", "id": 30234, "pid": 435, "tid": 435, "ts": 1776649790911531, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790911540, "dur": 15, + "args": { + "External id": 30235, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30235, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30235, "pid": 0, "tid": 7, "ts": 1776649790911540, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911532, "dur": 4, + "args": { + "External id": 30235, + "cbid": 211, "correlation": 30235 + } + }, + { + "ph": "s", "id": 30235, "pid": 435, "tid": 435, "ts": 1776649790911532, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790911576, "dur": 1, + "args": { + "External id": 3698, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30253, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30253, "pid": 0, "tid": 7, "ts": 1776649790911576, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911567, "dur": 7, + "args": { + "External id": 3698, + "cbid": 211, "correlation": 30253 + } + }, + { + "ph": "s", "id": 30253, "pid": 435, "tid": 435, "ts": 1776649790911567, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911579, "dur": 0, + "args": { + "External id": 30259, + "cbid": 200, "correlation": 30259 + } + }, + { + "ph": "f", "id": 30259, "pid": 435, "tid": 435, "ts": 1776649790911579, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911580, "dur": 0, + "args": { + "External id": 30260, + "cbid": 273, "correlation": 30260 + } + }, + { + "ph": "f", "id": 30260, "pid": 435, "tid": 435, "ts": 1776649790911580, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790911589, "dur": 15, + "args": { + "External id": 30261, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30261, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30261, "pid": 0, "tid": 7, "ts": 1776649790911589, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911581, "dur": 4, + "args": { + "External id": 30261, + "cbid": 211, "correlation": 30261 + } + }, + { + "ph": "s", "id": 30261, "pid": 435, "tid": 435, "ts": 1776649790911581, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790911624, "dur": 1, + "args": { + "External id": 3703, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30279, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30279, "pid": 0, "tid": 7, "ts": 1776649790911624, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911615, "dur": 6, + "args": { + "External id": 3703, + "cbid": 211, "correlation": 30279 + } + }, + { + "ph": "s", "id": 30279, "pid": 435, "tid": 435, "ts": 1776649790911615, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911627, "dur": 0, + "args": { + "External id": 30285, + "cbid": 200, "correlation": 30285 + } + }, + { + "ph": "f", "id": 30285, "pid": 435, "tid": 435, "ts": 1776649790911627, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911628, "dur": 0, + "args": { + "External id": 30286, + "cbid": 273, "correlation": 30286 + } + }, + { + "ph": "f", "id": 30286, "pid": 435, "tid": 435, "ts": 1776649790911628, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790911637, "dur": 17, + "args": { + "External id": 30287, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30287, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30287, "pid": 0, "tid": 7, "ts": 1776649790911637, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911629, "dur": 4, + "args": { + "External id": 30287, + "cbid": 211, "correlation": 30287 + } + }, + { + "ph": "s", "id": 30287, "pid": 435, "tid": 435, "ts": 1776649790911629, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790911672, "dur": 1, + "args": { + "External id": 3708, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30305, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30305, "pid": 0, "tid": 7, "ts": 1776649790911672, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911663, "dur": 6, + "args": { + "External id": 3708, + "cbid": 211, "correlation": 30305 + } + }, + { + "ph": "s", "id": 30305, "pid": 435, "tid": 435, "ts": 1776649790911663, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911675, "dur": 0, + "args": { + "External id": 30311, + "cbid": 200, "correlation": 30311 + } + }, + { + "ph": "f", "id": 30311, "pid": 435, "tid": 435, "ts": 1776649790911675, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911676, "dur": 0, + "args": { + "External id": 30312, + "cbid": 273, "correlation": 30312 + } + }, + { + "ph": "f", "id": 30312, "pid": 435, "tid": 435, "ts": 1776649790911676, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790911685, "dur": 16, + "args": { + "External id": 30313, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30313, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30313, "pid": 0, "tid": 7, "ts": 1776649790911685, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911677, "dur": 4, + "args": { + "External id": 30313, + "cbid": 211, "correlation": 30313 + } + }, + { + "ph": "s", "id": 30313, "pid": 435, "tid": 435, "ts": 1776649790911677, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790911720, "dur": 1, + "args": { + "External id": 3713, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30331, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30331, "pid": 0, "tid": 7, "ts": 1776649790911720, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911711, "dur": 6, + "args": { + "External id": 3713, + "cbid": 211, "correlation": 30331 + } + }, + { + "ph": "s", "id": 30331, "pid": 435, "tid": 435, "ts": 1776649790911711, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911723, "dur": 0, + "args": { + "External id": 30337, + "cbid": 200, "correlation": 30337 + } + }, + { + "ph": "f", "id": 30337, "pid": 435, "tid": 435, "ts": 1776649790911723, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911724, "dur": 0, + "args": { + "External id": 30338, + "cbid": 273, "correlation": 30338 + } + }, + { + "ph": "f", "id": 30338, "pid": 435, "tid": 435, "ts": 1776649790911724, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790911733, "dur": 15, + "args": { + "External id": 30339, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30339, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30339, "pid": 0, "tid": 7, "ts": 1776649790911733, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911725, "dur": 5, + "args": { + "External id": 30339, + "cbid": 211, "correlation": 30339 + } + }, + { + "ph": "s", "id": 30339, "pid": 435, "tid": 435, "ts": 1776649790911725, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790911769, "dur": 1, + "args": { + "External id": 3718, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30357, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30357, "pid": 0, "tid": 7, "ts": 1776649790911769, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911759, "dur": 6, + "args": { + "External id": 3718, + "cbid": 211, "correlation": 30357 + } + }, + { + "ph": "s", "id": 30357, "pid": 435, "tid": 435, "ts": 1776649790911759, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911771, "dur": 0, + "args": { + "External id": 30363, + "cbid": 200, "correlation": 30363 + } + }, + { + "ph": "f", "id": 30363, "pid": 435, "tid": 435, "ts": 1776649790911771, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911772, "dur": 0, + "args": { + "External id": 30364, + "cbid": 273, "correlation": 30364 + } + }, + { + "ph": "f", "id": 30364, "pid": 435, "tid": 435, "ts": 1776649790911772, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790911781, "dur": 16, + "args": { + "External id": 30365, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30365, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30365, "pid": 0, "tid": 7, "ts": 1776649790911781, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911773, "dur": 5, + "args": { + "External id": 30365, + "cbid": 211, "correlation": 30365 + } + }, + { + "ph": "s", "id": 30365, "pid": 435, "tid": 435, "ts": 1776649790911773, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790911817, "dur": 1, + "args": { + "External id": 3723, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30383, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30383, "pid": 0, "tid": 7, "ts": 1776649790911817, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911807, "dur": 6, + "args": { + "External id": 3723, + "cbid": 211, "correlation": 30383 + } + }, + { + "ph": "s", "id": 30383, "pid": 435, "tid": 435, "ts": 1776649790911807, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911820, "dur": 0, + "args": { + "External id": 30389, + "cbid": 200, "correlation": 30389 + } + }, + { + "ph": "f", "id": 30389, "pid": 435, "tid": 435, "ts": 1776649790911820, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911820, "dur": 0, + "args": { + "External id": 30390, + "cbid": 273, "correlation": 30390 + } + }, + { + "ph": "f", "id": 30390, "pid": 435, "tid": 435, "ts": 1776649790911820, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790911830, "dur": 16, + "args": { + "External id": 30391, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30391, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30391, "pid": 0, "tid": 7, "ts": 1776649790911830, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911822, "dur": 5, + "args": { + "External id": 30391, + "cbid": 211, "correlation": 30391 + } + }, + { + "ph": "s", "id": 30391, "pid": 435, "tid": 435, "ts": 1776649790911822, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790911865, "dur": 1, + "args": { + "External id": 3728, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30409, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30409, "pid": 0, "tid": 7, "ts": 1776649790911865, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911856, "dur": 6, + "args": { + "External id": 3728, + "cbid": 211, "correlation": 30409 + } + }, + { + "ph": "s", "id": 30409, "pid": 435, "tid": 435, "ts": 1776649790911856, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911868, "dur": 0, + "args": { + "External id": 30415, + "cbid": 200, "correlation": 30415 + } + }, + { + "ph": "f", "id": 30415, "pid": 435, "tid": 435, "ts": 1776649790911868, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911869, "dur": 0, + "args": { + "External id": 30416, + "cbid": 273, "correlation": 30416 + } + }, + { + "ph": "f", "id": 30416, "pid": 435, "tid": 435, "ts": 1776649790911869, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790911878, "dur": 15, + "args": { + "External id": 30417, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30417, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30417, "pid": 0, "tid": 7, "ts": 1776649790911878, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911870, "dur": 5, + "args": { + "External id": 30417, + "cbid": 211, "correlation": 30417 + } + }, + { + "ph": "s", "id": 30417, "pid": 435, "tid": 435, "ts": 1776649790911870, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790911914, "dur": 1, + "args": { + "External id": 3733, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30435, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30435, "pid": 0, "tid": 7, "ts": 1776649790911914, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911905, "dur": 6, + "args": { + "External id": 3733, + "cbid": 211, "correlation": 30435 + } + }, + { + "ph": "s", "id": 30435, "pid": 435, "tid": 435, "ts": 1776649790911905, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911917, "dur": 0, + "args": { + "External id": 30441, + "cbid": 200, "correlation": 30441 + } + }, + { + "ph": "f", "id": 30441, "pid": 435, "tid": 435, "ts": 1776649790911917, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911918, "dur": 0, + "args": { + "External id": 30442, + "cbid": 273, "correlation": 30442 + } + }, + { + "ph": "f", "id": 30442, "pid": 435, "tid": 435, "ts": 1776649790911918, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790911926, "dur": 16, + "args": { + "External id": 30443, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30443, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30443, "pid": 0, "tid": 7, "ts": 1776649790911926, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911919, "dur": 4, + "args": { + "External id": 30443, + "cbid": 211, "correlation": 30443 + } + }, + { + "ph": "s", "id": 30443, "pid": 435, "tid": 435, "ts": 1776649790911919, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790911962, "dur": 1, + "args": { + "External id": 3738, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30461, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30461, "pid": 0, "tid": 7, "ts": 1776649790911962, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911953, "dur": 6, + "args": { + "External id": 3738, + "cbid": 211, "correlation": 30461 + } + }, + { + "ph": "s", "id": 30461, "pid": 435, "tid": 435, "ts": 1776649790911953, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911965, "dur": 0, + "args": { + "External id": 30467, + "cbid": 200, "correlation": 30467 + } + }, + { + "ph": "f", "id": 30467, "pid": 435, "tid": 435, "ts": 1776649790911965, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790911966, "dur": 0, + "args": { + "External id": 30468, + "cbid": 273, "correlation": 30468 + } + }, + { + "ph": "f", "id": 30468, "pid": 435, "tid": 435, "ts": 1776649790911966, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790911975, "dur": 16, + "args": { + "External id": 30469, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30469, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30469, "pid": 0, "tid": 7, "ts": 1776649790911975, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790911967, "dur": 5, + "args": { + "External id": 30469, + "cbid": 211, "correlation": 30469 + } + }, + { + "ph": "s", "id": 30469, "pid": 435, "tid": 435, "ts": 1776649790911967, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790912010, "dur": 1, + "args": { + "External id": 3743, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30487, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30487, "pid": 0, "tid": 7, "ts": 1776649790912010, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912001, "dur": 6, + "args": { + "External id": 3743, + "cbid": 211, "correlation": 30487 + } + }, + { + "ph": "s", "id": 30487, "pid": 435, "tid": 435, "ts": 1776649790912001, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912013, "dur": 0, + "args": { + "External id": 30493, + "cbid": 200, "correlation": 30493 + } + }, + { + "ph": "f", "id": 30493, "pid": 435, "tid": 435, "ts": 1776649790912013, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912014, "dur": 0, + "args": { + "External id": 30494, + "cbid": 273, "correlation": 30494 + } + }, + { + "ph": "f", "id": 30494, "pid": 435, "tid": 435, "ts": 1776649790912014, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790912023, "dur": 17, + "args": { + "External id": 30495, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30495, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30495, "pid": 0, "tid": 7, "ts": 1776649790912023, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912015, "dur": 4, + "args": { + "External id": 30495, + "cbid": 211, "correlation": 30495 + } + }, + { + "ph": "s", "id": 30495, "pid": 435, "tid": 435, "ts": 1776649790912015, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790912059, "dur": 1, + "args": { + "External id": 3748, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30513, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30513, "pid": 0, "tid": 7, "ts": 1776649790912059, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912050, "dur": 6, + "args": { + "External id": 3748, + "cbid": 211, "correlation": 30513 + } + }, + { + "ph": "s", "id": 30513, "pid": 435, "tid": 435, "ts": 1776649790912050, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912062, "dur": 0, + "args": { + "External id": 30519, + "cbid": 200, "correlation": 30519 + } + }, + { + "ph": "f", "id": 30519, "pid": 435, "tid": 435, "ts": 1776649790912062, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912063, "dur": 0, + "args": { + "External id": 30520, + "cbid": 273, "correlation": 30520 + } + }, + { + "ph": "f", "id": 30520, "pid": 435, "tid": 435, "ts": 1776649790912063, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790912071, "dur": 16, + "args": { + "External id": 30521, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30521, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30521, "pid": 0, "tid": 7, "ts": 1776649790912071, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912064, "dur": 4, + "args": { + "External id": 30521, + "cbid": 211, "correlation": 30521 + } + }, + { + "ph": "s", "id": 30521, "pid": 435, "tid": 435, "ts": 1776649790912064, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790912106, "dur": 1, + "args": { + "External id": 3753, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30539, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30539, "pid": 0, "tid": 7, "ts": 1776649790912106, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912097, "dur": 7, + "args": { + "External id": 3753, + "cbid": 211, "correlation": 30539 + } + }, + { + "ph": "s", "id": 30539, "pid": 435, "tid": 435, "ts": 1776649790912097, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912110, "dur": 0, + "args": { + "External id": 30545, + "cbid": 200, "correlation": 30545 + } + }, + { + "ph": "f", "id": 30545, "pid": 435, "tid": 435, "ts": 1776649790912110, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912111, "dur": 0, + "args": { + "External id": 30546, + "cbid": 273, "correlation": 30546 + } + }, + { + "ph": "f", "id": 30546, "pid": 435, "tid": 435, "ts": 1776649790912111, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790912120, "dur": 16, + "args": { + "External id": 30547, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30547, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30547, "pid": 0, "tid": 7, "ts": 1776649790912120, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912112, "dur": 4, + "args": { + "External id": 30547, + "cbid": 211, "correlation": 30547 + } + }, + { + "ph": "s", "id": 30547, "pid": 435, "tid": 435, "ts": 1776649790912112, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790912156, "dur": 1, + "args": { + "External id": 3758, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30565, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30565, "pid": 0, "tid": 7, "ts": 1776649790912156, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912146, "dur": 6, + "args": { + "External id": 3758, + "cbid": 211, "correlation": 30565 + } + }, + { + "ph": "s", "id": 30565, "pid": 435, "tid": 435, "ts": 1776649790912146, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912159, "dur": 0, + "args": { + "External id": 30571, + "cbid": 200, "correlation": 30571 + } + }, + { + "ph": "f", "id": 30571, "pid": 435, "tid": 435, "ts": 1776649790912159, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912160, "dur": 0, + "args": { + "External id": 30572, + "cbid": 273, "correlation": 30572 + } + }, + { + "ph": "f", "id": 30572, "pid": 435, "tid": 435, "ts": 1776649790912160, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790912169, "dur": 16, + "args": { + "External id": 30573, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30573, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30573, "pid": 0, "tid": 7, "ts": 1776649790912169, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912161, "dur": 5, + "args": { + "External id": 30573, + "cbid": 211, "correlation": 30573 + } + }, + { + "ph": "s", "id": 30573, "pid": 435, "tid": 435, "ts": 1776649790912161, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790912204, "dur": 1, + "args": { + "External id": 3763, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30591, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30591, "pid": 0, "tid": 7, "ts": 1776649790912204, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912195, "dur": 6, + "args": { + "External id": 3763, + "cbid": 211, "correlation": 30591 + } + }, + { + "ph": "s", "id": 30591, "pid": 435, "tid": 435, "ts": 1776649790912195, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912207, "dur": 0, + "args": { + "External id": 30597, + "cbid": 200, "correlation": 30597 + } + }, + { + "ph": "f", "id": 30597, "pid": 435, "tid": 435, "ts": 1776649790912207, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912208, "dur": 0, + "args": { + "External id": 30598, + "cbid": 273, "correlation": 30598 + } + }, + { + "ph": "f", "id": 30598, "pid": 435, "tid": 435, "ts": 1776649790912208, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790912217, "dur": 16, + "args": { + "External id": 30599, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30599, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30599, "pid": 0, "tid": 7, "ts": 1776649790912217, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912209, "dur": 4, + "args": { + "External id": 30599, + "cbid": 211, "correlation": 30599 + } + }, + { + "ph": "s", "id": 30599, "pid": 435, "tid": 435, "ts": 1776649790912209, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790912276, "dur": 1, + "args": { + "External id": 3768, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30617, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30617, "pid": 0, "tid": 7, "ts": 1776649790912276, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912266, "dur": 7, + "args": { + "External id": 3768, + "cbid": 211, "correlation": 30617 + } + }, + { + "ph": "s", "id": 30617, "pid": 435, "tid": 435, "ts": 1776649790912266, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912279, "dur": 0, + "args": { + "External id": 30623, + "cbid": 200, "correlation": 30623 + } + }, + { + "ph": "f", "id": 30623, "pid": 435, "tid": 435, "ts": 1776649790912279, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912280, "dur": 0, + "args": { + "External id": 30624, + "cbid": 273, "correlation": 30624 + } + }, + { + "ph": "f", "id": 30624, "pid": 435, "tid": 435, "ts": 1776649790912280, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790912289, "dur": 16, + "args": { + "External id": 30625, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30625, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30625, "pid": 0, "tid": 7, "ts": 1776649790912289, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912281, "dur": 5, + "args": { + "External id": 30625, + "cbid": 211, "correlation": 30625 + } + }, + { + "ph": "s", "id": 30625, "pid": 435, "tid": 435, "ts": 1776649790912281, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790912329, "dur": 1, + "args": { + "External id": 3773, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30643, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30643, "pid": 0, "tid": 7, "ts": 1776649790912329, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912319, "dur": 6, + "args": { + "External id": 3773, + "cbid": 211, "correlation": 30643 + } + }, + { + "ph": "s", "id": 30643, "pid": 435, "tid": 435, "ts": 1776649790912319, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912332, "dur": 0, + "args": { + "External id": 30649, + "cbid": 200, "correlation": 30649 + } + }, + { + "ph": "f", "id": 30649, "pid": 435, "tid": 435, "ts": 1776649790912332, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912332, "dur": 0, + "args": { + "External id": 30650, + "cbid": 273, "correlation": 30650 + } + }, + { + "ph": "f", "id": 30650, "pid": 435, "tid": 435, "ts": 1776649790912332, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790912341, "dur": 15, + "args": { + "External id": 30651, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30651, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30651, "pid": 0, "tid": 7, "ts": 1776649790912341, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912334, "dur": 4, + "args": { + "External id": 30651, + "cbid": 211, "correlation": 30651 + } + }, + { + "ph": "s", "id": 30651, "pid": 435, "tid": 435, "ts": 1776649790912334, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790912377, "dur": 1, + "args": { + "External id": 3778, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30669, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30669, "pid": 0, "tid": 7, "ts": 1776649790912377, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912368, "dur": 6, + "args": { + "External id": 3778, + "cbid": 211, "correlation": 30669 + } + }, + { + "ph": "s", "id": 30669, "pid": 435, "tid": 435, "ts": 1776649790912368, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912380, "dur": 0, + "args": { + "External id": 30675, + "cbid": 200, "correlation": 30675 + } + }, + { + "ph": "f", "id": 30675, "pid": 435, "tid": 435, "ts": 1776649790912380, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912381, "dur": 0, + "args": { + "External id": 30676, + "cbid": 273, "correlation": 30676 + } + }, + { + "ph": "f", "id": 30676, "pid": 435, "tid": 435, "ts": 1776649790912381, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790912390, "dur": 16, + "args": { + "External id": 30677, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30677, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30677, "pid": 0, "tid": 7, "ts": 1776649790912390, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912382, "dur": 4, + "args": { + "External id": 30677, + "cbid": 211, "correlation": 30677 + } + }, + { + "ph": "s", "id": 30677, "pid": 435, "tid": 435, "ts": 1776649790912382, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790912426, "dur": 1, + "args": { + "External id": 3783, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30695, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30695, "pid": 0, "tid": 7, "ts": 1776649790912426, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912416, "dur": 6, + "args": { + "External id": 3783, + "cbid": 211, "correlation": 30695 + } + }, + { + "ph": "s", "id": 30695, "pid": 435, "tid": 435, "ts": 1776649790912416, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912429, "dur": 0, + "args": { + "External id": 30701, + "cbid": 200, "correlation": 30701 + } + }, + { + "ph": "f", "id": 30701, "pid": 435, "tid": 435, "ts": 1776649790912429, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912430, "dur": 0, + "args": { + "External id": 30702, + "cbid": 273, "correlation": 30702 + } + }, + { + "ph": "f", "id": 30702, "pid": 435, "tid": 435, "ts": 1776649790912430, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790912438, "dur": 15, + "args": { + "External id": 30703, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30703, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30703, "pid": 0, "tid": 7, "ts": 1776649790912438, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912431, "dur": 4, + "args": { + "External id": 30703, + "cbid": 211, "correlation": 30703 + } + }, + { + "ph": "s", "id": 30703, "pid": 435, "tid": 435, "ts": 1776649790912431, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790912475, "dur": 1, + "args": { + "External id": 3788, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30721, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30721, "pid": 0, "tid": 7, "ts": 1776649790912475, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912465, "dur": 6, + "args": { + "External id": 3788, + "cbid": 211, "correlation": 30721 + } + }, + { + "ph": "s", "id": 30721, "pid": 435, "tid": 435, "ts": 1776649790912465, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912478, "dur": 0, + "args": { + "External id": 30727, + "cbid": 200, "correlation": 30727 + } + }, + { + "ph": "f", "id": 30727, "pid": 435, "tid": 435, "ts": 1776649790912478, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912478, "dur": 0, + "args": { + "External id": 30728, + "cbid": 273, "correlation": 30728 + } + }, + { + "ph": "f", "id": 30728, "pid": 435, "tid": 435, "ts": 1776649790912478, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790912487, "dur": 19, + "args": { + "External id": 30729, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30729, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30729, "pid": 0, "tid": 7, "ts": 1776649790912487, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912480, "dur": 5, + "args": { + "External id": 30729, + "cbid": 211, "correlation": 30729 + } + }, + { + "ph": "s", "id": 30729, "pid": 435, "tid": 435, "ts": 1776649790912480, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790912523, "dur": 1, + "args": { + "External id": 3793, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30747, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30747, "pid": 0, "tid": 7, "ts": 1776649790912523, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912514, "dur": 6, + "args": { + "External id": 3793, + "cbid": 211, "correlation": 30747 + } + }, + { + "ph": "s", "id": 30747, "pid": 435, "tid": 435, "ts": 1776649790912514, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912527, "dur": 0, + "args": { + "External id": 30753, + "cbid": 200, "correlation": 30753 + } + }, + { + "ph": "f", "id": 30753, "pid": 435, "tid": 435, "ts": 1776649790912527, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912527, "dur": 0, + "args": { + "External id": 30754, + "cbid": 273, "correlation": 30754 + } + }, + { + "ph": "f", "id": 30754, "pid": 435, "tid": 435, "ts": 1776649790912527, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790912536, "dur": 16, + "args": { + "External id": 30755, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30755, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30755, "pid": 0, "tid": 7, "ts": 1776649790912536, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912528, "dur": 4, + "args": { + "External id": 30755, + "cbid": 211, "correlation": 30755 + } + }, + { + "ph": "s", "id": 30755, "pid": 435, "tid": 435, "ts": 1776649790912528, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790912571, "dur": 1, + "args": { + "External id": 3798, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30773, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30773, "pid": 0, "tid": 7, "ts": 1776649790912571, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912562, "dur": 6, + "args": { + "External id": 3798, + "cbid": 211, "correlation": 30773 + } + }, + { + "ph": "s", "id": 30773, "pid": 435, "tid": 435, "ts": 1776649790912562, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912574, "dur": 0, + "args": { + "External id": 30779, + "cbid": 200, "correlation": 30779 + } + }, + { + "ph": "f", "id": 30779, "pid": 435, "tid": 435, "ts": 1776649790912574, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912575, "dur": 0, + "args": { + "External id": 30780, + "cbid": 273, "correlation": 30780 + } + }, + { + "ph": "f", "id": 30780, "pid": 435, "tid": 435, "ts": 1776649790912575, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790912584, "dur": 15, + "args": { + "External id": 30781, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30781, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30781, "pid": 0, "tid": 7, "ts": 1776649790912584, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912576, "dur": 4, + "args": { + "External id": 30781, + "cbid": 211, "correlation": 30781 + } + }, + { + "ph": "s", "id": 30781, "pid": 435, "tid": 435, "ts": 1776649790912576, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790912619, "dur": 1, + "args": { + "External id": 3803, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30799, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30799, "pid": 0, "tid": 7, "ts": 1776649790912619, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912610, "dur": 6, + "args": { + "External id": 3803, + "cbid": 211, "correlation": 30799 + } + }, + { + "ph": "s", "id": 30799, "pid": 435, "tid": 435, "ts": 1776649790912610, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912623, "dur": 0, + "args": { + "External id": 30805, + "cbid": 200, "correlation": 30805 + } + }, + { + "ph": "f", "id": 30805, "pid": 435, "tid": 435, "ts": 1776649790912623, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912623, "dur": 0, + "args": { + "External id": 30806, + "cbid": 273, "correlation": 30806 + } + }, + { + "ph": "f", "id": 30806, "pid": 435, "tid": 435, "ts": 1776649790912623, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790912632, "dur": 15, + "args": { + "External id": 30807, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30807, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30807, "pid": 0, "tid": 7, "ts": 1776649790912632, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912624, "dur": 5, + "args": { + "External id": 30807, + "cbid": 211, "correlation": 30807 + } + }, + { + "ph": "s", "id": 30807, "pid": 435, "tid": 435, "ts": 1776649790912624, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790912668, "dur": 1, + "args": { + "External id": 3808, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30825, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30825, "pid": 0, "tid": 7, "ts": 1776649790912668, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912658, "dur": 7, + "args": { + "External id": 3808, + "cbid": 211, "correlation": 30825 + } + }, + { + "ph": "s", "id": 30825, "pid": 435, "tid": 435, "ts": 1776649790912658, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912671, "dur": 0, + "args": { + "External id": 30831, + "cbid": 200, "correlation": 30831 + } + }, + { + "ph": "f", "id": 30831, "pid": 435, "tid": 435, "ts": 1776649790912671, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912672, "dur": 0, + "args": { + "External id": 30832, + "cbid": 273, "correlation": 30832 + } + }, + { + "ph": "f", "id": 30832, "pid": 435, "tid": 435, "ts": 1776649790912672, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790912681, "dur": 15, + "args": { + "External id": 30833, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30833, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30833, "pid": 0, "tid": 7, "ts": 1776649790912681, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912673, "dur": 5, + "args": { + "External id": 30833, + "cbid": 211, "correlation": 30833 + } + }, + { + "ph": "s", "id": 30833, "pid": 435, "tid": 435, "ts": 1776649790912673, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790912716, "dur": 1, + "args": { + "External id": 3813, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30851, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30851, "pid": 0, "tid": 7, "ts": 1776649790912716, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912707, "dur": 6, + "args": { + "External id": 3813, + "cbid": 211, "correlation": 30851 + } + }, + { + "ph": "s", "id": 30851, "pid": 435, "tid": 435, "ts": 1776649790912707, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912720, "dur": 0, + "args": { + "External id": 30857, + "cbid": 200, "correlation": 30857 + } + }, + { + "ph": "f", "id": 30857, "pid": 435, "tid": 435, "ts": 1776649790912720, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912720, "dur": 0, + "args": { + "External id": 30858, + "cbid": 273, "correlation": 30858 + } + }, + { + "ph": "f", "id": 30858, "pid": 435, "tid": 435, "ts": 1776649790912720, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790912729, "dur": 15, + "args": { + "External id": 30859, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30859, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30859, "pid": 0, "tid": 7, "ts": 1776649790912729, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912721, "dur": 5, + "args": { + "External id": 30859, + "cbid": 211, "correlation": 30859 + } + }, + { + "ph": "s", "id": 30859, "pid": 435, "tid": 435, "ts": 1776649790912721, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790912765, "dur": 1, + "args": { + "External id": 3818, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30877, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30877, "pid": 0, "tid": 7, "ts": 1776649790912765, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912755, "dur": 6, + "args": { + "External id": 3818, + "cbid": 211, "correlation": 30877 + } + }, + { + "ph": "s", "id": 30877, "pid": 435, "tid": 435, "ts": 1776649790912755, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912768, "dur": 0, + "args": { + "External id": 30883, + "cbid": 200, "correlation": 30883 + } + }, + { + "ph": "f", "id": 30883, "pid": 435, "tid": 435, "ts": 1776649790912768, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912768, "dur": 0, + "args": { + "External id": 30884, + "cbid": 273, "correlation": 30884 + } + }, + { + "ph": "f", "id": 30884, "pid": 435, "tid": 435, "ts": 1776649790912768, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790912777, "dur": 15, + "args": { + "External id": 30885, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30885, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30885, "pid": 0, "tid": 7, "ts": 1776649790912777, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912769, "dur": 4, + "args": { + "External id": 30885, + "cbid": 211, "correlation": 30885 + } + }, + { + "ph": "s", "id": 30885, "pid": 435, "tid": 435, "ts": 1776649790912769, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790912813, "dur": 1, + "args": { + "External id": 3823, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30903, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30903, "pid": 0, "tid": 7, "ts": 1776649790912813, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912804, "dur": 6, + "args": { + "External id": 3823, + "cbid": 211, "correlation": 30903 + } + }, + { + "ph": "s", "id": 30903, "pid": 435, "tid": 435, "ts": 1776649790912804, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912816, "dur": 0, + "args": { + "External id": 30909, + "cbid": 200, "correlation": 30909 + } + }, + { + "ph": "f", "id": 30909, "pid": 435, "tid": 435, "ts": 1776649790912816, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912816, "dur": 0, + "args": { + "External id": 30910, + "cbid": 273, "correlation": 30910 + } + }, + { + "ph": "f", "id": 30910, "pid": 435, "tid": 435, "ts": 1776649790912816, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790912826, "dur": 16, + "args": { + "External id": 30911, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30911, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30911, "pid": 0, "tid": 7, "ts": 1776649790912826, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912818, "dur": 5, + "args": { + "External id": 30911, + "cbid": 211, "correlation": 30911 + } + }, + { + "ph": "s", "id": 30911, "pid": 435, "tid": 435, "ts": 1776649790912818, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790912861, "dur": 1, + "args": { + "External id": 3828, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30929, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30929, "pid": 0, "tid": 7, "ts": 1776649790912861, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912852, "dur": 6, + "args": { + "External id": 3828, + "cbid": 211, "correlation": 30929 + } + }, + { + "ph": "s", "id": 30929, "pid": 435, "tid": 435, "ts": 1776649790912852, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912864, "dur": 0, + "args": { + "External id": 30935, + "cbid": 200, "correlation": 30935 + } + }, + { + "ph": "f", "id": 30935, "pid": 435, "tid": 435, "ts": 1776649790912864, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912864, "dur": 0, + "args": { + "External id": 30936, + "cbid": 273, "correlation": 30936 + } + }, + { + "ph": "f", "id": 30936, "pid": 435, "tid": 435, "ts": 1776649790912864, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790912873, "dur": 18, + "args": { + "External id": 30937, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30937, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30937, "pid": 0, "tid": 7, "ts": 1776649790912873, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912866, "dur": 4, + "args": { + "External id": 30937, + "cbid": 211, "correlation": 30937 + } + }, + { + "ph": "s", "id": 30937, "pid": 435, "tid": 435, "ts": 1776649790912866, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790912909, "dur": 1, + "args": { + "External id": 3833, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30955, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30955, "pid": 0, "tid": 7, "ts": 1776649790912909, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912899, "dur": 6, + "args": { + "External id": 3833, + "cbid": 211, "correlation": 30955 + } + }, + { + "ph": "s", "id": 30955, "pid": 435, "tid": 435, "ts": 1776649790912899, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912912, "dur": 0, + "args": { + "External id": 30961, + "cbid": 200, "correlation": 30961 + } + }, + { + "ph": "f", "id": 30961, "pid": 435, "tid": 435, "ts": 1776649790912912, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912912, "dur": 0, + "args": { + "External id": 30962, + "cbid": 273, "correlation": 30962 + } + }, + { + "ph": "f", "id": 30962, "pid": 435, "tid": 435, "ts": 1776649790912912, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790912921, "dur": 16, + "args": { + "External id": 30963, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30963, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30963, "pid": 0, "tid": 7, "ts": 1776649790912921, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912914, "dur": 4, + "args": { + "External id": 30963, + "cbid": 211, "correlation": 30963 + } + }, + { + "ph": "s", "id": 30963, "pid": 435, "tid": 435, "ts": 1776649790912914, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790912957, "dur": 1, + "args": { + "External id": 3838, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30981, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30981, "pid": 0, "tid": 7, "ts": 1776649790912957, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912948, "dur": 6, + "args": { + "External id": 3838, + "cbid": 211, "correlation": 30981 + } + }, + { + "ph": "s", "id": 30981, "pid": 435, "tid": 435, "ts": 1776649790912948, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912960, "dur": 0, + "args": { + "External id": 30987, + "cbid": 200, "correlation": 30987 + } + }, + { + "ph": "f", "id": 30987, "pid": 435, "tid": 435, "ts": 1776649790912960, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790912961, "dur": 0, + "args": { + "External id": 30988, + "cbid": 273, "correlation": 30988 + } + }, + { + "ph": "f", "id": 30988, "pid": 435, "tid": 435, "ts": 1776649790912961, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790912970, "dur": 16, + "args": { + "External id": 30989, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 30989, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 30989, "pid": 0, "tid": 7, "ts": 1776649790912970, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912962, "dur": 4, + "args": { + "External id": 30989, + "cbid": 211, "correlation": 30989 + } + }, + { + "ph": "s", "id": 30989, "pid": 435, "tid": 435, "ts": 1776649790912962, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790913006, "dur": 1, + "args": { + "External id": 3843, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31007, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31007, "pid": 0, "tid": 7, "ts": 1776649790913006, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790912996, "dur": 6, + "args": { + "External id": 3843, + "cbid": 211, "correlation": 31007 + } + }, + { + "ph": "s", "id": 31007, "pid": 435, "tid": 435, "ts": 1776649790912996, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913008, "dur": 0, + "args": { + "External id": 31013, + "cbid": 200, "correlation": 31013 + } + }, + { + "ph": "f", "id": 31013, "pid": 435, "tid": 435, "ts": 1776649790913008, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913009, "dur": 0, + "args": { + "External id": 31014, + "cbid": 273, "correlation": 31014 + } + }, + { + "ph": "f", "id": 31014, "pid": 435, "tid": 435, "ts": 1776649790913009, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790913018, "dur": 15, + "args": { + "External id": 31015, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31015, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31015, "pid": 0, "tid": 7, "ts": 1776649790913018, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913010, "dur": 4, + "args": { + "External id": 31015, + "cbid": 211, "correlation": 31015 + } + }, + { + "ph": "s", "id": 31015, "pid": 435, "tid": 435, "ts": 1776649790913010, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790913053, "dur": 1, + "args": { + "External id": 3848, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31033, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31033, "pid": 0, "tid": 7, "ts": 1776649790913053, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913044, "dur": 6, + "args": { + "External id": 3848, + "cbid": 211, "correlation": 31033 + } + }, + { + "ph": "s", "id": 31033, "pid": 435, "tid": 435, "ts": 1776649790913044, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913057, "dur": 0, + "args": { + "External id": 31039, + "cbid": 200, "correlation": 31039 + } + }, + { + "ph": "f", "id": 31039, "pid": 435, "tid": 435, "ts": 1776649790913057, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913057, "dur": 0, + "args": { + "External id": 31040, + "cbid": 273, "correlation": 31040 + } + }, + { + "ph": "f", "id": 31040, "pid": 435, "tid": 435, "ts": 1776649790913057, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790913066, "dur": 15, + "args": { + "External id": 31041, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31041, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31041, "pid": 0, "tid": 7, "ts": 1776649790913066, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913058, "dur": 4, + "args": { + "External id": 31041, + "cbid": 211, "correlation": 31041 + } + }, + { + "ph": "s", "id": 31041, "pid": 435, "tid": 435, "ts": 1776649790913058, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790913102, "dur": 1, + "args": { + "External id": 3853, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31059, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31059, "pid": 0, "tid": 7, "ts": 1776649790913102, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913093, "dur": 6, + "args": { + "External id": 3853, + "cbid": 211, "correlation": 31059 + } + }, + { + "ph": "s", "id": 31059, "pid": 435, "tid": 435, "ts": 1776649790913093, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913105, "dur": 0, + "args": { + "External id": 31065, + "cbid": 200, "correlation": 31065 + } + }, + { + "ph": "f", "id": 31065, "pid": 435, "tid": 435, "ts": 1776649790913105, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913106, "dur": 0, + "args": { + "External id": 31066, + "cbid": 273, "correlation": 31066 + } + }, + { + "ph": "f", "id": 31066, "pid": 435, "tid": 435, "ts": 1776649790913106, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790913115, "dur": 16, + "args": { + "External id": 31067, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31067, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31067, "pid": 0, "tid": 7, "ts": 1776649790913115, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913107, "dur": 4, + "args": { + "External id": 31067, + "cbid": 211, "correlation": 31067 + } + }, + { + "ph": "s", "id": 31067, "pid": 435, "tid": 435, "ts": 1776649790913107, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790913138, "dur": 3, + "args": { + "External id": 31075, + "cbid": 135, "correlation": 31075 + } + }, + { + "ph": "f", "id": 31075, "pid": 435, "tid": 435, "ts": 1776649790913138, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790913153, "dur": 5, + "args": { + "External id": 31080, + "cbid": 165, "correlation": 31080 + } + }, + { + "ph": "s", "id": 31080, "pid": 435, "tid": 435, "ts": 1776649790913153, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649790913161, "dur": 4, + "args": { + "External id": 31083, + "cbid": 139, "correlation": 31083 + } + }, + { + "ph": "f", "id": 31083, "pid": 435, "tid": 435, "ts": 1776649790913161, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790913219, "dur": 1, + "args": { + "External id": 3858, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31111, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31111, "pid": 0, "tid": 7, "ts": 1776649790913219, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913207, "dur": 9, + "args": { + "External id": 3858, + "cbid": 211, "correlation": 31111 + } + }, + { + "ph": "s", "id": 31111, "pid": 435, "tid": 435, "ts": 1776649790913207, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913223, "dur": 0, + "args": { + "External id": 31117, + "cbid": 200, "correlation": 31117 + } + }, + { + "ph": "f", "id": 31117, "pid": 435, "tid": 435, "ts": 1776649790913223, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913224, "dur": 0, + "args": { + "External id": 31118, + "cbid": 273, "correlation": 31118 + } + }, + { + "ph": "f", "id": 31118, "pid": 435, "tid": 435, "ts": 1776649790913224, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790913233, "dur": 15, + "args": { + "External id": 31119, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31119, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31119, "pid": 0, "tid": 7, "ts": 1776649790913233, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913225, "dur": 5, + "args": { + "External id": 31119, + "cbid": 211, "correlation": 31119 + } + }, + { + "ph": "s", "id": 31119, "pid": 435, "tid": 435, "ts": 1776649790913225, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790913272, "dur": 1, + "args": { + "External id": 3863, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31137, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31137, "pid": 0, "tid": 7, "ts": 1776649790913272, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913262, "dur": 7, + "args": { + "External id": 3863, + "cbid": 211, "correlation": 31137 + } + }, + { + "ph": "s", "id": 31137, "pid": 435, "tid": 435, "ts": 1776649790913262, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913275, "dur": 0, + "args": { + "External id": 31143, + "cbid": 200, "correlation": 31143 + } + }, + { + "ph": "f", "id": 31143, "pid": 435, "tid": 435, "ts": 1776649790913275, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913276, "dur": 0, + "args": { + "External id": 31144, + "cbid": 273, "correlation": 31144 + } + }, + { + "ph": "f", "id": 31144, "pid": 435, "tid": 435, "ts": 1776649790913276, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790913285, "dur": 16, + "args": { + "External id": 31145, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31145, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31145, "pid": 0, "tid": 7, "ts": 1776649790913285, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913277, "dur": 5, + "args": { + "External id": 31145, + "cbid": 211, "correlation": 31145 + } + }, + { + "ph": "s", "id": 31145, "pid": 435, "tid": 435, "ts": 1776649790913277, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790913321, "dur": 1, + "args": { + "External id": 3868, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31163, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31163, "pid": 0, "tid": 7, "ts": 1776649790913321, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913312, "dur": 6, + "args": { + "External id": 3868, + "cbid": 211, "correlation": 31163 + } + }, + { + "ph": "s", "id": 31163, "pid": 435, "tid": 435, "ts": 1776649790913312, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913325, "dur": 0, + "args": { + "External id": 31169, + "cbid": 200, "correlation": 31169 + } + }, + { + "ph": "f", "id": 31169, "pid": 435, "tid": 435, "ts": 1776649790913325, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913325, "dur": 0, + "args": { + "External id": 31170, + "cbid": 273, "correlation": 31170 + } + }, + { + "ph": "f", "id": 31170, "pid": 435, "tid": 435, "ts": 1776649790913325, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790913334, "dur": 15, + "args": { + "External id": 31171, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31171, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31171, "pid": 0, "tid": 7, "ts": 1776649790913334, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913327, "dur": 4, + "args": { + "External id": 31171, + "cbid": 211, "correlation": 31171 + } + }, + { + "ph": "s", "id": 31171, "pid": 435, "tid": 435, "ts": 1776649790913327, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790913370, "dur": 1, + "args": { + "External id": 3873, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31189, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31189, "pid": 0, "tid": 7, "ts": 1776649790913370, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913361, "dur": 6, + "args": { + "External id": 3873, + "cbid": 211, "correlation": 31189 + } + }, + { + "ph": "s", "id": 31189, "pid": 435, "tid": 435, "ts": 1776649790913361, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913373, "dur": 0, + "args": { + "External id": 31195, + "cbid": 200, "correlation": 31195 + } + }, + { + "ph": "f", "id": 31195, "pid": 435, "tid": 435, "ts": 1776649790913373, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913374, "dur": 0, + "args": { + "External id": 31196, + "cbid": 273, "correlation": 31196 + } + }, + { + "ph": "f", "id": 31196, "pid": 435, "tid": 435, "ts": 1776649790913374, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790913383, "dur": 17, + "args": { + "External id": 31197, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31197, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31197, "pid": 0, "tid": 7, "ts": 1776649790913383, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913375, "dur": 4, + "args": { + "External id": 31197, + "cbid": 211, "correlation": 31197 + } + }, + { + "ph": "s", "id": 31197, "pid": 435, "tid": 435, "ts": 1776649790913375, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790913419, "dur": 1, + "args": { + "External id": 3878, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31215, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31215, "pid": 0, "tid": 7, "ts": 1776649790913419, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913409, "dur": 7, + "args": { + "External id": 3878, + "cbid": 211, "correlation": 31215 + } + }, + { + "ph": "s", "id": 31215, "pid": 435, "tid": 435, "ts": 1776649790913409, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913422, "dur": 0, + "args": { + "External id": 31221, + "cbid": 200, "correlation": 31221 + } + }, + { + "ph": "f", "id": 31221, "pid": 435, "tid": 435, "ts": 1776649790913422, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913423, "dur": 0, + "args": { + "External id": 31222, + "cbid": 273, "correlation": 31222 + } + }, + { + "ph": "f", "id": 31222, "pid": 435, "tid": 435, "ts": 1776649790913423, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790913432, "dur": 16, + "args": { + "External id": 31223, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31223, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31223, "pid": 0, "tid": 7, "ts": 1776649790913432, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913424, "dur": 5, + "args": { + "External id": 31223, + "cbid": 211, "correlation": 31223 + } + }, + { + "ph": "s", "id": 31223, "pid": 435, "tid": 435, "ts": 1776649790913424, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790913468, "dur": 1, + "args": { + "External id": 3883, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31241, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31241, "pid": 0, "tid": 7, "ts": 1776649790913468, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913459, "dur": 7, + "args": { + "External id": 3883, + "cbid": 211, "correlation": 31241 + } + }, + { + "ph": "s", "id": 31241, "pid": 435, "tid": 435, "ts": 1776649790913459, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913471, "dur": 0, + "args": { + "External id": 31247, + "cbid": 200, "correlation": 31247 + } + }, + { + "ph": "f", "id": 31247, "pid": 435, "tid": 435, "ts": 1776649790913471, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913472, "dur": 0, + "args": { + "External id": 31248, + "cbid": 273, "correlation": 31248 + } + }, + { + "ph": "f", "id": 31248, "pid": 435, "tid": 435, "ts": 1776649790913472, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790913481, "dur": 15, + "args": { + "External id": 31249, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31249, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31249, "pid": 0, "tid": 7, "ts": 1776649790913481, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913473, "dur": 4, + "args": { + "External id": 31249, + "cbid": 211, "correlation": 31249 + } + }, + { + "ph": "s", "id": 31249, "pid": 435, "tid": 435, "ts": 1776649790913473, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790913517, "dur": 1, + "args": { + "External id": 3888, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31267, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31267, "pid": 0, "tid": 7, "ts": 1776649790913517, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913507, "dur": 7, + "args": { + "External id": 3888, + "cbid": 211, "correlation": 31267 + } + }, + { + "ph": "s", "id": 31267, "pid": 435, "tid": 435, "ts": 1776649790913507, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913520, "dur": 0, + "args": { + "External id": 31273, + "cbid": 200, "correlation": 31273 + } + }, + { + "ph": "f", "id": 31273, "pid": 435, "tid": 435, "ts": 1776649790913520, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913521, "dur": 0, + "args": { + "External id": 31274, + "cbid": 273, "correlation": 31274 + } + }, + { + "ph": "f", "id": 31274, "pid": 435, "tid": 435, "ts": 1776649790913521, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790913529, "dur": 16, + "args": { + "External id": 31275, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31275, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31275, "pid": 0, "tid": 7, "ts": 1776649790913529, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913522, "dur": 5, + "args": { + "External id": 31275, + "cbid": 211, "correlation": 31275 + } + }, + { + "ph": "s", "id": 31275, "pid": 435, "tid": 435, "ts": 1776649790913522, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790913565, "dur": 1, + "args": { + "External id": 3893, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31293, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31293, "pid": 0, "tid": 7, "ts": 1776649790913565, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913556, "dur": 6, + "args": { + "External id": 3893, + "cbid": 211, "correlation": 31293 + } + }, + { + "ph": "s", "id": 31293, "pid": 435, "tid": 435, "ts": 1776649790913556, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913568, "dur": 0, + "args": { + "External id": 31299, + "cbid": 200, "correlation": 31299 + } + }, + { + "ph": "f", "id": 31299, "pid": 435, "tid": 435, "ts": 1776649790913568, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913569, "dur": 0, + "args": { + "External id": 31300, + "cbid": 273, "correlation": 31300 + } + }, + { + "ph": "f", "id": 31300, "pid": 435, "tid": 435, "ts": 1776649790913569, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790913578, "dur": 16, + "args": { + "External id": 31301, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31301, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31301, "pid": 0, "tid": 7, "ts": 1776649790913578, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913570, "dur": 4, + "args": { + "External id": 31301, + "cbid": 211, "correlation": 31301 + } + }, + { + "ph": "s", "id": 31301, "pid": 435, "tid": 435, "ts": 1776649790913570, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790913613, "dur": 1, + "args": { + "External id": 3898, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31319, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31319, "pid": 0, "tid": 7, "ts": 1776649790913613, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913604, "dur": 6, + "args": { + "External id": 3898, + "cbid": 211, "correlation": 31319 + } + }, + { + "ph": "s", "id": 31319, "pid": 435, "tid": 435, "ts": 1776649790913604, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913616, "dur": 0, + "args": { + "External id": 31325, + "cbid": 200, "correlation": 31325 + } + }, + { + "ph": "f", "id": 31325, "pid": 435, "tid": 435, "ts": 1776649790913616, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913617, "dur": 0, + "args": { + "External id": 31326, + "cbid": 273, "correlation": 31326 + } + }, + { + "ph": "f", "id": 31326, "pid": 435, "tid": 435, "ts": 1776649790913617, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790913626, "dur": 16, + "args": { + "External id": 31327, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31327, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31327, "pid": 0, "tid": 7, "ts": 1776649790913626, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913618, "dur": 4, + "args": { + "External id": 31327, + "cbid": 211, "correlation": 31327 + } + }, + { + "ph": "s", "id": 31327, "pid": 435, "tid": 435, "ts": 1776649790913618, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790913661, "dur": 1, + "args": { + "External id": 3903, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31345, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31345, "pid": 0, "tid": 7, "ts": 1776649790913661, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913652, "dur": 6, + "args": { + "External id": 3903, + "cbid": 211, "correlation": 31345 + } + }, + { + "ph": "s", "id": 31345, "pid": 435, "tid": 435, "ts": 1776649790913652, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913665, "dur": 0, + "args": { + "External id": 31351, + "cbid": 200, "correlation": 31351 + } + }, + { + "ph": "f", "id": 31351, "pid": 435, "tid": 435, "ts": 1776649790913665, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913665, "dur": 0, + "args": { + "External id": 31352, + "cbid": 273, "correlation": 31352 + } + }, + { + "ph": "f", "id": 31352, "pid": 435, "tid": 435, "ts": 1776649790913665, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790913674, "dur": 16, + "args": { + "External id": 31353, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31353, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31353, "pid": 0, "tid": 7, "ts": 1776649790913674, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913666, "dur": 5, + "args": { + "External id": 31353, + "cbid": 211, "correlation": 31353 + } + }, + { + "ph": "s", "id": 31353, "pid": 435, "tid": 435, "ts": 1776649790913666, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790913688, "dur": 6, + "args": { + "External id": 31356, + "cbid": 165, "correlation": 31356 + } + }, + { + "ph": "s", "id": 31356, "pid": 435, "tid": 435, "ts": 1776649790913688, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790913716, "dur": 3, + "args": { + "External id": 31365, + "cbid": 135, "correlation": 31365 + } + }, + { + "ph": "f", "id": 31365, "pid": 435, "tid": 435, "ts": 1776649790913716, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790913764, "dur": 1, + "args": { + "External id": 3908, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31385, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31385, "pid": 0, "tid": 7, "ts": 1776649790913764, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913754, "dur": 8, + "args": { + "External id": 3908, + "cbid": 211, "correlation": 31385 + } + }, + { + "ph": "s", "id": 31385, "pid": 435, "tid": 435, "ts": 1776649790913754, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913769, "dur": 0, + "args": { + "External id": 31391, + "cbid": 200, "correlation": 31391 + } + }, + { + "ph": "f", "id": 31391, "pid": 435, "tid": 435, "ts": 1776649790913769, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913770, "dur": 0, + "args": { + "External id": 31392, + "cbid": 273, "correlation": 31392 + } + }, + { + "ph": "f", "id": 31392, "pid": 435, "tid": 435, "ts": 1776649790913770, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790913779, "dur": 16, + "args": { + "External id": 31393, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31393, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31393, "pid": 0, "tid": 7, "ts": 1776649790913779, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913771, "dur": 5, + "args": { + "External id": 31393, + "cbid": 211, "correlation": 31393 + } + }, + { + "ph": "s", "id": 31393, "pid": 435, "tid": 435, "ts": 1776649790913771, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790913817, "dur": 1, + "args": { + "External id": 3913, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31411, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31411, "pid": 0, "tid": 7, "ts": 1776649790913817, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913807, "dur": 7, + "args": { + "External id": 3913, + "cbid": 211, "correlation": 31411 + } + }, + { + "ph": "s", "id": 31411, "pid": 435, "tid": 435, "ts": 1776649790913807, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913820, "dur": 0, + "args": { + "External id": 31417, + "cbid": 200, "correlation": 31417 + } + }, + { + "ph": "f", "id": 31417, "pid": 435, "tid": 435, "ts": 1776649790913820, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913821, "dur": 0, + "args": { + "External id": 31418, + "cbid": 273, "correlation": 31418 + } + }, + { + "ph": "f", "id": 31418, "pid": 435, "tid": 435, "ts": 1776649790913821, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790913829, "dur": 17, + "args": { + "External id": 31419, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31419, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31419, "pid": 0, "tid": 7, "ts": 1776649790913829, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913822, "dur": 5, + "args": { + "External id": 31419, + "cbid": 211, "correlation": 31419 + } + }, + { + "ph": "s", "id": 31419, "pid": 435, "tid": 435, "ts": 1776649790913822, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790913865, "dur": 1, + "args": { + "External id": 3918, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31437, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31437, "pid": 0, "tid": 7, "ts": 1776649790913865, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913856, "dur": 6, + "args": { + "External id": 3918, + "cbid": 211, "correlation": 31437 + } + }, + { + "ph": "s", "id": 31437, "pid": 435, "tid": 435, "ts": 1776649790913856, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913869, "dur": 0, + "args": { + "External id": 31443, + "cbid": 200, "correlation": 31443 + } + }, + { + "ph": "f", "id": 31443, "pid": 435, "tid": 435, "ts": 1776649790913869, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913869, "dur": 0, + "args": { + "External id": 31444, + "cbid": 273, "correlation": 31444 + } + }, + { + "ph": "f", "id": 31444, "pid": 435, "tid": 435, "ts": 1776649790913869, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790913878, "dur": 16, + "args": { + "External id": 31445, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31445, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31445, "pid": 0, "tid": 7, "ts": 1776649790913878, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913870, "dur": 5, + "args": { + "External id": 31445, + "cbid": 211, "correlation": 31445 + } + }, + { + "ph": "s", "id": 31445, "pid": 435, "tid": 435, "ts": 1776649790913870, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790913914, "dur": 1, + "args": { + "External id": 3923, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31463, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31463, "pid": 0, "tid": 7, "ts": 1776649790913914, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913905, "dur": 6, + "args": { + "External id": 3923, + "cbid": 211, "correlation": 31463 + } + }, + { + "ph": "s", "id": 31463, "pid": 435, "tid": 435, "ts": 1776649790913905, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913917, "dur": 0, + "args": { + "External id": 31469, + "cbid": 200, "correlation": 31469 + } + }, + { + "ph": "f", "id": 31469, "pid": 435, "tid": 435, "ts": 1776649790913917, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913918, "dur": 0, + "args": { + "External id": 31470, + "cbid": 273, "correlation": 31470 + } + }, + { + "ph": "f", "id": 31470, "pid": 435, "tid": 435, "ts": 1776649790913918, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790913927, "dur": 15, + "args": { + "External id": 31471, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31471, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31471, "pid": 0, "tid": 7, "ts": 1776649790913927, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913919, "dur": 5, + "args": { + "External id": 31471, + "cbid": 211, "correlation": 31471 + } + }, + { + "ph": "s", "id": 31471, "pid": 435, "tid": 435, "ts": 1776649790913919, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790913963, "dur": 1, + "args": { + "External id": 3928, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31489, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31489, "pid": 0, "tid": 7, "ts": 1776649790913963, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913953, "dur": 6, + "args": { + "External id": 3928, + "cbid": 211, "correlation": 31489 + } + }, + { + "ph": "s", "id": 31489, "pid": 435, "tid": 435, "ts": 1776649790913953, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913966, "dur": 0, + "args": { + "External id": 31495, + "cbid": 200, "correlation": 31495 + } + }, + { + "ph": "f", "id": 31495, "pid": 435, "tid": 435, "ts": 1776649790913966, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790913967, "dur": 0, + "args": { + "External id": 31496, + "cbid": 273, "correlation": 31496 + } + }, + { + "ph": "f", "id": 31496, "pid": 435, "tid": 435, "ts": 1776649790913967, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790913976, "dur": 16, + "args": { + "External id": 31497, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31497, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31497, "pid": 0, "tid": 7, "ts": 1776649790913976, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790913968, "dur": 5, + "args": { + "External id": 31497, + "cbid": 211, "correlation": 31497 + } + }, + { + "ph": "s", "id": 31497, "pid": 435, "tid": 435, "ts": 1776649790913968, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790914012, "dur": 1, + "args": { + "External id": 3933, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31515, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31515, "pid": 0, "tid": 7, "ts": 1776649790914012, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914002, "dur": 6, + "args": { + "External id": 3933, + "cbid": 211, "correlation": 31515 + } + }, + { + "ph": "s", "id": 31515, "pid": 435, "tid": 435, "ts": 1776649790914002, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914015, "dur": 0, + "args": { + "External id": 31521, + "cbid": 200, "correlation": 31521 + } + }, + { + "ph": "f", "id": 31521, "pid": 435, "tid": 435, "ts": 1776649790914015, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914015, "dur": 0, + "args": { + "External id": 31522, + "cbid": 273, "correlation": 31522 + } + }, + { + "ph": "f", "id": 31522, "pid": 435, "tid": 435, "ts": 1776649790914015, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790914024, "dur": 16, + "args": { + "External id": 31523, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31523, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31523, "pid": 0, "tid": 7, "ts": 1776649790914024, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914016, "dur": 5, + "args": { + "External id": 31523, + "cbid": 211, "correlation": 31523 + } + }, + { + "ph": "s", "id": 31523, "pid": 435, "tid": 435, "ts": 1776649790914016, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790914061, "dur": 1, + "args": { + "External id": 3938, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31541, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31541, "pid": 0, "tid": 7, "ts": 1776649790914061, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914051, "dur": 6, + "args": { + "External id": 3938, + "cbid": 211, "correlation": 31541 + } + }, + { + "ph": "s", "id": 31541, "pid": 435, "tid": 435, "ts": 1776649790914051, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914064, "dur": 0, + "args": { + "External id": 31547, + "cbid": 200, "correlation": 31547 + } + }, + { + "ph": "f", "id": 31547, "pid": 435, "tid": 435, "ts": 1776649790914064, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914064, "dur": 0, + "args": { + "External id": 31548, + "cbid": 273, "correlation": 31548 + } + }, + { + "ph": "f", "id": 31548, "pid": 435, "tid": 435, "ts": 1776649790914064, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790914073, "dur": 15, + "args": { + "External id": 31549, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31549, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31549, "pid": 0, "tid": 7, "ts": 1776649790914073, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914066, "dur": 4, + "args": { + "External id": 31549, + "cbid": 211, "correlation": 31549 + } + }, + { + "ph": "s", "id": 31549, "pid": 435, "tid": 435, "ts": 1776649790914066, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790914110, "dur": 1, + "args": { + "External id": 3943, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31567, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31567, "pid": 0, "tid": 7, "ts": 1776649790914110, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914100, "dur": 6, + "args": { + "External id": 3943, + "cbid": 211, "correlation": 31567 + } + }, + { + "ph": "s", "id": 31567, "pid": 435, "tid": 435, "ts": 1776649790914100, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914113, "dur": 0, + "args": { + "External id": 31573, + "cbid": 200, "correlation": 31573 + } + }, + { + "ph": "f", "id": 31573, "pid": 435, "tid": 435, "ts": 1776649790914113, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914114, "dur": 0, + "args": { + "External id": 31574, + "cbid": 273, "correlation": 31574 + } + }, + { + "ph": "f", "id": 31574, "pid": 435, "tid": 435, "ts": 1776649790914114, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790914123, "dur": 16, + "args": { + "External id": 31575, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31575, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31575, "pid": 0, "tid": 7, "ts": 1776649790914123, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914115, "dur": 5, + "args": { + "External id": 31575, + "cbid": 211, "correlation": 31575 + } + }, + { + "ph": "s", "id": 31575, "pid": 435, "tid": 435, "ts": 1776649790914115, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790914158, "dur": 1, + "args": { + "External id": 3948, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31593, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31593, "pid": 0, "tid": 7, "ts": 1776649790914158, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914149, "dur": 6, + "args": { + "External id": 3948, + "cbid": 211, "correlation": 31593 + } + }, + { + "ph": "s", "id": 31593, "pid": 435, "tid": 435, "ts": 1776649790914149, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914161, "dur": 0, + "args": { + "External id": 31599, + "cbid": 200, "correlation": 31599 + } + }, + { + "ph": "f", "id": 31599, "pid": 435, "tid": 435, "ts": 1776649790914161, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914162, "dur": 0, + "args": { + "External id": 31600, + "cbid": 273, "correlation": 31600 + } + }, + { + "ph": "f", "id": 31600, "pid": 435, "tid": 435, "ts": 1776649790914162, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790914171, "dur": 16, + "args": { + "External id": 31601, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31601, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31601, "pid": 0, "tid": 7, "ts": 1776649790914171, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914163, "dur": 4, + "args": { + "External id": 31601, + "cbid": 211, "correlation": 31601 + } + }, + { + "ph": "s", "id": 31601, "pid": 435, "tid": 435, "ts": 1776649790914163, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790914207, "dur": 1, + "args": { + "External id": 3953, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31619, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31619, "pid": 0, "tid": 7, "ts": 1776649790914207, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914197, "dur": 6, + "args": { + "External id": 3953, + "cbid": 211, "correlation": 31619 + } + }, + { + "ph": "s", "id": 31619, "pid": 435, "tid": 435, "ts": 1776649790914197, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914210, "dur": 0, + "args": { + "External id": 31625, + "cbid": 200, "correlation": 31625 + } + }, + { + "ph": "f", "id": 31625, "pid": 435, "tid": 435, "ts": 1776649790914210, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914211, "dur": 0, + "args": { + "External id": 31626, + "cbid": 273, "correlation": 31626 + } + }, + { + "ph": "f", "id": 31626, "pid": 435, "tid": 435, "ts": 1776649790914211, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790914219, "dur": 16, + "args": { + "External id": 31627, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31627, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31627, "pid": 0, "tid": 7, "ts": 1776649790914219, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914212, "dur": 5, + "args": { + "External id": 31627, + "cbid": 211, "correlation": 31627 + } + }, + { + "ph": "s", "id": 31627, "pid": 435, "tid": 435, "ts": 1776649790914212, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790914256, "dur": 1, + "args": { + "External id": 3958, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31645, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31645, "pid": 0, "tid": 7, "ts": 1776649790914256, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914246, "dur": 7, + "args": { + "External id": 3958, + "cbid": 211, "correlation": 31645 + } + }, + { + "ph": "s", "id": 31645, "pid": 435, "tid": 435, "ts": 1776649790914246, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914259, "dur": 0, + "args": { + "External id": 31651, + "cbid": 200, "correlation": 31651 + } + }, + { + "ph": "f", "id": 31651, "pid": 435, "tid": 435, "ts": 1776649790914259, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914259, "dur": 0, + "args": { + "External id": 31652, + "cbid": 273, "correlation": 31652 + } + }, + { + "ph": "f", "id": 31652, "pid": 435, "tid": 435, "ts": 1776649790914259, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790914268, "dur": 17, + "args": { + "External id": 31653, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31653, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31653, "pid": 0, "tid": 7, "ts": 1776649790914268, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914260, "dur": 4, + "args": { + "External id": 31653, + "cbid": 211, "correlation": 31653 + } + }, + { + "ph": "s", "id": 31653, "pid": 435, "tid": 435, "ts": 1776649790914260, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790914304, "dur": 1, + "args": { + "External id": 3963, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31671, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31671, "pid": 0, "tid": 7, "ts": 1776649790914304, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914295, "dur": 6, + "args": { + "External id": 3963, + "cbid": 211, "correlation": 31671 + } + }, + { + "ph": "s", "id": 31671, "pid": 435, "tid": 435, "ts": 1776649790914295, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914307, "dur": 0, + "args": { + "External id": 31677, + "cbid": 200, "correlation": 31677 + } + }, + { + "ph": "f", "id": 31677, "pid": 435, "tid": 435, "ts": 1776649790914307, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914308, "dur": 0, + "args": { + "External id": 31678, + "cbid": 273, "correlation": 31678 + } + }, + { + "ph": "f", "id": 31678, "pid": 435, "tid": 435, "ts": 1776649790914308, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790914317, "dur": 15, + "args": { + "External id": 31679, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31679, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31679, "pid": 0, "tid": 7, "ts": 1776649790914317, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914309, "dur": 5, + "args": { + "External id": 31679, + "cbid": 211, "correlation": 31679 + } + }, + { + "ph": "s", "id": 31679, "pid": 435, "tid": 435, "ts": 1776649790914309, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790914353, "dur": 1, + "args": { + "External id": 3968, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31697, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31697, "pid": 0, "tid": 7, "ts": 1776649790914353, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914343, "dur": 6, + "args": { + "External id": 3968, + "cbid": 211, "correlation": 31697 + } + }, + { + "ph": "s", "id": 31697, "pid": 435, "tid": 435, "ts": 1776649790914343, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914356, "dur": 0, + "args": { + "External id": 31703, + "cbid": 200, "correlation": 31703 + } + }, + { + "ph": "f", "id": 31703, "pid": 435, "tid": 435, "ts": 1776649790914356, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914356, "dur": 0, + "args": { + "External id": 31704, + "cbid": 273, "correlation": 31704 + } + }, + { + "ph": "f", "id": 31704, "pid": 435, "tid": 435, "ts": 1776649790914356, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790914365, "dur": 15, + "args": { + "External id": 31705, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31705, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31705, "pid": 0, "tid": 7, "ts": 1776649790914365, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914358, "dur": 4, + "args": { + "External id": 31705, + "cbid": 211, "correlation": 31705 + } + }, + { + "ph": "s", "id": 31705, "pid": 435, "tid": 435, "ts": 1776649790914358, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790914401, "dur": 1, + "args": { + "External id": 3973, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31723, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31723, "pid": 0, "tid": 7, "ts": 1776649790914401, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914392, "dur": 7, + "args": { + "External id": 3973, + "cbid": 211, "correlation": 31723 + } + }, + { + "ph": "s", "id": 31723, "pid": 435, "tid": 435, "ts": 1776649790914392, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914404, "dur": 0, + "args": { + "External id": 31729, + "cbid": 200, "correlation": 31729 + } + }, + { + "ph": "f", "id": 31729, "pid": 435, "tid": 435, "ts": 1776649790914404, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914405, "dur": 0, + "args": { + "External id": 31730, + "cbid": 273, "correlation": 31730 + } + }, + { + "ph": "f", "id": 31730, "pid": 435, "tid": 435, "ts": 1776649790914405, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790914414, "dur": 16, + "args": { + "External id": 31731, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31731, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31731, "pid": 0, "tid": 7, "ts": 1776649790914414, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914406, "dur": 4, + "args": { + "External id": 31731, + "cbid": 211, "correlation": 31731 + } + }, + { + "ph": "s", "id": 31731, "pid": 435, "tid": 435, "ts": 1776649790914406, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790914450, "dur": 1, + "args": { + "External id": 3978, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31749, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31749, "pid": 0, "tid": 7, "ts": 1776649790914450, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914441, "dur": 6, + "args": { + "External id": 3978, + "cbid": 211, "correlation": 31749 + } + }, + { + "ph": "s", "id": 31749, "pid": 435, "tid": 435, "ts": 1776649790914441, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914453, "dur": 0, + "args": { + "External id": 31755, + "cbid": 200, "correlation": 31755 + } + }, + { + "ph": "f", "id": 31755, "pid": 435, "tid": 435, "ts": 1776649790914453, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914454, "dur": 0, + "args": { + "External id": 31756, + "cbid": 273, "correlation": 31756 + } + }, + { + "ph": "f", "id": 31756, "pid": 435, "tid": 435, "ts": 1776649790914454, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790914463, "dur": 15, + "args": { + "External id": 31757, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31757, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31757, "pid": 0, "tid": 7, "ts": 1776649790914463, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914455, "dur": 4, + "args": { + "External id": 31757, + "cbid": 211, "correlation": 31757 + } + }, + { + "ph": "s", "id": 31757, "pid": 435, "tid": 435, "ts": 1776649790914455, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790914498, "dur": 1, + "args": { + "External id": 3983, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31775, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31775, "pid": 0, "tid": 7, "ts": 1776649790914498, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914489, "dur": 6, + "args": { + "External id": 3983, + "cbid": 211, "correlation": 31775 + } + }, + { + "ph": "s", "id": 31775, "pid": 435, "tid": 435, "ts": 1776649790914489, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914501, "dur": 0, + "args": { + "External id": 31781, + "cbid": 200, "correlation": 31781 + } + }, + { + "ph": "f", "id": 31781, "pid": 435, "tid": 435, "ts": 1776649790914501, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914502, "dur": 0, + "args": { + "External id": 31782, + "cbid": 273, "correlation": 31782 + } + }, + { + "ph": "f", "id": 31782, "pid": 435, "tid": 435, "ts": 1776649790914502, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790914511, "dur": 16, + "args": { + "External id": 31783, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31783, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31783, "pid": 0, "tid": 7, "ts": 1776649790914511, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914503, "dur": 4, + "args": { + "External id": 31783, + "cbid": 211, "correlation": 31783 + } + }, + { + "ph": "s", "id": 31783, "pid": 435, "tid": 435, "ts": 1776649790914503, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790914546, "dur": 1, + "args": { + "External id": 3988, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31801, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31801, "pid": 0, "tid": 7, "ts": 1776649790914546, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914537, "dur": 6, + "args": { + "External id": 3988, + "cbid": 211, "correlation": 31801 + } + }, + { + "ph": "s", "id": 31801, "pid": 435, "tid": 435, "ts": 1776649790914537, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914549, "dur": 0, + "args": { + "External id": 31807, + "cbid": 200, "correlation": 31807 + } + }, + { + "ph": "f", "id": 31807, "pid": 435, "tid": 435, "ts": 1776649790914549, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914550, "dur": 0, + "args": { + "External id": 31808, + "cbid": 273, "correlation": 31808 + } + }, + { + "ph": "f", "id": 31808, "pid": 435, "tid": 435, "ts": 1776649790914550, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790914559, "dur": 16, + "args": { + "External id": 31809, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31809, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31809, "pid": 0, "tid": 7, "ts": 1776649790914559, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914551, "dur": 5, + "args": { + "External id": 31809, + "cbid": 211, "correlation": 31809 + } + }, + { + "ph": "s", "id": 31809, "pid": 435, "tid": 435, "ts": 1776649790914551, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790914595, "dur": 1, + "args": { + "External id": 3993, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31827, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31827, "pid": 0, "tid": 7, "ts": 1776649790914595, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914585, "dur": 7, + "args": { + "External id": 3993, + "cbid": 211, "correlation": 31827 + } + }, + { + "ph": "s", "id": 31827, "pid": 435, "tid": 435, "ts": 1776649790914585, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914598, "dur": 0, + "args": { + "External id": 31833, + "cbid": 200, "correlation": 31833 + } + }, + { + "ph": "f", "id": 31833, "pid": 435, "tid": 435, "ts": 1776649790914598, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914599, "dur": 0, + "args": { + "External id": 31834, + "cbid": 273, "correlation": 31834 + } + }, + { + "ph": "f", "id": 31834, "pid": 435, "tid": 435, "ts": 1776649790914599, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790914607, "dur": 15, + "args": { + "External id": 31835, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31835, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31835, "pid": 0, "tid": 7, "ts": 1776649790914607, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914600, "dur": 4, + "args": { + "External id": 31835, + "cbid": 211, "correlation": 31835 + } + }, + { + "ph": "s", "id": 31835, "pid": 435, "tid": 435, "ts": 1776649790914600, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790914643, "dur": 1, + "args": { + "External id": 3998, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31853, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31853, "pid": 0, "tid": 7, "ts": 1776649790914643, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914634, "dur": 6, + "args": { + "External id": 3998, + "cbid": 211, "correlation": 31853 + } + }, + { + "ph": "s", "id": 31853, "pid": 435, "tid": 435, "ts": 1776649790914634, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914646, "dur": 0, + "args": { + "External id": 31859, + "cbid": 200, "correlation": 31859 + } + }, + { + "ph": "f", "id": 31859, "pid": 435, "tid": 435, "ts": 1776649790914646, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914647, "dur": 0, + "args": { + "External id": 31860, + "cbid": 273, "correlation": 31860 + } + }, + { + "ph": "f", "id": 31860, "pid": 435, "tid": 435, "ts": 1776649790914647, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790914656, "dur": 18, + "args": { + "External id": 31861, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31861, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31861, "pid": 0, "tid": 7, "ts": 1776649790914656, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914648, "dur": 4, + "args": { + "External id": 31861, + "cbid": 211, "correlation": 31861 + } + }, + { + "ph": "s", "id": 31861, "pid": 435, "tid": 435, "ts": 1776649790914648, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790914692, "dur": 1, + "args": { + "External id": 4003, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31879, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31879, "pid": 0, "tid": 7, "ts": 1776649790914692, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914682, "dur": 6, + "args": { + "External id": 4003, + "cbid": 211, "correlation": 31879 + } + }, + { + "ph": "s", "id": 31879, "pid": 435, "tid": 435, "ts": 1776649790914682, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914695, "dur": 0, + "args": { + "External id": 31885, + "cbid": 200, "correlation": 31885 + } + }, + { + "ph": "f", "id": 31885, "pid": 435, "tid": 435, "ts": 1776649790914695, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914695, "dur": 0, + "args": { + "External id": 31886, + "cbid": 273, "correlation": 31886 + } + }, + { + "ph": "f", "id": 31886, "pid": 435, "tid": 435, "ts": 1776649790914695, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790914704, "dur": 16, + "args": { + "External id": 31887, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31887, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31887, "pid": 0, "tid": 7, "ts": 1776649790914704, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914696, "dur": 5, + "args": { + "External id": 31887, + "cbid": 211, "correlation": 31887 + } + }, + { + "ph": "s", "id": 31887, "pid": 435, "tid": 435, "ts": 1776649790914696, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790914740, "dur": 1, + "args": { + "External id": 4008, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31905, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31905, "pid": 0, "tid": 7, "ts": 1776649790914740, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914731, "dur": 6, + "args": { + "External id": 4008, + "cbid": 211, "correlation": 31905 + } + }, + { + "ph": "s", "id": 31905, "pid": 435, "tid": 435, "ts": 1776649790914731, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914743, "dur": 0, + "args": { + "External id": 31911, + "cbid": 200, "correlation": 31911 + } + }, + { + "ph": "f", "id": 31911, "pid": 435, "tid": 435, "ts": 1776649790914743, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914744, "dur": 0, + "args": { + "External id": 31912, + "cbid": 273, "correlation": 31912 + } + }, + { + "ph": "f", "id": 31912, "pid": 435, "tid": 435, "ts": 1776649790914744, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790914752, "dur": 16, + "args": { + "External id": 31913, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31913, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31913, "pid": 0, "tid": 7, "ts": 1776649790914752, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914745, "dur": 4, + "args": { + "External id": 31913, + "cbid": 211, "correlation": 31913 + } + }, + { + "ph": "s", "id": 31913, "pid": 435, "tid": 435, "ts": 1776649790914745, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790914788, "dur": 1, + "args": { + "External id": 4013, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31931, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31931, "pid": 0, "tid": 7, "ts": 1776649790914788, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914779, "dur": 6, + "args": { + "External id": 4013, + "cbid": 211, "correlation": 31931 + } + }, + { + "ph": "s", "id": 31931, "pid": 435, "tid": 435, "ts": 1776649790914779, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914791, "dur": 0, + "args": { + "External id": 31937, + "cbid": 200, "correlation": 31937 + } + }, + { + "ph": "f", "id": 31937, "pid": 435, "tid": 435, "ts": 1776649790914791, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914792, "dur": 0, + "args": { + "External id": 31938, + "cbid": 273, "correlation": 31938 + } + }, + { + "ph": "f", "id": 31938, "pid": 435, "tid": 435, "ts": 1776649790914792, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790914801, "dur": 15, + "args": { + "External id": 31939, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31939, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31939, "pid": 0, "tid": 7, "ts": 1776649790914801, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914793, "dur": 5, + "args": { + "External id": 31939, + "cbid": 211, "correlation": 31939 + } + }, + { + "ph": "s", "id": 31939, "pid": 435, "tid": 435, "ts": 1776649790914793, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790914836, "dur": 1, + "args": { + "External id": 4018, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31957, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31957, "pid": 0, "tid": 7, "ts": 1776649790914836, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914827, "dur": 6, + "args": { + "External id": 4018, + "cbid": 211, "correlation": 31957 + } + }, + { + "ph": "s", "id": 31957, "pid": 435, "tid": 435, "ts": 1776649790914827, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914839, "dur": 0, + "args": { + "External id": 31963, + "cbid": 200, "correlation": 31963 + } + }, + { + "ph": "f", "id": 31963, "pid": 435, "tid": 435, "ts": 1776649790914839, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914840, "dur": 0, + "args": { + "External id": 31964, + "cbid": 273, "correlation": 31964 + } + }, + { + "ph": "f", "id": 31964, "pid": 435, "tid": 435, "ts": 1776649790914840, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790914849, "dur": 16, + "args": { + "External id": 31965, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31965, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31965, "pid": 0, "tid": 7, "ts": 1776649790914849, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914841, "dur": 5, + "args": { + "External id": 31965, + "cbid": 211, "correlation": 31965 + } + }, + { + "ph": "s", "id": 31965, "pid": 435, "tid": 435, "ts": 1776649790914841, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790914884, "dur": 1, + "args": { + "External id": 4023, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31983, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31983, "pid": 0, "tid": 7, "ts": 1776649790914884, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914875, "dur": 6, + "args": { + "External id": 4023, + "cbid": 211, "correlation": 31983 + } + }, + { + "ph": "s", "id": 31983, "pid": 435, "tid": 435, "ts": 1776649790914875, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914887, "dur": 0, + "args": { + "External id": 31989, + "cbid": 200, "correlation": 31989 + } + }, + { + "ph": "f", "id": 31989, "pid": 435, "tid": 435, "ts": 1776649790914887, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914888, "dur": 0, + "args": { + "External id": 31990, + "cbid": 273, "correlation": 31990 + } + }, + { + "ph": "f", "id": 31990, "pid": 435, "tid": 435, "ts": 1776649790914888, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790914897, "dur": 15, + "args": { + "External id": 31991, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 31991, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 31991, "pid": 0, "tid": 7, "ts": 1776649790914897, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914889, "dur": 5, + "args": { + "External id": 31991, + "cbid": 211, "correlation": 31991 + } + }, + { + "ph": "s", "id": 31991, "pid": 435, "tid": 435, "ts": 1776649790914889, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790914933, "dur": 1, + "args": { + "External id": 4028, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32009, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32009, "pid": 0, "tid": 7, "ts": 1776649790914933, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914924, "dur": 6, + "args": { + "External id": 4028, + "cbid": 211, "correlation": 32009 + } + }, + { + "ph": "s", "id": 32009, "pid": 435, "tid": 435, "ts": 1776649790914924, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914936, "dur": 0, + "args": { + "External id": 32015, + "cbid": 200, "correlation": 32015 + } + }, + { + "ph": "f", "id": 32015, "pid": 435, "tid": 435, "ts": 1776649790914936, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914936, "dur": 0, + "args": { + "External id": 32016, + "cbid": 273, "correlation": 32016 + } + }, + { + "ph": "f", "id": 32016, "pid": 435, "tid": 435, "ts": 1776649790914936, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790914945, "dur": 15, + "args": { + "External id": 32017, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32017, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32017, "pid": 0, "tid": 7, "ts": 1776649790914945, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914937, "dur": 4, + "args": { + "External id": 32017, + "cbid": 211, "correlation": 32017 + } + }, + { + "ph": "s", "id": 32017, "pid": 435, "tid": 435, "ts": 1776649790914937, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790914981, "dur": 1, + "args": { + "External id": 4033, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32035, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32035, "pid": 0, "tid": 7, "ts": 1776649790914981, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914971, "dur": 6, + "args": { + "External id": 4033, + "cbid": 211, "correlation": 32035 + } + }, + { + "ph": "s", "id": 32035, "pid": 435, "tid": 435, "ts": 1776649790914971, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914984, "dur": 0, + "args": { + "External id": 32041, + "cbid": 200, "correlation": 32041 + } + }, + { + "ph": "f", "id": 32041, "pid": 435, "tid": 435, "ts": 1776649790914984, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790914984, "dur": 0, + "args": { + "External id": 32042, + "cbid": 273, "correlation": 32042 + } + }, + { + "ph": "f", "id": 32042, "pid": 435, "tid": 435, "ts": 1776649790914984, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790914993, "dur": 15, + "args": { + "External id": 32043, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32043, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32043, "pid": 0, "tid": 7, "ts": 1776649790914993, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790914985, "dur": 4, + "args": { + "External id": 32043, + "cbid": 211, "correlation": 32043 + } + }, + { + "ph": "s", "id": 32043, "pid": 435, "tid": 435, "ts": 1776649790914985, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790915029, "dur": 1, + "args": { + "External id": 4038, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32061, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32061, "pid": 0, "tid": 7, "ts": 1776649790915029, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915019, "dur": 6, + "args": { + "External id": 4038, + "cbid": 211, "correlation": 32061 + } + }, + { + "ph": "s", "id": 32061, "pid": 435, "tid": 435, "ts": 1776649790915019, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915032, "dur": 0, + "args": { + "External id": 32067, + "cbid": 200, "correlation": 32067 + } + }, + { + "ph": "f", "id": 32067, "pid": 435, "tid": 435, "ts": 1776649790915032, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915032, "dur": 0, + "args": { + "External id": 32068, + "cbid": 273, "correlation": 32068 + } + }, + { + "ph": "f", "id": 32068, "pid": 435, "tid": 435, "ts": 1776649790915032, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790915041, "dur": 16, + "args": { + "External id": 32069, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32069, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32069, "pid": 0, "tid": 7, "ts": 1776649790915041, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915033, "dur": 5, + "args": { + "External id": 32069, + "cbid": 211, "correlation": 32069 + } + }, + { + "ph": "s", "id": 32069, "pid": 435, "tid": 435, "ts": 1776649790915033, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790915076, "dur": 1, + "args": { + "External id": 4043, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32087, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32087, "pid": 0, "tid": 7, "ts": 1776649790915076, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915067, "dur": 6, + "args": { + "External id": 4043, + "cbid": 211, "correlation": 32087 + } + }, + { + "ph": "s", "id": 32087, "pid": 435, "tid": 435, "ts": 1776649790915067, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915080, "dur": 0, + "args": { + "External id": 32093, + "cbid": 200, "correlation": 32093 + } + }, + { + "ph": "f", "id": 32093, "pid": 435, "tid": 435, "ts": 1776649790915080, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915081, "dur": 1, + "args": { + "External id": 32094, + "cbid": 273, "correlation": 32094 + } + }, + { + "ph": "f", "id": 32094, "pid": 435, "tid": 435, "ts": 1776649790915081, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790915091, "dur": 18, + "args": { + "External id": 32095, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32095, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32095, "pid": 0, "tid": 7, "ts": 1776649790915091, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915083, "dur": 5, + "args": { + "External id": 32095, + "cbid": 211, "correlation": 32095 + } + }, + { + "ph": "s", "id": 32095, "pid": 435, "tid": 435, "ts": 1776649790915083, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790915130, "dur": 1, + "args": { + "External id": 4048, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32113, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32113, "pid": 0, "tid": 7, "ts": 1776649790915130, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915121, "dur": 7, + "args": { + "External id": 4048, + "cbid": 211, "correlation": 32113 + } + }, + { + "ph": "s", "id": 32113, "pid": 435, "tid": 435, "ts": 1776649790915121, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915133, "dur": 0, + "args": { + "External id": 32119, + "cbid": 200, "correlation": 32119 + } + }, + { + "ph": "f", "id": 32119, "pid": 435, "tid": 435, "ts": 1776649790915133, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915134, "dur": 0, + "args": { + "External id": 32120, + "cbid": 273, "correlation": 32120 + } + }, + { + "ph": "f", "id": 32120, "pid": 435, "tid": 435, "ts": 1776649790915134, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790915143, "dur": 15, + "args": { + "External id": 32121, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32121, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32121, "pid": 0, "tid": 7, "ts": 1776649790915143, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915135, "dur": 4, + "args": { + "External id": 32121, + "cbid": 211, "correlation": 32121 + } + }, + { + "ph": "s", "id": 32121, "pid": 435, "tid": 435, "ts": 1776649790915135, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790915179, "dur": 1, + "args": { + "External id": 4053, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32139, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32139, "pid": 0, "tid": 7, "ts": 1776649790915179, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915169, "dur": 6, + "args": { + "External id": 4053, + "cbid": 211, "correlation": 32139 + } + }, + { + "ph": "s", "id": 32139, "pid": 435, "tid": 435, "ts": 1776649790915169, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915182, "dur": 0, + "args": { + "External id": 32145, + "cbid": 200, "correlation": 32145 + } + }, + { + "ph": "f", "id": 32145, "pid": 435, "tid": 435, "ts": 1776649790915182, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915182, "dur": 0, + "args": { + "External id": 32146, + "cbid": 273, "correlation": 32146 + } + }, + { + "ph": "f", "id": 32146, "pid": 435, "tid": 435, "ts": 1776649790915182, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790915191, "dur": 16, + "args": { + "External id": 32147, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32147, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32147, "pid": 0, "tid": 7, "ts": 1776649790915191, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915183, "dur": 5, + "args": { + "External id": 32147, + "cbid": 211, "correlation": 32147 + } + }, + { + "ph": "s", "id": 32147, "pid": 435, "tid": 435, "ts": 1776649790915183, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790915227, "dur": 1, + "args": { + "External id": 4058, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32165, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32165, "pid": 0, "tid": 7, "ts": 1776649790915227, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915218, "dur": 6, + "args": { + "External id": 4058, + "cbid": 211, "correlation": 32165 + } + }, + { + "ph": "s", "id": 32165, "pid": 435, "tid": 435, "ts": 1776649790915218, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915230, "dur": 0, + "args": { + "External id": 32171, + "cbid": 200, "correlation": 32171 + } + }, + { + "ph": "f", "id": 32171, "pid": 435, "tid": 435, "ts": 1776649790915230, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915231, "dur": 0, + "args": { + "External id": 32172, + "cbid": 273, "correlation": 32172 + } + }, + { + "ph": "f", "id": 32172, "pid": 435, "tid": 435, "ts": 1776649790915231, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790915240, "dur": 16, + "args": { + "External id": 32173, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32173, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32173, "pid": 0, "tid": 7, "ts": 1776649790915240, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915232, "dur": 4, + "args": { + "External id": 32173, + "cbid": 211, "correlation": 32173 + } + }, + { + "ph": "s", "id": 32173, "pid": 435, "tid": 435, "ts": 1776649790915232, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790915275, "dur": 1, + "args": { + "External id": 4063, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32191, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32191, "pid": 0, "tid": 7, "ts": 1776649790915275, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915266, "dur": 6, + "args": { + "External id": 4063, + "cbid": 211, "correlation": 32191 + } + }, + { + "ph": "s", "id": 32191, "pid": 435, "tid": 435, "ts": 1776649790915266, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915278, "dur": 0, + "args": { + "External id": 32197, + "cbid": 200, "correlation": 32197 + } + }, + { + "ph": "f", "id": 32197, "pid": 435, "tid": 435, "ts": 1776649790915278, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915279, "dur": 0, + "args": { + "External id": 32198, + "cbid": 273, "correlation": 32198 + } + }, + { + "ph": "f", "id": 32198, "pid": 435, "tid": 435, "ts": 1776649790915279, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790915288, "dur": 16, + "args": { + "External id": 32199, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32199, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32199, "pid": 0, "tid": 7, "ts": 1776649790915288, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915280, "dur": 5, + "args": { + "External id": 32199, + "cbid": 211, "correlation": 32199 + } + }, + { + "ph": "s", "id": 32199, "pid": 435, "tid": 435, "ts": 1776649790915280, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790915324, "dur": 1, + "args": { + "External id": 4068, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32217, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32217, "pid": 0, "tid": 7, "ts": 1776649790915324, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915315, "dur": 7, + "args": { + "External id": 4068, + "cbid": 211, "correlation": 32217 + } + }, + { + "ph": "s", "id": 32217, "pid": 435, "tid": 435, "ts": 1776649790915315, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915327, "dur": 0, + "args": { + "External id": 32223, + "cbid": 200, "correlation": 32223 + } + }, + { + "ph": "f", "id": 32223, "pid": 435, "tid": 435, "ts": 1776649790915327, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915328, "dur": 0, + "args": { + "External id": 32224, + "cbid": 273, "correlation": 32224 + } + }, + { + "ph": "f", "id": 32224, "pid": 435, "tid": 435, "ts": 1776649790915328, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790915337, "dur": 15, + "args": { + "External id": 32225, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32225, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32225, "pid": 0, "tid": 7, "ts": 1776649790915337, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915329, "dur": 5, + "args": { + "External id": 32225, + "cbid": 211, "correlation": 32225 + } + }, + { + "ph": "s", "id": 32225, "pid": 435, "tid": 435, "ts": 1776649790915329, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790915373, "dur": 1, + "args": { + "External id": 4073, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32243, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32243, "pid": 0, "tid": 7, "ts": 1776649790915373, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915363, "dur": 6, + "args": { + "External id": 4073, + "cbid": 211, "correlation": 32243 + } + }, + { + "ph": "s", "id": 32243, "pid": 435, "tid": 435, "ts": 1776649790915363, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915376, "dur": 0, + "args": { + "External id": 32249, + "cbid": 200, "correlation": 32249 + } + }, + { + "ph": "f", "id": 32249, "pid": 435, "tid": 435, "ts": 1776649790915376, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915377, "dur": 0, + "args": { + "External id": 32250, + "cbid": 273, "correlation": 32250 + } + }, + { + "ph": "f", "id": 32250, "pid": 435, "tid": 435, "ts": 1776649790915377, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790915386, "dur": 15, + "args": { + "External id": 32251, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32251, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32251, "pid": 0, "tid": 7, "ts": 1776649790915386, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915378, "dur": 5, + "args": { + "External id": 32251, + "cbid": 211, "correlation": 32251 + } + }, + { + "ph": "s", "id": 32251, "pid": 435, "tid": 435, "ts": 1776649790915378, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790915421, "dur": 1, + "args": { + "External id": 4078, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32269, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32269, "pid": 0, "tid": 7, "ts": 1776649790915421, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915412, "dur": 6, + "args": { + "External id": 4078, + "cbid": 211, "correlation": 32269 + } + }, + { + "ph": "s", "id": 32269, "pid": 435, "tid": 435, "ts": 1776649790915412, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915424, "dur": 0, + "args": { + "External id": 32275, + "cbid": 200, "correlation": 32275 + } + }, + { + "ph": "f", "id": 32275, "pid": 435, "tid": 435, "ts": 1776649790915424, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915425, "dur": 0, + "args": { + "External id": 32276, + "cbid": 273, "correlation": 32276 + } + }, + { + "ph": "f", "id": 32276, "pid": 435, "tid": 435, "ts": 1776649790915425, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790915434, "dur": 15, + "args": { + "External id": 32277, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32277, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32277, "pid": 0, "tid": 7, "ts": 1776649790915434, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915426, "dur": 5, + "args": { + "External id": 32277, + "cbid": 211, "correlation": 32277 + } + }, + { + "ph": "s", "id": 32277, "pid": 435, "tid": 435, "ts": 1776649790915426, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790915469, "dur": 1, + "args": { + "External id": 4083, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32295, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32295, "pid": 0, "tid": 7, "ts": 1776649790915469, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915460, "dur": 6, + "args": { + "External id": 4083, + "cbid": 211, "correlation": 32295 + } + }, + { + "ph": "s", "id": 32295, "pid": 435, "tid": 435, "ts": 1776649790915460, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915472, "dur": 0, + "args": { + "External id": 32301, + "cbid": 200, "correlation": 32301 + } + }, + { + "ph": "f", "id": 32301, "pid": 435, "tid": 435, "ts": 1776649790915472, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915473, "dur": 0, + "args": { + "External id": 32302, + "cbid": 273, "correlation": 32302 + } + }, + { + "ph": "f", "id": 32302, "pid": 435, "tid": 435, "ts": 1776649790915473, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790915482, "dur": 17, + "args": { + "External id": 32303, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32303, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32303, "pid": 0, "tid": 7, "ts": 1776649790915482, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915474, "dur": 4, + "args": { + "External id": 32303, + "cbid": 211, "correlation": 32303 + } + }, + { + "ph": "s", "id": 32303, "pid": 435, "tid": 435, "ts": 1776649790915474, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790915516, "dur": 1, + "args": { + "External id": 4088, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32321, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32321, "pid": 0, "tid": 7, "ts": 1776649790915516, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915507, "dur": 6, + "args": { + "External id": 4088, + "cbid": 211, "correlation": 32321 + } + }, + { + "ph": "s", "id": 32321, "pid": 435, "tid": 435, "ts": 1776649790915507, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915519, "dur": 0, + "args": { + "External id": 32327, + "cbid": 200, "correlation": 32327 + } + }, + { + "ph": "f", "id": 32327, "pid": 435, "tid": 435, "ts": 1776649790915519, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915520, "dur": 0, + "args": { + "External id": 32328, + "cbid": 273, "correlation": 32328 + } + }, + { + "ph": "f", "id": 32328, "pid": 435, "tid": 435, "ts": 1776649790915520, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790915529, "dur": 17, + "args": { + "External id": 32329, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32329, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32329, "pid": 0, "tid": 7, "ts": 1776649790915529, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915521, "dur": 5, + "args": { + "External id": 32329, + "cbid": 211, "correlation": 32329 + } + }, + { + "ph": "s", "id": 32329, "pid": 435, "tid": 435, "ts": 1776649790915521, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790915565, "dur": 1, + "args": { + "External id": 4093, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32347, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32347, "pid": 0, "tid": 7, "ts": 1776649790915565, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915555, "dur": 6, + "args": { + "External id": 4093, + "cbid": 211, "correlation": 32347 + } + }, + { + "ph": "s", "id": 32347, "pid": 435, "tid": 435, "ts": 1776649790915555, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915568, "dur": 0, + "args": { + "External id": 32353, + "cbid": 200, "correlation": 32353 + } + }, + { + "ph": "f", "id": 32353, "pid": 435, "tid": 435, "ts": 1776649790915568, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915568, "dur": 0, + "args": { + "External id": 32354, + "cbid": 273, "correlation": 32354 + } + }, + { + "ph": "f", "id": 32354, "pid": 435, "tid": 435, "ts": 1776649790915568, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790915577, "dur": 15, + "args": { + "External id": 32355, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32355, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32355, "pid": 0, "tid": 7, "ts": 1776649790915577, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915569, "dur": 5, + "args": { + "External id": 32355, + "cbid": 211, "correlation": 32355 + } + }, + { + "ph": "s", "id": 32355, "pid": 435, "tid": 435, "ts": 1776649790915569, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790915623, "dur": 1, + "args": { + "External id": 4098, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32373, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32373, "pid": 0, "tid": 7, "ts": 1776649790915623, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915614, "dur": 6, + "args": { + "External id": 4098, + "cbid": 211, "correlation": 32373 + } + }, + { + "ph": "s", "id": 32373, "pid": 435, "tid": 435, "ts": 1776649790915614, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915626, "dur": 0, + "args": { + "External id": 32379, + "cbid": 200, "correlation": 32379 + } + }, + { + "ph": "f", "id": 32379, "pid": 435, "tid": 435, "ts": 1776649790915626, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915627, "dur": 0, + "args": { + "External id": 32380, + "cbid": 273, "correlation": 32380 + } + }, + { + "ph": "f", "id": 32380, "pid": 435, "tid": 435, "ts": 1776649790915627, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790915636, "dur": 16, + "args": { + "External id": 32381, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32381, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32381, "pid": 0, "tid": 7, "ts": 1776649790915636, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915628, "dur": 4, + "args": { + "External id": 32381, + "cbid": 211, "correlation": 32381 + } + }, + { + "ph": "s", "id": 32381, "pid": 435, "tid": 435, "ts": 1776649790915628, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790915672, "dur": 1, + "args": { + "External id": 4103, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32399, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32399, "pid": 0, "tid": 7, "ts": 1776649790915672, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915663, "dur": 6, + "args": { + "External id": 4103, + "cbid": 211, "correlation": 32399 + } + }, + { + "ph": "s", "id": 32399, "pid": 435, "tid": 435, "ts": 1776649790915663, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915675, "dur": 0, + "args": { + "External id": 32405, + "cbid": 200, "correlation": 32405 + } + }, + { + "ph": "f", "id": 32405, "pid": 435, "tid": 435, "ts": 1776649790915675, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915676, "dur": 0, + "args": { + "External id": 32406, + "cbid": 273, "correlation": 32406 + } + }, + { + "ph": "f", "id": 32406, "pid": 435, "tid": 435, "ts": 1776649790915676, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790915685, "dur": 16, + "args": { + "External id": 32407, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32407, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32407, "pid": 0, "tid": 7, "ts": 1776649790915685, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915677, "dur": 5, + "args": { + "External id": 32407, + "cbid": 211, "correlation": 32407 + } + }, + { + "ph": "s", "id": 32407, "pid": 435, "tid": 435, "ts": 1776649790915677, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790915720, "dur": 1, + "args": { + "External id": 4108, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32425, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32425, "pid": 0, "tid": 7, "ts": 1776649790915720, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915711, "dur": 6, + "args": { + "External id": 4108, + "cbid": 211, "correlation": 32425 + } + }, + { + "ph": "s", "id": 32425, "pid": 435, "tid": 435, "ts": 1776649790915711, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915724, "dur": 0, + "args": { + "External id": 32431, + "cbid": 200, "correlation": 32431 + } + }, + { + "ph": "f", "id": 32431, "pid": 435, "tid": 435, "ts": 1776649790915724, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915724, "dur": 0, + "args": { + "External id": 32432, + "cbid": 273, "correlation": 32432 + } + }, + { + "ph": "f", "id": 32432, "pid": 435, "tid": 435, "ts": 1776649790915724, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790915734, "dur": 15, + "args": { + "External id": 32433, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32433, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32433, "pid": 0, "tid": 7, "ts": 1776649790915734, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915726, "dur": 5, + "args": { + "External id": 32433, + "cbid": 211, "correlation": 32433 + } + }, + { + "ph": "s", "id": 32433, "pid": 435, "tid": 435, "ts": 1776649790915726, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790915770, "dur": 1, + "args": { + "External id": 4113, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32451, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32451, "pid": 0, "tid": 7, "ts": 1776649790915770, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915761, "dur": 6, + "args": { + "External id": 4113, + "cbid": 211, "correlation": 32451 + } + }, + { + "ph": "s", "id": 32451, "pid": 435, "tid": 435, "ts": 1776649790915761, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915773, "dur": 0, + "args": { + "External id": 32457, + "cbid": 200, "correlation": 32457 + } + }, + { + "ph": "f", "id": 32457, "pid": 435, "tid": 435, "ts": 1776649790915773, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915774, "dur": 0, + "args": { + "External id": 32458, + "cbid": 273, "correlation": 32458 + } + }, + { + "ph": "f", "id": 32458, "pid": 435, "tid": 435, "ts": 1776649790915774, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790915783, "dur": 15, + "args": { + "External id": 32459, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32459, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32459, "pid": 0, "tid": 7, "ts": 1776649790915783, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915775, "dur": 5, + "args": { + "External id": 32459, + "cbid": 211, "correlation": 32459 + } + }, + { + "ph": "s", "id": 32459, "pid": 435, "tid": 435, "ts": 1776649790915775, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790915818, "dur": 1, + "args": { + "External id": 4118, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32477, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32477, "pid": 0, "tid": 7, "ts": 1776649790915818, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915809, "dur": 6, + "args": { + "External id": 4118, + "cbid": 211, "correlation": 32477 + } + }, + { + "ph": "s", "id": 32477, "pid": 435, "tid": 435, "ts": 1776649790915809, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915822, "dur": 0, + "args": { + "External id": 32483, + "cbid": 200, "correlation": 32483 + } + }, + { + "ph": "f", "id": 32483, "pid": 435, "tid": 435, "ts": 1776649790915822, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915822, "dur": 0, + "args": { + "External id": 32484, + "cbid": 273, "correlation": 32484 + } + }, + { + "ph": "f", "id": 32484, "pid": 435, "tid": 435, "ts": 1776649790915822, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790915831, "dur": 15, + "args": { + "External id": 32485, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32485, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32485, "pid": 0, "tid": 7, "ts": 1776649790915831, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915823, "dur": 4, + "args": { + "External id": 32485, + "cbid": 211, "correlation": 32485 + } + }, + { + "ph": "s", "id": 32485, "pid": 435, "tid": 435, "ts": 1776649790915823, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790915867, "dur": 1, + "args": { + "External id": 4123, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32503, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32503, "pid": 0, "tid": 7, "ts": 1776649790915867, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915858, "dur": 6, + "args": { + "External id": 4123, + "cbid": 211, "correlation": 32503 + } + }, + { + "ph": "s", "id": 32503, "pid": 435, "tid": 435, "ts": 1776649790915858, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915870, "dur": 0, + "args": { + "External id": 32509, + "cbid": 200, "correlation": 32509 + } + }, + { + "ph": "f", "id": 32509, "pid": 435, "tid": 435, "ts": 1776649790915870, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915871, "dur": 0, + "args": { + "External id": 32510, + "cbid": 273, "correlation": 32510 + } + }, + { + "ph": "f", "id": 32510, "pid": 435, "tid": 435, "ts": 1776649790915871, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790915880, "dur": 15, + "args": { + "External id": 32511, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32511, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32511, "pid": 0, "tid": 7, "ts": 1776649790915880, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915872, "dur": 5, + "args": { + "External id": 32511, + "cbid": 211, "correlation": 32511 + } + }, + { + "ph": "s", "id": 32511, "pid": 435, "tid": 435, "ts": 1776649790915872, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790915915, "dur": 1, + "args": { + "External id": 4128, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32529, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32529, "pid": 0, "tid": 7, "ts": 1776649790915915, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915906, "dur": 6, + "args": { + "External id": 4128, + "cbid": 211, "correlation": 32529 + } + }, + { + "ph": "s", "id": 32529, "pid": 435, "tid": 435, "ts": 1776649790915906, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915919, "dur": 0, + "args": { + "External id": 32535, + "cbid": 200, "correlation": 32535 + } + }, + { + "ph": "f", "id": 32535, "pid": 435, "tid": 435, "ts": 1776649790915919, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915919, "dur": 0, + "args": { + "External id": 32536, + "cbid": 273, "correlation": 32536 + } + }, + { + "ph": "f", "id": 32536, "pid": 435, "tid": 435, "ts": 1776649790915919, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790915928, "dur": 17, + "args": { + "External id": 32537, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32537, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32537, "pid": 0, "tid": 7, "ts": 1776649790915928, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915920, "dur": 4, + "args": { + "External id": 32537, + "cbid": 211, "correlation": 32537 + } + }, + { + "ph": "s", "id": 32537, "pid": 435, "tid": 435, "ts": 1776649790915920, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790915964, "dur": 1, + "args": { + "External id": 4133, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32555, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32555, "pid": 0, "tid": 7, "ts": 1776649790915964, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915955, "dur": 6, + "args": { + "External id": 4133, + "cbid": 211, "correlation": 32555 + } + }, + { + "ph": "s", "id": 32555, "pid": 435, "tid": 435, "ts": 1776649790915955, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915967, "dur": 0, + "args": { + "External id": 32561, + "cbid": 200, "correlation": 32561 + } + }, + { + "ph": "f", "id": 32561, "pid": 435, "tid": 435, "ts": 1776649790915967, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790915968, "dur": 0, + "args": { + "External id": 32562, + "cbid": 273, "correlation": 32562 + } + }, + { + "ph": "f", "id": 32562, "pid": 435, "tid": 435, "ts": 1776649790915968, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790915977, "dur": 16, + "args": { + "External id": 32563, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32563, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32563, "pid": 0, "tid": 7, "ts": 1776649790915977, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790915969, "dur": 5, + "args": { + "External id": 32563, + "cbid": 211, "correlation": 32563 + } + }, + { + "ph": "s", "id": 32563, "pid": 435, "tid": 435, "ts": 1776649790915969, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790916013, "dur": 1, + "args": { + "External id": 4138, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32581, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32581, "pid": 0, "tid": 7, "ts": 1776649790916013, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916003, "dur": 6, + "args": { + "External id": 4138, + "cbid": 211, "correlation": 32581 + } + }, + { + "ph": "s", "id": 32581, "pid": 435, "tid": 435, "ts": 1776649790916003, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916016, "dur": 0, + "args": { + "External id": 32587, + "cbid": 200, "correlation": 32587 + } + }, + { + "ph": "f", "id": 32587, "pid": 435, "tid": 435, "ts": 1776649790916016, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916017, "dur": 0, + "args": { + "External id": 32588, + "cbid": 273, "correlation": 32588 + } + }, + { + "ph": "f", "id": 32588, "pid": 435, "tid": 435, "ts": 1776649790916017, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790916026, "dur": 16, + "args": { + "External id": 32589, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32589, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32589, "pid": 0, "tid": 7, "ts": 1776649790916026, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916018, "dur": 4, + "args": { + "External id": 32589, + "cbid": 211, "correlation": 32589 + } + }, + { + "ph": "s", "id": 32589, "pid": 435, "tid": 435, "ts": 1776649790916018, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790916062, "dur": 1, + "args": { + "External id": 4143, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32607, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32607, "pid": 0, "tid": 7, "ts": 1776649790916062, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916052, "dur": 7, + "args": { + "External id": 4143, + "cbid": 211, "correlation": 32607 + } + }, + { + "ph": "s", "id": 32607, "pid": 435, "tid": 435, "ts": 1776649790916052, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916065, "dur": 0, + "args": { + "External id": 32613, + "cbid": 200, "correlation": 32613 + } + }, + { + "ph": "f", "id": 32613, "pid": 435, "tid": 435, "ts": 1776649790916065, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916066, "dur": 0, + "args": { + "External id": 32614, + "cbid": 273, "correlation": 32614 + } + }, + { + "ph": "f", "id": 32614, "pid": 435, "tid": 435, "ts": 1776649790916066, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790916075, "dur": 16, + "args": { + "External id": 32615, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32615, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32615, "pid": 0, "tid": 7, "ts": 1776649790916075, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916067, "dur": 4, + "args": { + "External id": 32615, + "cbid": 211, "correlation": 32615 + } + }, + { + "ph": "s", "id": 32615, "pid": 435, "tid": 435, "ts": 1776649790916067, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790916110, "dur": 1, + "args": { + "External id": 4148, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32633, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32633, "pid": 0, "tid": 7, "ts": 1776649790916110, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916101, "dur": 6, + "args": { + "External id": 4148, + "cbid": 211, "correlation": 32633 + } + }, + { + "ph": "s", "id": 32633, "pid": 435, "tid": 435, "ts": 1776649790916101, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916114, "dur": 0, + "args": { + "External id": 32639, + "cbid": 200, "correlation": 32639 + } + }, + { + "ph": "f", "id": 32639, "pid": 435, "tid": 435, "ts": 1776649790916114, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916114, "dur": 0, + "args": { + "External id": 32640, + "cbid": 273, "correlation": 32640 + } + }, + { + "ph": "f", "id": 32640, "pid": 435, "tid": 435, "ts": 1776649790916114, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790916123, "dur": 15, + "args": { + "External id": 32641, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32641, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32641, "pid": 0, "tid": 7, "ts": 1776649790916123, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916115, "dur": 4, + "args": { + "External id": 32641, + "cbid": 211, "correlation": 32641 + } + }, + { + "ph": "s", "id": 32641, "pid": 435, "tid": 435, "ts": 1776649790916115, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790916158, "dur": 1, + "args": { + "External id": 4153, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32659, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32659, "pid": 0, "tid": 7, "ts": 1776649790916158, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916149, "dur": 6, + "args": { + "External id": 4153, + "cbid": 211, "correlation": 32659 + } + }, + { + "ph": "s", "id": 32659, "pid": 435, "tid": 435, "ts": 1776649790916149, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916161, "dur": 0, + "args": { + "External id": 32665, + "cbid": 200, "correlation": 32665 + } + }, + { + "ph": "f", "id": 32665, "pid": 435, "tid": 435, "ts": 1776649790916161, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916162, "dur": 0, + "args": { + "External id": 32666, + "cbid": 273, "correlation": 32666 + } + }, + { + "ph": "f", "id": 32666, "pid": 435, "tid": 435, "ts": 1776649790916162, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790916171, "dur": 16, + "args": { + "External id": 32667, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32667, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32667, "pid": 0, "tid": 7, "ts": 1776649790916171, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916163, "dur": 5, + "args": { + "External id": 32667, + "cbid": 211, "correlation": 32667 + } + }, + { + "ph": "s", "id": 32667, "pid": 435, "tid": 435, "ts": 1776649790916163, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790916193, "dur": 3, + "args": { + "External id": 32675, + "cbid": 135, "correlation": 32675 + } + }, + { + "ph": "f", "id": 32675, "pid": 435, "tid": 435, "ts": 1776649790916193, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790916207, "dur": 5, + "args": { + "External id": 32680, + "cbid": 165, "correlation": 32680 + } + }, + { + "ph": "s", "id": 32680, "pid": 435, "tid": 435, "ts": 1776649790916207, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649790916233, "dur": 5, + "args": { + "External id": 32683, + "cbid": 139, "correlation": 32683 + } + }, + { + "ph": "f", "id": 32683, "pid": 435, "tid": 435, "ts": 1776649790916233, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790916293, "dur": 1, + "args": { + "External id": 4158, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32711, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32711, "pid": 0, "tid": 7, "ts": 1776649790916293, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916281, "dur": 9, + "args": { + "External id": 4158, + "cbid": 211, "correlation": 32711 + } + }, + { + "ph": "s", "id": 32711, "pid": 435, "tid": 435, "ts": 1776649790916281, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916298, "dur": 0, + "args": { + "External id": 32717, + "cbid": 200, "correlation": 32717 + } + }, + { + "ph": "f", "id": 32717, "pid": 435, "tid": 435, "ts": 1776649790916298, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916298, "dur": 1, + "args": { + "External id": 32718, + "cbid": 273, "correlation": 32718 + } + }, + { + "ph": "f", "id": 32718, "pid": 435, "tid": 435, "ts": 1776649790916298, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790916309, "dur": 16, + "args": { + "External id": 32719, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32719, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32719, "pid": 0, "tid": 7, "ts": 1776649790916309, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916300, "dur": 5, + "args": { + "External id": 32719, + "cbid": 211, "correlation": 32719 + } + }, + { + "ph": "s", "id": 32719, "pid": 435, "tid": 435, "ts": 1776649790916300, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790916346, "dur": 1, + "args": { + "External id": 4163, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32737, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32737, "pid": 0, "tid": 7, "ts": 1776649790916346, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916336, "dur": 6, + "args": { + "External id": 4163, + "cbid": 211, "correlation": 32737 + } + }, + { + "ph": "s", "id": 32737, "pid": 435, "tid": 435, "ts": 1776649790916336, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916349, "dur": 0, + "args": { + "External id": 32743, + "cbid": 200, "correlation": 32743 + } + }, + { + "ph": "f", "id": 32743, "pid": 435, "tid": 435, "ts": 1776649790916349, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916350, "dur": 0, + "args": { + "External id": 32744, + "cbid": 273, "correlation": 32744 + } + }, + { + "ph": "f", "id": 32744, "pid": 435, "tid": 435, "ts": 1776649790916350, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790916359, "dur": 15, + "args": { + "External id": 32745, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32745, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32745, "pid": 0, "tid": 7, "ts": 1776649790916359, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916351, "dur": 4, + "args": { + "External id": 32745, + "cbid": 211, "correlation": 32745 + } + }, + { + "ph": "s", "id": 32745, "pid": 435, "tid": 435, "ts": 1776649790916351, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790916395, "dur": 1, + "args": { + "External id": 4168, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32763, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32763, "pid": 0, "tid": 7, "ts": 1776649790916395, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916386, "dur": 6, + "args": { + "External id": 4168, + "cbid": 211, "correlation": 32763 + } + }, + { + "ph": "s", "id": 32763, "pid": 435, "tid": 435, "ts": 1776649790916386, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916398, "dur": 0, + "args": { + "External id": 32769, + "cbid": 200, "correlation": 32769 + } + }, + { + "ph": "f", "id": 32769, "pid": 435, "tid": 435, "ts": 1776649790916398, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916399, "dur": 0, + "args": { + "External id": 32770, + "cbid": 273, "correlation": 32770 + } + }, + { + "ph": "f", "id": 32770, "pid": 435, "tid": 435, "ts": 1776649790916399, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790916408, "dur": 16, + "args": { + "External id": 32771, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32771, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32771, "pid": 0, "tid": 7, "ts": 1776649790916408, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916400, "dur": 4, + "args": { + "External id": 32771, + "cbid": 211, "correlation": 32771 + } + }, + { + "ph": "s", "id": 32771, "pid": 435, "tid": 435, "ts": 1776649790916400, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790916443, "dur": 1, + "args": { + "External id": 4173, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32789, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32789, "pid": 0, "tid": 7, "ts": 1776649790916443, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916434, "dur": 6, + "args": { + "External id": 4173, + "cbid": 211, "correlation": 32789 + } + }, + { + "ph": "s", "id": 32789, "pid": 435, "tid": 435, "ts": 1776649790916434, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916447, "dur": 0, + "args": { + "External id": 32795, + "cbid": 200, "correlation": 32795 + } + }, + { + "ph": "f", "id": 32795, "pid": 435, "tid": 435, "ts": 1776649790916447, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916447, "dur": 0, + "args": { + "External id": 32796, + "cbid": 273, "correlation": 32796 + } + }, + { + "ph": "f", "id": 32796, "pid": 435, "tid": 435, "ts": 1776649790916447, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790916456, "dur": 16, + "args": { + "External id": 32797, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32797, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32797, "pid": 0, "tid": 7, "ts": 1776649790916456, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916448, "dur": 4, + "args": { + "External id": 32797, + "cbid": 211, "correlation": 32797 + } + }, + { + "ph": "s", "id": 32797, "pid": 435, "tid": 435, "ts": 1776649790916448, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790916492, "dur": 1, + "args": { + "External id": 4178, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32815, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32815, "pid": 0, "tid": 7, "ts": 1776649790916492, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916482, "dur": 7, + "args": { + "External id": 4178, + "cbid": 211, "correlation": 32815 + } + }, + { + "ph": "s", "id": 32815, "pid": 435, "tid": 435, "ts": 1776649790916482, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916495, "dur": 0, + "args": { + "External id": 32821, + "cbid": 200, "correlation": 32821 + } + }, + { + "ph": "f", "id": 32821, "pid": 435, "tid": 435, "ts": 1776649790916495, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916496, "dur": 0, + "args": { + "External id": 32822, + "cbid": 273, "correlation": 32822 + } + }, + { + "ph": "f", "id": 32822, "pid": 435, "tid": 435, "ts": 1776649790916496, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790916505, "dur": 16, + "args": { + "External id": 32823, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32823, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32823, "pid": 0, "tid": 7, "ts": 1776649790916505, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916497, "dur": 4, + "args": { + "External id": 32823, + "cbid": 211, "correlation": 32823 + } + }, + { + "ph": "s", "id": 32823, "pid": 435, "tid": 435, "ts": 1776649790916497, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790916540, "dur": 1, + "args": { + "External id": 4183, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32841, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32841, "pid": 0, "tid": 7, "ts": 1776649790916540, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916531, "dur": 6, + "args": { + "External id": 4183, + "cbid": 211, "correlation": 32841 + } + }, + { + "ph": "s", "id": 32841, "pid": 435, "tid": 435, "ts": 1776649790916531, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916544, "dur": 0, + "args": { + "External id": 32847, + "cbid": 200, "correlation": 32847 + } + }, + { + "ph": "f", "id": 32847, "pid": 435, "tid": 435, "ts": 1776649790916544, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916544, "dur": 0, + "args": { + "External id": 32848, + "cbid": 273, "correlation": 32848 + } + }, + { + "ph": "f", "id": 32848, "pid": 435, "tid": 435, "ts": 1776649790916544, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790916553, "dur": 16, + "args": { + "External id": 32849, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32849, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32849, "pid": 0, "tid": 7, "ts": 1776649790916553, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916545, "dur": 5, + "args": { + "External id": 32849, + "cbid": 211, "correlation": 32849 + } + }, + { + "ph": "s", "id": 32849, "pid": 435, "tid": 435, "ts": 1776649790916545, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790916589, "dur": 1, + "args": { + "External id": 4188, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32867, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32867, "pid": 0, "tid": 7, "ts": 1776649790916589, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916579, "dur": 6, + "args": { + "External id": 4188, + "cbid": 211, "correlation": 32867 + } + }, + { + "ph": "s", "id": 32867, "pid": 435, "tid": 435, "ts": 1776649790916579, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916592, "dur": 0, + "args": { + "External id": 32873, + "cbid": 200, "correlation": 32873 + } + }, + { + "ph": "f", "id": 32873, "pid": 435, "tid": 435, "ts": 1776649790916592, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916592, "dur": 0, + "args": { + "External id": 32874, + "cbid": 273, "correlation": 32874 + } + }, + { + "ph": "f", "id": 32874, "pid": 435, "tid": 435, "ts": 1776649790916592, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790916601, "dur": 16, + "args": { + "External id": 32875, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32875, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32875, "pid": 0, "tid": 7, "ts": 1776649790916601, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916594, "dur": 4, + "args": { + "External id": 32875, + "cbid": 211, "correlation": 32875 + } + }, + { + "ph": "s", "id": 32875, "pid": 435, "tid": 435, "ts": 1776649790916594, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790916637, "dur": 1, + "args": { + "External id": 4193, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32893, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32893, "pid": 0, "tid": 7, "ts": 1776649790916637, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916628, "dur": 6, + "args": { + "External id": 4193, + "cbid": 211, "correlation": 32893 + } + }, + { + "ph": "s", "id": 32893, "pid": 435, "tid": 435, "ts": 1776649790916628, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916640, "dur": 0, + "args": { + "External id": 32899, + "cbid": 200, "correlation": 32899 + } + }, + { + "ph": "f", "id": 32899, "pid": 435, "tid": 435, "ts": 1776649790916640, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916641, "dur": 0, + "args": { + "External id": 32900, + "cbid": 273, "correlation": 32900 + } + }, + { + "ph": "f", "id": 32900, "pid": 435, "tid": 435, "ts": 1776649790916641, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790916650, "dur": 16, + "args": { + "External id": 32901, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32901, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32901, "pid": 0, "tid": 7, "ts": 1776649790916650, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916642, "dur": 4, + "args": { + "External id": 32901, + "cbid": 211, "correlation": 32901 + } + }, + { + "ph": "s", "id": 32901, "pid": 435, "tid": 435, "ts": 1776649790916642, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790916685, "dur": 1, + "args": { + "External id": 4198, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32919, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32919, "pid": 0, "tid": 7, "ts": 1776649790916685, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916676, "dur": 6, + "args": { + "External id": 4198, + "cbid": 211, "correlation": 32919 + } + }, + { + "ph": "s", "id": 32919, "pid": 435, "tid": 435, "ts": 1776649790916676, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916688, "dur": 0, + "args": { + "External id": 32925, + "cbid": 200, "correlation": 32925 + } + }, + { + "ph": "f", "id": 32925, "pid": 435, "tid": 435, "ts": 1776649790916688, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916689, "dur": 0, + "args": { + "External id": 32926, + "cbid": 273, "correlation": 32926 + } + }, + { + "ph": "f", "id": 32926, "pid": 435, "tid": 435, "ts": 1776649790916689, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790916698, "dur": 15, + "args": { + "External id": 32927, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32927, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32927, "pid": 0, "tid": 7, "ts": 1776649790916698, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916690, "dur": 4, + "args": { + "External id": 32927, + "cbid": 211, "correlation": 32927 + } + }, + { + "ph": "s", "id": 32927, "pid": 435, "tid": 435, "ts": 1776649790916690, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790916733, "dur": 1, + "args": { + "External id": 4203, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32945, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32945, "pid": 0, "tid": 7, "ts": 1776649790916733, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916724, "dur": 6, + "args": { + "External id": 4203, + "cbid": 211, "correlation": 32945 + } + }, + { + "ph": "s", "id": 32945, "pid": 435, "tid": 435, "ts": 1776649790916724, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916736, "dur": 0, + "args": { + "External id": 32951, + "cbid": 200, "correlation": 32951 + } + }, + { + "ph": "f", "id": 32951, "pid": 435, "tid": 435, "ts": 1776649790916736, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916737, "dur": 0, + "args": { + "External id": 32952, + "cbid": 273, "correlation": 32952 + } + }, + { + "ph": "f", "id": 32952, "pid": 435, "tid": 435, "ts": 1776649790916737, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790916746, "dur": 15, + "args": { + "External id": 32953, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32953, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32953, "pid": 0, "tid": 7, "ts": 1776649790916746, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916738, "dur": 4, + "args": { + "External id": 32953, + "cbid": 211, "correlation": 32953 + } + }, + { + "ph": "s", "id": 32953, "pid": 435, "tid": 435, "ts": 1776649790916738, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790916761, "dur": 6, + "args": { + "External id": 32956, + "cbid": 165, "correlation": 32956 + } + }, + { + "ph": "s", "id": 32956, "pid": 435, "tid": 435, "ts": 1776649790916761, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790916788, "dur": 3, + "args": { + "External id": 32965, + "cbid": 135, "correlation": 32965 + } + }, + { + "ph": "f", "id": 32965, "pid": 435, "tid": 435, "ts": 1776649790916788, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790916836, "dur": 1, + "args": { + "External id": 4208, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32985, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32985, "pid": 0, "tid": 7, "ts": 1776649790916836, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916825, "dur": 8, + "args": { + "External id": 4208, + "cbid": 211, "correlation": 32985 + } + }, + { + "ph": "s", "id": 32985, "pid": 435, "tid": 435, "ts": 1776649790916825, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916840, "dur": 0, + "args": { + "External id": 32991, + "cbid": 200, "correlation": 32991 + } + }, + { + "ph": "f", "id": 32991, "pid": 435, "tid": 435, "ts": 1776649790916840, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916841, "dur": 0, + "args": { + "External id": 32992, + "cbid": 273, "correlation": 32992 + } + }, + { + "ph": "f", "id": 32992, "pid": 435, "tid": 435, "ts": 1776649790916841, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790916850, "dur": 17, + "args": { + "External id": 32993, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 32993, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 32993, "pid": 0, "tid": 7, "ts": 1776649790916850, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916842, "dur": 5, + "args": { + "External id": 32993, + "cbid": 211, "correlation": 32993 + } + }, + { + "ph": "s", "id": 32993, "pid": 435, "tid": 435, "ts": 1776649790916842, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790916888, "dur": 1, + "args": { + "External id": 4213, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33011, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33011, "pid": 0, "tid": 7, "ts": 1776649790916888, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916879, "dur": 6, + "args": { + "External id": 4213, + "cbid": 211, "correlation": 33011 + } + }, + { + "ph": "s", "id": 33011, "pid": 435, "tid": 435, "ts": 1776649790916879, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916891, "dur": 0, + "args": { + "External id": 33017, + "cbid": 200, "correlation": 33017 + } + }, + { + "ph": "f", "id": 33017, "pid": 435, "tid": 435, "ts": 1776649790916891, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916892, "dur": 0, + "args": { + "External id": 33018, + "cbid": 273, "correlation": 33018 + } + }, + { + "ph": "f", "id": 33018, "pid": 435, "tid": 435, "ts": 1776649790916892, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790916901, "dur": 16, + "args": { + "External id": 33019, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33019, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33019, "pid": 0, "tid": 7, "ts": 1776649790916901, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916893, "dur": 5, + "args": { + "External id": 33019, + "cbid": 211, "correlation": 33019 + } + }, + { + "ph": "s", "id": 33019, "pid": 435, "tid": 435, "ts": 1776649790916893, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790916937, "dur": 1, + "args": { + "External id": 4218, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33037, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33037, "pid": 0, "tid": 7, "ts": 1776649790916937, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916928, "dur": 7, + "args": { + "External id": 4218, + "cbid": 211, "correlation": 33037 + } + }, + { + "ph": "s", "id": 33037, "pid": 435, "tid": 435, "ts": 1776649790916928, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916941, "dur": 0, + "args": { + "External id": 33043, + "cbid": 200, "correlation": 33043 + } + }, + { + "ph": "f", "id": 33043, "pid": 435, "tid": 435, "ts": 1776649790916941, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916941, "dur": 0, + "args": { + "External id": 33044, + "cbid": 273, "correlation": 33044 + } + }, + { + "ph": "f", "id": 33044, "pid": 435, "tid": 435, "ts": 1776649790916941, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790916950, "dur": 16, + "args": { + "External id": 33045, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33045, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33045, "pid": 0, "tid": 7, "ts": 1776649790916950, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916942, "dur": 5, + "args": { + "External id": 33045, + "cbid": 211, "correlation": 33045 + } + }, + { + "ph": "s", "id": 33045, "pid": 435, "tid": 435, "ts": 1776649790916942, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790916986, "dur": 1, + "args": { + "External id": 4223, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33063, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33063, "pid": 0, "tid": 7, "ts": 1776649790916986, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916976, "dur": 6, + "args": { + "External id": 4223, + "cbid": 211, "correlation": 33063 + } + }, + { + "ph": "s", "id": 33063, "pid": 435, "tid": 435, "ts": 1776649790916976, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916989, "dur": 0, + "args": { + "External id": 33069, + "cbid": 200, "correlation": 33069 + } + }, + { + "ph": "f", "id": 33069, "pid": 435, "tid": 435, "ts": 1776649790916989, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790916990, "dur": 0, + "args": { + "External id": 33070, + "cbid": 273, "correlation": 33070 + } + }, + { + "ph": "f", "id": 33070, "pid": 435, "tid": 435, "ts": 1776649790916990, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790916999, "dur": 15, + "args": { + "External id": 33071, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33071, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33071, "pid": 0, "tid": 7, "ts": 1776649790916999, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790916991, "dur": 4, + "args": { + "External id": 33071, + "cbid": 211, "correlation": 33071 + } + }, + { + "ph": "s", "id": 33071, "pid": 435, "tid": 435, "ts": 1776649790916991, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790917034, "dur": 1, + "args": { + "External id": 4228, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33089, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33089, "pid": 0, "tid": 7, "ts": 1776649790917034, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917025, "dur": 6, + "args": { + "External id": 4228, + "cbid": 211, "correlation": 33089 + } + }, + { + "ph": "s", "id": 33089, "pid": 435, "tid": 435, "ts": 1776649790917025, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917037, "dur": 0, + "args": { + "External id": 33095, + "cbid": 200, "correlation": 33095 + } + }, + { + "ph": "f", "id": 33095, "pid": 435, "tid": 435, "ts": 1776649790917037, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917038, "dur": 0, + "args": { + "External id": 33096, + "cbid": 273, "correlation": 33096 + } + }, + { + "ph": "f", "id": 33096, "pid": 435, "tid": 435, "ts": 1776649790917038, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790917047, "dur": 16, + "args": { + "External id": 33097, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33097, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33097, "pid": 0, "tid": 7, "ts": 1776649790917047, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917039, "dur": 4, + "args": { + "External id": 33097, + "cbid": 211, "correlation": 33097 + } + }, + { + "ph": "s", "id": 33097, "pid": 435, "tid": 435, "ts": 1776649790917039, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790917083, "dur": 1, + "args": { + "External id": 4233, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33115, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33115, "pid": 0, "tid": 7, "ts": 1776649790917083, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917073, "dur": 7, + "args": { + "External id": 4233, + "cbid": 211, "correlation": 33115 + } + }, + { + "ph": "s", "id": 33115, "pid": 435, "tid": 435, "ts": 1776649790917073, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917086, "dur": 0, + "args": { + "External id": 33121, + "cbid": 200, "correlation": 33121 + } + }, + { + "ph": "f", "id": 33121, "pid": 435, "tid": 435, "ts": 1776649790917086, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917086, "dur": 0, + "args": { + "External id": 33122, + "cbid": 273, "correlation": 33122 + } + }, + { + "ph": "f", "id": 33122, "pid": 435, "tid": 435, "ts": 1776649790917086, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790917095, "dur": 16, + "args": { + "External id": 33123, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33123, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33123, "pid": 0, "tid": 7, "ts": 1776649790917095, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917088, "dur": 4, + "args": { + "External id": 33123, + "cbid": 211, "correlation": 33123 + } + }, + { + "ph": "s", "id": 33123, "pid": 435, "tid": 435, "ts": 1776649790917088, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790917131, "dur": 1, + "args": { + "External id": 4238, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33141, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33141, "pid": 0, "tid": 7, "ts": 1776649790917131, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917122, "dur": 6, + "args": { + "External id": 4238, + "cbid": 211, "correlation": 33141 + } + }, + { + "ph": "s", "id": 33141, "pid": 435, "tid": 435, "ts": 1776649790917122, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917134, "dur": 0, + "args": { + "External id": 33147, + "cbid": 200, "correlation": 33147 + } + }, + { + "ph": "f", "id": 33147, "pid": 435, "tid": 435, "ts": 1776649790917134, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917135, "dur": 0, + "args": { + "External id": 33148, + "cbid": 273, "correlation": 33148 + } + }, + { + "ph": "f", "id": 33148, "pid": 435, "tid": 435, "ts": 1776649790917135, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790917144, "dur": 16, + "args": { + "External id": 33149, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33149, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33149, "pid": 0, "tid": 7, "ts": 1776649790917144, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917136, "dur": 4, + "args": { + "External id": 33149, + "cbid": 211, "correlation": 33149 + } + }, + { + "ph": "s", "id": 33149, "pid": 435, "tid": 435, "ts": 1776649790917136, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790917179, "dur": 1, + "args": { + "External id": 4243, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33167, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33167, "pid": 0, "tid": 7, "ts": 1776649790917179, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917170, "dur": 6, + "args": { + "External id": 4243, + "cbid": 211, "correlation": 33167 + } + }, + { + "ph": "s", "id": 33167, "pid": 435, "tid": 435, "ts": 1776649790917170, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917182, "dur": 0, + "args": { + "External id": 33173, + "cbid": 200, "correlation": 33173 + } + }, + { + "ph": "f", "id": 33173, "pid": 435, "tid": 435, "ts": 1776649790917182, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917183, "dur": 0, + "args": { + "External id": 33174, + "cbid": 273, "correlation": 33174 + } + }, + { + "ph": "f", "id": 33174, "pid": 435, "tid": 435, "ts": 1776649790917183, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790917192, "dur": 15, + "args": { + "External id": 33175, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33175, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33175, "pid": 0, "tid": 7, "ts": 1776649790917192, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917184, "dur": 4, + "args": { + "External id": 33175, + "cbid": 211, "correlation": 33175 + } + }, + { + "ph": "s", "id": 33175, "pid": 435, "tid": 435, "ts": 1776649790917184, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790917227, "dur": 1, + "args": { + "External id": 4248, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33193, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33193, "pid": 0, "tid": 7, "ts": 1776649790917227, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917218, "dur": 6, + "args": { + "External id": 4248, + "cbid": 211, "correlation": 33193 + } + }, + { + "ph": "s", "id": 33193, "pid": 435, "tid": 435, "ts": 1776649790917218, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917230, "dur": 0, + "args": { + "External id": 33199, + "cbid": 200, "correlation": 33199 + } + }, + { + "ph": "f", "id": 33199, "pid": 435, "tid": 435, "ts": 1776649790917230, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917231, "dur": 0, + "args": { + "External id": 33200, + "cbid": 273, "correlation": 33200 + } + }, + { + "ph": "f", "id": 33200, "pid": 435, "tid": 435, "ts": 1776649790917231, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790917240, "dur": 16, + "args": { + "External id": 33201, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33201, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33201, "pid": 0, "tid": 7, "ts": 1776649790917240, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917232, "dur": 5, + "args": { + "External id": 33201, + "cbid": 211, "correlation": 33201 + } + }, + { + "ph": "s", "id": 33201, "pid": 435, "tid": 435, "ts": 1776649790917232, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790917275, "dur": 1, + "args": { + "External id": 4253, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33219, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33219, "pid": 0, "tid": 7, "ts": 1776649790917275, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917266, "dur": 6, + "args": { + "External id": 4253, + "cbid": 211, "correlation": 33219 + } + }, + { + "ph": "s", "id": 33219, "pid": 435, "tid": 435, "ts": 1776649790917266, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917279, "dur": 0, + "args": { + "External id": 33225, + "cbid": 200, "correlation": 33225 + } + }, + { + "ph": "f", "id": 33225, "pid": 435, "tid": 435, "ts": 1776649790917279, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917279, "dur": 0, + "args": { + "External id": 33226, + "cbid": 273, "correlation": 33226 + } + }, + { + "ph": "f", "id": 33226, "pid": 435, "tid": 435, "ts": 1776649790917279, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790917288, "dur": 16, + "args": { + "External id": 33227, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33227, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33227, "pid": 0, "tid": 7, "ts": 1776649790917288, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917281, "dur": 4, + "args": { + "External id": 33227, + "cbid": 211, "correlation": 33227 + } + }, + { + "ph": "s", "id": 33227, "pid": 435, "tid": 435, "ts": 1776649790917281, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790917324, "dur": 1, + "args": { + "External id": 4258, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33245, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33245, "pid": 0, "tid": 7, "ts": 1776649790917324, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917315, "dur": 6, + "args": { + "External id": 4258, + "cbid": 211, "correlation": 33245 + } + }, + { + "ph": "s", "id": 33245, "pid": 435, "tid": 435, "ts": 1776649790917315, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917327, "dur": 0, + "args": { + "External id": 33251, + "cbid": 200, "correlation": 33251 + } + }, + { + "ph": "f", "id": 33251, "pid": 435, "tid": 435, "ts": 1776649790917327, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917328, "dur": 0, + "args": { + "External id": 33252, + "cbid": 273, "correlation": 33252 + } + }, + { + "ph": "f", "id": 33252, "pid": 435, "tid": 435, "ts": 1776649790917328, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790917337, "dur": 15, + "args": { + "External id": 33253, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33253, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33253, "pid": 0, "tid": 7, "ts": 1776649790917337, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917329, "dur": 5, + "args": { + "External id": 33253, + "cbid": 211, "correlation": 33253 + } + }, + { + "ph": "s", "id": 33253, "pid": 435, "tid": 435, "ts": 1776649790917329, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790917373, "dur": 1, + "args": { + "External id": 4263, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33271, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33271, "pid": 0, "tid": 7, "ts": 1776649790917373, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917363, "dur": 6, + "args": { + "External id": 4263, + "cbid": 211, "correlation": 33271 + } + }, + { + "ph": "s", "id": 33271, "pid": 435, "tid": 435, "ts": 1776649790917363, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917376, "dur": 0, + "args": { + "External id": 33277, + "cbid": 200, "correlation": 33277 + } + }, + { + "ph": "f", "id": 33277, "pid": 435, "tid": 435, "ts": 1776649790917376, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917377, "dur": 0, + "args": { + "External id": 33278, + "cbid": 273, "correlation": 33278 + } + }, + { + "ph": "f", "id": 33278, "pid": 435, "tid": 435, "ts": 1776649790917377, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790917386, "dur": 15, + "args": { + "External id": 33279, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33279, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33279, "pid": 0, "tid": 7, "ts": 1776649790917386, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917378, "dur": 4, + "args": { + "External id": 33279, + "cbid": 211, "correlation": 33279 + } + }, + { + "ph": "s", "id": 33279, "pid": 435, "tid": 435, "ts": 1776649790917378, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790917421, "dur": 1, + "args": { + "External id": 4268, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33297, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33297, "pid": 0, "tid": 7, "ts": 1776649790917421, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917412, "dur": 7, + "args": { + "External id": 4268, + "cbid": 211, "correlation": 33297 + } + }, + { + "ph": "s", "id": 33297, "pid": 435, "tid": 435, "ts": 1776649790917412, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917424, "dur": 0, + "args": { + "External id": 33303, + "cbid": 200, "correlation": 33303 + } + }, + { + "ph": "f", "id": 33303, "pid": 435, "tid": 435, "ts": 1776649790917424, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917425, "dur": 0, + "args": { + "External id": 33304, + "cbid": 273, "correlation": 33304 + } + }, + { + "ph": "f", "id": 33304, "pid": 435, "tid": 435, "ts": 1776649790917425, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790917434, "dur": 15, + "args": { + "External id": 33305, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33305, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33305, "pid": 0, "tid": 7, "ts": 1776649790917434, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917426, "dur": 4, + "args": { + "External id": 33305, + "cbid": 211, "correlation": 33305 + } + }, + { + "ph": "s", "id": 33305, "pid": 435, "tid": 435, "ts": 1776649790917426, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790917470, "dur": 1, + "args": { + "External id": 4273, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33323, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33323, "pid": 0, "tid": 7, "ts": 1776649790917470, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917460, "dur": 6, + "args": { + "External id": 4273, + "cbid": 211, "correlation": 33323 + } + }, + { + "ph": "s", "id": 33323, "pid": 435, "tid": 435, "ts": 1776649790917460, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917473, "dur": 0, + "args": { + "External id": 33329, + "cbid": 200, "correlation": 33329 + } + }, + { + "ph": "f", "id": 33329, "pid": 435, "tid": 435, "ts": 1776649790917473, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917473, "dur": 0, + "args": { + "External id": 33330, + "cbid": 273, "correlation": 33330 + } + }, + { + "ph": "f", "id": 33330, "pid": 435, "tid": 435, "ts": 1776649790917473, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790917482, "dur": 16, + "args": { + "External id": 33331, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33331, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33331, "pid": 0, "tid": 7, "ts": 1776649790917482, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917474, "dur": 4, + "args": { + "External id": 33331, + "cbid": 211, "correlation": 33331 + } + }, + { + "ph": "s", "id": 33331, "pid": 435, "tid": 435, "ts": 1776649790917474, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790917517, "dur": 1, + "args": { + "External id": 4278, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33349, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33349, "pid": 0, "tid": 7, "ts": 1776649790917517, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917508, "dur": 6, + "args": { + "External id": 4278, + "cbid": 211, "correlation": 33349 + } + }, + { + "ph": "s", "id": 33349, "pid": 435, "tid": 435, "ts": 1776649790917508, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917520, "dur": 0, + "args": { + "External id": 33355, + "cbid": 200, "correlation": 33355 + } + }, + { + "ph": "f", "id": 33355, "pid": 435, "tid": 435, "ts": 1776649790917520, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917521, "dur": 0, + "args": { + "External id": 33356, + "cbid": 273, "correlation": 33356 + } + }, + { + "ph": "f", "id": 33356, "pid": 435, "tid": 435, "ts": 1776649790917521, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790917531, "dur": 16, + "args": { + "External id": 33357, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33357, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33357, "pid": 0, "tid": 7, "ts": 1776649790917531, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917522, "dur": 5, + "args": { + "External id": 33357, + "cbid": 211, "correlation": 33357 + } + }, + { + "ph": "s", "id": 33357, "pid": 435, "tid": 435, "ts": 1776649790917522, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790917568, "dur": 1, + "args": { + "External id": 4283, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33375, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33375, "pid": 0, "tid": 7, "ts": 1776649790917568, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917558, "dur": 6, + "args": { + "External id": 4283, + "cbid": 211, "correlation": 33375 + } + }, + { + "ph": "s", "id": 33375, "pid": 435, "tid": 435, "ts": 1776649790917558, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917571, "dur": 0, + "args": { + "External id": 33381, + "cbid": 200, "correlation": 33381 + } + }, + { + "ph": "f", "id": 33381, "pid": 435, "tid": 435, "ts": 1776649790917571, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917571, "dur": 0, + "args": { + "External id": 33382, + "cbid": 273, "correlation": 33382 + } + }, + { + "ph": "f", "id": 33382, "pid": 435, "tid": 435, "ts": 1776649790917571, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790917581, "dur": 15, + "args": { + "External id": 33383, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33383, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33383, "pid": 0, "tid": 7, "ts": 1776649790917581, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917573, "dur": 5, + "args": { + "External id": 33383, + "cbid": 211, "correlation": 33383 + } + }, + { + "ph": "s", "id": 33383, "pid": 435, "tid": 435, "ts": 1776649790917573, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790917616, "dur": 1, + "args": { + "External id": 4288, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33401, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33401, "pid": 0, "tid": 7, "ts": 1776649790917616, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917607, "dur": 6, + "args": { + "External id": 4288, + "cbid": 211, "correlation": 33401 + } + }, + { + "ph": "s", "id": 33401, "pid": 435, "tid": 435, "ts": 1776649790917607, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917619, "dur": 0, + "args": { + "External id": 33407, + "cbid": 200, "correlation": 33407 + } + }, + { + "ph": "f", "id": 33407, "pid": 435, "tid": 435, "ts": 1776649790917619, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917620, "dur": 0, + "args": { + "External id": 33408, + "cbid": 273, "correlation": 33408 + } + }, + { + "ph": "f", "id": 33408, "pid": 435, "tid": 435, "ts": 1776649790917620, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790917629, "dur": 16, + "args": { + "External id": 33409, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33409, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33409, "pid": 0, "tid": 7, "ts": 1776649790917629, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917621, "dur": 5, + "args": { + "External id": 33409, + "cbid": 211, "correlation": 33409 + } + }, + { + "ph": "s", "id": 33409, "pid": 435, "tid": 435, "ts": 1776649790917621, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790917664, "dur": 1, + "args": { + "External id": 4293, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33427, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33427, "pid": 0, "tid": 7, "ts": 1776649790917664, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917655, "dur": 7, + "args": { + "External id": 4293, + "cbid": 211, "correlation": 33427 + } + }, + { + "ph": "s", "id": 33427, "pid": 435, "tid": 435, "ts": 1776649790917655, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917668, "dur": 0, + "args": { + "External id": 33433, + "cbid": 200, "correlation": 33433 + } + }, + { + "ph": "f", "id": 33433, "pid": 435, "tid": 435, "ts": 1776649790917668, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917668, "dur": 0, + "args": { + "External id": 33434, + "cbid": 273, "correlation": 33434 + } + }, + { + "ph": "f", "id": 33434, "pid": 435, "tid": 435, "ts": 1776649790917668, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790917677, "dur": 16, + "args": { + "External id": 33435, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33435, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33435, "pid": 0, "tid": 7, "ts": 1776649790917677, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917669, "dur": 4, + "args": { + "External id": 33435, + "cbid": 211, "correlation": 33435 + } + }, + { + "ph": "s", "id": 33435, "pid": 435, "tid": 435, "ts": 1776649790917669, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790917713, "dur": 1, + "args": { + "External id": 4298, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33453, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33453, "pid": 0, "tid": 7, "ts": 1776649790917713, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917703, "dur": 6, + "args": { + "External id": 4298, + "cbid": 211, "correlation": 33453 + } + }, + { + "ph": "s", "id": 33453, "pid": 435, "tid": 435, "ts": 1776649790917703, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917716, "dur": 0, + "args": { + "External id": 33459, + "cbid": 200, "correlation": 33459 + } + }, + { + "ph": "f", "id": 33459, "pid": 435, "tid": 435, "ts": 1776649790917716, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917717, "dur": 0, + "args": { + "External id": 33460, + "cbid": 273, "correlation": 33460 + } + }, + { + "ph": "f", "id": 33460, "pid": 435, "tid": 435, "ts": 1776649790917717, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790917726, "dur": 16, + "args": { + "External id": 33461, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33461, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33461, "pid": 0, "tid": 7, "ts": 1776649790917726, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917718, "dur": 5, + "args": { + "External id": 33461, + "cbid": 211, "correlation": 33461 + } + }, + { + "ph": "s", "id": 33461, "pid": 435, "tid": 435, "ts": 1776649790917718, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790917762, "dur": 1, + "args": { + "External id": 4303, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33479, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33479, "pid": 0, "tid": 7, "ts": 1776649790917762, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917752, "dur": 6, + "args": { + "External id": 4303, + "cbid": 211, "correlation": 33479 + } + }, + { + "ph": "s", "id": 33479, "pid": 435, "tid": 435, "ts": 1776649790917752, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917765, "dur": 0, + "args": { + "External id": 33485, + "cbid": 200, "correlation": 33485 + } + }, + { + "ph": "f", "id": 33485, "pid": 435, "tid": 435, "ts": 1776649790917765, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917766, "dur": 0, + "args": { + "External id": 33486, + "cbid": 273, "correlation": 33486 + } + }, + { + "ph": "f", "id": 33486, "pid": 435, "tid": 435, "ts": 1776649790917766, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790917775, "dur": 16, + "args": { + "External id": 33487, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33487, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33487, "pid": 0, "tid": 7, "ts": 1776649790917775, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917767, "dur": 4, + "args": { + "External id": 33487, + "cbid": 211, "correlation": 33487 + } + }, + { + "ph": "s", "id": 33487, "pid": 435, "tid": 435, "ts": 1776649790917767, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790917811, "dur": 1, + "args": { + "External id": 4308, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33505, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33505, "pid": 0, "tid": 7, "ts": 1776649790917811, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917801, "dur": 6, + "args": { + "External id": 4308, + "cbid": 211, "correlation": 33505 + } + }, + { + "ph": "s", "id": 33505, "pid": 435, "tid": 435, "ts": 1776649790917801, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917814, "dur": 0, + "args": { + "External id": 33511, + "cbid": 200, "correlation": 33511 + } + }, + { + "ph": "f", "id": 33511, "pid": 435, "tid": 435, "ts": 1776649790917814, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917814, "dur": 0, + "args": { + "External id": 33512, + "cbid": 273, "correlation": 33512 + } + }, + { + "ph": "f", "id": 33512, "pid": 435, "tid": 435, "ts": 1776649790917814, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790917823, "dur": 15, + "args": { + "External id": 33513, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33513, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33513, "pid": 0, "tid": 7, "ts": 1776649790917823, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917815, "dur": 4, + "args": { + "External id": 33513, + "cbid": 211, "correlation": 33513 + } + }, + { + "ph": "s", "id": 33513, "pid": 435, "tid": 435, "ts": 1776649790917815, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790917859, "dur": 1, + "args": { + "External id": 4313, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33531, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33531, "pid": 0, "tid": 7, "ts": 1776649790917859, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917850, "dur": 6, + "args": { + "External id": 4313, + "cbid": 211, "correlation": 33531 + } + }, + { + "ph": "s", "id": 33531, "pid": 435, "tid": 435, "ts": 1776649790917850, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917863, "dur": 0, + "args": { + "External id": 33537, + "cbid": 200, "correlation": 33537 + } + }, + { + "ph": "f", "id": 33537, "pid": 435, "tid": 435, "ts": 1776649790917863, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917863, "dur": 0, + "args": { + "External id": 33538, + "cbid": 273, "correlation": 33538 + } + }, + { + "ph": "f", "id": 33538, "pid": 435, "tid": 435, "ts": 1776649790917863, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790917872, "dur": 16, + "args": { + "External id": 33539, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33539, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33539, "pid": 0, "tid": 7, "ts": 1776649790917872, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917864, "dur": 5, + "args": { + "External id": 33539, + "cbid": 211, "correlation": 33539 + } + }, + { + "ph": "s", "id": 33539, "pid": 435, "tid": 435, "ts": 1776649790917864, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790917910, "dur": 1, + "args": { + "External id": 4318, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33557, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33557, "pid": 0, "tid": 7, "ts": 1776649790917910, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917898, "dur": 8, + "args": { + "External id": 4318, + "cbid": 211, "correlation": 33557 + } + }, + { + "ph": "s", "id": 33557, "pid": 435, "tid": 435, "ts": 1776649790917898, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917913, "dur": 0, + "args": { + "External id": 33563, + "cbid": 200, "correlation": 33563 + } + }, + { + "ph": "f", "id": 33563, "pid": 435, "tid": 435, "ts": 1776649790917913, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917913, "dur": 0, + "args": { + "External id": 33564, + "cbid": 273, "correlation": 33564 + } + }, + { + "ph": "f", "id": 33564, "pid": 435, "tid": 435, "ts": 1776649790917913, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790917922, "dur": 15, + "args": { + "External id": 33565, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33565, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33565, "pid": 0, "tid": 7, "ts": 1776649790917922, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917914, "dur": 5, + "args": { + "External id": 33565, + "cbid": 211, "correlation": 33565 + } + }, + { + "ph": "s", "id": 33565, "pid": 435, "tid": 435, "ts": 1776649790917914, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790917958, "dur": 1, + "args": { + "External id": 4323, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33583, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33583, "pid": 0, "tid": 7, "ts": 1776649790917958, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917948, "dur": 6, + "args": { + "External id": 4323, + "cbid": 211, "correlation": 33583 + } + }, + { + "ph": "s", "id": 33583, "pid": 435, "tid": 435, "ts": 1776649790917948, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917961, "dur": 0, + "args": { + "External id": 33589, + "cbid": 200, "correlation": 33589 + } + }, + { + "ph": "f", "id": 33589, "pid": 435, "tid": 435, "ts": 1776649790917961, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790917962, "dur": 0, + "args": { + "External id": 33590, + "cbid": 273, "correlation": 33590 + } + }, + { + "ph": "f", "id": 33590, "pid": 435, "tid": 435, "ts": 1776649790917962, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790917971, "dur": 16, + "args": { + "External id": 33591, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33591, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33591, "pid": 0, "tid": 7, "ts": 1776649790917971, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917963, "dur": 4, + "args": { + "External id": 33591, + "cbid": 211, "correlation": 33591 + } + }, + { + "ph": "s", "id": 33591, "pid": 435, "tid": 435, "ts": 1776649790917963, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790918007, "dur": 1, + "args": { + "External id": 4328, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33609, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33609, "pid": 0, "tid": 7, "ts": 1776649790918007, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790917997, "dur": 6, + "args": { + "External id": 4328, + "cbid": 211, "correlation": 33609 + } + }, + { + "ph": "s", "id": 33609, "pid": 435, "tid": 435, "ts": 1776649790917997, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918010, "dur": 0, + "args": { + "External id": 33615, + "cbid": 200, "correlation": 33615 + } + }, + { + "ph": "f", "id": 33615, "pid": 435, "tid": 435, "ts": 1776649790918010, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918010, "dur": 0, + "args": { + "External id": 33616, + "cbid": 273, "correlation": 33616 + } + }, + { + "ph": "f", "id": 33616, "pid": 435, "tid": 435, "ts": 1776649790918010, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790918019, "dur": 16, + "args": { + "External id": 33617, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33617, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33617, "pid": 0, "tid": 7, "ts": 1776649790918019, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918012, "dur": 5, + "args": { + "External id": 33617, + "cbid": 211, "correlation": 33617 + } + }, + { + "ph": "s", "id": 33617, "pid": 435, "tid": 435, "ts": 1776649790918012, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790918055, "dur": 1, + "args": { + "External id": 4333, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33635, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33635, "pid": 0, "tid": 7, "ts": 1776649790918055, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918046, "dur": 6, + "args": { + "External id": 4333, + "cbid": 211, "correlation": 33635 + } + }, + { + "ph": "s", "id": 33635, "pid": 435, "tid": 435, "ts": 1776649790918046, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918059, "dur": 0, + "args": { + "External id": 33641, + "cbid": 200, "correlation": 33641 + } + }, + { + "ph": "f", "id": 33641, "pid": 435, "tid": 435, "ts": 1776649790918059, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918060, "dur": 0, + "args": { + "External id": 33642, + "cbid": 273, "correlation": 33642 + } + }, + { + "ph": "f", "id": 33642, "pid": 435, "tid": 435, "ts": 1776649790918060, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790918069, "dur": 15, + "args": { + "External id": 33643, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33643, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33643, "pid": 0, "tid": 7, "ts": 1776649790918069, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918061, "dur": 5, + "args": { + "External id": 33643, + "cbid": 211, "correlation": 33643 + } + }, + { + "ph": "s", "id": 33643, "pid": 435, "tid": 435, "ts": 1776649790918061, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790918109, "dur": 1, + "args": { + "External id": 4338, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33661, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33661, "pid": 0, "tid": 7, "ts": 1776649790918109, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918098, "dur": 7, + "args": { + "External id": 4338, + "cbid": 211, "correlation": 33661 + } + }, + { + "ph": "s", "id": 33661, "pid": 435, "tid": 435, "ts": 1776649790918098, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918112, "dur": 0, + "args": { + "External id": 33667, + "cbid": 200, "correlation": 33667 + } + }, + { + "ph": "f", "id": 33667, "pid": 435, "tid": 435, "ts": 1776649790918112, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918113, "dur": 0, + "args": { + "External id": 33668, + "cbid": 273, "correlation": 33668 + } + }, + { + "ph": "f", "id": 33668, "pid": 435, "tid": 435, "ts": 1776649790918113, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790918122, "dur": 17, + "args": { + "External id": 33669, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33669, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33669, "pid": 0, "tid": 7, "ts": 1776649790918122, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918114, "dur": 5, + "args": { + "External id": 33669, + "cbid": 211, "correlation": 33669 + } + }, + { + "ph": "s", "id": 33669, "pid": 435, "tid": 435, "ts": 1776649790918114, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790918157, "dur": 1, + "args": { + "External id": 4343, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33687, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33687, "pid": 0, "tid": 7, "ts": 1776649790918157, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918148, "dur": 6, + "args": { + "External id": 4343, + "cbid": 211, "correlation": 33687 + } + }, + { + "ph": "s", "id": 33687, "pid": 435, "tid": 435, "ts": 1776649790918148, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918160, "dur": 0, + "args": { + "External id": 33693, + "cbid": 200, "correlation": 33693 + } + }, + { + "ph": "f", "id": 33693, "pid": 435, "tid": 435, "ts": 1776649790918160, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918161, "dur": 0, + "args": { + "External id": 33694, + "cbid": 273, "correlation": 33694 + } + }, + { + "ph": "f", "id": 33694, "pid": 435, "tid": 435, "ts": 1776649790918161, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790918170, "dur": 16, + "args": { + "External id": 33695, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33695, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33695, "pid": 0, "tid": 7, "ts": 1776649790918170, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918162, "dur": 4, + "args": { + "External id": 33695, + "cbid": 211, "correlation": 33695 + } + }, + { + "ph": "s", "id": 33695, "pid": 435, "tid": 435, "ts": 1776649790918162, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790918206, "dur": 1, + "args": { + "External id": 4348, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33713, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33713, "pid": 0, "tid": 7, "ts": 1776649790918206, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918197, "dur": 6, + "args": { + "External id": 4348, + "cbid": 211, "correlation": 33713 + } + }, + { + "ph": "s", "id": 33713, "pid": 435, "tid": 435, "ts": 1776649790918197, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918209, "dur": 0, + "args": { + "External id": 33719, + "cbid": 200, "correlation": 33719 + } + }, + { + "ph": "f", "id": 33719, "pid": 435, "tid": 435, "ts": 1776649790918209, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918210, "dur": 0, + "args": { + "External id": 33720, + "cbid": 273, "correlation": 33720 + } + }, + { + "ph": "f", "id": 33720, "pid": 435, "tid": 435, "ts": 1776649790918210, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790918219, "dur": 16, + "args": { + "External id": 33721, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33721, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33721, "pid": 0, "tid": 7, "ts": 1776649790918219, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918211, "dur": 4, + "args": { + "External id": 33721, + "cbid": 211, "correlation": 33721 + } + }, + { + "ph": "s", "id": 33721, "pid": 435, "tid": 435, "ts": 1776649790918211, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790918254, "dur": 1, + "args": { + "External id": 4353, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33739, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33739, "pid": 0, "tid": 7, "ts": 1776649790918254, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918245, "dur": 6, + "args": { + "External id": 4353, + "cbid": 211, "correlation": 33739 + } + }, + { + "ph": "s", "id": 33739, "pid": 435, "tid": 435, "ts": 1776649790918245, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918257, "dur": 0, + "args": { + "External id": 33745, + "cbid": 200, "correlation": 33745 + } + }, + { + "ph": "f", "id": 33745, "pid": 435, "tid": 435, "ts": 1776649790918257, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918258, "dur": 0, + "args": { + "External id": 33746, + "cbid": 273, "correlation": 33746 + } + }, + { + "ph": "f", "id": 33746, "pid": 435, "tid": 435, "ts": 1776649790918258, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790918267, "dur": 16, + "args": { + "External id": 33747, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33747, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33747, "pid": 0, "tid": 7, "ts": 1776649790918267, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918259, "dur": 4, + "args": { + "External id": 33747, + "cbid": 211, "correlation": 33747 + } + }, + { + "ph": "s", "id": 33747, "pid": 435, "tid": 435, "ts": 1776649790918259, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790918302, "dur": 1, + "args": { + "External id": 4358, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33765, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33765, "pid": 0, "tid": 7, "ts": 1776649790918302, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918292, "dur": 6, + "args": { + "External id": 4358, + "cbid": 211, "correlation": 33765 + } + }, + { + "ph": "s", "id": 33765, "pid": 435, "tid": 435, "ts": 1776649790918292, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918305, "dur": 0, + "args": { + "External id": 33771, + "cbid": 200, "correlation": 33771 + } + }, + { + "ph": "f", "id": 33771, "pid": 435, "tid": 435, "ts": 1776649790918305, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918306, "dur": 0, + "args": { + "External id": 33772, + "cbid": 273, "correlation": 33772 + } + }, + { + "ph": "f", "id": 33772, "pid": 435, "tid": 435, "ts": 1776649790918306, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790918314, "dur": 16, + "args": { + "External id": 33773, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33773, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33773, "pid": 0, "tid": 7, "ts": 1776649790918314, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918307, "dur": 4, + "args": { + "External id": 33773, + "cbid": 211, "correlation": 33773 + } + }, + { + "ph": "s", "id": 33773, "pid": 435, "tid": 435, "ts": 1776649790918307, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790918350, "dur": 1, + "args": { + "External id": 4363, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33791, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33791, "pid": 0, "tid": 7, "ts": 1776649790918350, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918341, "dur": 6, + "args": { + "External id": 4363, + "cbid": 211, "correlation": 33791 + } + }, + { + "ph": "s", "id": 33791, "pid": 435, "tid": 435, "ts": 1776649790918341, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918353, "dur": 0, + "args": { + "External id": 33797, + "cbid": 200, "correlation": 33797 + } + }, + { + "ph": "f", "id": 33797, "pid": 435, "tid": 435, "ts": 1776649790918353, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918354, "dur": 0, + "args": { + "External id": 33798, + "cbid": 273, "correlation": 33798 + } + }, + { + "ph": "f", "id": 33798, "pid": 435, "tid": 435, "ts": 1776649790918354, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790918363, "dur": 16, + "args": { + "External id": 33799, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33799, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33799, "pid": 0, "tid": 7, "ts": 1776649790918363, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918355, "dur": 5, + "args": { + "External id": 33799, + "cbid": 211, "correlation": 33799 + } + }, + { + "ph": "s", "id": 33799, "pid": 435, "tid": 435, "ts": 1776649790918355, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790918398, "dur": 1, + "args": { + "External id": 4368, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33817, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33817, "pid": 0, "tid": 7, "ts": 1776649790918398, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918389, "dur": 6, + "args": { + "External id": 4368, + "cbid": 211, "correlation": 33817 + } + }, + { + "ph": "s", "id": 33817, "pid": 435, "tid": 435, "ts": 1776649790918389, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918401, "dur": 0, + "args": { + "External id": 33823, + "cbid": 200, "correlation": 33823 + } + }, + { + "ph": "f", "id": 33823, "pid": 435, "tid": 435, "ts": 1776649790918401, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918402, "dur": 0, + "args": { + "External id": 33824, + "cbid": 273, "correlation": 33824 + } + }, + { + "ph": "f", "id": 33824, "pid": 435, "tid": 435, "ts": 1776649790918402, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790918411, "dur": 16, + "args": { + "External id": 33825, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33825, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33825, "pid": 0, "tid": 7, "ts": 1776649790918411, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918403, "dur": 4, + "args": { + "External id": 33825, + "cbid": 211, "correlation": 33825 + } + }, + { + "ph": "s", "id": 33825, "pid": 435, "tid": 435, "ts": 1776649790918403, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790918446, "dur": 1, + "args": { + "External id": 4373, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33843, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33843, "pid": 0, "tid": 7, "ts": 1776649790918446, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918437, "dur": 6, + "args": { + "External id": 4373, + "cbid": 211, "correlation": 33843 + } + }, + { + "ph": "s", "id": 33843, "pid": 435, "tid": 435, "ts": 1776649790918437, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918449, "dur": 0, + "args": { + "External id": 33849, + "cbid": 200, "correlation": 33849 + } + }, + { + "ph": "f", "id": 33849, "pid": 435, "tid": 435, "ts": 1776649790918449, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918450, "dur": 0, + "args": { + "External id": 33850, + "cbid": 273, "correlation": 33850 + } + }, + { + "ph": "f", "id": 33850, "pid": 435, "tid": 435, "ts": 1776649790918450, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790918459, "dur": 15, + "args": { + "External id": 33851, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33851, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33851, "pid": 0, "tid": 7, "ts": 1776649790918459, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918451, "dur": 5, + "args": { + "External id": 33851, + "cbid": 211, "correlation": 33851 + } + }, + { + "ph": "s", "id": 33851, "pid": 435, "tid": 435, "ts": 1776649790918451, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790918495, "dur": 1, + "args": { + "External id": 4378, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33869, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33869, "pid": 0, "tid": 7, "ts": 1776649790918495, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918485, "dur": 6, + "args": { + "External id": 4378, + "cbid": 211, "correlation": 33869 + } + }, + { + "ph": "s", "id": 33869, "pid": 435, "tid": 435, "ts": 1776649790918485, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918498, "dur": 0, + "args": { + "External id": 33875, + "cbid": 200, "correlation": 33875 + } + }, + { + "ph": "f", "id": 33875, "pid": 435, "tid": 435, "ts": 1776649790918498, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918498, "dur": 0, + "args": { + "External id": 33876, + "cbid": 273, "correlation": 33876 + } + }, + { + "ph": "f", "id": 33876, "pid": 435, "tid": 435, "ts": 1776649790918498, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790918507, "dur": 16, + "args": { + "External id": 33877, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33877, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33877, "pid": 0, "tid": 7, "ts": 1776649790918507, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918499, "dur": 5, + "args": { + "External id": 33877, + "cbid": 211, "correlation": 33877 + } + }, + { + "ph": "s", "id": 33877, "pid": 435, "tid": 435, "ts": 1776649790918499, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790918543, "dur": 1, + "args": { + "External id": 4383, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33895, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33895, "pid": 0, "tid": 7, "ts": 1776649790918543, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918533, "dur": 6, + "args": { + "External id": 4383, + "cbid": 211, "correlation": 33895 + } + }, + { + "ph": "s", "id": 33895, "pid": 435, "tid": 435, "ts": 1776649790918533, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918546, "dur": 0, + "args": { + "External id": 33901, + "cbid": 200, "correlation": 33901 + } + }, + { + "ph": "f", "id": 33901, "pid": 435, "tid": 435, "ts": 1776649790918546, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918546, "dur": 0, + "args": { + "External id": 33902, + "cbid": 273, "correlation": 33902 + } + }, + { + "ph": "f", "id": 33902, "pid": 435, "tid": 435, "ts": 1776649790918546, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790918555, "dur": 17, + "args": { + "External id": 33903, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33903, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33903, "pid": 0, "tid": 7, "ts": 1776649790918555, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918547, "dur": 4, + "args": { + "External id": 33903, + "cbid": 211, "correlation": 33903 + } + }, + { + "ph": "s", "id": 33903, "pid": 435, "tid": 435, "ts": 1776649790918547, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790918591, "dur": 1, + "args": { + "External id": 4388, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33921, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33921, "pid": 0, "tid": 7, "ts": 1776649790918591, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918582, "dur": 6, + "args": { + "External id": 4388, + "cbid": 211, "correlation": 33921 + } + }, + { + "ph": "s", "id": 33921, "pid": 435, "tid": 435, "ts": 1776649790918582, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918594, "dur": 0, + "args": { + "External id": 33927, + "cbid": 200, "correlation": 33927 + } + }, + { + "ph": "f", "id": 33927, "pid": 435, "tid": 435, "ts": 1776649790918594, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918595, "dur": 0, + "args": { + "External id": 33928, + "cbid": 273, "correlation": 33928 + } + }, + { + "ph": "f", "id": 33928, "pid": 435, "tid": 435, "ts": 1776649790918595, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790918604, "dur": 16, + "args": { + "External id": 33929, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33929, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33929, "pid": 0, "tid": 7, "ts": 1776649790918604, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918596, "dur": 5, + "args": { + "External id": 33929, + "cbid": 211, "correlation": 33929 + } + }, + { + "ph": "s", "id": 33929, "pid": 435, "tid": 435, "ts": 1776649790918596, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790918639, "dur": 1, + "args": { + "External id": 4393, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33947, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33947, "pid": 0, "tid": 7, "ts": 1776649790918639, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918630, "dur": 6, + "args": { + "External id": 4393, + "cbid": 211, "correlation": 33947 + } + }, + { + "ph": "s", "id": 33947, "pid": 435, "tid": 435, "ts": 1776649790918630, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918642, "dur": 0, + "args": { + "External id": 33953, + "cbid": 200, "correlation": 33953 + } + }, + { + "ph": "f", "id": 33953, "pid": 435, "tid": 435, "ts": 1776649790918642, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918643, "dur": 0, + "args": { + "External id": 33954, + "cbid": 273, "correlation": 33954 + } + }, + { + "ph": "f", "id": 33954, "pid": 435, "tid": 435, "ts": 1776649790918643, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790918652, "dur": 15, + "args": { + "External id": 33955, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33955, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33955, "pid": 0, "tid": 7, "ts": 1776649790918652, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918644, "dur": 4, + "args": { + "External id": 33955, + "cbid": 211, "correlation": 33955 + } + }, + { + "ph": "s", "id": 33955, "pid": 435, "tid": 435, "ts": 1776649790918644, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790918687, "dur": 1, + "args": { + "External id": 4398, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33973, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33973, "pid": 0, "tid": 7, "ts": 1776649790918687, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918677, "dur": 6, + "args": { + "External id": 4398, + "cbid": 211, "correlation": 33973 + } + }, + { + "ph": "s", "id": 33973, "pid": 435, "tid": 435, "ts": 1776649790918677, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918690, "dur": 0, + "args": { + "External id": 33979, + "cbid": 200, "correlation": 33979 + } + }, + { + "ph": "f", "id": 33979, "pid": 435, "tid": 435, "ts": 1776649790918690, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918690, "dur": 0, + "args": { + "External id": 33980, + "cbid": 273, "correlation": 33980 + } + }, + { + "ph": "f", "id": 33980, "pid": 435, "tid": 435, "ts": 1776649790918690, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790918699, "dur": 16, + "args": { + "External id": 33981, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33981, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33981, "pid": 0, "tid": 7, "ts": 1776649790918699, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918692, "dur": 4, + "args": { + "External id": 33981, + "cbid": 211, "correlation": 33981 + } + }, + { + "ph": "s", "id": 33981, "pid": 435, "tid": 435, "ts": 1776649790918692, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790918735, "dur": 1, + "args": { + "External id": 4403, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 33999, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 33999, "pid": 0, "tid": 7, "ts": 1776649790918735, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918725, "dur": 6, + "args": { + "External id": 4403, + "cbid": 211, "correlation": 33999 + } + }, + { + "ph": "s", "id": 33999, "pid": 435, "tid": 435, "ts": 1776649790918725, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918738, "dur": 0, + "args": { + "External id": 34005, + "cbid": 200, "correlation": 34005 + } + }, + { + "ph": "f", "id": 34005, "pid": 435, "tid": 435, "ts": 1776649790918738, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918739, "dur": 0, + "args": { + "External id": 34006, + "cbid": 273, "correlation": 34006 + } + }, + { + "ph": "f", "id": 34006, "pid": 435, "tid": 435, "ts": 1776649790918739, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790918748, "dur": 15, + "args": { + "External id": 34007, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34007, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34007, "pid": 0, "tid": 7, "ts": 1776649790918748, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918740, "dur": 4, + "args": { + "External id": 34007, + "cbid": 211, "correlation": 34007 + } + }, + { + "ph": "s", "id": 34007, "pid": 435, "tid": 435, "ts": 1776649790918740, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790918783, "dur": 1, + "args": { + "External id": 4408, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34025, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34025, "pid": 0, "tid": 7, "ts": 1776649790918783, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918773, "dur": 6, + "args": { + "External id": 4408, + "cbid": 211, "correlation": 34025 + } + }, + { + "ph": "s", "id": 34025, "pid": 435, "tid": 435, "ts": 1776649790918773, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918786, "dur": 0, + "args": { + "External id": 34031, + "cbid": 200, "correlation": 34031 + } + }, + { + "ph": "f", "id": 34031, "pid": 435, "tid": 435, "ts": 1776649790918786, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918787, "dur": 0, + "args": { + "External id": 34032, + "cbid": 273, "correlation": 34032 + } + }, + { + "ph": "f", "id": 34032, "pid": 435, "tid": 435, "ts": 1776649790918787, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790918796, "dur": 16, + "args": { + "External id": 34033, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34033, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34033, "pid": 0, "tid": 7, "ts": 1776649790918796, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918788, "dur": 4, + "args": { + "External id": 34033, + "cbid": 211, "correlation": 34033 + } + }, + { + "ph": "s", "id": 34033, "pid": 435, "tid": 435, "ts": 1776649790918788, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790918831, "dur": 1, + "args": { + "External id": 4413, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34051, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34051, "pid": 0, "tid": 7, "ts": 1776649790918831, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918822, "dur": 6, + "args": { + "External id": 4413, + "cbid": 211, "correlation": 34051 + } + }, + { + "ph": "s", "id": 34051, "pid": 435, "tid": 435, "ts": 1776649790918822, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918834, "dur": 0, + "args": { + "External id": 34057, + "cbid": 200, "correlation": 34057 + } + }, + { + "ph": "f", "id": 34057, "pid": 435, "tid": 435, "ts": 1776649790918834, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918835, "dur": 0, + "args": { + "External id": 34058, + "cbid": 273, "correlation": 34058 + } + }, + { + "ph": "f", "id": 34058, "pid": 435, "tid": 435, "ts": 1776649790918835, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790918844, "dur": 16, + "args": { + "External id": 34059, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34059, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34059, "pid": 0, "tid": 7, "ts": 1776649790918844, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918836, "dur": 5, + "args": { + "External id": 34059, + "cbid": 211, "correlation": 34059 + } + }, + { + "ph": "s", "id": 34059, "pid": 435, "tid": 435, "ts": 1776649790918836, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790918879, "dur": 1, + "args": { + "External id": 4418, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34077, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34077, "pid": 0, "tid": 7, "ts": 1776649790918879, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918870, "dur": 6, + "args": { + "External id": 4418, + "cbid": 211, "correlation": 34077 + } + }, + { + "ph": "s", "id": 34077, "pid": 435, "tid": 435, "ts": 1776649790918870, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918883, "dur": 0, + "args": { + "External id": 34083, + "cbid": 200, "correlation": 34083 + } + }, + { + "ph": "f", "id": 34083, "pid": 435, "tid": 435, "ts": 1776649790918883, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918883, "dur": 0, + "args": { + "External id": 34084, + "cbid": 273, "correlation": 34084 + } + }, + { + "ph": "f", "id": 34084, "pid": 435, "tid": 435, "ts": 1776649790918883, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790918892, "dur": 16, + "args": { + "External id": 34085, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34085, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34085, "pid": 0, "tid": 7, "ts": 1776649790918892, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918884, "dur": 5, + "args": { + "External id": 34085, + "cbid": 211, "correlation": 34085 + } + }, + { + "ph": "s", "id": 34085, "pid": 435, "tid": 435, "ts": 1776649790918884, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790918927, "dur": 1, + "args": { + "External id": 4423, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34103, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34103, "pid": 0, "tid": 7, "ts": 1776649790918927, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918918, "dur": 6, + "args": { + "External id": 4423, + "cbid": 211, "correlation": 34103 + } + }, + { + "ph": "s", "id": 34103, "pid": 435, "tid": 435, "ts": 1776649790918918, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918931, "dur": 0, + "args": { + "External id": 34109, + "cbid": 200, "correlation": 34109 + } + }, + { + "ph": "f", "id": 34109, "pid": 435, "tid": 435, "ts": 1776649790918931, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918931, "dur": 0, + "args": { + "External id": 34110, + "cbid": 273, "correlation": 34110 + } + }, + { + "ph": "f", "id": 34110, "pid": 435, "tid": 435, "ts": 1776649790918931, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790918941, "dur": 17, + "args": { + "External id": 34111, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34111, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34111, "pid": 0, "tid": 7, "ts": 1776649790918941, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918933, "dur": 5, + "args": { + "External id": 34111, + "cbid": 211, "correlation": 34111 + } + }, + { + "ph": "s", "id": 34111, "pid": 435, "tid": 435, "ts": 1776649790918933, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790918976, "dur": 1, + "args": { + "External id": 4428, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34129, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34129, "pid": 0, "tid": 7, "ts": 1776649790918976, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918967, "dur": 6, + "args": { + "External id": 4428, + "cbid": 211, "correlation": 34129 + } + }, + { + "ph": "s", "id": 34129, "pid": 435, "tid": 435, "ts": 1776649790918967, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918979, "dur": 0, + "args": { + "External id": 34135, + "cbid": 200, "correlation": 34135 + } + }, + { + "ph": "f", "id": 34135, "pid": 435, "tid": 435, "ts": 1776649790918979, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790918980, "dur": 0, + "args": { + "External id": 34136, + "cbid": 273, "correlation": 34136 + } + }, + { + "ph": "f", "id": 34136, "pid": 435, "tid": 435, "ts": 1776649790918980, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790918989, "dur": 17, + "args": { + "External id": 34137, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34137, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34137, "pid": 0, "tid": 7, "ts": 1776649790918989, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790918981, "dur": 5, + "args": { + "External id": 34137, + "cbid": 211, "correlation": 34137 + } + }, + { + "ph": "s", "id": 34137, "pid": 435, "tid": 435, "ts": 1776649790918981, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790919025, "dur": 1, + "args": { + "External id": 4433, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34155, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34155, "pid": 0, "tid": 7, "ts": 1776649790919025, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919015, "dur": 7, + "args": { + "External id": 4433, + "cbid": 211, "correlation": 34155 + } + }, + { + "ph": "s", "id": 34155, "pid": 435, "tid": 435, "ts": 1776649790919015, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919028, "dur": 0, + "args": { + "External id": 34161, + "cbid": 200, "correlation": 34161 + } + }, + { + "ph": "f", "id": 34161, "pid": 435, "tid": 435, "ts": 1776649790919028, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919028, "dur": 0, + "args": { + "External id": 34162, + "cbid": 273, "correlation": 34162 + } + }, + { + "ph": "f", "id": 34162, "pid": 435, "tid": 435, "ts": 1776649790919028, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790919037, "dur": 15, + "args": { + "External id": 34163, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34163, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34163, "pid": 0, "tid": 7, "ts": 1776649790919037, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919029, "dur": 5, + "args": { + "External id": 34163, + "cbid": 211, "correlation": 34163 + } + }, + { + "ph": "s", "id": 34163, "pid": 435, "tid": 435, "ts": 1776649790919029, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790919073, "dur": 1, + "args": { + "External id": 4438, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34181, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34181, "pid": 0, "tid": 7, "ts": 1776649790919073, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919064, "dur": 6, + "args": { + "External id": 4438, + "cbid": 211, "correlation": 34181 + } + }, + { + "ph": "s", "id": 34181, "pid": 435, "tid": 435, "ts": 1776649790919064, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919076, "dur": 0, + "args": { + "External id": 34187, + "cbid": 200, "correlation": 34187 + } + }, + { + "ph": "f", "id": 34187, "pid": 435, "tid": 435, "ts": 1776649790919076, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919077, "dur": 0, + "args": { + "External id": 34188, + "cbid": 273, "correlation": 34188 + } + }, + { + "ph": "f", "id": 34188, "pid": 435, "tid": 435, "ts": 1776649790919077, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790919085, "dur": 16, + "args": { + "External id": 34189, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34189, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34189, "pid": 0, "tid": 7, "ts": 1776649790919085, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919078, "dur": 4, + "args": { + "External id": 34189, + "cbid": 211, "correlation": 34189 + } + }, + { + "ph": "s", "id": 34189, "pid": 435, "tid": 435, "ts": 1776649790919078, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790919121, "dur": 1, + "args": { + "External id": 4443, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34207, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34207, "pid": 0, "tid": 7, "ts": 1776649790919121, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919112, "dur": 6, + "args": { + "External id": 4443, + "cbid": 211, "correlation": 34207 + } + }, + { + "ph": "s", "id": 34207, "pid": 435, "tid": 435, "ts": 1776649790919112, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919124, "dur": 0, + "args": { + "External id": 34213, + "cbid": 200, "correlation": 34213 + } + }, + { + "ph": "f", "id": 34213, "pid": 435, "tid": 435, "ts": 1776649790919124, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919125, "dur": 0, + "args": { + "External id": 34214, + "cbid": 273, "correlation": 34214 + } + }, + { + "ph": "f", "id": 34214, "pid": 435, "tid": 435, "ts": 1776649790919125, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790919134, "dur": 16, + "args": { + "External id": 34215, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34215, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34215, "pid": 0, "tid": 7, "ts": 1776649790919134, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919126, "dur": 4, + "args": { + "External id": 34215, + "cbid": 211, "correlation": 34215 + } + }, + { + "ph": "s", "id": 34215, "pid": 435, "tid": 435, "ts": 1776649790919126, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790919169, "dur": 1, + "args": { + "External id": 4448, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34233, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34233, "pid": 0, "tid": 7, "ts": 1776649790919169, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919160, "dur": 6, + "args": { + "External id": 4448, + "cbid": 211, "correlation": 34233 + } + }, + { + "ph": "s", "id": 34233, "pid": 435, "tid": 435, "ts": 1776649790919160, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919172, "dur": 0, + "args": { + "External id": 34239, + "cbid": 200, "correlation": 34239 + } + }, + { + "ph": "f", "id": 34239, "pid": 435, "tid": 435, "ts": 1776649790919172, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919173, "dur": 0, + "args": { + "External id": 34240, + "cbid": 273, "correlation": 34240 + } + }, + { + "ph": "f", "id": 34240, "pid": 435, "tid": 435, "ts": 1776649790919173, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790919181, "dur": 15, + "args": { + "External id": 34241, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34241, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34241, "pid": 0, "tid": 7, "ts": 1776649790919181, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919174, "dur": 5, + "args": { + "External id": 34241, + "cbid": 211, "correlation": 34241 + } + }, + { + "ph": "s", "id": 34241, "pid": 435, "tid": 435, "ts": 1776649790919174, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790919217, "dur": 1, + "args": { + "External id": 4453, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34259, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34259, "pid": 0, "tid": 7, "ts": 1776649790919217, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919208, "dur": 6, + "args": { + "External id": 4453, + "cbid": 211, "correlation": 34259 + } + }, + { + "ph": "s", "id": 34259, "pid": 435, "tid": 435, "ts": 1776649790919208, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919220, "dur": 0, + "args": { + "External id": 34265, + "cbid": 200, "correlation": 34265 + } + }, + { + "ph": "f", "id": 34265, "pid": 435, "tid": 435, "ts": 1776649790919220, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919221, "dur": 0, + "args": { + "External id": 34266, + "cbid": 273, "correlation": 34266 + } + }, + { + "ph": "f", "id": 34266, "pid": 435, "tid": 435, "ts": 1776649790919221, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790919230, "dur": 15, + "args": { + "External id": 34267, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34267, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34267, "pid": 0, "tid": 7, "ts": 1776649790919230, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919222, "dur": 4, + "args": { + "External id": 34267, + "cbid": 211, "correlation": 34267 + } + }, + { + "ph": "s", "id": 34267, "pid": 435, "tid": 435, "ts": 1776649790919222, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790919251, "dur": 3, + "args": { + "External id": 34275, + "cbid": 135, "correlation": 34275 + } + }, + { + "ph": "f", "id": 34275, "pid": 435, "tid": 435, "ts": 1776649790919251, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790919265, "dur": 5, + "args": { + "External id": 34280, + "cbid": 165, "correlation": 34280 + } + }, + { + "ph": "s", "id": 34280, "pid": 435, "tid": 435, "ts": 1776649790919265, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649790919273, "dur": 4, + "args": { + "External id": 34283, + "cbid": 139, "correlation": 34283 + } + }, + { + "ph": "f", "id": 34283, "pid": 435, "tid": 435, "ts": 1776649790919273, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790919336, "dur": 1, + "args": { + "External id": 4458, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34311, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34311, "pid": 0, "tid": 7, "ts": 1776649790919336, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919325, "dur": 9, + "args": { + "External id": 4458, + "cbid": 211, "correlation": 34311 + } + }, + { + "ph": "s", "id": 34311, "pid": 435, "tid": 435, "ts": 1776649790919325, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919341, "dur": 0, + "args": { + "External id": 34317, + "cbid": 200, "correlation": 34317 + } + }, + { + "ph": "f", "id": 34317, "pid": 435, "tid": 435, "ts": 1776649790919341, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919342, "dur": 1, + "args": { + "External id": 34318, + "cbid": 273, "correlation": 34318 + } + }, + { + "ph": "f", "id": 34318, "pid": 435, "tid": 435, "ts": 1776649790919342, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790919352, "dur": 16, + "args": { + "External id": 34319, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34319, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34319, "pid": 0, "tid": 7, "ts": 1776649790919352, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919344, "dur": 5, + "args": { + "External id": 34319, + "cbid": 211, "correlation": 34319 + } + }, + { + "ph": "s", "id": 34319, "pid": 435, "tid": 435, "ts": 1776649790919344, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790919390, "dur": 1, + "args": { + "External id": 4463, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34337, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34337, "pid": 0, "tid": 7, "ts": 1776649790919390, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919380, "dur": 7, + "args": { + "External id": 4463, + "cbid": 211, "correlation": 34337 + } + }, + { + "ph": "s", "id": 34337, "pid": 435, "tid": 435, "ts": 1776649790919380, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919393, "dur": 0, + "args": { + "External id": 34343, + "cbid": 200, "correlation": 34343 + } + }, + { + "ph": "f", "id": 34343, "pid": 435, "tid": 435, "ts": 1776649790919393, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919394, "dur": 0, + "args": { + "External id": 34344, + "cbid": 273, "correlation": 34344 + } + }, + { + "ph": "f", "id": 34344, "pid": 435, "tid": 435, "ts": 1776649790919394, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790919403, "dur": 16, + "args": { + "External id": 34345, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34345, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34345, "pid": 0, "tid": 7, "ts": 1776649790919403, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919395, "dur": 5, + "args": { + "External id": 34345, + "cbid": 211, "correlation": 34345 + } + }, + { + "ph": "s", "id": 34345, "pid": 435, "tid": 435, "ts": 1776649790919395, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790919440, "dur": 1, + "args": { + "External id": 4468, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34363, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34363, "pid": 0, "tid": 7, "ts": 1776649790919440, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919430, "dur": 6, + "args": { + "External id": 4468, + "cbid": 211, "correlation": 34363 + } + }, + { + "ph": "s", "id": 34363, "pid": 435, "tid": 435, "ts": 1776649790919430, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919443, "dur": 0, + "args": { + "External id": 34369, + "cbid": 200, "correlation": 34369 + } + }, + { + "ph": "f", "id": 34369, "pid": 435, "tid": 435, "ts": 1776649790919443, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919444, "dur": 0, + "args": { + "External id": 34370, + "cbid": 273, "correlation": 34370 + } + }, + { + "ph": "f", "id": 34370, "pid": 435, "tid": 435, "ts": 1776649790919444, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790919453, "dur": 17, + "args": { + "External id": 34371, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34371, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34371, "pid": 0, "tid": 7, "ts": 1776649790919453, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919445, "dur": 4, + "args": { + "External id": 34371, + "cbid": 211, "correlation": 34371 + } + }, + { + "ph": "s", "id": 34371, "pid": 435, "tid": 435, "ts": 1776649790919445, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790919489, "dur": 1, + "args": { + "External id": 4473, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34389, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34389, "pid": 0, "tid": 7, "ts": 1776649790919489, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919480, "dur": 6, + "args": { + "External id": 4473, + "cbid": 211, "correlation": 34389 + } + }, + { + "ph": "s", "id": 34389, "pid": 435, "tid": 435, "ts": 1776649790919480, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919492, "dur": 0, + "args": { + "External id": 34395, + "cbid": 200, "correlation": 34395 + } + }, + { + "ph": "f", "id": 34395, "pid": 435, "tid": 435, "ts": 1776649790919492, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919493, "dur": 0, + "args": { + "External id": 34396, + "cbid": 273, "correlation": 34396 + } + }, + { + "ph": "f", "id": 34396, "pid": 435, "tid": 435, "ts": 1776649790919493, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790919502, "dur": 16, + "args": { + "External id": 34397, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34397, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34397, "pid": 0, "tid": 7, "ts": 1776649790919502, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919494, "dur": 4, + "args": { + "External id": 34397, + "cbid": 211, "correlation": 34397 + } + }, + { + "ph": "s", "id": 34397, "pid": 435, "tid": 435, "ts": 1776649790919494, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790919537, "dur": 1, + "args": { + "External id": 4478, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34415, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34415, "pid": 0, "tid": 7, "ts": 1776649790919537, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919528, "dur": 6, + "args": { + "External id": 4478, + "cbid": 211, "correlation": 34415 + } + }, + { + "ph": "s", "id": 34415, "pid": 435, "tid": 435, "ts": 1776649790919528, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919541, "dur": 0, + "args": { + "External id": 34421, + "cbid": 200, "correlation": 34421 + } + }, + { + "ph": "f", "id": 34421, "pid": 435, "tid": 435, "ts": 1776649790919541, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919541, "dur": 0, + "args": { + "External id": 34422, + "cbid": 273, "correlation": 34422 + } + }, + { + "ph": "f", "id": 34422, "pid": 435, "tid": 435, "ts": 1776649790919541, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790919550, "dur": 16, + "args": { + "External id": 34423, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34423, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34423, "pid": 0, "tid": 7, "ts": 1776649790919550, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919542, "dur": 4, + "args": { + "External id": 34423, + "cbid": 211, "correlation": 34423 + } + }, + { + "ph": "s", "id": 34423, "pid": 435, "tid": 435, "ts": 1776649790919542, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790919586, "dur": 1, + "args": { + "External id": 4483, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34441, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34441, "pid": 0, "tid": 7, "ts": 1776649790919586, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919577, "dur": 6, + "args": { + "External id": 4483, + "cbid": 211, "correlation": 34441 + } + }, + { + "ph": "s", "id": 34441, "pid": 435, "tid": 435, "ts": 1776649790919577, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919590, "dur": 0, + "args": { + "External id": 34447, + "cbid": 200, "correlation": 34447 + } + }, + { + "ph": "f", "id": 34447, "pid": 435, "tid": 435, "ts": 1776649790919590, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919590, "dur": 0, + "args": { + "External id": 34448, + "cbid": 273, "correlation": 34448 + } + }, + { + "ph": "f", "id": 34448, "pid": 435, "tid": 435, "ts": 1776649790919590, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790919599, "dur": 16, + "args": { + "External id": 34449, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34449, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34449, "pid": 0, "tid": 7, "ts": 1776649790919599, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919591, "dur": 4, + "args": { + "External id": 34449, + "cbid": 211, "correlation": 34449 + } + }, + { + "ph": "s", "id": 34449, "pid": 435, "tid": 435, "ts": 1776649790919591, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790919635, "dur": 1, + "args": { + "External id": 4488, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34467, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34467, "pid": 0, "tid": 7, "ts": 1776649790919635, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919626, "dur": 6, + "args": { + "External id": 4488, + "cbid": 211, "correlation": 34467 + } + }, + { + "ph": "s", "id": 34467, "pid": 435, "tid": 435, "ts": 1776649790919626, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919638, "dur": 0, + "args": { + "External id": 34473, + "cbid": 200, "correlation": 34473 + } + }, + { + "ph": "f", "id": 34473, "pid": 435, "tid": 435, "ts": 1776649790919638, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919639, "dur": 0, + "args": { + "External id": 34474, + "cbid": 273, "correlation": 34474 + } + }, + { + "ph": "f", "id": 34474, "pid": 435, "tid": 435, "ts": 1776649790919639, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790919647, "dur": 16, + "args": { + "External id": 34475, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34475, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34475, "pid": 0, "tid": 7, "ts": 1776649790919647, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919640, "dur": 4, + "args": { + "External id": 34475, + "cbid": 211, "correlation": 34475 + } + }, + { + "ph": "s", "id": 34475, "pid": 435, "tid": 435, "ts": 1776649790919640, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790919682, "dur": 1, + "args": { + "External id": 4493, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34493, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34493, "pid": 0, "tid": 7, "ts": 1776649790919682, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919673, "dur": 6, + "args": { + "External id": 4493, + "cbid": 211, "correlation": 34493 + } + }, + { + "ph": "s", "id": 34493, "pid": 435, "tid": 435, "ts": 1776649790919673, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919686, "dur": 0, + "args": { + "External id": 34499, + "cbid": 200, "correlation": 34499 + } + }, + { + "ph": "f", "id": 34499, "pid": 435, "tid": 435, "ts": 1776649790919686, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919686, "dur": 0, + "args": { + "External id": 34500, + "cbid": 273, "correlation": 34500 + } + }, + { + "ph": "f", "id": 34500, "pid": 435, "tid": 435, "ts": 1776649790919686, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790919696, "dur": 15, + "args": { + "External id": 34501, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34501, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34501, "pid": 0, "tid": 7, "ts": 1776649790919696, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919688, "dur": 4, + "args": { + "External id": 34501, + "cbid": 211, "correlation": 34501 + } + }, + { + "ph": "s", "id": 34501, "pid": 435, "tid": 435, "ts": 1776649790919688, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790919731, "dur": 1, + "args": { + "External id": 4498, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34519, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34519, "pid": 0, "tid": 7, "ts": 1776649790919731, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919722, "dur": 6, + "args": { + "External id": 4498, + "cbid": 211, "correlation": 34519 + } + }, + { + "ph": "s", "id": 34519, "pid": 435, "tid": 435, "ts": 1776649790919722, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919734, "dur": 0, + "args": { + "External id": 34525, + "cbid": 200, "correlation": 34525 + } + }, + { + "ph": "f", "id": 34525, "pid": 435, "tid": 435, "ts": 1776649790919734, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919735, "dur": 0, + "args": { + "External id": 34526, + "cbid": 273, "correlation": 34526 + } + }, + { + "ph": "f", "id": 34526, "pid": 435, "tid": 435, "ts": 1776649790919735, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790919744, "dur": 16, + "args": { + "External id": 34527, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34527, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34527, "pid": 0, "tid": 7, "ts": 1776649790919744, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919736, "dur": 4, + "args": { + "External id": 34527, + "cbid": 211, "correlation": 34527 + } + }, + { + "ph": "s", "id": 34527, "pid": 435, "tid": 435, "ts": 1776649790919736, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790919779, "dur": 1, + "args": { + "External id": 4503, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34545, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34545, "pid": 0, "tid": 7, "ts": 1776649790919779, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919770, "dur": 6, + "args": { + "External id": 4503, + "cbid": 211, "correlation": 34545 + } + }, + { + "ph": "s", "id": 34545, "pid": 435, "tid": 435, "ts": 1776649790919770, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919782, "dur": 0, + "args": { + "External id": 34551, + "cbid": 200, "correlation": 34551 + } + }, + { + "ph": "f", "id": 34551, "pid": 435, "tid": 435, "ts": 1776649790919782, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919783, "dur": 0, + "args": { + "External id": 34552, + "cbid": 273, "correlation": 34552 + } + }, + { + "ph": "f", "id": 34552, "pid": 435, "tid": 435, "ts": 1776649790919783, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790919792, "dur": 16, + "args": { + "External id": 34553, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34553, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34553, "pid": 0, "tid": 7, "ts": 1776649790919792, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919784, "dur": 5, + "args": { + "External id": 34553, + "cbid": 211, "correlation": 34553 + } + }, + { + "ph": "s", "id": 34553, "pid": 435, "tid": 435, "ts": 1776649790919784, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790919806, "dur": 6, + "args": { + "External id": 34556, + "cbid": 165, "correlation": 34556 + } + }, + { + "ph": "s", "id": 34556, "pid": 435, "tid": 435, "ts": 1776649790919806, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790919834, "dur": 3, + "args": { + "External id": 34565, + "cbid": 135, "correlation": 34565 + } + }, + { + "ph": "f", "id": 34565, "pid": 435, "tid": 435, "ts": 1776649790919834, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790919882, "dur": 1, + "args": { + "External id": 4508, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34585, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34585, "pid": 0, "tid": 7, "ts": 1776649790919882, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919871, "dur": 8, + "args": { + "External id": 4508, + "cbid": 211, "correlation": 34585 + } + }, + { + "ph": "s", "id": 34585, "pid": 435, "tid": 435, "ts": 1776649790919871, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919886, "dur": 0, + "args": { + "External id": 34591, + "cbid": 200, "correlation": 34591 + } + }, + { + "ph": "f", "id": 34591, "pid": 435, "tid": 435, "ts": 1776649790919886, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919887, "dur": 0, + "args": { + "External id": 34592, + "cbid": 273, "correlation": 34592 + } + }, + { + "ph": "f", "id": 34592, "pid": 435, "tid": 435, "ts": 1776649790919887, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790919897, "dur": 17, + "args": { + "External id": 34593, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34593, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34593, "pid": 0, "tid": 7, "ts": 1776649790919897, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919889, "dur": 5, + "args": { + "External id": 34593, + "cbid": 211, "correlation": 34593 + } + }, + { + "ph": "s", "id": 34593, "pid": 435, "tid": 435, "ts": 1776649790919889, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790919934, "dur": 1, + "args": { + "External id": 4513, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34611, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34611, "pid": 0, "tid": 7, "ts": 1776649790919934, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919925, "dur": 6, + "args": { + "External id": 4513, + "cbid": 211, "correlation": 34611 + } + }, + { + "ph": "s", "id": 34611, "pid": 435, "tid": 435, "ts": 1776649790919925, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919937, "dur": 0, + "args": { + "External id": 34617, + "cbid": 200, "correlation": 34617 + } + }, + { + "ph": "f", "id": 34617, "pid": 435, "tid": 435, "ts": 1776649790919937, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919938, "dur": 0, + "args": { + "External id": 34618, + "cbid": 273, "correlation": 34618 + } + }, + { + "ph": "f", "id": 34618, "pid": 435, "tid": 435, "ts": 1776649790919938, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790919947, "dur": 16, + "args": { + "External id": 34619, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34619, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34619, "pid": 0, "tid": 7, "ts": 1776649790919947, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919939, "dur": 4, + "args": { + "External id": 34619, + "cbid": 211, "correlation": 34619 + } + }, + { + "ph": "s", "id": 34619, "pid": 435, "tid": 435, "ts": 1776649790919939, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790919983, "dur": 1, + "args": { + "External id": 4518, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34637, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34637, "pid": 0, "tid": 7, "ts": 1776649790919983, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919974, "dur": 6, + "args": { + "External id": 4518, + "cbid": 211, "correlation": 34637 + } + }, + { + "ph": "s", "id": 34637, "pid": 435, "tid": 435, "ts": 1776649790919974, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919987, "dur": 0, + "args": { + "External id": 34643, + "cbid": 200, "correlation": 34643 + } + }, + { + "ph": "f", "id": 34643, "pid": 435, "tid": 435, "ts": 1776649790919987, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790919987, "dur": 0, + "args": { + "External id": 34644, + "cbid": 273, "correlation": 34644 + } + }, + { + "ph": "f", "id": 34644, "pid": 435, "tid": 435, "ts": 1776649790919987, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790919996, "dur": 16, + "args": { + "External id": 34645, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34645, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34645, "pid": 0, "tid": 7, "ts": 1776649790919996, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790919988, "dur": 4, + "args": { + "External id": 34645, + "cbid": 211, "correlation": 34645 + } + }, + { + "ph": "s", "id": 34645, "pid": 435, "tid": 435, "ts": 1776649790919988, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790920032, "dur": 1, + "args": { + "External id": 4523, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34663, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34663, "pid": 0, "tid": 7, "ts": 1776649790920032, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920022, "dur": 6, + "args": { + "External id": 4523, + "cbid": 211, "correlation": 34663 + } + }, + { + "ph": "s", "id": 34663, "pid": 435, "tid": 435, "ts": 1776649790920022, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920035, "dur": 0, + "args": { + "External id": 34669, + "cbid": 200, "correlation": 34669 + } + }, + { + "ph": "f", "id": 34669, "pid": 435, "tid": 435, "ts": 1776649790920035, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920036, "dur": 0, + "args": { + "External id": 34670, + "cbid": 273, "correlation": 34670 + } + }, + { + "ph": "f", "id": 34670, "pid": 435, "tid": 435, "ts": 1776649790920036, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790920045, "dur": 15, + "args": { + "External id": 34671, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34671, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34671, "pid": 0, "tid": 7, "ts": 1776649790920045, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920037, "dur": 5, + "args": { + "External id": 34671, + "cbid": 211, "correlation": 34671 + } + }, + { + "ph": "s", "id": 34671, "pid": 435, "tid": 435, "ts": 1776649790920037, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790920081, "dur": 1, + "args": { + "External id": 4528, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34689, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34689, "pid": 0, "tid": 7, "ts": 1776649790920081, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920071, "dur": 6, + "args": { + "External id": 4528, + "cbid": 211, "correlation": 34689 + } + }, + { + "ph": "s", "id": 34689, "pid": 435, "tid": 435, "ts": 1776649790920071, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920084, "dur": 0, + "args": { + "External id": 34695, + "cbid": 200, "correlation": 34695 + } + }, + { + "ph": "f", "id": 34695, "pid": 435, "tid": 435, "ts": 1776649790920084, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920085, "dur": 0, + "args": { + "External id": 34696, + "cbid": 273, "correlation": 34696 + } + }, + { + "ph": "f", "id": 34696, "pid": 435, "tid": 435, "ts": 1776649790920085, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790920093, "dur": 15, + "args": { + "External id": 34697, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34697, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34697, "pid": 0, "tid": 7, "ts": 1776649790920093, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920086, "dur": 4, + "args": { + "External id": 34697, + "cbid": 211, "correlation": 34697 + } + }, + { + "ph": "s", "id": 34697, "pid": 435, "tid": 435, "ts": 1776649790920086, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790920129, "dur": 1, + "args": { + "External id": 4533, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34715, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34715, "pid": 0, "tid": 7, "ts": 1776649790920129, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920120, "dur": 6, + "args": { + "External id": 4533, + "cbid": 211, "correlation": 34715 + } + }, + { + "ph": "s", "id": 34715, "pid": 435, "tid": 435, "ts": 1776649790920120, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920132, "dur": 0, + "args": { + "External id": 34721, + "cbid": 200, "correlation": 34721 + } + }, + { + "ph": "f", "id": 34721, "pid": 435, "tid": 435, "ts": 1776649790920132, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920133, "dur": 0, + "args": { + "External id": 34722, + "cbid": 273, "correlation": 34722 + } + }, + { + "ph": "f", "id": 34722, "pid": 435, "tid": 435, "ts": 1776649790920133, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790920142, "dur": 16, + "args": { + "External id": 34723, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34723, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34723, "pid": 0, "tid": 7, "ts": 1776649790920142, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920134, "dur": 5, + "args": { + "External id": 34723, + "cbid": 211, "correlation": 34723 + } + }, + { + "ph": "s", "id": 34723, "pid": 435, "tid": 435, "ts": 1776649790920134, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790920177, "dur": 1, + "args": { + "External id": 4538, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34741, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34741, "pid": 0, "tid": 7, "ts": 1776649790920177, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920168, "dur": 6, + "args": { + "External id": 4538, + "cbid": 211, "correlation": 34741 + } + }, + { + "ph": "s", "id": 34741, "pid": 435, "tid": 435, "ts": 1776649790920168, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920180, "dur": 0, + "args": { + "External id": 34747, + "cbid": 200, "correlation": 34747 + } + }, + { + "ph": "f", "id": 34747, "pid": 435, "tid": 435, "ts": 1776649790920180, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920181, "dur": 0, + "args": { + "External id": 34748, + "cbid": 273, "correlation": 34748 + } + }, + { + "ph": "f", "id": 34748, "pid": 435, "tid": 435, "ts": 1776649790920181, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790920190, "dur": 16, + "args": { + "External id": 34749, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34749, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34749, "pid": 0, "tid": 7, "ts": 1776649790920190, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920182, "dur": 5, + "args": { + "External id": 34749, + "cbid": 211, "correlation": 34749 + } + }, + { + "ph": "s", "id": 34749, "pid": 435, "tid": 435, "ts": 1776649790920182, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790920246, "dur": 1, + "args": { + "External id": 4543, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34767, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34767, "pid": 0, "tid": 7, "ts": 1776649790920246, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920236, "dur": 7, + "args": { + "External id": 4543, + "cbid": 211, "correlation": 34767 + } + }, + { + "ph": "s", "id": 34767, "pid": 435, "tid": 435, "ts": 1776649790920236, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920250, "dur": 0, + "args": { + "External id": 34773, + "cbid": 200, "correlation": 34773 + } + }, + { + "ph": "f", "id": 34773, "pid": 435, "tid": 435, "ts": 1776649790920250, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920251, "dur": 0, + "args": { + "External id": 34774, + "cbid": 273, "correlation": 34774 + } + }, + { + "ph": "f", "id": 34774, "pid": 435, "tid": 435, "ts": 1776649790920251, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790920260, "dur": 15, + "args": { + "External id": 34775, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34775, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34775, "pid": 0, "tid": 7, "ts": 1776649790920260, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920252, "dur": 5, + "args": { + "External id": 34775, + "cbid": 211, "correlation": 34775 + } + }, + { + "ph": "s", "id": 34775, "pid": 435, "tid": 435, "ts": 1776649790920252, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790920299, "dur": 1, + "args": { + "External id": 4548, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34793, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34793, "pid": 0, "tid": 7, "ts": 1776649790920299, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920289, "dur": 7, + "args": { + "External id": 4548, + "cbid": 211, "correlation": 34793 + } + }, + { + "ph": "s", "id": 34793, "pid": 435, "tid": 435, "ts": 1776649790920289, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920303, "dur": 0, + "args": { + "External id": 34799, + "cbid": 200, "correlation": 34799 + } + }, + { + "ph": "f", "id": 34799, "pid": 435, "tid": 435, "ts": 1776649790920303, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920303, "dur": 0, + "args": { + "External id": 34800, + "cbid": 273, "correlation": 34800 + } + }, + { + "ph": "f", "id": 34800, "pid": 435, "tid": 435, "ts": 1776649790920303, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790920312, "dur": 15, + "args": { + "External id": 34801, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34801, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34801, "pid": 0, "tid": 7, "ts": 1776649790920312, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920304, "dur": 4, + "args": { + "External id": 34801, + "cbid": 211, "correlation": 34801 + } + }, + { + "ph": "s", "id": 34801, "pid": 435, "tid": 435, "ts": 1776649790920304, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790920351, "dur": 1, + "args": { + "External id": 4553, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34819, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34819, "pid": 0, "tid": 7, "ts": 1776649790920351, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920342, "dur": 7, + "args": { + "External id": 4553, + "cbid": 211, "correlation": 34819 + } + }, + { + "ph": "s", "id": 34819, "pid": 435, "tid": 435, "ts": 1776649790920342, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920355, "dur": 0, + "args": { + "External id": 34825, + "cbid": 200, "correlation": 34825 + } + }, + { + "ph": "f", "id": 34825, "pid": 435, "tid": 435, "ts": 1776649790920355, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920355, "dur": 0, + "args": { + "External id": 34826, + "cbid": 273, "correlation": 34826 + } + }, + { + "ph": "f", "id": 34826, "pid": 435, "tid": 435, "ts": 1776649790920355, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790920364, "dur": 16, + "args": { + "External id": 34827, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34827, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34827, "pid": 0, "tid": 7, "ts": 1776649790920364, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920356, "dur": 5, + "args": { + "External id": 34827, + "cbid": 211, "correlation": 34827 + } + }, + { + "ph": "s", "id": 34827, "pid": 435, "tid": 435, "ts": 1776649790920356, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790920400, "dur": 1, + "args": { + "External id": 4558, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34845, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34845, "pid": 0, "tid": 7, "ts": 1776649790920400, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920390, "dur": 6, + "args": { + "External id": 4558, + "cbid": 211, "correlation": 34845 + } + }, + { + "ph": "s", "id": 34845, "pid": 435, "tid": 435, "ts": 1776649790920390, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920403, "dur": 0, + "args": { + "External id": 34851, + "cbid": 200, "correlation": 34851 + } + }, + { + "ph": "f", "id": 34851, "pid": 435, "tid": 435, "ts": 1776649790920403, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920404, "dur": 0, + "args": { + "External id": 34852, + "cbid": 273, "correlation": 34852 + } + }, + { + "ph": "f", "id": 34852, "pid": 435, "tid": 435, "ts": 1776649790920404, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790920412, "dur": 15, + "args": { + "External id": 34853, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34853, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34853, "pid": 0, "tid": 7, "ts": 1776649790920412, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920405, "dur": 4, + "args": { + "External id": 34853, + "cbid": 211, "correlation": 34853 + } + }, + { + "ph": "s", "id": 34853, "pid": 435, "tid": 435, "ts": 1776649790920405, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790920449, "dur": 1, + "args": { + "External id": 4563, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34871, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34871, "pid": 0, "tid": 7, "ts": 1776649790920449, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920439, "dur": 6, + "args": { + "External id": 4563, + "cbid": 211, "correlation": 34871 + } + }, + { + "ph": "s", "id": 34871, "pid": 435, "tid": 435, "ts": 1776649790920439, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920452, "dur": 0, + "args": { + "External id": 34877, + "cbid": 200, "correlation": 34877 + } + }, + { + "ph": "f", "id": 34877, "pid": 435, "tid": 435, "ts": 1776649790920452, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920453, "dur": 0, + "args": { + "External id": 34878, + "cbid": 273, "correlation": 34878 + } + }, + { + "ph": "f", "id": 34878, "pid": 435, "tid": 435, "ts": 1776649790920453, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790920461, "dur": 16, + "args": { + "External id": 34879, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34879, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34879, "pid": 0, "tid": 7, "ts": 1776649790920461, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920454, "dur": 4, + "args": { + "External id": 34879, + "cbid": 211, "correlation": 34879 + } + }, + { + "ph": "s", "id": 34879, "pid": 435, "tid": 435, "ts": 1776649790920454, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790920497, "dur": 1, + "args": { + "External id": 4568, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34897, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34897, "pid": 0, "tid": 7, "ts": 1776649790920497, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920488, "dur": 6, + "args": { + "External id": 4568, + "cbid": 211, "correlation": 34897 + } + }, + { + "ph": "s", "id": 34897, "pid": 435, "tid": 435, "ts": 1776649790920488, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920500, "dur": 0, + "args": { + "External id": 34903, + "cbid": 200, "correlation": 34903 + } + }, + { + "ph": "f", "id": 34903, "pid": 435, "tid": 435, "ts": 1776649790920500, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920501, "dur": 0, + "args": { + "External id": 34904, + "cbid": 273, "correlation": 34904 + } + }, + { + "ph": "f", "id": 34904, "pid": 435, "tid": 435, "ts": 1776649790920501, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790920510, "dur": 15, + "args": { + "External id": 34905, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34905, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34905, "pid": 0, "tid": 7, "ts": 1776649790920510, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920502, "dur": 4, + "args": { + "External id": 34905, + "cbid": 211, "correlation": 34905 + } + }, + { + "ph": "s", "id": 34905, "pid": 435, "tid": 435, "ts": 1776649790920502, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790920546, "dur": 1, + "args": { + "External id": 4573, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34923, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34923, "pid": 0, "tid": 7, "ts": 1776649790920546, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920536, "dur": 6, + "args": { + "External id": 4573, + "cbid": 211, "correlation": 34923 + } + }, + { + "ph": "s", "id": 34923, "pid": 435, "tid": 435, "ts": 1776649790920536, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920549, "dur": 0, + "args": { + "External id": 34929, + "cbid": 200, "correlation": 34929 + } + }, + { + "ph": "f", "id": 34929, "pid": 435, "tid": 435, "ts": 1776649790920549, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920549, "dur": 0, + "args": { + "External id": 34930, + "cbid": 273, "correlation": 34930 + } + }, + { + "ph": "f", "id": 34930, "pid": 435, "tid": 435, "ts": 1776649790920549, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790920559, "dur": 15, + "args": { + "External id": 34931, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34931, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34931, "pid": 0, "tid": 7, "ts": 1776649790920559, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920551, "dur": 5, + "args": { + "External id": 34931, + "cbid": 211, "correlation": 34931 + } + }, + { + "ph": "s", "id": 34931, "pid": 435, "tid": 435, "ts": 1776649790920551, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790920594, "dur": 1, + "args": { + "External id": 4578, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34949, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34949, "pid": 0, "tid": 7, "ts": 1776649790920594, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920585, "dur": 6, + "args": { + "External id": 4578, + "cbid": 211, "correlation": 34949 + } + }, + { + "ph": "s", "id": 34949, "pid": 435, "tid": 435, "ts": 1776649790920585, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920597, "dur": 0, + "args": { + "External id": 34955, + "cbid": 200, "correlation": 34955 + } + }, + { + "ph": "f", "id": 34955, "pid": 435, "tid": 435, "ts": 1776649790920597, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920598, "dur": 0, + "args": { + "External id": 34956, + "cbid": 273, "correlation": 34956 + } + }, + { + "ph": "f", "id": 34956, "pid": 435, "tid": 435, "ts": 1776649790920598, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790920607, "dur": 15, + "args": { + "External id": 34957, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34957, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34957, "pid": 0, "tid": 7, "ts": 1776649790920607, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920599, "dur": 4, + "args": { + "External id": 34957, + "cbid": 211, "correlation": 34957 + } + }, + { + "ph": "s", "id": 34957, "pid": 435, "tid": 435, "ts": 1776649790920599, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790920643, "dur": 1, + "args": { + "External id": 4583, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34975, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34975, "pid": 0, "tid": 7, "ts": 1776649790920643, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920633, "dur": 6, + "args": { + "External id": 4583, + "cbid": 211, "correlation": 34975 + } + }, + { + "ph": "s", "id": 34975, "pid": 435, "tid": 435, "ts": 1776649790920633, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920646, "dur": 0, + "args": { + "External id": 34981, + "cbid": 200, "correlation": 34981 + } + }, + { + "ph": "f", "id": 34981, "pid": 435, "tid": 435, "ts": 1776649790920646, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920646, "dur": 0, + "args": { + "External id": 34982, + "cbid": 273, "correlation": 34982 + } + }, + { + "ph": "f", "id": 34982, "pid": 435, "tid": 435, "ts": 1776649790920646, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790920655, "dur": 15, + "args": { + "External id": 34983, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 34983, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 34983, "pid": 0, "tid": 7, "ts": 1776649790920655, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920647, "dur": 5, + "args": { + "External id": 34983, + "cbid": 211, "correlation": 34983 + } + }, + { + "ph": "s", "id": 34983, "pid": 435, "tid": 435, "ts": 1776649790920647, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790920691, "dur": 1, + "args": { + "External id": 4588, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35001, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35001, "pid": 0, "tid": 7, "ts": 1776649790920691, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920681, "dur": 6, + "args": { + "External id": 4588, + "cbid": 211, "correlation": 35001 + } + }, + { + "ph": "s", "id": 35001, "pid": 435, "tid": 435, "ts": 1776649790920681, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920694, "dur": 0, + "args": { + "External id": 35007, + "cbid": 200, "correlation": 35007 + } + }, + { + "ph": "f", "id": 35007, "pid": 435, "tid": 435, "ts": 1776649790920694, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920695, "dur": 0, + "args": { + "External id": 35008, + "cbid": 273, "correlation": 35008 + } + }, + { + "ph": "f", "id": 35008, "pid": 435, "tid": 435, "ts": 1776649790920695, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790920704, "dur": 15, + "args": { + "External id": 35009, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35009, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35009, "pid": 0, "tid": 7, "ts": 1776649790920704, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920696, "dur": 4, + "args": { + "External id": 35009, + "cbid": 211, "correlation": 35009 + } + }, + { + "ph": "s", "id": 35009, "pid": 435, "tid": 435, "ts": 1776649790920696, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790920740, "dur": 1, + "args": { + "External id": 4593, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35027, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35027, "pid": 0, "tid": 7, "ts": 1776649790920740, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920731, "dur": 6, + "args": { + "External id": 4593, + "cbid": 211, "correlation": 35027 + } + }, + { + "ph": "s", "id": 35027, "pid": 435, "tid": 435, "ts": 1776649790920731, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920743, "dur": 0, + "args": { + "External id": 35033, + "cbid": 200, "correlation": 35033 + } + }, + { + "ph": "f", "id": 35033, "pid": 435, "tid": 435, "ts": 1776649790920743, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920744, "dur": 0, + "args": { + "External id": 35034, + "cbid": 273, "correlation": 35034 + } + }, + { + "ph": "f", "id": 35034, "pid": 435, "tid": 435, "ts": 1776649790920744, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790920753, "dur": 16, + "args": { + "External id": 35035, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35035, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35035, "pid": 0, "tid": 7, "ts": 1776649790920753, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920745, "dur": 5, + "args": { + "External id": 35035, + "cbid": 211, "correlation": 35035 + } + }, + { + "ph": "s", "id": 35035, "pid": 435, "tid": 435, "ts": 1776649790920745, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790920789, "dur": 1, + "args": { + "External id": 4598, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35053, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35053, "pid": 0, "tid": 7, "ts": 1776649790920789, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920780, "dur": 7, + "args": { + "External id": 4598, + "cbid": 211, "correlation": 35053 + } + }, + { + "ph": "s", "id": 35053, "pid": 435, "tid": 435, "ts": 1776649790920780, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920792, "dur": 0, + "args": { + "External id": 35059, + "cbid": 200, "correlation": 35059 + } + }, + { + "ph": "f", "id": 35059, "pid": 435, "tid": 435, "ts": 1776649790920792, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920793, "dur": 0, + "args": { + "External id": 35060, + "cbid": 273, "correlation": 35060 + } + }, + { + "ph": "f", "id": 35060, "pid": 435, "tid": 435, "ts": 1776649790920793, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790920802, "dur": 16, + "args": { + "External id": 35061, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35061, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35061, "pid": 0, "tid": 7, "ts": 1776649790920802, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920794, "dur": 4, + "args": { + "External id": 35061, + "cbid": 211, "correlation": 35061 + } + }, + { + "ph": "s", "id": 35061, "pid": 435, "tid": 435, "ts": 1776649790920794, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790920838, "dur": 1, + "args": { + "External id": 4603, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35079, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35079, "pid": 0, "tid": 7, "ts": 1776649790920838, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920829, "dur": 7, + "args": { + "External id": 4603, + "cbid": 211, "correlation": 35079 + } + }, + { + "ph": "s", "id": 35079, "pid": 435, "tid": 435, "ts": 1776649790920829, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920841, "dur": 0, + "args": { + "External id": 35085, + "cbid": 200, "correlation": 35085 + } + }, + { + "ph": "f", "id": 35085, "pid": 435, "tid": 435, "ts": 1776649790920841, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920842, "dur": 0, + "args": { + "External id": 35086, + "cbid": 273, "correlation": 35086 + } + }, + { + "ph": "f", "id": 35086, "pid": 435, "tid": 435, "ts": 1776649790920842, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790920851, "dur": 15, + "args": { + "External id": 35087, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35087, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35087, "pid": 0, "tid": 7, "ts": 1776649790920851, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920843, "dur": 5, + "args": { + "External id": 35087, + "cbid": 211, "correlation": 35087 + } + }, + { + "ph": "s", "id": 35087, "pid": 435, "tid": 435, "ts": 1776649790920843, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790920887, "dur": 1, + "args": { + "External id": 4608, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35105, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35105, "pid": 0, "tid": 7, "ts": 1776649790920887, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920877, "dur": 6, + "args": { + "External id": 4608, + "cbid": 211, "correlation": 35105 + } + }, + { + "ph": "s", "id": 35105, "pid": 435, "tid": 435, "ts": 1776649790920877, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920890, "dur": 0, + "args": { + "External id": 35111, + "cbid": 200, "correlation": 35111 + } + }, + { + "ph": "f", "id": 35111, "pid": 435, "tid": 435, "ts": 1776649790920890, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920890, "dur": 0, + "args": { + "External id": 35112, + "cbid": 273, "correlation": 35112 + } + }, + { + "ph": "f", "id": 35112, "pid": 435, "tid": 435, "ts": 1776649790920890, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790920899, "dur": 15, + "args": { + "External id": 35113, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35113, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35113, "pid": 0, "tid": 7, "ts": 1776649790920899, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920891, "dur": 4, + "args": { + "External id": 35113, + "cbid": 211, "correlation": 35113 + } + }, + { + "ph": "s", "id": 35113, "pid": 435, "tid": 435, "ts": 1776649790920891, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790920945, "dur": 1, + "args": { + "External id": 4613, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35131, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35131, "pid": 0, "tid": 7, "ts": 1776649790920945, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920935, "dur": 6, + "args": { + "External id": 4613, + "cbid": 211, "correlation": 35131 + } + }, + { + "ph": "s", "id": 35131, "pid": 435, "tid": 435, "ts": 1776649790920935, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920948, "dur": 0, + "args": { + "External id": 35137, + "cbid": 200, "correlation": 35137 + } + }, + { + "ph": "f", "id": 35137, "pid": 435, "tid": 435, "ts": 1776649790920948, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920948, "dur": 0, + "args": { + "External id": 35138, + "cbid": 273, "correlation": 35138 + } + }, + { + "ph": "f", "id": 35138, "pid": 435, "tid": 435, "ts": 1776649790920948, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790920957, "dur": 15, + "args": { + "External id": 35139, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35139, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35139, "pid": 0, "tid": 7, "ts": 1776649790920957, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920950, "dur": 4, + "args": { + "External id": 35139, + "cbid": 211, "correlation": 35139 + } + }, + { + "ph": "s", "id": 35139, "pid": 435, "tid": 435, "ts": 1776649790920950, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790920993, "dur": 1, + "args": { + "External id": 4618, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35157, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35157, "pid": 0, "tid": 7, "ts": 1776649790920993, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920984, "dur": 6, + "args": { + "External id": 4618, + "cbid": 211, "correlation": 35157 + } + }, + { + "ph": "s", "id": 35157, "pid": 435, "tid": 435, "ts": 1776649790920984, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920996, "dur": 0, + "args": { + "External id": 35163, + "cbid": 200, "correlation": 35163 + } + }, + { + "ph": "f", "id": 35163, "pid": 435, "tid": 435, "ts": 1776649790920996, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790920997, "dur": 0, + "args": { + "External id": 35164, + "cbid": 273, "correlation": 35164 + } + }, + { + "ph": "f", "id": 35164, "pid": 435, "tid": 435, "ts": 1776649790920997, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790921006, "dur": 16, + "args": { + "External id": 35165, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35165, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35165, "pid": 0, "tid": 7, "ts": 1776649790921006, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790920998, "dur": 4, + "args": { + "External id": 35165, + "cbid": 211, "correlation": 35165 + } + }, + { + "ph": "s", "id": 35165, "pid": 435, "tid": 435, "ts": 1776649790920998, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790921042, "dur": 1, + "args": { + "External id": 4623, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35183, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35183, "pid": 0, "tid": 7, "ts": 1776649790921042, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921032, "dur": 6, + "args": { + "External id": 4623, + "cbid": 211, "correlation": 35183 + } + }, + { + "ph": "s", "id": 35183, "pid": 435, "tid": 435, "ts": 1776649790921032, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921045, "dur": 0, + "args": { + "External id": 35189, + "cbid": 200, "correlation": 35189 + } + }, + { + "ph": "f", "id": 35189, "pid": 435, "tid": 435, "ts": 1776649790921045, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921045, "dur": 0, + "args": { + "External id": 35190, + "cbid": 273, "correlation": 35190 + } + }, + { + "ph": "f", "id": 35190, "pid": 435, "tid": 435, "ts": 1776649790921045, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790921054, "dur": 15, + "args": { + "External id": 35191, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35191, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35191, "pid": 0, "tid": 7, "ts": 1776649790921054, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921047, "dur": 5, + "args": { + "External id": 35191, + "cbid": 211, "correlation": 35191 + } + }, + { + "ph": "s", "id": 35191, "pid": 435, "tid": 435, "ts": 1776649790921047, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790921090, "dur": 1, + "args": { + "External id": 4628, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35209, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35209, "pid": 0, "tid": 7, "ts": 1776649790921090, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921080, "dur": 6, + "args": { + "External id": 4628, + "cbid": 211, "correlation": 35209 + } + }, + { + "ph": "s", "id": 35209, "pid": 435, "tid": 435, "ts": 1776649790921080, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921093, "dur": 0, + "args": { + "External id": 35215, + "cbid": 200, "correlation": 35215 + } + }, + { + "ph": "f", "id": 35215, "pid": 435, "tid": 435, "ts": 1776649790921093, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921094, "dur": 0, + "args": { + "External id": 35216, + "cbid": 273, "correlation": 35216 + } + }, + { + "ph": "f", "id": 35216, "pid": 435, "tid": 435, "ts": 1776649790921094, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790921102, "dur": 16, + "args": { + "External id": 35217, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35217, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35217, "pid": 0, "tid": 7, "ts": 1776649790921102, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921095, "dur": 4, + "args": { + "External id": 35217, + "cbid": 211, "correlation": 35217 + } + }, + { + "ph": "s", "id": 35217, "pid": 435, "tid": 435, "ts": 1776649790921095, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790921138, "dur": 1, + "args": { + "External id": 4633, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35235, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35235, "pid": 0, "tid": 7, "ts": 1776649790921138, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921129, "dur": 6, + "args": { + "External id": 4633, + "cbid": 211, "correlation": 35235 + } + }, + { + "ph": "s", "id": 35235, "pid": 435, "tid": 435, "ts": 1776649790921129, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921141, "dur": 0, + "args": { + "External id": 35241, + "cbid": 200, "correlation": 35241 + } + }, + { + "ph": "f", "id": 35241, "pid": 435, "tid": 435, "ts": 1776649790921141, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921142, "dur": 0, + "args": { + "External id": 35242, + "cbid": 273, "correlation": 35242 + } + }, + { + "ph": "f", "id": 35242, "pid": 435, "tid": 435, "ts": 1776649790921142, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790921151, "dur": 16, + "args": { + "External id": 35243, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35243, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35243, "pid": 0, "tid": 7, "ts": 1776649790921151, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921143, "dur": 4, + "args": { + "External id": 35243, + "cbid": 211, "correlation": 35243 + } + }, + { + "ph": "s", "id": 35243, "pid": 435, "tid": 435, "ts": 1776649790921143, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790921186, "dur": 1, + "args": { + "External id": 4638, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35261, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35261, "pid": 0, "tid": 7, "ts": 1776649790921186, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921177, "dur": 6, + "args": { + "External id": 4638, + "cbid": 211, "correlation": 35261 + } + }, + { + "ph": "s", "id": 35261, "pid": 435, "tid": 435, "ts": 1776649790921177, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921189, "dur": 0, + "args": { + "External id": 35267, + "cbid": 200, "correlation": 35267 + } + }, + { + "ph": "f", "id": 35267, "pid": 435, "tid": 435, "ts": 1776649790921189, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921190, "dur": 0, + "args": { + "External id": 35268, + "cbid": 273, "correlation": 35268 + } + }, + { + "ph": "f", "id": 35268, "pid": 435, "tid": 435, "ts": 1776649790921190, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790921199, "dur": 17, + "args": { + "External id": 35269, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35269, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35269, "pid": 0, "tid": 7, "ts": 1776649790921199, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921191, "dur": 4, + "args": { + "External id": 35269, + "cbid": 211, "correlation": 35269 + } + }, + { + "ph": "s", "id": 35269, "pid": 435, "tid": 435, "ts": 1776649790921191, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790921235, "dur": 1, + "args": { + "External id": 4643, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35287, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35287, "pid": 0, "tid": 7, "ts": 1776649790921235, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921225, "dur": 7, + "args": { + "External id": 4643, + "cbid": 211, "correlation": 35287 + } + }, + { + "ph": "s", "id": 35287, "pid": 435, "tid": 435, "ts": 1776649790921225, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921238, "dur": 0, + "args": { + "External id": 35293, + "cbid": 200, "correlation": 35293 + } + }, + { + "ph": "f", "id": 35293, "pid": 435, "tid": 435, "ts": 1776649790921238, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921238, "dur": 0, + "args": { + "External id": 35294, + "cbid": 273, "correlation": 35294 + } + }, + { + "ph": "f", "id": 35294, "pid": 435, "tid": 435, "ts": 1776649790921238, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790921247, "dur": 16, + "args": { + "External id": 35295, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35295, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35295, "pid": 0, "tid": 7, "ts": 1776649790921247, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921239, "dur": 5, + "args": { + "External id": 35295, + "cbid": 211, "correlation": 35295 + } + }, + { + "ph": "s", "id": 35295, "pid": 435, "tid": 435, "ts": 1776649790921239, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790921283, "dur": 1, + "args": { + "External id": 4648, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35313, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35313, "pid": 0, "tid": 7, "ts": 1776649790921283, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921273, "dur": 6, + "args": { + "External id": 4648, + "cbid": 211, "correlation": 35313 + } + }, + { + "ph": "s", "id": 35313, "pid": 435, "tid": 435, "ts": 1776649790921273, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921286, "dur": 0, + "args": { + "External id": 35319, + "cbid": 200, "correlation": 35319 + } + }, + { + "ph": "f", "id": 35319, "pid": 435, "tid": 435, "ts": 1776649790921286, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921287, "dur": 0, + "args": { + "External id": 35320, + "cbid": 273, "correlation": 35320 + } + }, + { + "ph": "f", "id": 35320, "pid": 435, "tid": 435, "ts": 1776649790921287, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790921296, "dur": 15, + "args": { + "External id": 35321, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35321, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35321, "pid": 0, "tid": 7, "ts": 1776649790921296, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921288, "dur": 4, + "args": { + "External id": 35321, + "cbid": 211, "correlation": 35321 + } + }, + { + "ph": "s", "id": 35321, "pid": 435, "tid": 435, "ts": 1776649790921288, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790921331, "dur": 1, + "args": { + "External id": 4653, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35339, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35339, "pid": 0, "tid": 7, "ts": 1776649790921331, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921322, "dur": 6, + "args": { + "External id": 4653, + "cbid": 211, "correlation": 35339 + } + }, + { + "ph": "s", "id": 35339, "pid": 435, "tid": 435, "ts": 1776649790921322, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921334, "dur": 0, + "args": { + "External id": 35345, + "cbid": 200, "correlation": 35345 + } + }, + { + "ph": "f", "id": 35345, "pid": 435, "tid": 435, "ts": 1776649790921334, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921335, "dur": 0, + "args": { + "External id": 35346, + "cbid": 273, "correlation": 35346 + } + }, + { + "ph": "f", "id": 35346, "pid": 435, "tid": 435, "ts": 1776649790921335, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790921344, "dur": 16, + "args": { + "External id": 35347, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35347, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35347, "pid": 0, "tid": 7, "ts": 1776649790921344, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921336, "dur": 5, + "args": { + "External id": 35347, + "cbid": 211, "correlation": 35347 + } + }, + { + "ph": "s", "id": 35347, "pid": 435, "tid": 435, "ts": 1776649790921336, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790921379, "dur": 1, + "args": { + "External id": 4658, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35365, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35365, "pid": 0, "tid": 7, "ts": 1776649790921379, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921369, "dur": 6, + "args": { + "External id": 4658, + "cbid": 211, "correlation": 35365 + } + }, + { + "ph": "s", "id": 35365, "pid": 435, "tid": 435, "ts": 1776649790921369, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921382, "dur": 0, + "args": { + "External id": 35371, + "cbid": 200, "correlation": 35371 + } + }, + { + "ph": "f", "id": 35371, "pid": 435, "tid": 435, "ts": 1776649790921382, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921383, "dur": 0, + "args": { + "External id": 35372, + "cbid": 273, "correlation": 35372 + } + }, + { + "ph": "f", "id": 35372, "pid": 435, "tid": 435, "ts": 1776649790921383, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790921392, "dur": 15, + "args": { + "External id": 35373, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35373, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35373, "pid": 0, "tid": 7, "ts": 1776649790921392, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921384, "dur": 4, + "args": { + "External id": 35373, + "cbid": 211, "correlation": 35373 + } + }, + { + "ph": "s", "id": 35373, "pid": 435, "tid": 435, "ts": 1776649790921384, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790921427, "dur": 1, + "args": { + "External id": 4663, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35391, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35391, "pid": 0, "tid": 7, "ts": 1776649790921427, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921418, "dur": 6, + "args": { + "External id": 4663, + "cbid": 211, "correlation": 35391 + } + }, + { + "ph": "s", "id": 35391, "pid": 435, "tid": 435, "ts": 1776649790921418, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921430, "dur": 0, + "args": { + "External id": 35397, + "cbid": 200, "correlation": 35397 + } + }, + { + "ph": "f", "id": 35397, "pid": 435, "tid": 435, "ts": 1776649790921430, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921431, "dur": 0, + "args": { + "External id": 35398, + "cbid": 273, "correlation": 35398 + } + }, + { + "ph": "f", "id": 35398, "pid": 435, "tid": 435, "ts": 1776649790921431, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790921440, "dur": 16, + "args": { + "External id": 35399, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35399, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35399, "pid": 0, "tid": 7, "ts": 1776649790921440, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921432, "dur": 5, + "args": { + "External id": 35399, + "cbid": 211, "correlation": 35399 + } + }, + { + "ph": "s", "id": 35399, "pid": 435, "tid": 435, "ts": 1776649790921432, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790921475, "dur": 1, + "args": { + "External id": 4668, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35417, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35417, "pid": 0, "tid": 7, "ts": 1776649790921475, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921466, "dur": 6, + "args": { + "External id": 4668, + "cbid": 211, "correlation": 35417 + } + }, + { + "ph": "s", "id": 35417, "pid": 435, "tid": 435, "ts": 1776649790921466, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921478, "dur": 0, + "args": { + "External id": 35423, + "cbid": 200, "correlation": 35423 + } + }, + { + "ph": "f", "id": 35423, "pid": 435, "tid": 435, "ts": 1776649790921478, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921479, "dur": 0, + "args": { + "External id": 35424, + "cbid": 273, "correlation": 35424 + } + }, + { + "ph": "f", "id": 35424, "pid": 435, "tid": 435, "ts": 1776649790921479, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790921488, "dur": 16, + "args": { + "External id": 35425, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35425, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35425, "pid": 0, "tid": 7, "ts": 1776649790921488, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921480, "dur": 4, + "args": { + "External id": 35425, + "cbid": 211, "correlation": 35425 + } + }, + { + "ph": "s", "id": 35425, "pid": 435, "tid": 435, "ts": 1776649790921480, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790921523, "dur": 1, + "args": { + "External id": 4673, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35443, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35443, "pid": 0, "tid": 7, "ts": 1776649790921523, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921513, "dur": 6, + "args": { + "External id": 4673, + "cbid": 211, "correlation": 35443 + } + }, + { + "ph": "s", "id": 35443, "pid": 435, "tid": 435, "ts": 1776649790921513, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921526, "dur": 0, + "args": { + "External id": 35449, + "cbid": 200, "correlation": 35449 + } + }, + { + "ph": "f", "id": 35449, "pid": 435, "tid": 435, "ts": 1776649790921526, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921526, "dur": 0, + "args": { + "External id": 35450, + "cbid": 273, "correlation": 35450 + } + }, + { + "ph": "f", "id": 35450, "pid": 435, "tid": 435, "ts": 1776649790921526, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790921535, "dur": 16, + "args": { + "External id": 35451, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35451, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35451, "pid": 0, "tid": 7, "ts": 1776649790921535, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921528, "dur": 5, + "args": { + "External id": 35451, + "cbid": 211, "correlation": 35451 + } + }, + { + "ph": "s", "id": 35451, "pid": 435, "tid": 435, "ts": 1776649790921528, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790921571, "dur": 1, + "args": { + "External id": 4678, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35469, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35469, "pid": 0, "tid": 7, "ts": 1776649790921571, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921561, "dur": 6, + "args": { + "External id": 4678, + "cbid": 211, "correlation": 35469 + } + }, + { + "ph": "s", "id": 35469, "pid": 435, "tid": 435, "ts": 1776649790921561, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921574, "dur": 0, + "args": { + "External id": 35475, + "cbid": 200, "correlation": 35475 + } + }, + { + "ph": "f", "id": 35475, "pid": 435, "tid": 435, "ts": 1776649790921574, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921574, "dur": 0, + "args": { + "External id": 35476, + "cbid": 273, "correlation": 35476 + } + }, + { + "ph": "f", "id": 35476, "pid": 435, "tid": 435, "ts": 1776649790921574, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790921583, "dur": 16, + "args": { + "External id": 35477, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35477, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35477, "pid": 0, "tid": 7, "ts": 1776649790921583, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921576, "dur": 4, + "args": { + "External id": 35477, + "cbid": 211, "correlation": 35477 + } + }, + { + "ph": "s", "id": 35477, "pid": 435, "tid": 435, "ts": 1776649790921576, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790921619, "dur": 1, + "args": { + "External id": 4683, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35495, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35495, "pid": 0, "tid": 7, "ts": 1776649790921619, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921609, "dur": 6, + "args": { + "External id": 4683, + "cbid": 211, "correlation": 35495 + } + }, + { + "ph": "s", "id": 35495, "pid": 435, "tid": 435, "ts": 1776649790921609, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921622, "dur": 0, + "args": { + "External id": 35501, + "cbid": 200, "correlation": 35501 + } + }, + { + "ph": "f", "id": 35501, "pid": 435, "tid": 435, "ts": 1776649790921622, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921622, "dur": 0, + "args": { + "External id": 35502, + "cbid": 273, "correlation": 35502 + } + }, + { + "ph": "f", "id": 35502, "pid": 435, "tid": 435, "ts": 1776649790921622, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790921632, "dur": 16, + "args": { + "External id": 35503, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35503, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35503, "pid": 0, "tid": 7, "ts": 1776649790921632, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921624, "dur": 5, + "args": { + "External id": 35503, + "cbid": 211, "correlation": 35503 + } + }, + { + "ph": "s", "id": 35503, "pid": 435, "tid": 435, "ts": 1776649790921624, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790921667, "dur": 1, + "args": { + "External id": 4688, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35521, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35521, "pid": 0, "tid": 7, "ts": 1776649790921667, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921657, "dur": 6, + "args": { + "External id": 4688, + "cbid": 211, "correlation": 35521 + } + }, + { + "ph": "s", "id": 35521, "pid": 435, "tid": 435, "ts": 1776649790921657, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921670, "dur": 0, + "args": { + "External id": 35527, + "cbid": 200, "correlation": 35527 + } + }, + { + "ph": "f", "id": 35527, "pid": 435, "tid": 435, "ts": 1776649790921670, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921670, "dur": 0, + "args": { + "External id": 35528, + "cbid": 273, "correlation": 35528 + } + }, + { + "ph": "f", "id": 35528, "pid": 435, "tid": 435, "ts": 1776649790921670, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790921679, "dur": 16, + "args": { + "External id": 35529, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35529, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35529, "pid": 0, "tid": 7, "ts": 1776649790921679, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921672, "dur": 4, + "args": { + "External id": 35529, + "cbid": 211, "correlation": 35529 + } + }, + { + "ph": "s", "id": 35529, "pid": 435, "tid": 435, "ts": 1776649790921672, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790921715, "dur": 1, + "args": { + "External id": 4693, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35547, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35547, "pid": 0, "tid": 7, "ts": 1776649790921715, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921706, "dur": 6, + "args": { + "External id": 4693, + "cbid": 211, "correlation": 35547 + } + }, + { + "ph": "s", "id": 35547, "pid": 435, "tid": 435, "ts": 1776649790921706, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921718, "dur": 0, + "args": { + "External id": 35553, + "cbid": 200, "correlation": 35553 + } + }, + { + "ph": "f", "id": 35553, "pid": 435, "tid": 435, "ts": 1776649790921718, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921718, "dur": 0, + "args": { + "External id": 35554, + "cbid": 273, "correlation": 35554 + } + }, + { + "ph": "f", "id": 35554, "pid": 435, "tid": 435, "ts": 1776649790921718, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790921727, "dur": 16, + "args": { + "External id": 35555, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35555, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35555, "pid": 0, "tid": 7, "ts": 1776649790921727, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921720, "dur": 4, + "args": { + "External id": 35555, + "cbid": 211, "correlation": 35555 + } + }, + { + "ph": "s", "id": 35555, "pid": 435, "tid": 435, "ts": 1776649790921720, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790921762, "dur": 1, + "args": { + "External id": 4698, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35573, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35573, "pid": 0, "tid": 7, "ts": 1776649790921762, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921753, "dur": 6, + "args": { + "External id": 4698, + "cbid": 211, "correlation": 35573 + } + }, + { + "ph": "s", "id": 35573, "pid": 435, "tid": 435, "ts": 1776649790921753, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921766, "dur": 0, + "args": { + "External id": 35579, + "cbid": 200, "correlation": 35579 + } + }, + { + "ph": "f", "id": 35579, "pid": 435, "tid": 435, "ts": 1776649790921766, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921766, "dur": 0, + "args": { + "External id": 35580, + "cbid": 273, "correlation": 35580 + } + }, + { + "ph": "f", "id": 35580, "pid": 435, "tid": 435, "ts": 1776649790921766, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790921775, "dur": 16, + "args": { + "External id": 35581, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35581, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35581, "pid": 0, "tid": 7, "ts": 1776649790921775, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921767, "dur": 4, + "args": { + "External id": 35581, + "cbid": 211, "correlation": 35581 + } + }, + { + "ph": "s", "id": 35581, "pid": 435, "tid": 435, "ts": 1776649790921767, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790921811, "dur": 1, + "args": { + "External id": 4703, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35599, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35599, "pid": 0, "tid": 7, "ts": 1776649790921811, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921801, "dur": 6, + "args": { + "External id": 4703, + "cbid": 211, "correlation": 35599 + } + }, + { + "ph": "s", "id": 35599, "pid": 435, "tid": 435, "ts": 1776649790921801, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921814, "dur": 0, + "args": { + "External id": 35605, + "cbid": 200, "correlation": 35605 + } + }, + { + "ph": "f", "id": 35605, "pid": 435, "tid": 435, "ts": 1776649790921814, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921815, "dur": 0, + "args": { + "External id": 35606, + "cbid": 273, "correlation": 35606 + } + }, + { + "ph": "f", "id": 35606, "pid": 435, "tid": 435, "ts": 1776649790921815, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790921824, "dur": 15, + "args": { + "External id": 35607, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35607, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35607, "pid": 0, "tid": 7, "ts": 1776649790921824, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921816, "dur": 5, + "args": { + "External id": 35607, + "cbid": 211, "correlation": 35607 + } + }, + { + "ph": "s", "id": 35607, "pid": 435, "tid": 435, "ts": 1776649790921816, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790921859, "dur": 1, + "args": { + "External id": 4708, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35625, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35625, "pid": 0, "tid": 7, "ts": 1776649790921859, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921850, "dur": 6, + "args": { + "External id": 4708, + "cbid": 211, "correlation": 35625 + } + }, + { + "ph": "s", "id": 35625, "pid": 435, "tid": 435, "ts": 1776649790921850, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921863, "dur": 0, + "args": { + "External id": 35631, + "cbid": 200, "correlation": 35631 + } + }, + { + "ph": "f", "id": 35631, "pid": 435, "tid": 435, "ts": 1776649790921863, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921863, "dur": 0, + "args": { + "External id": 35632, + "cbid": 273, "correlation": 35632 + } + }, + { + "ph": "f", "id": 35632, "pid": 435, "tid": 435, "ts": 1776649790921863, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790921872, "dur": 15, + "args": { + "External id": 35633, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35633, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35633, "pid": 0, "tid": 7, "ts": 1776649790921872, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921865, "dur": 5, + "args": { + "External id": 35633, + "cbid": 211, "correlation": 35633 + } + }, + { + "ph": "s", "id": 35633, "pid": 435, "tid": 435, "ts": 1776649790921865, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790921908, "dur": 1, + "args": { + "External id": 4713, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35651, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35651, "pid": 0, "tid": 7, "ts": 1776649790921908, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921899, "dur": 6, + "args": { + "External id": 4713, + "cbid": 211, "correlation": 35651 + } + }, + { + "ph": "s", "id": 35651, "pid": 435, "tid": 435, "ts": 1776649790921899, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921912, "dur": 0, + "args": { + "External id": 35657, + "cbid": 200, "correlation": 35657 + } + }, + { + "ph": "f", "id": 35657, "pid": 435, "tid": 435, "ts": 1776649790921912, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921912, "dur": 0, + "args": { + "External id": 35658, + "cbid": 273, "correlation": 35658 + } + }, + { + "ph": "f", "id": 35658, "pid": 435, "tid": 435, "ts": 1776649790921912, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790921921, "dur": 16, + "args": { + "External id": 35659, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35659, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35659, "pid": 0, "tid": 7, "ts": 1776649790921921, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921913, "dur": 5, + "args": { + "External id": 35659, + "cbid": 211, "correlation": 35659 + } + }, + { + "ph": "s", "id": 35659, "pid": 435, "tid": 435, "ts": 1776649790921913, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790921957, "dur": 1, + "args": { + "External id": 4718, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35677, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35677, "pid": 0, "tid": 7, "ts": 1776649790921957, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921947, "dur": 7, + "args": { + "External id": 4718, + "cbid": 211, "correlation": 35677 + } + }, + { + "ph": "s", "id": 35677, "pid": 435, "tid": 435, "ts": 1776649790921947, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921960, "dur": 0, + "args": { + "External id": 35683, + "cbid": 200, "correlation": 35683 + } + }, + { + "ph": "f", "id": 35683, "pid": 435, "tid": 435, "ts": 1776649790921960, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790921961, "dur": 0, + "args": { + "External id": 35684, + "cbid": 273, "correlation": 35684 + } + }, + { + "ph": "f", "id": 35684, "pid": 435, "tid": 435, "ts": 1776649790921961, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790921970, "dur": 15, + "args": { + "External id": 35685, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35685, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35685, "pid": 0, "tid": 7, "ts": 1776649790921970, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921962, "dur": 5, + "args": { + "External id": 35685, + "cbid": 211, "correlation": 35685 + } + }, + { + "ph": "s", "id": 35685, "pid": 435, "tid": 435, "ts": 1776649790921962, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790922005, "dur": 1, + "args": { + "External id": 4723, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35703, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35703, "pid": 0, "tid": 7, "ts": 1776649790922005, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790921996, "dur": 6, + "args": { + "External id": 4723, + "cbid": 211, "correlation": 35703 + } + }, + { + "ph": "s", "id": 35703, "pid": 435, "tid": 435, "ts": 1776649790921996, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790922008, "dur": 0, + "args": { + "External id": 35709, + "cbid": 200, "correlation": 35709 + } + }, + { + "ph": "f", "id": 35709, "pid": 435, "tid": 435, "ts": 1776649790922008, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790922009, "dur": 0, + "args": { + "External id": 35710, + "cbid": 273, "correlation": 35710 + } + }, + { + "ph": "f", "id": 35710, "pid": 435, "tid": 435, "ts": 1776649790922009, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790922018, "dur": 16, + "args": { + "External id": 35711, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35711, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35711, "pid": 0, "tid": 7, "ts": 1776649790922018, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790922010, "dur": 5, + "args": { + "External id": 35711, + "cbid": 211, "correlation": 35711 + } + }, + { + "ph": "s", "id": 35711, "pid": 435, "tid": 435, "ts": 1776649790922010, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790922053, "dur": 1, + "args": { + "External id": 4728, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35729, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35729, "pid": 0, "tid": 7, "ts": 1776649790922053, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790922044, "dur": 6, + "args": { + "External id": 4728, + "cbid": 211, "correlation": 35729 + } + }, + { + "ph": "s", "id": 35729, "pid": 435, "tid": 435, "ts": 1776649790922044, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790922056, "dur": 0, + "args": { + "External id": 35735, + "cbid": 200, "correlation": 35735 + } + }, + { + "ph": "f", "id": 35735, "pid": 435, "tid": 435, "ts": 1776649790922056, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790922057, "dur": 0, + "args": { + "External id": 35736, + "cbid": 273, "correlation": 35736 + } + }, + { + "ph": "f", "id": 35736, "pid": 435, "tid": 435, "ts": 1776649790922057, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790922066, "dur": 16, + "args": { + "External id": 35737, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35737, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35737, "pid": 0, "tid": 7, "ts": 1776649790922066, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790922058, "dur": 4, + "args": { + "External id": 35737, + "cbid": 211, "correlation": 35737 + } + }, + { + "ph": "s", "id": 35737, "pid": 435, "tid": 435, "ts": 1776649790922058, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790922101, "dur": 1, + "args": { + "External id": 4733, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35755, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35755, "pid": 0, "tid": 7, "ts": 1776649790922101, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790922092, "dur": 7, + "args": { + "External id": 4733, + "cbid": 211, "correlation": 35755 + } + }, + { + "ph": "s", "id": 35755, "pid": 435, "tid": 435, "ts": 1776649790922092, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790922105, "dur": 0, + "args": { + "External id": 35761, + "cbid": 200, "correlation": 35761 + } + }, + { + "ph": "f", "id": 35761, "pid": 435, "tid": 435, "ts": 1776649790922105, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790922105, "dur": 0, + "args": { + "External id": 35762, + "cbid": 273, "correlation": 35762 + } + }, + { + "ph": "f", "id": 35762, "pid": 435, "tid": 435, "ts": 1776649790922105, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790922114, "dur": 15, + "args": { + "External id": 35763, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35763, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35763, "pid": 0, "tid": 7, "ts": 1776649790922114, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790922106, "dur": 5, + "args": { + "External id": 35763, + "cbid": 211, "correlation": 35763 + } + }, + { + "ph": "s", "id": 35763, "pid": 435, "tid": 435, "ts": 1776649790922106, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790922150, "dur": 1, + "args": { + "External id": 4738, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35781, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35781, "pid": 0, "tid": 7, "ts": 1776649790922150, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790922140, "dur": 6, + "args": { + "External id": 4738, + "cbid": 211, "correlation": 35781 + } + }, + { + "ph": "s", "id": 35781, "pid": 435, "tid": 435, "ts": 1776649790922140, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790922153, "dur": 0, + "args": { + "External id": 35787, + "cbid": 200, "correlation": 35787 + } + }, + { + "ph": "f", "id": 35787, "pid": 435, "tid": 435, "ts": 1776649790922153, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790922153, "dur": 0, + "args": { + "External id": 35788, + "cbid": 273, "correlation": 35788 + } + }, + { + "ph": "f", "id": 35788, "pid": 435, "tid": 435, "ts": 1776649790922153, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790922162, "dur": 15, + "args": { + "External id": 35789, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35789, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35789, "pid": 0, "tid": 7, "ts": 1776649790922162, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790922154, "dur": 4, + "args": { + "External id": 35789, + "cbid": 211, "correlation": 35789 + } + }, + { + "ph": "s", "id": 35789, "pid": 435, "tid": 435, "ts": 1776649790922154, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790922198, "dur": 1, + "args": { + "External id": 4743, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35807, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35807, "pid": 0, "tid": 7, "ts": 1776649790922198, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790922188, "dur": 6, + "args": { + "External id": 4743, + "cbid": 211, "correlation": 35807 + } + }, + { + "ph": "s", "id": 35807, "pid": 435, "tid": 435, "ts": 1776649790922188, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790922201, "dur": 0, + "args": { + "External id": 35813, + "cbid": 200, "correlation": 35813 + } + }, + { + "ph": "f", "id": 35813, "pid": 435, "tid": 435, "ts": 1776649790922201, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790922202, "dur": 0, + "args": { + "External id": 35814, + "cbid": 273, "correlation": 35814 + } + }, + { + "ph": "f", "id": 35814, "pid": 435, "tid": 435, "ts": 1776649790922202, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790922211, "dur": 16, + "args": { + "External id": 35815, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35815, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35815, "pid": 0, "tid": 7, "ts": 1776649790922211, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790922203, "dur": 5, + "args": { + "External id": 35815, + "cbid": 211, "correlation": 35815 + } + }, + { + "ph": "s", "id": 35815, "pid": 435, "tid": 435, "ts": 1776649790922203, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790922246, "dur": 1, + "args": { + "External id": 4748, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35833, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35833, "pid": 0, "tid": 7, "ts": 1776649790922246, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790922237, "dur": 6, + "args": { + "External id": 4748, + "cbid": 211, "correlation": 35833 + } + }, + { + "ph": "s", "id": 35833, "pid": 435, "tid": 435, "ts": 1776649790922237, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790922249, "dur": 0, + "args": { + "External id": 35839, + "cbid": 200, "correlation": 35839 + } + }, + { + "ph": "f", "id": 35839, "pid": 435, "tid": 435, "ts": 1776649790922249, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790922250, "dur": 0, + "args": { + "External id": 35840, + "cbid": 273, "correlation": 35840 + } + }, + { + "ph": "f", "id": 35840, "pid": 435, "tid": 435, "ts": 1776649790922250, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790922259, "dur": 16, + "args": { + "External id": 35841, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35841, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35841, "pid": 0, "tid": 7, "ts": 1776649790922259, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790922251, "dur": 4, + "args": { + "External id": 35841, + "cbid": 211, "correlation": 35841 + } + }, + { + "ph": "s", "id": 35841, "pid": 435, "tid": 435, "ts": 1776649790922251, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790922295, "dur": 1, + "args": { + "External id": 4753, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35859, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35859, "pid": 0, "tid": 7, "ts": 1776649790922295, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790922285, "dur": 6, + "args": { + "External id": 4753, + "cbid": 211, "correlation": 35859 + } + }, + { + "ph": "s", "id": 35859, "pid": 435, "tid": 435, "ts": 1776649790922285, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790922298, "dur": 0, + "args": { + "External id": 35865, + "cbid": 200, "correlation": 35865 + } + }, + { + "ph": "f", "id": 35865, "pid": 435, "tid": 435, "ts": 1776649790922298, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649790922298, "dur": 0, + "args": { + "External id": 35866, + "cbid": 273, "correlation": 35866 + } + }, + { + "ph": "f", "id": 35866, "pid": 435, "tid": 435, "ts": 1776649790922298, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649790922307, "dur": 15, + "args": { + "External id": 35867, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35867, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 35867, "pid": 0, "tid": 7, "ts": 1776649790922307, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790922299, "dur": 4, + "args": { + "External id": 35867, + "cbid": 211, "correlation": 35867 + } + }, + { + "ph": "s", "id": 35867, "pid": 435, "tid": 435, "ts": 1776649790922299, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790922329, "dur": 3, + "args": { + "External id": 35875, + "cbid": 135, "correlation": 35875 + } + }, + { + "ph": "f", "id": 35875, "pid": 435, "tid": 435, "ts": 1776649790922329, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790922344, "dur": 5, + "args": { + "External id": 35880, + "cbid": 165, "correlation": 35880 + } + }, + { + "ph": "s", "id": 35880, "pid": 435, "tid": 435, "ts": 1776649790922344, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649790922353, "dur": 4, + "args": { + "External id": 35883, + "cbid": 139, "correlation": 35883 + } + }, + { + "ph": "f", "id": 35883, "pid": 435, "tid": 435, "ts": 1776649790922353, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649790922481, "dur": 2, + "args": { + "External id": 4756, + "cbid": 317, "correlation": 35908 + } + }, + { + "ph": "f", "id": 35908, "pid": 435, "tid": 435, "ts": 1776649790922481, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649790922485, "dur": 0, + "args": { + "External id": 4756, + "cbid": 317, "correlation": 35912 + } + }, + { + "ph": "f", "id": 35912, "pid": 435, "tid": 435, "ts": 1776649790922485, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::distribution_elementwise_grid_stride_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::(anonymous namespace)::distribution_nullary_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2} const&, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(int, float)#1}>(int, at::PhiloxCudaState, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::(anonymous namespace)::distribution_nullary_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2} const&, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(int, float)#1})", "pid": 0, "tid": 7, + "ts": 1776649790922501, "dur": 4, + "args": { + "External id": 4756, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35914, + "registers per thread": 38, + "shared memory": 0, + "blocks per SM": 1.185185, + "warps per SM": 9.481482, + "grid": [128, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 15 + } + }, + { + "ph": "f", "id": 35914, "pid": 0, "tid": 7, "ts": 1776649790922501, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790922488, "dur": 12, + "args": { + "External id": 4756, + "cbid": 211, "correlation": 35914 + } + }, + { + "ph": "s", "id": 35914, "pid": 435, "tid": 435, "ts": 1776649790922488, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649790922514, "dur": 0, + "args": { + "External id": 35921, + "cbid": 317, "correlation": 35921 + } + }, + { + "ph": "f", "id": 35921, "pid": 435, "tid": 435, "ts": 1776649790922514, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649790922521, "dur": 0, + "args": { + "External id": 35925, + "cbid": 317, "correlation": 35925 + } + }, + { + "ph": "f", "id": 35925, "pid": 435, "tid": 435, "ts": 1776649790922521, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649790922543, "dur": 0, + "args": { + "External id": 4759, + "cbid": 317, "correlation": 35940 + } + }, + { + "ph": "f", "id": 35940, "pid": 435, "tid": 435, "ts": 1776649790922543, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649790922545, "dur": 0, + "args": { + "External id": 4759, + "cbid": 317, "correlation": 35944 + } + }, + { + "ph": "f", "id": 35944, "pid": 435, "tid": 435, "ts": 1776649790922545, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::distribution_elementwise_grid_stride_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::(anonymous namespace)::distribution_nullary_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2} const&, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(int, float)#1}>(int, at::PhiloxCudaState, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::(anonymous namespace)::distribution_nullary_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2} const&, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(int, float)#1})", "pid": 0, "tid": 7, + "ts": 1776649790922557, "dur": 85, + "args": { + "External id": 4759, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35946, + "registers per thread": 38, + "shared memory": 0, + "blocks per SM": 8.000000, + "warps per SM": 64.000000, + "grid": [864, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 75 + } + }, + { + "ph": "f", "id": 35946, "pid": 0, "tid": 7, "ts": 1776649790922557, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790922547, "dur": 7, + "args": { + "External id": 4759, + "cbid": 211, "correlation": 35946 + } + }, + { + "ph": "s", "id": 35946, "pid": 435, "tid": 435, "ts": 1776649790922547, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor >, at::detail::Array >(int, at::native::AUnaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790922643, "dur": 38, + "args": { + "External id": 4760, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35956, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 35956, "pid": 0, "tid": 7, "ts": 1776649790922643, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790922613, "dur": 10, + "args": { + "External id": 4760, + "cbid": 211, "correlation": 35956 + } + }, + { + "ph": "s", "id": 35956, "pid": 435, "tid": 435, "ts": 1776649790922613, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AbsFunctor, at::detail::Array >(int, at::native::AbsFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790922704, "dur": 41, + "args": { + "External id": 4765, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35976, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 35976, "pid": 0, "tid": 7, "ts": 1776649790922704, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790922693, "dur": 9, + "args": { + "External id": 4765, + "cbid": 211, "correlation": 35976 + } + }, + { + "ph": "s", "id": 35976, "pid": 435, "tid": 435, "ts": 1776649790922693, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<128, 4, at::native::ReduceOp >, unsigned int, c10::Half, 4> >(at::native::ReduceOp >, unsigned int, c10::Half, 4>)", "pid": 0, "tid": 7, + "ts": 1776649790922756, "dur": 46, + "args": { + "External id": 4767, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35989, + "registers per thread": 49, + "shared memory": 1040, + "blocks per SM": 9.481482, + "warps per SM": 37.925926, + "grid": [1024, 1, 1], + "block": [32, 4, 1], + "est. achieved occupancy %": 56 + } + }, + { + "ph": "f", "id": 35989, "pid": 0, "tid": 7, "ts": 1776649790922756, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790922744, "dur": 10, + "args": { + "External id": 4767, + "cbid": 211, "correlation": 35989 + } + }, + { + "ph": "s", "id": 35989, "pid": 435, "tid": 435, "ts": 1776649790922744, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BUnaryFunctor >, at::detail::Array >(int, at::native::BUnaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790922804, "dur": 2, + "args": { + "External id": 4769, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 35999, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 2.370370, + "warps per SM": 9.481482, + "grid": [256, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 15 + } + }, + { + "ph": "f", "id": 35999, "pid": 0, "tid": 7, "ts": 1776649790922804, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790922783, "dur": 8, + "args": { + "External id": 4769, + "cbid": 211, "correlation": 35999 + } + }, + { + "ph": "s", "id": 35999, "pid": 435, "tid": 435, "ts": 1776649790922783, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#8}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#8}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790922837, "dur": 3, + "args": { + "External id": 4770, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36009, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 2.370370, + "warps per SM": 9.481482, + "grid": [256, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 15 + } + }, + { + "ph": "f", "id": 36009, "pid": 0, "tid": 7, "ts": 1776649790922837, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790922826, "dur": 9, + "args": { + "External id": 4770, + "cbid": 211, "correlation": 36009 + } + }, + { + "ph": "s", "id": 36009, "pid": 435, "tid": 435, "ts": 1776649790922826, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649790922885, "dur": 99, + "args": { + "External id": 4775, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36019, + "registers per thread": 20, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 36019, "pid": 0, "tid": 7, "ts": 1776649790922885, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790922875, "dur": 8, + "args": { + "External id": 4775, + "cbid": 211, "correlation": 36019 + } + }, + { + "ph": "s", "id": 36019, "pid": 435, "tid": 435, "ts": 1776649790922875, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790922985, "dur": 39, + "args": { + "External id": 4776, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36029, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 36029, "pid": 0, "tid": 7, "ts": 1776649790922985, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790922904, "dur": 7, + "args": { + "External id": 4776, + "cbid": 211, "correlation": 36029 + } + }, + { + "ph": "s", "id": 36029, "pid": 435, "tid": 435, "ts": 1776649790922904, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#8}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#8}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790923026, "dur": 42, + "args": { + "External id": 4777, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36039, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 36039, "pid": 0, "tid": 7, "ts": 1776649790923026, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790922928, "dur": 6, + "args": { + "External id": 4777, + "cbid": 211, "correlation": 36039 + } + }, + { + "ph": "s", "id": 36039, "pid": 435, "tid": 435, "ts": 1776649790922928, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649790923070, "dur": 95, + "args": { + "External id": 4782, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36052, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 36052, "pid": 0, "tid": 7, "ts": 1776649790923070, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790922968, "dur": 9, + "args": { + "External id": 4782, + "cbid": 211, "correlation": 36052 + } + }, + { + "ph": "s", "id": 36052, "pid": 435, "tid": 435, "ts": 1776649790922968, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, at::detail::Array >(int, at::native::CUDAFunctorOnSelf_add, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790923167, "dur": 33, + "args": { + "External id": 4783, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36063, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 36063, "pid": 0, "tid": 7, "ts": 1776649790923167, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790923011, "dur": 9, + "args": { + "External id": 4783, + "cbid": 211, "correlation": 36063 + } + }, + { + "ph": "s", "id": 36063, "pid": 435, "tid": 435, "ts": 1776649790923011, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#1}::operator()() const::{lambda(unsigned char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649790923201, "dur": 57, + "args": { + "External id": 4787, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36076, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 36076, "pid": 0, "tid": 7, "ts": 1776649790923201, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790923042, "dur": 8, + "args": { + "External id": 4787, + "cbid": 211, "correlation": 36076 + } + }, + { + "ph": "s", "id": 36076, "pid": 435, "tid": 435, "ts": 1776649790923042, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast >(at::TensorIteratorBase&, at::native::BUnaryFunctor const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast >(at::TensorIteratorBase&, at::native::BUnaryFunctor const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649790923260, "dur": 36, + "args": { + "External id": 4798, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36096, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 36096, "pid": 0, "tid": 7, "ts": 1776649790923260, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790923141, "dur": 11, + "args": { + "External id": 4798, + "cbid": 211, "correlation": 36096 + } + }, + { + "ph": "s", "id": 36096, "pid": 435, "tid": 435, "ts": 1776649790923141, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649790923298, "dur": 39, + "args": { + "External id": 4801, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36106, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 36106, "pid": 0, "tid": 7, "ts": 1776649790923298, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790923181, "dur": 7, + "args": { + "External id": 4801, + "cbid": 211, "correlation": 36106 + } + }, + { + "ph": "s", "id": 36106, "pid": 435, "tid": 435, "ts": 1776649790923181, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor >, at::detail::Array >(int, at::native::AUnaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790923339, "dur": 18, + "args": { + "External id": 4805, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36116, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 36116, "pid": 0, "tid": 7, "ts": 1776649790923339, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790923235, "dur": 8, + "args": { + "External id": 4805, + "cbid": 211, "correlation": 36116 + } + }, + { + "ph": "s", "id": 36116, "pid": 435, "tid": 435, "ts": 1776649790923235, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649790923358, "dur": 30, + "args": { + "External id": 4809, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36129, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 36129, "pid": 0, "tid": 7, "ts": 1776649790923358, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790923268, "dur": 7, + "args": { + "External id": 4809, + "cbid": 211, "correlation": 36129 + } + }, + { + "ph": "s", "id": 36129, "pid": 435, "tid": 435, "ts": 1776649790923268, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BUnaryFunctor, at::detail::Array >(int, at::native::BUnaryFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790923390, "dur": 17, + "args": { + "External id": 4810, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36143, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 36143, "pid": 0, "tid": 7, "ts": 1776649790923390, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790923301, "dur": 8, + "args": { + "External id": 4810, + "cbid": 211, "correlation": 36143 + } + }, + { + "ph": "s", "id": 36143, "pid": 435, "tid": 435, "ts": 1776649790923301, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649790923409, "dur": 29, + "args": { + "External id": 4815, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36156, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 36156, "pid": 0, "tid": 7, "ts": 1776649790923409, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790923329, "dur": 6, + "args": { + "External id": 4815, + "cbid": 211, "correlation": 36156 + } + }, + { + "ph": "s", "id": 36156, "pid": 435, "tid": 435, "ts": 1776649790923329, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy_contig, unsigned int, 3, 128, 1>(at::native::(anonymous namespace)::OpaqueType<1u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata, unsigned int, 128, 1>, at::native::(anonymous namespace)::TensorSizeStride, int, unsigned int)", "pid": 0, "tid": 7, + "ts": 1776649790923440, "dur": 64, + "args": { + "External id": 4821, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36168, + "registers per thread": 21, + "shared memory": 0, + "blocks per SM": 4.000000, + "warps per SM": 64.000000, + "grid": [216, 2, 1], + "block": [512, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 36168, "pid": 0, "tid": 7, "ts": 1776649790923440, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790923392, "dur": 10, + "args": { + "External id": 4821, + "cbid": 211, "correlation": 36168 + } + }, + { + "ph": "s", "id": 36168, "pid": 435, "tid": 435, "ts": 1776649790923392, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649790923505, "dur": 72, + "args": { + "External id": 4832, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36181, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 36181, "pid": 0, "tid": 7, "ts": 1776649790923505, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790923469, "dur": 9, + "args": { + "External id": 4832, + "cbid": 211, "correlation": 36181 + } + }, + { + "ph": "s", "id": 36181, "pid": 435, "tid": 435, "ts": 1776649790923469, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, at::detail::Array >(int, at::native::CUDAFunctorOnSelf_add, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790923579, "dur": 33, + "args": { + "External id": 4835, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36192, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 36192, "pid": 0, "tid": 7, "ts": 1776649790923579, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790923511, "dur": 7, + "args": { + "External id": 4835, + "cbid": 211, "correlation": 36192 + } + }, + { + "ph": "s", "id": 36192, "pid": 435, "tid": 435, "ts": 1776649790923511, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649790923613, "dur": 91, + "args": { + "External id": 4839, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36205, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 36205, "pid": 0, "tid": 7, "ts": 1776649790923613, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790923541, "dur": 7, + "args": { + "External id": 4839, + "cbid": 211, "correlation": 36205 + } + }, + { + "ph": "s", "id": 36205, "pid": 435, "tid": 435, "ts": 1776649790923541, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor >, at::detail::Array >(int, at::native::BinaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790923706, "dur": 60, + "args": { + "External id": 4840, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36216, + "registers per thread": 22, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 36216, "pid": 0, "tid": 7, "ts": 1776649790923706, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790923568, "dur": 7, + "args": { + "External id": 4840, + "cbid": 211, "correlation": 36216 + } + }, + { + "ph": "s", "id": 36216, "pid": 435, "tid": 435, "ts": 1776649790923568, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790923758, "dur": 6, + "args": { + "External id": 4842, + "cbid": 210, "correlation": 36243 + } + }, + { + "ph": "f", "id": 36243, "pid": 435, "tid": 435, "ts": 1776649790923758, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790923775, "dur": 0, + "args": { + "External id": 4842, + "cbid": 210, "correlation": 36244 + } + }, + { + "ph": "f", "id": 36244, "pid": 435, "tid": 435, "ts": 1776649790923775, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790923794, "dur": 31, + "args": { + "External id": 4842, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36245, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36245, "pid": 0, "tid": 7, "ts": 1776649790923794, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790923781, "dur": 12, + "args": { + "External id": 4842, + "cbid": 211, "correlation": 36245 + } + }, + { + "ph": "s", "id": 36245, "pid": 435, "tid": 435, "ts": 1776649790923781, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790923824, "dur": 0, + "args": { + "External id": 4844, + "cbid": 210, "correlation": 36272 + } + }, + { + "ph": "f", "id": 36272, "pid": 435, "tid": 435, "ts": 1776649790923824, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790923834, "dur": 23, + "args": { + "External id": 4844, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36273, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36273, "pid": 0, "tid": 7, "ts": 1776649790923834, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790923825, "dur": 6, + "args": { + "External id": 4844, + "cbid": 211, "correlation": 36273 + } + }, + { + "ph": "s", "id": 36273, "pid": 435, "tid": 435, "ts": 1776649790923825, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790923856, "dur": 0, + "args": { + "External id": 4846, + "cbid": 210, "correlation": 36300 + } + }, + { + "ph": "f", "id": 36300, "pid": 435, "tid": 435, "ts": 1776649790923856, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790923866, "dur": 23, + "args": { + "External id": 4846, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36301, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36301, "pid": 0, "tid": 7, "ts": 1776649790923866, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790923857, "dur": 6, + "args": { + "External id": 4846, + "cbid": 211, "correlation": 36301 + } + }, + { + "ph": "s", "id": 36301, "pid": 435, "tid": 435, "ts": 1776649790923857, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790923887, "dur": 0, + "args": { + "External id": 4848, + "cbid": 210, "correlation": 36328 + } + }, + { + "ph": "f", "id": 36328, "pid": 435, "tid": 435, "ts": 1776649790923887, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790923897, "dur": 23, + "args": { + "External id": 4848, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36329, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36329, "pid": 0, "tid": 7, "ts": 1776649790923897, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790923888, "dur": 6, + "args": { + "External id": 4848, + "cbid": 211, "correlation": 36329 + } + }, + { + "ph": "s", "id": 36329, "pid": 435, "tid": 435, "ts": 1776649790923888, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790923918, "dur": 0, + "args": { + "External id": 4850, + "cbid": 210, "correlation": 36356 + } + }, + { + "ph": "f", "id": 36356, "pid": 435, "tid": 435, "ts": 1776649790923918, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790923928, "dur": 23, + "args": { + "External id": 4850, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36357, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36357, "pid": 0, "tid": 7, "ts": 1776649790923928, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790923919, "dur": 6, + "args": { + "External id": 4850, + "cbid": 211, "correlation": 36357 + } + }, + { + "ph": "s", "id": 36357, "pid": 435, "tid": 435, "ts": 1776649790923919, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790923948, "dur": 0, + "args": { + "External id": 4852, + "cbid": 210, "correlation": 36384 + } + }, + { + "ph": "f", "id": 36384, "pid": 435, "tid": 435, "ts": 1776649790923948, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790923958, "dur": 23, + "args": { + "External id": 4852, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36385, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36385, "pid": 0, "tid": 7, "ts": 1776649790923958, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790923949, "dur": 6, + "args": { + "External id": 4852, + "cbid": 211, "correlation": 36385 + } + }, + { + "ph": "s", "id": 36385, "pid": 435, "tid": 435, "ts": 1776649790923949, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790923979, "dur": 0, + "args": { + "External id": 4854, + "cbid": 210, "correlation": 36412 + } + }, + { + "ph": "f", "id": 36412, "pid": 435, "tid": 435, "ts": 1776649790923979, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790923989, "dur": 23, + "args": { + "External id": 4854, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36413, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36413, "pid": 0, "tid": 7, "ts": 1776649790923989, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790923980, "dur": 6, + "args": { + "External id": 4854, + "cbid": 211, "correlation": 36413 + } + }, + { + "ph": "s", "id": 36413, "pid": 435, "tid": 435, "ts": 1776649790923980, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924009, "dur": 0, + "args": { + "External id": 4856, + "cbid": 210, "correlation": 36440 + } + }, + { + "ph": "f", "id": 36440, "pid": 435, "tid": 435, "ts": 1776649790924009, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924019, "dur": 23, + "args": { + "External id": 4856, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36441, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36441, "pid": 0, "tid": 7, "ts": 1776649790924019, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924010, "dur": 6, + "args": { + "External id": 4856, + "cbid": 211, "correlation": 36441 + } + }, + { + "ph": "s", "id": 36441, "pid": 435, "tid": 435, "ts": 1776649790924010, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924039, "dur": 0, + "args": { + "External id": 4858, + "cbid": 210, "correlation": 36468 + } + }, + { + "ph": "f", "id": 36468, "pid": 435, "tid": 435, "ts": 1776649790924039, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924050, "dur": 23, + "args": { + "External id": 4858, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36469, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36469, "pid": 0, "tid": 7, "ts": 1776649790924050, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924041, "dur": 6, + "args": { + "External id": 4858, + "cbid": 211, "correlation": 36469 + } + }, + { + "ph": "s", "id": 36469, "pid": 435, "tid": 435, "ts": 1776649790924041, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924070, "dur": 0, + "args": { + "External id": 4860, + "cbid": 210, "correlation": 36496 + } + }, + { + "ph": "f", "id": 36496, "pid": 435, "tid": 435, "ts": 1776649790924070, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924080, "dur": 23, + "args": { + "External id": 4860, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36497, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36497, "pid": 0, "tid": 7, "ts": 1776649790924080, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924071, "dur": 6, + "args": { + "External id": 4860, + "cbid": 211, "correlation": 36497 + } + }, + { + "ph": "s", "id": 36497, "pid": 435, "tid": 435, "ts": 1776649790924071, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790924100, "dur": 8, + "args": { + "External id": 36503, + "cbid": 165, "correlation": 36503 + } + }, + { + "ph": "s", "id": 36503, "pid": 435, "tid": 435, "ts": 1776649790924100, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790924133, "dur": 4, + "args": { + "External id": 36512, + "cbid": 135, "correlation": 36512 + } + }, + { + "ph": "f", "id": 36512, "pid": 435, "tid": 435, "ts": 1776649790924133, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924166, "dur": 1, + "args": { + "External id": 4862, + "cbid": 210, "correlation": 36538 + } + }, + { + "ph": "f", "id": 36538, "pid": 435, "tid": 435, "ts": 1776649790924166, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924179, "dur": 23, + "args": { + "External id": 4862, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36539, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36539, "pid": 0, "tid": 7, "ts": 1776649790924179, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924168, "dur": 8, + "args": { + "External id": 4862, + "cbid": 211, "correlation": 36539 + } + }, + { + "ph": "s", "id": 36539, "pid": 435, "tid": 435, "ts": 1776649790924168, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924202, "dur": 0, + "args": { + "External id": 4864, + "cbid": 210, "correlation": 36566 + } + }, + { + "ph": "f", "id": 36566, "pid": 435, "tid": 435, "ts": 1776649790924202, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924213, "dur": 23, + "args": { + "External id": 4864, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36567, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36567, "pid": 0, "tid": 7, "ts": 1776649790924213, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924204, "dur": 6, + "args": { + "External id": 4864, + "cbid": 211, "correlation": 36567 + } + }, + { + "ph": "s", "id": 36567, "pid": 435, "tid": 435, "ts": 1776649790924204, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924255, "dur": 1, + "args": { + "External id": 4866, + "cbid": 210, "correlation": 36594 + } + }, + { + "ph": "f", "id": 36594, "pid": 435, "tid": 435, "ts": 1776649790924255, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924267, "dur": 23, + "args": { + "External id": 4866, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36595, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36595, "pid": 0, "tid": 7, "ts": 1776649790924267, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924257, "dur": 6, + "args": { + "External id": 4866, + "cbid": 211, "correlation": 36595 + } + }, + { + "ph": "s", "id": 36595, "pid": 435, "tid": 435, "ts": 1776649790924257, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924288, "dur": 0, + "args": { + "External id": 4868, + "cbid": 210, "correlation": 36622 + } + }, + { + "ph": "f", "id": 36622, "pid": 435, "tid": 435, "ts": 1776649790924288, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924298, "dur": 23, + "args": { + "External id": 4868, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36623, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36623, "pid": 0, "tid": 7, "ts": 1776649790924298, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924290, "dur": 6, + "args": { + "External id": 4868, + "cbid": 211, "correlation": 36623 + } + }, + { + "ph": "s", "id": 36623, "pid": 435, "tid": 435, "ts": 1776649790924290, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924319, "dur": 0, + "args": { + "External id": 4870, + "cbid": 210, "correlation": 36650 + } + }, + { + "ph": "f", "id": 36650, "pid": 435, "tid": 435, "ts": 1776649790924319, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924329, "dur": 23, + "args": { + "External id": 4870, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36651, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36651, "pid": 0, "tid": 7, "ts": 1776649790924329, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924320, "dur": 5, + "args": { + "External id": 4870, + "cbid": 211, "correlation": 36651 + } + }, + { + "ph": "s", "id": 36651, "pid": 435, "tid": 435, "ts": 1776649790924320, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924349, "dur": 0, + "args": { + "External id": 4872, + "cbid": 210, "correlation": 36678 + } + }, + { + "ph": "f", "id": 36678, "pid": 435, "tid": 435, "ts": 1776649790924349, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924359, "dur": 23, + "args": { + "External id": 4872, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36679, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36679, "pid": 0, "tid": 7, "ts": 1776649790924359, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924350, "dur": 6, + "args": { + "External id": 4872, + "cbid": 211, "correlation": 36679 + } + }, + { + "ph": "s", "id": 36679, "pid": 435, "tid": 435, "ts": 1776649790924350, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924379, "dur": 0, + "args": { + "External id": 4874, + "cbid": 210, "correlation": 36706 + } + }, + { + "ph": "f", "id": 36706, "pid": 435, "tid": 435, "ts": 1776649790924379, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924389, "dur": 23, + "args": { + "External id": 4874, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36707, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36707, "pid": 0, "tid": 7, "ts": 1776649790924389, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924380, "dur": 6, + "args": { + "External id": 4874, + "cbid": 211, "correlation": 36707 + } + }, + { + "ph": "s", "id": 36707, "pid": 435, "tid": 435, "ts": 1776649790924380, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924409, "dur": 0, + "args": { + "External id": 4876, + "cbid": 210, "correlation": 36734 + } + }, + { + "ph": "f", "id": 36734, "pid": 435, "tid": 435, "ts": 1776649790924409, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924419, "dur": 23, + "args": { + "External id": 4876, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36735, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36735, "pid": 0, "tid": 7, "ts": 1776649790924419, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924411, "dur": 6, + "args": { + "External id": 4876, + "cbid": 211, "correlation": 36735 + } + }, + { + "ph": "s", "id": 36735, "pid": 435, "tid": 435, "ts": 1776649790924411, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924440, "dur": 0, + "args": { + "External id": 4878, + "cbid": 210, "correlation": 36762 + } + }, + { + "ph": "f", "id": 36762, "pid": 435, "tid": 435, "ts": 1776649790924440, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924450, "dur": 23, + "args": { + "External id": 4878, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36763, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36763, "pid": 0, "tid": 7, "ts": 1776649790924450, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924441, "dur": 6, + "args": { + "External id": 4878, + "cbid": 211, "correlation": 36763 + } + }, + { + "ph": "s", "id": 36763, "pid": 435, "tid": 435, "ts": 1776649790924441, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924470, "dur": 0, + "args": { + "External id": 4880, + "cbid": 210, "correlation": 36790 + } + }, + { + "ph": "f", "id": 36790, "pid": 435, "tid": 435, "ts": 1776649790924470, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924482, "dur": 23, + "args": { + "External id": 4880, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36791, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36791, "pid": 0, "tid": 7, "ts": 1776649790924482, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924471, "dur": 6, + "args": { + "External id": 4880, + "cbid": 211, "correlation": 36791 + } + }, + { + "ph": "s", "id": 36791, "pid": 435, "tid": 435, "ts": 1776649790924471, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924501, "dur": 0, + "args": { + "External id": 4882, + "cbid": 210, "correlation": 36818 + } + }, + { + "ph": "f", "id": 36818, "pid": 435, "tid": 435, "ts": 1776649790924501, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924513, "dur": 23, + "args": { + "External id": 4882, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36819, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36819, "pid": 0, "tid": 7, "ts": 1776649790924513, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924502, "dur": 6, + "args": { + "External id": 4882, + "cbid": 211, "correlation": 36819 + } + }, + { + "ph": "s", "id": 36819, "pid": 435, "tid": 435, "ts": 1776649790924502, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924531, "dur": 0, + "args": { + "External id": 4884, + "cbid": 210, "correlation": 36846 + } + }, + { + "ph": "f", "id": 36846, "pid": 435, "tid": 435, "ts": 1776649790924531, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924544, "dur": 23, + "args": { + "External id": 4884, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36847, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36847, "pid": 0, "tid": 7, "ts": 1776649790924544, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924533, "dur": 6, + "args": { + "External id": 4884, + "cbid": 211, "correlation": 36847 + } + }, + { + "ph": "s", "id": 36847, "pid": 435, "tid": 435, "ts": 1776649790924533, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924562, "dur": 0, + "args": { + "External id": 4886, + "cbid": 210, "correlation": 36874 + } + }, + { + "ph": "f", "id": 36874, "pid": 435, "tid": 435, "ts": 1776649790924562, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924574, "dur": 23, + "args": { + "External id": 4886, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36875, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36875, "pid": 0, "tid": 7, "ts": 1776649790924574, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924563, "dur": 6, + "args": { + "External id": 4886, + "cbid": 211, "correlation": 36875 + } + }, + { + "ph": "s", "id": 36875, "pid": 435, "tid": 435, "ts": 1776649790924563, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924592, "dur": 0, + "args": { + "External id": 4888, + "cbid": 210, "correlation": 36902 + } + }, + { + "ph": "f", "id": 36902, "pid": 435, "tid": 435, "ts": 1776649790924592, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924604, "dur": 23, + "args": { + "External id": 4888, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36903, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36903, "pid": 0, "tid": 7, "ts": 1776649790924604, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924594, "dur": 6, + "args": { + "External id": 4888, + "cbid": 211, "correlation": 36903 + } + }, + { + "ph": "s", "id": 36903, "pid": 435, "tid": 435, "ts": 1776649790924594, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924623, "dur": 0, + "args": { + "External id": 4890, + "cbid": 210, "correlation": 36930 + } + }, + { + "ph": "f", "id": 36930, "pid": 435, "tid": 435, "ts": 1776649790924623, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924634, "dur": 23, + "args": { + "External id": 4890, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36931, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36931, "pid": 0, "tid": 7, "ts": 1776649790924634, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924624, "dur": 6, + "args": { + "External id": 4890, + "cbid": 211, "correlation": 36931 + } + }, + { + "ph": "s", "id": 36931, "pid": 435, "tid": 435, "ts": 1776649790924624, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924653, "dur": 0, + "args": { + "External id": 4892, + "cbid": 210, "correlation": 36958 + } + }, + { + "ph": "f", "id": 36958, "pid": 435, "tid": 435, "ts": 1776649790924653, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924664, "dur": 23, + "args": { + "External id": 4892, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36959, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36959, "pid": 0, "tid": 7, "ts": 1776649790924664, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924654, "dur": 6, + "args": { + "External id": 4892, + "cbid": 211, "correlation": 36959 + } + }, + { + "ph": "s", "id": 36959, "pid": 435, "tid": 435, "ts": 1776649790924654, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924683, "dur": 0, + "args": { + "External id": 4894, + "cbid": 210, "correlation": 36986 + } + }, + { + "ph": "f", "id": 36986, "pid": 435, "tid": 435, "ts": 1776649790924683, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924695, "dur": 23, + "args": { + "External id": 4894, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 36987, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 36987, "pid": 0, "tid": 7, "ts": 1776649790924695, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924684, "dur": 6, + "args": { + "External id": 4894, + "cbid": 211, "correlation": 36987 + } + }, + { + "ph": "s", "id": 36987, "pid": 435, "tid": 435, "ts": 1776649790924684, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924713, "dur": 0, + "args": { + "External id": 4896, + "cbid": 210, "correlation": 37014 + } + }, + { + "ph": "f", "id": 37014, "pid": 435, "tid": 435, "ts": 1776649790924713, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924724, "dur": 23, + "args": { + "External id": 4896, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37015, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37015, "pid": 0, "tid": 7, "ts": 1776649790924724, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924714, "dur": 5, + "args": { + "External id": 4896, + "cbid": 211, "correlation": 37015 + } + }, + { + "ph": "s", "id": 37015, "pid": 435, "tid": 435, "ts": 1776649790924714, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924743, "dur": 0, + "args": { + "External id": 4898, + "cbid": 210, "correlation": 37042 + } + }, + { + "ph": "f", "id": 37042, "pid": 435, "tid": 435, "ts": 1776649790924743, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924755, "dur": 23, + "args": { + "External id": 4898, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37043, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37043, "pid": 0, "tid": 7, "ts": 1776649790924755, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924744, "dur": 6, + "args": { + "External id": 4898, + "cbid": 211, "correlation": 37043 + } + }, + { + "ph": "s", "id": 37043, "pid": 435, "tid": 435, "ts": 1776649790924744, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924773, "dur": 0, + "args": { + "External id": 4900, + "cbid": 210, "correlation": 37070 + } + }, + { + "ph": "f", "id": 37070, "pid": 435, "tid": 435, "ts": 1776649790924773, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924785, "dur": 23, + "args": { + "External id": 4900, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37071, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37071, "pid": 0, "tid": 7, "ts": 1776649790924785, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924774, "dur": 5, + "args": { + "External id": 4900, + "cbid": 211, "correlation": 37071 + } + }, + { + "ph": "s", "id": 37071, "pid": 435, "tid": 435, "ts": 1776649790924774, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924804, "dur": 0, + "args": { + "External id": 4902, + "cbid": 210, "correlation": 37098 + } + }, + { + "ph": "f", "id": 37098, "pid": 435, "tid": 435, "ts": 1776649790924804, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924816, "dur": 23, + "args": { + "External id": 4902, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37099, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37099, "pid": 0, "tid": 7, "ts": 1776649790924816, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924805, "dur": 6, + "args": { + "External id": 4902, + "cbid": 211, "correlation": 37099 + } + }, + { + "ph": "s", "id": 37099, "pid": 435, "tid": 435, "ts": 1776649790924805, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924834, "dur": 0, + "args": { + "External id": 4904, + "cbid": 210, "correlation": 37126 + } + }, + { + "ph": "f", "id": 37126, "pid": 435, "tid": 435, "ts": 1776649790924834, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924845, "dur": 23, + "args": { + "External id": 4904, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37127, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37127, "pid": 0, "tid": 7, "ts": 1776649790924845, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924835, "dur": 5, + "args": { + "External id": 4904, + "cbid": 211, "correlation": 37127 + } + }, + { + "ph": "s", "id": 37127, "pid": 435, "tid": 435, "ts": 1776649790924835, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924863, "dur": 0, + "args": { + "External id": 4906, + "cbid": 210, "correlation": 37154 + } + }, + { + "ph": "f", "id": 37154, "pid": 435, "tid": 435, "ts": 1776649790924863, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924875, "dur": 23, + "args": { + "External id": 4906, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37155, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37155, "pid": 0, "tid": 7, "ts": 1776649790924875, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924865, "dur": 6, + "args": { + "External id": 4906, + "cbid": 211, "correlation": 37155 + } + }, + { + "ph": "s", "id": 37155, "pid": 435, "tid": 435, "ts": 1776649790924865, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924893, "dur": 0, + "args": { + "External id": 4908, + "cbid": 210, "correlation": 37182 + } + }, + { + "ph": "f", "id": 37182, "pid": 435, "tid": 435, "ts": 1776649790924893, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924904, "dur": 23, + "args": { + "External id": 4908, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37183, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37183, "pid": 0, "tid": 7, "ts": 1776649790924904, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924894, "dur": 5, + "args": { + "External id": 4908, + "cbid": 211, "correlation": 37183 + } + }, + { + "ph": "s", "id": 37183, "pid": 435, "tid": 435, "ts": 1776649790924894, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924923, "dur": 0, + "args": { + "External id": 4910, + "cbid": 210, "correlation": 37210 + } + }, + { + "ph": "f", "id": 37210, "pid": 435, "tid": 435, "ts": 1776649790924923, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924934, "dur": 23, + "args": { + "External id": 4910, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37211, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37211, "pid": 0, "tid": 7, "ts": 1776649790924934, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924924, "dur": 5, + "args": { + "External id": 4910, + "cbid": 211, "correlation": 37211 + } + }, + { + "ph": "s", "id": 37211, "pid": 435, "tid": 435, "ts": 1776649790924924, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924952, "dur": 0, + "args": { + "External id": 4912, + "cbid": 210, "correlation": 37238 + } + }, + { + "ph": "f", "id": 37238, "pid": 435, "tid": 435, "ts": 1776649790924952, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924964, "dur": 23, + "args": { + "External id": 4912, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37239, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37239, "pid": 0, "tid": 7, "ts": 1776649790924964, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924954, "dur": 5, + "args": { + "External id": 4912, + "cbid": 211, "correlation": 37239 + } + }, + { + "ph": "s", "id": 37239, "pid": 435, "tid": 435, "ts": 1776649790924954, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790924982, "dur": 0, + "args": { + "External id": 4914, + "cbid": 210, "correlation": 37266 + } + }, + { + "ph": "f", "id": 37266, "pid": 435, "tid": 435, "ts": 1776649790924982, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790924994, "dur": 23, + "args": { + "External id": 4914, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37267, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37267, "pid": 0, "tid": 7, "ts": 1776649790924994, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790924983, "dur": 6, + "args": { + "External id": 4914, + "cbid": 211, "correlation": 37267 + } + }, + { + "ph": "s", "id": 37267, "pid": 435, "tid": 435, "ts": 1776649790924983, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925012, "dur": 0, + "args": { + "External id": 4916, + "cbid": 210, "correlation": 37294 + } + }, + { + "ph": "f", "id": 37294, "pid": 435, "tid": 435, "ts": 1776649790925012, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925023, "dur": 23, + "args": { + "External id": 4916, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37295, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37295, "pid": 0, "tid": 7, "ts": 1776649790925023, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925013, "dur": 5, + "args": { + "External id": 4916, + "cbid": 211, "correlation": 37295 + } + }, + { + "ph": "s", "id": 37295, "pid": 435, "tid": 435, "ts": 1776649790925013, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925042, "dur": 0, + "args": { + "External id": 4918, + "cbid": 210, "correlation": 37322 + } + }, + { + "ph": "f", "id": 37322, "pid": 435, "tid": 435, "ts": 1776649790925042, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925053, "dur": 23, + "args": { + "External id": 4918, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37323, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37323, "pid": 0, "tid": 7, "ts": 1776649790925053, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925043, "dur": 6, + "args": { + "External id": 4918, + "cbid": 211, "correlation": 37323 + } + }, + { + "ph": "s", "id": 37323, "pid": 435, "tid": 435, "ts": 1776649790925043, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925072, "dur": 0, + "args": { + "External id": 4920, + "cbid": 210, "correlation": 37350 + } + }, + { + "ph": "f", "id": 37350, "pid": 435, "tid": 435, "ts": 1776649790925072, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925083, "dur": 23, + "args": { + "External id": 4920, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37351, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37351, "pid": 0, "tid": 7, "ts": 1776649790925083, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925073, "dur": 5, + "args": { + "External id": 4920, + "cbid": 211, "correlation": 37351 + } + }, + { + "ph": "s", "id": 37351, "pid": 435, "tid": 435, "ts": 1776649790925073, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925102, "dur": 0, + "args": { + "External id": 4922, + "cbid": 210, "correlation": 37378 + } + }, + { + "ph": "f", "id": 37378, "pid": 435, "tid": 435, "ts": 1776649790925102, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925113, "dur": 23, + "args": { + "External id": 4922, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37379, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37379, "pid": 0, "tid": 7, "ts": 1776649790925113, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925103, "dur": 5, + "args": { + "External id": 4922, + "cbid": 211, "correlation": 37379 + } + }, + { + "ph": "s", "id": 37379, "pid": 435, "tid": 435, "ts": 1776649790925103, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925132, "dur": 0, + "args": { + "External id": 4924, + "cbid": 210, "correlation": 37406 + } + }, + { + "ph": "f", "id": 37406, "pid": 435, "tid": 435, "ts": 1776649790925132, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925143, "dur": 23, + "args": { + "External id": 4924, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37407, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37407, "pid": 0, "tid": 7, "ts": 1776649790925143, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925133, "dur": 5, + "args": { + "External id": 4924, + "cbid": 211, "correlation": 37407 + } + }, + { + "ph": "s", "id": 37407, "pid": 435, "tid": 435, "ts": 1776649790925133, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925161, "dur": 0, + "args": { + "External id": 4926, + "cbid": 210, "correlation": 37434 + } + }, + { + "ph": "f", "id": 37434, "pid": 435, "tid": 435, "ts": 1776649790925161, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925173, "dur": 23, + "args": { + "External id": 4926, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37435, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37435, "pid": 0, "tid": 7, "ts": 1776649790925173, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925162, "dur": 5, + "args": { + "External id": 4926, + "cbid": 211, "correlation": 37435 + } + }, + { + "ph": "s", "id": 37435, "pid": 435, "tid": 435, "ts": 1776649790925162, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925191, "dur": 0, + "args": { + "External id": 4928, + "cbid": 210, "correlation": 37462 + } + }, + { + "ph": "f", "id": 37462, "pid": 435, "tid": 435, "ts": 1776649790925191, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925202, "dur": 23, + "args": { + "External id": 4928, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37463, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37463, "pid": 0, "tid": 7, "ts": 1776649790925202, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925192, "dur": 5, + "args": { + "External id": 4928, + "cbid": 211, "correlation": 37463 + } + }, + { + "ph": "s", "id": 37463, "pid": 435, "tid": 435, "ts": 1776649790925192, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925221, "dur": 0, + "args": { + "External id": 4930, + "cbid": 210, "correlation": 37490 + } + }, + { + "ph": "f", "id": 37490, "pid": 435, "tid": 435, "ts": 1776649790925221, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925232, "dur": 23, + "args": { + "External id": 4930, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37491, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37491, "pid": 0, "tid": 7, "ts": 1776649790925232, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925222, "dur": 6, + "args": { + "External id": 4930, + "cbid": 211, "correlation": 37491 + } + }, + { + "ph": "s", "id": 37491, "pid": 435, "tid": 435, "ts": 1776649790925222, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925250, "dur": 0, + "args": { + "External id": 4932, + "cbid": 210, "correlation": 37518 + } + }, + { + "ph": "f", "id": 37518, "pid": 435, "tid": 435, "ts": 1776649790925250, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925261, "dur": 23, + "args": { + "External id": 4932, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37519, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37519, "pid": 0, "tid": 7, "ts": 1776649790925261, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925251, "dur": 5, + "args": { + "External id": 4932, + "cbid": 211, "correlation": 37519 + } + }, + { + "ph": "s", "id": 37519, "pid": 435, "tid": 435, "ts": 1776649790925251, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925280, "dur": 0, + "args": { + "External id": 4934, + "cbid": 210, "correlation": 37546 + } + }, + { + "ph": "f", "id": 37546, "pid": 435, "tid": 435, "ts": 1776649790925280, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925291, "dur": 23, + "args": { + "External id": 4934, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37547, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37547, "pid": 0, "tid": 7, "ts": 1776649790925291, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925281, "dur": 6, + "args": { + "External id": 4934, + "cbid": 211, "correlation": 37547 + } + }, + { + "ph": "s", "id": 37547, "pid": 435, "tid": 435, "ts": 1776649790925281, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925309, "dur": 0, + "args": { + "External id": 4936, + "cbid": 210, "correlation": 37574 + } + }, + { + "ph": "f", "id": 37574, "pid": 435, "tid": 435, "ts": 1776649790925309, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925321, "dur": 23, + "args": { + "External id": 4936, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37575, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37575, "pid": 0, "tid": 7, "ts": 1776649790925321, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925311, "dur": 5, + "args": { + "External id": 4936, + "cbid": 211, "correlation": 37575 + } + }, + { + "ph": "s", "id": 37575, "pid": 435, "tid": 435, "ts": 1776649790925311, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925339, "dur": 0, + "args": { + "External id": 4938, + "cbid": 210, "correlation": 37602 + } + }, + { + "ph": "f", "id": 37602, "pid": 435, "tid": 435, "ts": 1776649790925339, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925351, "dur": 23, + "args": { + "External id": 4938, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37603, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37603, "pid": 0, "tid": 7, "ts": 1776649790925351, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925340, "dur": 6, + "args": { + "External id": 4938, + "cbid": 211, "correlation": 37603 + } + }, + { + "ph": "s", "id": 37603, "pid": 435, "tid": 435, "ts": 1776649790925340, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925369, "dur": 0, + "args": { + "External id": 4940, + "cbid": 210, "correlation": 37630 + } + }, + { + "ph": "f", "id": 37630, "pid": 435, "tid": 435, "ts": 1776649790925369, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925380, "dur": 23, + "args": { + "External id": 4940, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37631, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37631, "pid": 0, "tid": 7, "ts": 1776649790925380, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925370, "dur": 5, + "args": { + "External id": 4940, + "cbid": 211, "correlation": 37631 + } + }, + { + "ph": "s", "id": 37631, "pid": 435, "tid": 435, "ts": 1776649790925370, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925399, "dur": 0, + "args": { + "External id": 4942, + "cbid": 210, "correlation": 37658 + } + }, + { + "ph": "f", "id": 37658, "pid": 435, "tid": 435, "ts": 1776649790925399, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925410, "dur": 23, + "args": { + "External id": 4942, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37659, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37659, "pid": 0, "tid": 7, "ts": 1776649790925410, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925400, "dur": 6, + "args": { + "External id": 4942, + "cbid": 211, "correlation": 37659 + } + }, + { + "ph": "s", "id": 37659, "pid": 435, "tid": 435, "ts": 1776649790925400, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925429, "dur": 0, + "args": { + "External id": 4944, + "cbid": 210, "correlation": 37686 + } + }, + { + "ph": "f", "id": 37686, "pid": 435, "tid": 435, "ts": 1776649790925429, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925440, "dur": 23, + "args": { + "External id": 4944, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37687, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37687, "pid": 0, "tid": 7, "ts": 1776649790925440, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925430, "dur": 5, + "args": { + "External id": 4944, + "cbid": 211, "correlation": 37687 + } + }, + { + "ph": "s", "id": 37687, "pid": 435, "tid": 435, "ts": 1776649790925430, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925458, "dur": 0, + "args": { + "External id": 4946, + "cbid": 210, "correlation": 37714 + } + }, + { + "ph": "f", "id": 37714, "pid": 435, "tid": 435, "ts": 1776649790925458, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925470, "dur": 23, + "args": { + "External id": 4946, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37715, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37715, "pid": 0, "tid": 7, "ts": 1776649790925470, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925460, "dur": 5, + "args": { + "External id": 4946, + "cbid": 211, "correlation": 37715 + } + }, + { + "ph": "s", "id": 37715, "pid": 435, "tid": 435, "ts": 1776649790925460, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925488, "dur": 0, + "args": { + "External id": 4948, + "cbid": 210, "correlation": 37742 + } + }, + { + "ph": "f", "id": 37742, "pid": 435, "tid": 435, "ts": 1776649790925488, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925499, "dur": 23, + "args": { + "External id": 4948, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37743, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37743, "pid": 0, "tid": 7, "ts": 1776649790925499, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925489, "dur": 5, + "args": { + "External id": 4948, + "cbid": 211, "correlation": 37743 + } + }, + { + "ph": "s", "id": 37743, "pid": 435, "tid": 435, "ts": 1776649790925489, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925518, "dur": 0, + "args": { + "External id": 4950, + "cbid": 210, "correlation": 37770 + } + }, + { + "ph": "f", "id": 37770, "pid": 435, "tid": 435, "ts": 1776649790925518, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925529, "dur": 22, + "args": { + "External id": 4950, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37771, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37771, "pid": 0, "tid": 7, "ts": 1776649790925529, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925519, "dur": 5, + "args": { + "External id": 4950, + "cbid": 211, "correlation": 37771 + } + }, + { + "ph": "s", "id": 37771, "pid": 435, "tid": 435, "ts": 1776649790925519, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925547, "dur": 0, + "args": { + "External id": 4952, + "cbid": 210, "correlation": 37798 + } + }, + { + "ph": "f", "id": 37798, "pid": 435, "tid": 435, "ts": 1776649790925547, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925559, "dur": 23, + "args": { + "External id": 4952, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37799, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37799, "pid": 0, "tid": 7, "ts": 1776649790925559, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925549, "dur": 5, + "args": { + "External id": 4952, + "cbid": 211, "correlation": 37799 + } + }, + { + "ph": "s", "id": 37799, "pid": 435, "tid": 435, "ts": 1776649790925549, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925577, "dur": 0, + "args": { + "External id": 4954, + "cbid": 210, "correlation": 37826 + } + }, + { + "ph": "f", "id": 37826, "pid": 435, "tid": 435, "ts": 1776649790925577, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925588, "dur": 23, + "args": { + "External id": 4954, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37827, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37827, "pid": 0, "tid": 7, "ts": 1776649790925588, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925578, "dur": 5, + "args": { + "External id": 4954, + "cbid": 211, "correlation": 37827 + } + }, + { + "ph": "s", "id": 37827, "pid": 435, "tid": 435, "ts": 1776649790925578, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925607, "dur": 0, + "args": { + "External id": 4956, + "cbid": 210, "correlation": 37854 + } + }, + { + "ph": "f", "id": 37854, "pid": 435, "tid": 435, "ts": 1776649790925607, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925618, "dur": 23, + "args": { + "External id": 4956, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37855, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37855, "pid": 0, "tid": 7, "ts": 1776649790925618, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925608, "dur": 5, + "args": { + "External id": 4956, + "cbid": 211, "correlation": 37855 + } + }, + { + "ph": "s", "id": 37855, "pid": 435, "tid": 435, "ts": 1776649790925608, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925636, "dur": 0, + "args": { + "External id": 4958, + "cbid": 210, "correlation": 37882 + } + }, + { + "ph": "f", "id": 37882, "pid": 435, "tid": 435, "ts": 1776649790925636, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925648, "dur": 23, + "args": { + "External id": 4958, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37883, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37883, "pid": 0, "tid": 7, "ts": 1776649790925648, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925638, "dur": 5, + "args": { + "External id": 4958, + "cbid": 211, "correlation": 37883 + } + }, + { + "ph": "s", "id": 37883, "pid": 435, "tid": 435, "ts": 1776649790925638, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925666, "dur": 0, + "args": { + "External id": 4960, + "cbid": 210, "correlation": 37910 + } + }, + { + "ph": "f", "id": 37910, "pid": 435, "tid": 435, "ts": 1776649790925666, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925677, "dur": 23, + "args": { + "External id": 4960, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37911, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37911, "pid": 0, "tid": 7, "ts": 1776649790925677, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925667, "dur": 5, + "args": { + "External id": 4960, + "cbid": 211, "correlation": 37911 + } + }, + { + "ph": "s", "id": 37911, "pid": 435, "tid": 435, "ts": 1776649790925667, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790925702, "dur": 3, + "args": { + "External id": 37922, + "cbid": 135, "correlation": 37922 + } + }, + { + "ph": "f", "id": 37922, "pid": 435, "tid": 435, "ts": 1776649790925702, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790925717, "dur": 5, + "args": { + "External id": 37927, + "cbid": 165, "correlation": 37927 + } + }, + { + "ph": "s", "id": 37927, "pid": 435, "tid": 435, "ts": 1776649790925717, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649790925726, "dur": 5, + "args": { + "External id": 37930, + "cbid": 139, "correlation": 37930 + } + }, + { + "ph": "f", "id": 37930, "pid": 435, "tid": 435, "ts": 1776649790925726, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925764, "dur": 1, + "args": { + "External id": 4962, + "cbid": 210, "correlation": 37964 + } + }, + { + "ph": "f", "id": 37964, "pid": 435, "tid": 435, "ts": 1776649790925764, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925776, "dur": 23, + "args": { + "External id": 4962, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37965, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37965, "pid": 0, "tid": 7, "ts": 1776649790925776, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925766, "dur": 8, + "args": { + "External id": 4962, + "cbid": 211, "correlation": 37965 + } + }, + { + "ph": "s", "id": 37965, "pid": 435, "tid": 435, "ts": 1776649790925766, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925800, "dur": 0, + "args": { + "External id": 4964, + "cbid": 210, "correlation": 37992 + } + }, + { + "ph": "f", "id": 37992, "pid": 435, "tid": 435, "ts": 1776649790925800, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925810, "dur": 23, + "args": { + "External id": 4964, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 37993, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 37993, "pid": 0, "tid": 7, "ts": 1776649790925810, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925801, "dur": 6, + "args": { + "External id": 4964, + "cbid": 211, "correlation": 37993 + } + }, + { + "ph": "s", "id": 37993, "pid": 435, "tid": 435, "ts": 1776649790925801, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925831, "dur": 0, + "args": { + "External id": 4966, + "cbid": 210, "correlation": 38020 + } + }, + { + "ph": "f", "id": 38020, "pid": 435, "tid": 435, "ts": 1776649790925831, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925841, "dur": 23, + "args": { + "External id": 4966, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38021, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38021, "pid": 0, "tid": 7, "ts": 1776649790925841, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925832, "dur": 6, + "args": { + "External id": 4966, + "cbid": 211, "correlation": 38021 + } + }, + { + "ph": "s", "id": 38021, "pid": 435, "tid": 435, "ts": 1776649790925832, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925861, "dur": 0, + "args": { + "External id": 4968, + "cbid": 210, "correlation": 38048 + } + }, + { + "ph": "f", "id": 38048, "pid": 435, "tid": 435, "ts": 1776649790925861, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925871, "dur": 23, + "args": { + "External id": 4968, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38049, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38049, "pid": 0, "tid": 7, "ts": 1776649790925871, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925862, "dur": 6, + "args": { + "External id": 4968, + "cbid": 211, "correlation": 38049 + } + }, + { + "ph": "s", "id": 38049, "pid": 435, "tid": 435, "ts": 1776649790925862, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925891, "dur": 0, + "args": { + "External id": 4970, + "cbid": 210, "correlation": 38076 + } + }, + { + "ph": "f", "id": 38076, "pid": 435, "tid": 435, "ts": 1776649790925891, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925901, "dur": 23, + "args": { + "External id": 4970, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38077, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38077, "pid": 0, "tid": 7, "ts": 1776649790925901, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925893, "dur": 6, + "args": { + "External id": 4970, + "cbid": 211, "correlation": 38077 + } + }, + { + "ph": "s", "id": 38077, "pid": 435, "tid": 435, "ts": 1776649790925893, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925922, "dur": 0, + "args": { + "External id": 4972, + "cbid": 210, "correlation": 38104 + } + }, + { + "ph": "f", "id": 38104, "pid": 435, "tid": 435, "ts": 1776649790925922, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925932, "dur": 23, + "args": { + "External id": 4972, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38105, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38105, "pid": 0, "tid": 7, "ts": 1776649790925932, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925923, "dur": 6, + "args": { + "External id": 4972, + "cbid": 211, "correlation": 38105 + } + }, + { + "ph": "s", "id": 38105, "pid": 435, "tid": 435, "ts": 1776649790925923, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925953, "dur": 0, + "args": { + "External id": 4974, + "cbid": 210, "correlation": 38132 + } + }, + { + "ph": "f", "id": 38132, "pid": 435, "tid": 435, "ts": 1776649790925953, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925963, "dur": 23, + "args": { + "External id": 4974, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38133, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38133, "pid": 0, "tid": 7, "ts": 1776649790925963, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925954, "dur": 5, + "args": { + "External id": 4974, + "cbid": 211, "correlation": 38133 + } + }, + { + "ph": "s", "id": 38133, "pid": 435, "tid": 435, "ts": 1776649790925954, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790925983, "dur": 0, + "args": { + "External id": 4976, + "cbid": 210, "correlation": 38160 + } + }, + { + "ph": "f", "id": 38160, "pid": 435, "tid": 435, "ts": 1776649790925983, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790925993, "dur": 23, + "args": { + "External id": 4976, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38161, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38161, "pid": 0, "tid": 7, "ts": 1776649790925993, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790925984, "dur": 6, + "args": { + "External id": 4976, + "cbid": 211, "correlation": 38161 + } + }, + { + "ph": "s", "id": 38161, "pid": 435, "tid": 435, "ts": 1776649790925984, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926013, "dur": 0, + "args": { + "External id": 4978, + "cbid": 210, "correlation": 38188 + } + }, + { + "ph": "f", "id": 38188, "pid": 435, "tid": 435, "ts": 1776649790926013, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926023, "dur": 23, + "args": { + "External id": 4978, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38189, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38189, "pid": 0, "tid": 7, "ts": 1776649790926023, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926014, "dur": 6, + "args": { + "External id": 4978, + "cbid": 211, "correlation": 38189 + } + }, + { + "ph": "s", "id": 38189, "pid": 435, "tid": 435, "ts": 1776649790926014, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926044, "dur": 0, + "args": { + "External id": 4980, + "cbid": 210, "correlation": 38216 + } + }, + { + "ph": "f", "id": 38216, "pid": 435, "tid": 435, "ts": 1776649790926044, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926054, "dur": 23, + "args": { + "External id": 4980, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38217, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38217, "pid": 0, "tid": 7, "ts": 1776649790926054, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926045, "dur": 6, + "args": { + "External id": 4980, + "cbid": 211, "correlation": 38217 + } + }, + { + "ph": "s", "id": 38217, "pid": 435, "tid": 435, "ts": 1776649790926045, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790926070, "dur": 10, + "args": { + "External id": 38223, + "cbid": 165, "correlation": 38223 + } + }, + { + "ph": "s", "id": 38223, "pid": 435, "tid": 435, "ts": 1776649790926070, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790926102, "dur": 3, + "args": { + "External id": 38232, + "cbid": 135, "correlation": 38232 + } + }, + { + "ph": "f", "id": 38232, "pid": 435, "tid": 435, "ts": 1776649790926102, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926131, "dur": 1, + "args": { + "External id": 4982, + "cbid": 210, "correlation": 38258 + } + }, + { + "ph": "f", "id": 38258, "pid": 435, "tid": 435, "ts": 1776649790926131, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926144, "dur": 23, + "args": { + "External id": 4982, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38259, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38259, "pid": 0, "tid": 7, "ts": 1776649790926144, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926133, "dur": 7, + "args": { + "External id": 4982, + "cbid": 211, "correlation": 38259 + } + }, + { + "ph": "s", "id": 38259, "pid": 435, "tid": 435, "ts": 1776649790926133, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926166, "dur": 0, + "args": { + "External id": 4984, + "cbid": 210, "correlation": 38286 + } + }, + { + "ph": "f", "id": 38286, "pid": 435, "tid": 435, "ts": 1776649790926166, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926177, "dur": 23, + "args": { + "External id": 4984, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38287, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38287, "pid": 0, "tid": 7, "ts": 1776649790926177, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926168, "dur": 6, + "args": { + "External id": 4984, + "cbid": 211, "correlation": 38287 + } + }, + { + "ph": "s", "id": 38287, "pid": 435, "tid": 435, "ts": 1776649790926168, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926197, "dur": 0, + "args": { + "External id": 4986, + "cbid": 210, "correlation": 38314 + } + }, + { + "ph": "f", "id": 38314, "pid": 435, "tid": 435, "ts": 1776649790926197, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926207, "dur": 23, + "args": { + "External id": 4986, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38315, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38315, "pid": 0, "tid": 7, "ts": 1776649790926207, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926198, "dur": 6, + "args": { + "External id": 4986, + "cbid": 211, "correlation": 38315 + } + }, + { + "ph": "s", "id": 38315, "pid": 435, "tid": 435, "ts": 1776649790926198, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926227, "dur": 0, + "args": { + "External id": 4988, + "cbid": 210, "correlation": 38342 + } + }, + { + "ph": "f", "id": 38342, "pid": 435, "tid": 435, "ts": 1776649790926227, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926238, "dur": 23, + "args": { + "External id": 4988, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38343, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38343, "pid": 0, "tid": 7, "ts": 1776649790926238, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926229, "dur": 6, + "args": { + "External id": 4988, + "cbid": 211, "correlation": 38343 + } + }, + { + "ph": "s", "id": 38343, "pid": 435, "tid": 435, "ts": 1776649790926229, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926258, "dur": 0, + "args": { + "External id": 4990, + "cbid": 210, "correlation": 38370 + } + }, + { + "ph": "f", "id": 38370, "pid": 435, "tid": 435, "ts": 1776649790926258, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926268, "dur": 23, + "args": { + "External id": 4990, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38371, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38371, "pid": 0, "tid": 7, "ts": 1776649790926268, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926259, "dur": 6, + "args": { + "External id": 4990, + "cbid": 211, "correlation": 38371 + } + }, + { + "ph": "s", "id": 38371, "pid": 435, "tid": 435, "ts": 1776649790926259, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926288, "dur": 0, + "args": { + "External id": 4992, + "cbid": 210, "correlation": 38398 + } + }, + { + "ph": "f", "id": 38398, "pid": 435, "tid": 435, "ts": 1776649790926288, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926299, "dur": 23, + "args": { + "External id": 4992, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38399, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38399, "pid": 0, "tid": 7, "ts": 1776649790926299, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926290, "dur": 6, + "args": { + "External id": 4992, + "cbid": 211, "correlation": 38399 + } + }, + { + "ph": "s", "id": 38399, "pid": 435, "tid": 435, "ts": 1776649790926290, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926319, "dur": 0, + "args": { + "External id": 4994, + "cbid": 210, "correlation": 38426 + } + }, + { + "ph": "f", "id": 38426, "pid": 435, "tid": 435, "ts": 1776649790926319, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926329, "dur": 23, + "args": { + "External id": 4994, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38427, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38427, "pid": 0, "tid": 7, "ts": 1776649790926329, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926320, "dur": 6, + "args": { + "External id": 4994, + "cbid": 211, "correlation": 38427 + } + }, + { + "ph": "s", "id": 38427, "pid": 435, "tid": 435, "ts": 1776649790926320, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926349, "dur": 0, + "args": { + "External id": 4996, + "cbid": 210, "correlation": 38454 + } + }, + { + "ph": "f", "id": 38454, "pid": 435, "tid": 435, "ts": 1776649790926349, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926359, "dur": 23, + "args": { + "External id": 4996, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38455, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38455, "pid": 0, "tid": 7, "ts": 1776649790926359, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926350, "dur": 6, + "args": { + "External id": 4996, + "cbid": 211, "correlation": 38455 + } + }, + { + "ph": "s", "id": 38455, "pid": 435, "tid": 435, "ts": 1776649790926350, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926379, "dur": 0, + "args": { + "External id": 4998, + "cbid": 210, "correlation": 38482 + } + }, + { + "ph": "f", "id": 38482, "pid": 435, "tid": 435, "ts": 1776649790926379, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926391, "dur": 23, + "args": { + "External id": 4998, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38483, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38483, "pid": 0, "tid": 7, "ts": 1776649790926391, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926380, "dur": 6, + "args": { + "External id": 4998, + "cbid": 211, "correlation": 38483 + } + }, + { + "ph": "s", "id": 38483, "pid": 435, "tid": 435, "ts": 1776649790926380, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926409, "dur": 0, + "args": { + "External id": 5000, + "cbid": 210, "correlation": 38510 + } + }, + { + "ph": "f", "id": 38510, "pid": 435, "tid": 435, "ts": 1776649790926409, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926420, "dur": 23, + "args": { + "External id": 5000, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38511, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38511, "pid": 0, "tid": 7, "ts": 1776649790926420, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926410, "dur": 5, + "args": { + "External id": 5000, + "cbid": 211, "correlation": 38511 + } + }, + { + "ph": "s", "id": 38511, "pid": 435, "tid": 435, "ts": 1776649790926410, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926439, "dur": 0, + "args": { + "External id": 5002, + "cbid": 210, "correlation": 38538 + } + }, + { + "ph": "f", "id": 38538, "pid": 435, "tid": 435, "ts": 1776649790926439, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926450, "dur": 23, + "args": { + "External id": 5002, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38539, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38539, "pid": 0, "tid": 7, "ts": 1776649790926450, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926440, "dur": 6, + "args": { + "External id": 5002, + "cbid": 211, "correlation": 38539 + } + }, + { + "ph": "s", "id": 38539, "pid": 435, "tid": 435, "ts": 1776649790926440, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926469, "dur": 0, + "args": { + "External id": 5004, + "cbid": 210, "correlation": 38566 + } + }, + { + "ph": "f", "id": 38566, "pid": 435, "tid": 435, "ts": 1776649790926469, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926480, "dur": 23, + "args": { + "External id": 5004, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38567, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38567, "pid": 0, "tid": 7, "ts": 1776649790926480, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926470, "dur": 5, + "args": { + "External id": 5004, + "cbid": 211, "correlation": 38567 + } + }, + { + "ph": "s", "id": 38567, "pid": 435, "tid": 435, "ts": 1776649790926470, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926499, "dur": 0, + "args": { + "External id": 5006, + "cbid": 210, "correlation": 38594 + } + }, + { + "ph": "f", "id": 38594, "pid": 435, "tid": 435, "ts": 1776649790926499, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926510, "dur": 23, + "args": { + "External id": 5006, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38595, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38595, "pid": 0, "tid": 7, "ts": 1776649790926510, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926500, "dur": 6, + "args": { + "External id": 5006, + "cbid": 211, "correlation": 38595 + } + }, + { + "ph": "s", "id": 38595, "pid": 435, "tid": 435, "ts": 1776649790926500, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926529, "dur": 0, + "args": { + "External id": 5008, + "cbid": 210, "correlation": 38622 + } + }, + { + "ph": "f", "id": 38622, "pid": 435, "tid": 435, "ts": 1776649790926529, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926540, "dur": 23, + "args": { + "External id": 5008, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38623, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38623, "pid": 0, "tid": 7, "ts": 1776649790926540, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926530, "dur": 5, + "args": { + "External id": 5008, + "cbid": 211, "correlation": 38623 + } + }, + { + "ph": "s", "id": 38623, "pid": 435, "tid": 435, "ts": 1776649790926530, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926559, "dur": 0, + "args": { + "External id": 5010, + "cbid": 210, "correlation": 38650 + } + }, + { + "ph": "f", "id": 38650, "pid": 435, "tid": 435, "ts": 1776649790926559, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926570, "dur": 23, + "args": { + "External id": 5010, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38651, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38651, "pid": 0, "tid": 7, "ts": 1776649790926570, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926560, "dur": 6, + "args": { + "External id": 5010, + "cbid": 211, "correlation": 38651 + } + }, + { + "ph": "s", "id": 38651, "pid": 435, "tid": 435, "ts": 1776649790926560, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926588, "dur": 0, + "args": { + "External id": 5012, + "cbid": 210, "correlation": 38678 + } + }, + { + "ph": "f", "id": 38678, "pid": 435, "tid": 435, "ts": 1776649790926588, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926600, "dur": 23, + "args": { + "External id": 5012, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38679, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38679, "pid": 0, "tid": 7, "ts": 1776649790926600, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926590, "dur": 5, + "args": { + "External id": 5012, + "cbid": 211, "correlation": 38679 + } + }, + { + "ph": "s", "id": 38679, "pid": 435, "tid": 435, "ts": 1776649790926590, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926618, "dur": 0, + "args": { + "External id": 5014, + "cbid": 210, "correlation": 38706 + } + }, + { + "ph": "f", "id": 38706, "pid": 435, "tid": 435, "ts": 1776649790926618, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926630, "dur": 23, + "args": { + "External id": 5014, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38707, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38707, "pid": 0, "tid": 7, "ts": 1776649790926630, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926620, "dur": 6, + "args": { + "External id": 5014, + "cbid": 211, "correlation": 38707 + } + }, + { + "ph": "s", "id": 38707, "pid": 435, "tid": 435, "ts": 1776649790926620, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926649, "dur": 0, + "args": { + "External id": 5016, + "cbid": 210, "correlation": 38734 + } + }, + { + "ph": "f", "id": 38734, "pid": 435, "tid": 435, "ts": 1776649790926649, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926660, "dur": 23, + "args": { + "External id": 5016, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38735, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38735, "pid": 0, "tid": 7, "ts": 1776649790926660, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926650, "dur": 5, + "args": { + "External id": 5016, + "cbid": 211, "correlation": 38735 + } + }, + { + "ph": "s", "id": 38735, "pid": 435, "tid": 435, "ts": 1776649790926650, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926678, "dur": 0, + "args": { + "External id": 5018, + "cbid": 210, "correlation": 38762 + } + }, + { + "ph": "f", "id": 38762, "pid": 435, "tid": 435, "ts": 1776649790926678, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926690, "dur": 23, + "args": { + "External id": 5018, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38763, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38763, "pid": 0, "tid": 7, "ts": 1776649790926690, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926680, "dur": 6, + "args": { + "External id": 5018, + "cbid": 211, "correlation": 38763 + } + }, + { + "ph": "s", "id": 38763, "pid": 435, "tid": 435, "ts": 1776649790926680, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926708, "dur": 0, + "args": { + "External id": 5020, + "cbid": 210, "correlation": 38790 + } + }, + { + "ph": "f", "id": 38790, "pid": 435, "tid": 435, "ts": 1776649790926708, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926719, "dur": 23, + "args": { + "External id": 5020, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38791, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38791, "pid": 0, "tid": 7, "ts": 1776649790926719, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926709, "dur": 5, + "args": { + "External id": 5020, + "cbid": 211, "correlation": 38791 + } + }, + { + "ph": "s", "id": 38791, "pid": 435, "tid": 435, "ts": 1776649790926709, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926738, "dur": 0, + "args": { + "External id": 5022, + "cbid": 210, "correlation": 38818 + } + }, + { + "ph": "f", "id": 38818, "pid": 435, "tid": 435, "ts": 1776649790926738, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926750, "dur": 23, + "args": { + "External id": 5022, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38819, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38819, "pid": 0, "tid": 7, "ts": 1776649790926750, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926739, "dur": 6, + "args": { + "External id": 5022, + "cbid": 211, "correlation": 38819 + } + }, + { + "ph": "s", "id": 38819, "pid": 435, "tid": 435, "ts": 1776649790926739, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926768, "dur": 0, + "args": { + "External id": 5024, + "cbid": 210, "correlation": 38846 + } + }, + { + "ph": "f", "id": 38846, "pid": 435, "tid": 435, "ts": 1776649790926768, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926779, "dur": 23, + "args": { + "External id": 5024, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38847, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38847, "pid": 0, "tid": 7, "ts": 1776649790926779, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926769, "dur": 5, + "args": { + "External id": 5024, + "cbid": 211, "correlation": 38847 + } + }, + { + "ph": "s", "id": 38847, "pid": 435, "tid": 435, "ts": 1776649790926769, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926798, "dur": 0, + "args": { + "External id": 5026, + "cbid": 210, "correlation": 38874 + } + }, + { + "ph": "f", "id": 38874, "pid": 435, "tid": 435, "ts": 1776649790926798, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926809, "dur": 23, + "args": { + "External id": 5026, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38875, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38875, "pid": 0, "tid": 7, "ts": 1776649790926809, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926799, "dur": 6, + "args": { + "External id": 5026, + "cbid": 211, "correlation": 38875 + } + }, + { + "ph": "s", "id": 38875, "pid": 435, "tid": 435, "ts": 1776649790926799, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926828, "dur": 0, + "args": { + "External id": 5028, + "cbid": 210, "correlation": 38902 + } + }, + { + "ph": "f", "id": 38902, "pid": 435, "tid": 435, "ts": 1776649790926828, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926839, "dur": 23, + "args": { + "External id": 5028, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38903, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38903, "pid": 0, "tid": 7, "ts": 1776649790926839, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926829, "dur": 5, + "args": { + "External id": 5028, + "cbid": 211, "correlation": 38903 + } + }, + { + "ph": "s", "id": 38903, "pid": 435, "tid": 435, "ts": 1776649790926829, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926858, "dur": 0, + "args": { + "External id": 5030, + "cbid": 210, "correlation": 38930 + } + }, + { + "ph": "f", "id": 38930, "pid": 435, "tid": 435, "ts": 1776649790926858, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926870, "dur": 23, + "args": { + "External id": 5030, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38931, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38931, "pid": 0, "tid": 7, "ts": 1776649790926870, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926860, "dur": 6, + "args": { + "External id": 5030, + "cbid": 211, "correlation": 38931 + } + }, + { + "ph": "s", "id": 38931, "pid": 435, "tid": 435, "ts": 1776649790926860, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926889, "dur": 0, + "args": { + "External id": 5032, + "cbid": 210, "correlation": 38958 + } + }, + { + "ph": "f", "id": 38958, "pid": 435, "tid": 435, "ts": 1776649790926889, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926900, "dur": 23, + "args": { + "External id": 5032, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38959, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38959, "pid": 0, "tid": 7, "ts": 1776649790926900, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926890, "dur": 6, + "args": { + "External id": 5032, + "cbid": 211, "correlation": 38959 + } + }, + { + "ph": "s", "id": 38959, "pid": 435, "tid": 435, "ts": 1776649790926890, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926918, "dur": 0, + "args": { + "External id": 5034, + "cbid": 210, "correlation": 38986 + } + }, + { + "ph": "f", "id": 38986, "pid": 435, "tid": 435, "ts": 1776649790926918, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926930, "dur": 23, + "args": { + "External id": 5034, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 38987, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 38987, "pid": 0, "tid": 7, "ts": 1776649790926930, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926920, "dur": 6, + "args": { + "External id": 5034, + "cbid": 211, "correlation": 38987 + } + }, + { + "ph": "s", "id": 38987, "pid": 435, "tid": 435, "ts": 1776649790926920, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926948, "dur": 0, + "args": { + "External id": 5036, + "cbid": 210, "correlation": 39014 + } + }, + { + "ph": "f", "id": 39014, "pid": 435, "tid": 435, "ts": 1776649790926948, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926960, "dur": 23, + "args": { + "External id": 5036, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39015, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39015, "pid": 0, "tid": 7, "ts": 1776649790926960, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926949, "dur": 5, + "args": { + "External id": 5036, + "cbid": 211, "correlation": 39015 + } + }, + { + "ph": "s", "id": 39015, "pid": 435, "tid": 435, "ts": 1776649790926949, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790926978, "dur": 0, + "args": { + "External id": 5038, + "cbid": 210, "correlation": 39042 + } + }, + { + "ph": "f", "id": 39042, "pid": 435, "tid": 435, "ts": 1776649790926978, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790926989, "dur": 23, + "args": { + "External id": 5038, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39043, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39043, "pid": 0, "tid": 7, "ts": 1776649790926989, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790926979, "dur": 6, + "args": { + "External id": 5038, + "cbid": 211, "correlation": 39043 + } + }, + { + "ph": "s", "id": 39043, "pid": 435, "tid": 435, "ts": 1776649790926979, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927008, "dur": 0, + "args": { + "External id": 5040, + "cbid": 210, "correlation": 39070 + } + }, + { + "ph": "f", "id": 39070, "pid": 435, "tid": 435, "ts": 1776649790927008, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927020, "dur": 23, + "args": { + "External id": 5040, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39071, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39071, "pid": 0, "tid": 7, "ts": 1776649790927020, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927009, "dur": 5, + "args": { + "External id": 5040, + "cbid": 211, "correlation": 39071 + } + }, + { + "ph": "s", "id": 39071, "pid": 435, "tid": 435, "ts": 1776649790927009, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927038, "dur": 0, + "args": { + "External id": 5042, + "cbid": 210, "correlation": 39098 + } + }, + { + "ph": "f", "id": 39098, "pid": 435, "tid": 435, "ts": 1776649790927038, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927050, "dur": 23, + "args": { + "External id": 5042, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39099, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39099, "pid": 0, "tid": 7, "ts": 1776649790927050, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927039, "dur": 6, + "args": { + "External id": 5042, + "cbid": 211, "correlation": 39099 + } + }, + { + "ph": "s", "id": 39099, "pid": 435, "tid": 435, "ts": 1776649790927039, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927068, "dur": 0, + "args": { + "External id": 5044, + "cbid": 210, "correlation": 39126 + } + }, + { + "ph": "f", "id": 39126, "pid": 435, "tid": 435, "ts": 1776649790927068, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927080, "dur": 23, + "args": { + "External id": 5044, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39127, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39127, "pid": 0, "tid": 7, "ts": 1776649790927080, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927070, "dur": 6, + "args": { + "External id": 5044, + "cbid": 211, "correlation": 39127 + } + }, + { + "ph": "s", "id": 39127, "pid": 435, "tid": 435, "ts": 1776649790927070, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927098, "dur": 0, + "args": { + "External id": 5046, + "cbid": 210, "correlation": 39154 + } + }, + { + "ph": "f", "id": 39154, "pid": 435, "tid": 435, "ts": 1776649790927098, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927110, "dur": 23, + "args": { + "External id": 5046, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39155, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39155, "pid": 0, "tid": 7, "ts": 1776649790927110, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927100, "dur": 6, + "args": { + "External id": 5046, + "cbid": 211, "correlation": 39155 + } + }, + { + "ph": "s", "id": 39155, "pid": 435, "tid": 435, "ts": 1776649790927100, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927128, "dur": 0, + "args": { + "External id": 5048, + "cbid": 210, "correlation": 39182 + } + }, + { + "ph": "f", "id": 39182, "pid": 435, "tid": 435, "ts": 1776649790927128, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927140, "dur": 23, + "args": { + "External id": 5048, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39183, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39183, "pid": 0, "tid": 7, "ts": 1776649790927140, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927130, "dur": 6, + "args": { + "External id": 5048, + "cbid": 211, "correlation": 39183 + } + }, + { + "ph": "s", "id": 39183, "pid": 435, "tid": 435, "ts": 1776649790927130, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927158, "dur": 0, + "args": { + "External id": 5050, + "cbid": 210, "correlation": 39210 + } + }, + { + "ph": "f", "id": 39210, "pid": 435, "tid": 435, "ts": 1776649790927158, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927170, "dur": 23, + "args": { + "External id": 5050, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39211, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39211, "pid": 0, "tid": 7, "ts": 1776649790927170, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927160, "dur": 6, + "args": { + "External id": 5050, + "cbid": 211, "correlation": 39211 + } + }, + { + "ph": "s", "id": 39211, "pid": 435, "tid": 435, "ts": 1776649790927160, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927188, "dur": 0, + "args": { + "External id": 5052, + "cbid": 210, "correlation": 39238 + } + }, + { + "ph": "f", "id": 39238, "pid": 435, "tid": 435, "ts": 1776649790927188, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927200, "dur": 23, + "args": { + "External id": 5052, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39239, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39239, "pid": 0, "tid": 7, "ts": 1776649790927200, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927190, "dur": 6, + "args": { + "External id": 5052, + "cbid": 211, "correlation": 39239 + } + }, + { + "ph": "s", "id": 39239, "pid": 435, "tid": 435, "ts": 1776649790927190, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927218, "dur": 0, + "args": { + "External id": 5054, + "cbid": 210, "correlation": 39266 + } + }, + { + "ph": "f", "id": 39266, "pid": 435, "tid": 435, "ts": 1776649790927218, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927230, "dur": 23, + "args": { + "External id": 5054, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39267, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39267, "pid": 0, "tid": 7, "ts": 1776649790927230, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927219, "dur": 6, + "args": { + "External id": 5054, + "cbid": 211, "correlation": 39267 + } + }, + { + "ph": "s", "id": 39267, "pid": 435, "tid": 435, "ts": 1776649790927219, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927248, "dur": 0, + "args": { + "External id": 5056, + "cbid": 210, "correlation": 39294 + } + }, + { + "ph": "f", "id": 39294, "pid": 435, "tid": 435, "ts": 1776649790927248, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927260, "dur": 23, + "args": { + "External id": 5056, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39295, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39295, "pid": 0, "tid": 7, "ts": 1776649790927260, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927250, "dur": 6, + "args": { + "External id": 5056, + "cbid": 211, "correlation": 39295 + } + }, + { + "ph": "s", "id": 39295, "pid": 435, "tid": 435, "ts": 1776649790927250, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927279, "dur": 0, + "args": { + "External id": 5058, + "cbid": 210, "correlation": 39322 + } + }, + { + "ph": "f", "id": 39322, "pid": 435, "tid": 435, "ts": 1776649790927279, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927290, "dur": 23, + "args": { + "External id": 5058, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39323, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39323, "pid": 0, "tid": 7, "ts": 1776649790927290, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927280, "dur": 5, + "args": { + "External id": 5058, + "cbid": 211, "correlation": 39323 + } + }, + { + "ph": "s", "id": 39323, "pid": 435, "tid": 435, "ts": 1776649790927280, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927308, "dur": 0, + "args": { + "External id": 5060, + "cbid": 210, "correlation": 39350 + } + }, + { + "ph": "f", "id": 39350, "pid": 435, "tid": 435, "ts": 1776649790927308, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927320, "dur": 23, + "args": { + "External id": 5060, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39351, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39351, "pid": 0, "tid": 7, "ts": 1776649790927320, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927309, "dur": 6, + "args": { + "External id": 5060, + "cbid": 211, "correlation": 39351 + } + }, + { + "ph": "s", "id": 39351, "pid": 435, "tid": 435, "ts": 1776649790927309, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927338, "dur": 0, + "args": { + "External id": 5062, + "cbid": 210, "correlation": 39378 + } + }, + { + "ph": "f", "id": 39378, "pid": 435, "tid": 435, "ts": 1776649790927338, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927350, "dur": 23, + "args": { + "External id": 5062, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39379, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39379, "pid": 0, "tid": 7, "ts": 1776649790927350, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927340, "dur": 6, + "args": { + "External id": 5062, + "cbid": 211, "correlation": 39379 + } + }, + { + "ph": "s", "id": 39379, "pid": 435, "tid": 435, "ts": 1776649790927340, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927368, "dur": 0, + "args": { + "External id": 5064, + "cbid": 210, "correlation": 39406 + } + }, + { + "ph": "f", "id": 39406, "pid": 435, "tid": 435, "ts": 1776649790927368, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927380, "dur": 23, + "args": { + "External id": 5064, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39407, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39407, "pid": 0, "tid": 7, "ts": 1776649790927380, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927369, "dur": 6, + "args": { + "External id": 5064, + "cbid": 211, "correlation": 39407 + } + }, + { + "ph": "s", "id": 39407, "pid": 435, "tid": 435, "ts": 1776649790927369, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927398, "dur": 0, + "args": { + "External id": 5066, + "cbid": 210, "correlation": 39434 + } + }, + { + "ph": "f", "id": 39434, "pid": 435, "tid": 435, "ts": 1776649790927398, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927410, "dur": 23, + "args": { + "External id": 5066, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39435, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39435, "pid": 0, "tid": 7, "ts": 1776649790927410, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927399, "dur": 5, + "args": { + "External id": 5066, + "cbid": 211, "correlation": 39435 + } + }, + { + "ph": "s", "id": 39435, "pid": 435, "tid": 435, "ts": 1776649790927399, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927427, "dur": 0, + "args": { + "External id": 5068, + "cbid": 210, "correlation": 39462 + } + }, + { + "ph": "f", "id": 39462, "pid": 435, "tid": 435, "ts": 1776649790927427, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927439, "dur": 23, + "args": { + "External id": 5068, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39463, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39463, "pid": 0, "tid": 7, "ts": 1776649790927439, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927429, "dur": 5, + "args": { + "External id": 5068, + "cbid": 211, "correlation": 39463 + } + }, + { + "ph": "s", "id": 39463, "pid": 435, "tid": 435, "ts": 1776649790927429, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927457, "dur": 0, + "args": { + "External id": 5070, + "cbid": 210, "correlation": 39490 + } + }, + { + "ph": "f", "id": 39490, "pid": 435, "tid": 435, "ts": 1776649790927457, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927469, "dur": 23, + "args": { + "External id": 5070, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39491, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39491, "pid": 0, "tid": 7, "ts": 1776649790927469, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927458, "dur": 6, + "args": { + "External id": 5070, + "cbid": 211, "correlation": 39491 + } + }, + { + "ph": "s", "id": 39491, "pid": 435, "tid": 435, "ts": 1776649790927458, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927488, "dur": 0, + "args": { + "External id": 5072, + "cbid": 210, "correlation": 39518 + } + }, + { + "ph": "f", "id": 39518, "pid": 435, "tid": 435, "ts": 1776649790927488, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927499, "dur": 23, + "args": { + "External id": 5072, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39519, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39519, "pid": 0, "tid": 7, "ts": 1776649790927499, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927489, "dur": 6, + "args": { + "External id": 5072, + "cbid": 211, "correlation": 39519 + } + }, + { + "ph": "s", "id": 39519, "pid": 435, "tid": 435, "ts": 1776649790927489, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927518, "dur": 0, + "args": { + "External id": 5074, + "cbid": 210, "correlation": 39546 + } + }, + { + "ph": "f", "id": 39546, "pid": 435, "tid": 435, "ts": 1776649790927518, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927530, "dur": 23, + "args": { + "External id": 5074, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39547, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39547, "pid": 0, "tid": 7, "ts": 1776649790927530, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927520, "dur": 6, + "args": { + "External id": 5074, + "cbid": 211, "correlation": 39547 + } + }, + { + "ph": "s", "id": 39547, "pid": 435, "tid": 435, "ts": 1776649790927520, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927548, "dur": 0, + "args": { + "External id": 5076, + "cbid": 210, "correlation": 39574 + } + }, + { + "ph": "f", "id": 39574, "pid": 435, "tid": 435, "ts": 1776649790927548, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927560, "dur": 23, + "args": { + "External id": 5076, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39575, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39575, "pid": 0, "tid": 7, "ts": 1776649790927560, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927549, "dur": 6, + "args": { + "External id": 5076, + "cbid": 211, "correlation": 39575 + } + }, + { + "ph": "s", "id": 39575, "pid": 435, "tid": 435, "ts": 1776649790927549, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927578, "dur": 0, + "args": { + "External id": 5078, + "cbid": 210, "correlation": 39602 + } + }, + { + "ph": "f", "id": 39602, "pid": 435, "tid": 435, "ts": 1776649790927578, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927590, "dur": 23, + "args": { + "External id": 5078, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39603, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39603, "pid": 0, "tid": 7, "ts": 1776649790927590, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927579, "dur": 6, + "args": { + "External id": 5078, + "cbid": 211, "correlation": 39603 + } + }, + { + "ph": "s", "id": 39603, "pid": 435, "tid": 435, "ts": 1776649790927579, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927608, "dur": 0, + "args": { + "External id": 5080, + "cbid": 210, "correlation": 39630 + } + }, + { + "ph": "f", "id": 39630, "pid": 435, "tid": 435, "ts": 1776649790927608, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927619, "dur": 23, + "args": { + "External id": 5080, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39631, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39631, "pid": 0, "tid": 7, "ts": 1776649790927619, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927609, "dur": 5, + "args": { + "External id": 5080, + "cbid": 211, "correlation": 39631 + } + }, + { + "ph": "s", "id": 39631, "pid": 435, "tid": 435, "ts": 1776649790927609, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790927642, "dur": 4, + "args": { + "External id": 39642, + "cbid": 135, "correlation": 39642 + } + }, + { + "ph": "f", "id": 39642, "pid": 435, "tid": 435, "ts": 1776649790927642, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790927657, "dur": 5, + "args": { + "External id": 39647, + "cbid": 165, "correlation": 39647 + } + }, + { + "ph": "s", "id": 39647, "pid": 435, "tid": 435, "ts": 1776649790927657, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649790927665, "dur": 4, + "args": { + "External id": 39650, + "cbid": 139, "correlation": 39650 + } + }, + { + "ph": "f", "id": 39650, "pid": 435, "tid": 435, "ts": 1776649790927665, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927702, "dur": 1, + "args": { + "External id": 5082, + "cbid": 210, "correlation": 39684 + } + }, + { + "ph": "f", "id": 39684, "pid": 435, "tid": 435, "ts": 1776649790927702, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927714, "dur": 23, + "args": { + "External id": 5082, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39685, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39685, "pid": 0, "tid": 7, "ts": 1776649790927714, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927703, "dur": 7, + "args": { + "External id": 5082, + "cbid": 211, "correlation": 39685 + } + }, + { + "ph": "s", "id": 39685, "pid": 435, "tid": 435, "ts": 1776649790927703, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927737, "dur": 0, + "args": { + "External id": 5084, + "cbid": 210, "correlation": 39712 + } + }, + { + "ph": "f", "id": 39712, "pid": 435, "tid": 435, "ts": 1776649790927737, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927748, "dur": 23, + "args": { + "External id": 5084, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39713, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39713, "pid": 0, "tid": 7, "ts": 1776649790927748, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927738, "dur": 6, + "args": { + "External id": 5084, + "cbid": 211, "correlation": 39713 + } + }, + { + "ph": "s", "id": 39713, "pid": 435, "tid": 435, "ts": 1776649790927738, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927768, "dur": 0, + "args": { + "External id": 5086, + "cbid": 210, "correlation": 39740 + } + }, + { + "ph": "f", "id": 39740, "pid": 435, "tid": 435, "ts": 1776649790927768, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927778, "dur": 23, + "args": { + "External id": 5086, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39741, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39741, "pid": 0, "tid": 7, "ts": 1776649790927778, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927769, "dur": 6, + "args": { + "External id": 5086, + "cbid": 211, "correlation": 39741 + } + }, + { + "ph": "s", "id": 39741, "pid": 435, "tid": 435, "ts": 1776649790927769, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927798, "dur": 0, + "args": { + "External id": 5088, + "cbid": 210, "correlation": 39768 + } + }, + { + "ph": "f", "id": 39768, "pid": 435, "tid": 435, "ts": 1776649790927798, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927808, "dur": 23, + "args": { + "External id": 5088, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39769, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39769, "pid": 0, "tid": 7, "ts": 1776649790927808, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927799, "dur": 6, + "args": { + "External id": 5088, + "cbid": 211, "correlation": 39769 + } + }, + { + "ph": "s", "id": 39769, "pid": 435, "tid": 435, "ts": 1776649790927799, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927828, "dur": 0, + "args": { + "External id": 5090, + "cbid": 210, "correlation": 39796 + } + }, + { + "ph": "f", "id": 39796, "pid": 435, "tid": 435, "ts": 1776649790927828, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927838, "dur": 23, + "args": { + "External id": 5090, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39797, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39797, "pid": 0, "tid": 7, "ts": 1776649790927838, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927829, "dur": 6, + "args": { + "External id": 5090, + "cbid": 211, "correlation": 39797 + } + }, + { + "ph": "s", "id": 39797, "pid": 435, "tid": 435, "ts": 1776649790927829, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927859, "dur": 0, + "args": { + "External id": 5092, + "cbid": 210, "correlation": 39824 + } + }, + { + "ph": "f", "id": 39824, "pid": 435, "tid": 435, "ts": 1776649790927859, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927869, "dur": 23, + "args": { + "External id": 5092, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39825, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39825, "pid": 0, "tid": 7, "ts": 1776649790927869, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927860, "dur": 6, + "args": { + "External id": 5092, + "cbid": 211, "correlation": 39825 + } + }, + { + "ph": "s", "id": 39825, "pid": 435, "tid": 435, "ts": 1776649790927860, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927889, "dur": 0, + "args": { + "External id": 5094, + "cbid": 210, "correlation": 39852 + } + }, + { + "ph": "f", "id": 39852, "pid": 435, "tid": 435, "ts": 1776649790927889, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927899, "dur": 23, + "args": { + "External id": 5094, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39853, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39853, "pid": 0, "tid": 7, "ts": 1776649790927899, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927891, "dur": 5, + "args": { + "External id": 5094, + "cbid": 211, "correlation": 39853 + } + }, + { + "ph": "s", "id": 39853, "pid": 435, "tid": 435, "ts": 1776649790927891, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927919, "dur": 0, + "args": { + "External id": 5096, + "cbid": 210, "correlation": 39880 + } + }, + { + "ph": "f", "id": 39880, "pid": 435, "tid": 435, "ts": 1776649790927919, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927929, "dur": 23, + "args": { + "External id": 5096, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39881, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39881, "pid": 0, "tid": 7, "ts": 1776649790927929, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927920, "dur": 6, + "args": { + "External id": 5096, + "cbid": 211, "correlation": 39881 + } + }, + { + "ph": "s", "id": 39881, "pid": 435, "tid": 435, "ts": 1776649790927920, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927949, "dur": 0, + "args": { + "External id": 5098, + "cbid": 210, "correlation": 39908 + } + }, + { + "ph": "f", "id": 39908, "pid": 435, "tid": 435, "ts": 1776649790927949, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927959, "dur": 23, + "args": { + "External id": 5098, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39909, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39909, "pid": 0, "tid": 7, "ts": 1776649790927959, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927950, "dur": 6, + "args": { + "External id": 5098, + "cbid": 211, "correlation": 39909 + } + }, + { + "ph": "s", "id": 39909, "pid": 435, "tid": 435, "ts": 1776649790927950, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790927979, "dur": 0, + "args": { + "External id": 5100, + "cbid": 210, "correlation": 39936 + } + }, + { + "ph": "f", "id": 39936, "pid": 435, "tid": 435, "ts": 1776649790927979, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790927989, "dur": 23, + "args": { + "External id": 5100, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39937, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39937, "pid": 0, "tid": 7, "ts": 1776649790927989, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790927980, "dur": 6, + "args": { + "External id": 5100, + "cbid": 211, "correlation": 39937 + } + }, + { + "ph": "s", "id": 39937, "pid": 435, "tid": 435, "ts": 1776649790927980, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790928005, "dur": 11, + "args": { + "External id": 39943, + "cbid": 165, "correlation": 39943 + } + }, + { + "ph": "s", "id": 39943, "pid": 435, "tid": 435, "ts": 1776649790928005, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790928037, "dur": 3, + "args": { + "External id": 39952, + "cbid": 135, "correlation": 39952 + } + }, + { + "ph": "f", "id": 39952, "pid": 435, "tid": 435, "ts": 1776649790928037, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928066, "dur": 1, + "args": { + "External id": 5102, + "cbid": 210, "correlation": 39978 + } + }, + { + "ph": "f", "id": 39978, "pid": 435, "tid": 435, "ts": 1776649790928066, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928078, "dur": 23, + "args": { + "External id": 5102, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 39979, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 39979, "pid": 0, "tid": 7, "ts": 1776649790928078, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928068, "dur": 7, + "args": { + "External id": 5102, + "cbid": 211, "correlation": 39979 + } + }, + { + "ph": "s", "id": 39979, "pid": 435, "tid": 435, "ts": 1776649790928068, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928101, "dur": 0, + "args": { + "External id": 5104, + "cbid": 210, "correlation": 40006 + } + }, + { + "ph": "f", "id": 40006, "pid": 435, "tid": 435, "ts": 1776649790928101, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928111, "dur": 23, + "args": { + "External id": 5104, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40007, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40007, "pid": 0, "tid": 7, "ts": 1776649790928111, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928102, "dur": 6, + "args": { + "External id": 5104, + "cbid": 211, "correlation": 40007 + } + }, + { + "ph": "s", "id": 40007, "pid": 435, "tid": 435, "ts": 1776649790928102, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928131, "dur": 0, + "args": { + "External id": 5106, + "cbid": 210, "correlation": 40034 + } + }, + { + "ph": "f", "id": 40034, "pid": 435, "tid": 435, "ts": 1776649790928131, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928142, "dur": 23, + "args": { + "External id": 5106, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40035, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40035, "pid": 0, "tid": 7, "ts": 1776649790928142, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928133, "dur": 6, + "args": { + "External id": 5106, + "cbid": 211, "correlation": 40035 + } + }, + { + "ph": "s", "id": 40035, "pid": 435, "tid": 435, "ts": 1776649790928133, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928162, "dur": 0, + "args": { + "External id": 5108, + "cbid": 210, "correlation": 40062 + } + }, + { + "ph": "f", "id": 40062, "pid": 435, "tid": 435, "ts": 1776649790928162, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928172, "dur": 23, + "args": { + "External id": 5108, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40063, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40063, "pid": 0, "tid": 7, "ts": 1776649790928172, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928163, "dur": 6, + "args": { + "External id": 5108, + "cbid": 211, "correlation": 40063 + } + }, + { + "ph": "s", "id": 40063, "pid": 435, "tid": 435, "ts": 1776649790928163, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928193, "dur": 0, + "args": { + "External id": 5110, + "cbid": 210, "correlation": 40090 + } + }, + { + "ph": "f", "id": 40090, "pid": 435, "tid": 435, "ts": 1776649790928193, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928203, "dur": 23, + "args": { + "External id": 5110, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40091, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40091, "pid": 0, "tid": 7, "ts": 1776649790928203, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928194, "dur": 6, + "args": { + "External id": 5110, + "cbid": 211, "correlation": 40091 + } + }, + { + "ph": "s", "id": 40091, "pid": 435, "tid": 435, "ts": 1776649790928194, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928262, "dur": 1, + "args": { + "External id": 5112, + "cbid": 210, "correlation": 40118 + } + }, + { + "ph": "f", "id": 40118, "pid": 435, "tid": 435, "ts": 1776649790928262, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928273, "dur": 23, + "args": { + "External id": 5112, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40119, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40119, "pid": 0, "tid": 7, "ts": 1776649790928273, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928264, "dur": 7, + "args": { + "External id": 5112, + "cbid": 211, "correlation": 40119 + } + }, + { + "ph": "s", "id": 40119, "pid": 435, "tid": 435, "ts": 1776649790928264, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928298, "dur": 0, + "args": { + "External id": 5114, + "cbid": 210, "correlation": 40146 + } + }, + { + "ph": "f", "id": 40146, "pid": 435, "tid": 435, "ts": 1776649790928298, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928308, "dur": 23, + "args": { + "External id": 5114, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40147, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40147, "pid": 0, "tid": 7, "ts": 1776649790928308, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928299, "dur": 6, + "args": { + "External id": 5114, + "cbid": 211, "correlation": 40147 + } + }, + { + "ph": "s", "id": 40147, "pid": 435, "tid": 435, "ts": 1776649790928299, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928328, "dur": 0, + "args": { + "External id": 5116, + "cbid": 210, "correlation": 40174 + } + }, + { + "ph": "f", "id": 40174, "pid": 435, "tid": 435, "ts": 1776649790928328, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928338, "dur": 23, + "args": { + "External id": 5116, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40175, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40175, "pid": 0, "tid": 7, "ts": 1776649790928338, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928329, "dur": 6, + "args": { + "External id": 5116, + "cbid": 211, "correlation": 40175 + } + }, + { + "ph": "s", "id": 40175, "pid": 435, "tid": 435, "ts": 1776649790928329, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928358, "dur": 0, + "args": { + "External id": 5118, + "cbid": 210, "correlation": 40202 + } + }, + { + "ph": "f", "id": 40202, "pid": 435, "tid": 435, "ts": 1776649790928358, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928369, "dur": 23, + "args": { + "External id": 5118, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40203, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40203, "pid": 0, "tid": 7, "ts": 1776649790928369, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928359, "dur": 6, + "args": { + "External id": 5118, + "cbid": 211, "correlation": 40203 + } + }, + { + "ph": "s", "id": 40203, "pid": 435, "tid": 435, "ts": 1776649790928359, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928388, "dur": 0, + "args": { + "External id": 5120, + "cbid": 210, "correlation": 40230 + } + }, + { + "ph": "f", "id": 40230, "pid": 435, "tid": 435, "ts": 1776649790928388, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928399, "dur": 23, + "args": { + "External id": 5120, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40231, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40231, "pid": 0, "tid": 7, "ts": 1776649790928399, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928390, "dur": 6, + "args": { + "External id": 5120, + "cbid": 211, "correlation": 40231 + } + }, + { + "ph": "s", "id": 40231, "pid": 435, "tid": 435, "ts": 1776649790928390, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928428, "dur": 0, + "args": { + "External id": 5122, + "cbid": 210, "correlation": 40258 + } + }, + { + "ph": "f", "id": 40258, "pid": 435, "tid": 435, "ts": 1776649790928428, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928438, "dur": 23, + "args": { + "External id": 5122, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40259, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40259, "pid": 0, "tid": 7, "ts": 1776649790928438, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928429, "dur": 6, + "args": { + "External id": 5122, + "cbid": 211, "correlation": 40259 + } + }, + { + "ph": "s", "id": 40259, "pid": 435, "tid": 435, "ts": 1776649790928429, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928458, "dur": 0, + "args": { + "External id": 5124, + "cbid": 210, "correlation": 40286 + } + }, + { + "ph": "f", "id": 40286, "pid": 435, "tid": 435, "ts": 1776649790928458, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928469, "dur": 23, + "args": { + "External id": 5124, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40287, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40287, "pid": 0, "tid": 7, "ts": 1776649790928469, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928460, "dur": 6, + "args": { + "External id": 5124, + "cbid": 211, "correlation": 40287 + } + }, + { + "ph": "s", "id": 40287, "pid": 435, "tid": 435, "ts": 1776649790928460, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928488, "dur": 0, + "args": { + "External id": 5126, + "cbid": 210, "correlation": 40314 + } + }, + { + "ph": "f", "id": 40314, "pid": 435, "tid": 435, "ts": 1776649790928488, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928500, "dur": 23, + "args": { + "External id": 5126, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40315, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40315, "pid": 0, "tid": 7, "ts": 1776649790928500, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928489, "dur": 6, + "args": { + "External id": 5126, + "cbid": 211, "correlation": 40315 + } + }, + { + "ph": "s", "id": 40315, "pid": 435, "tid": 435, "ts": 1776649790928489, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928518, "dur": 0, + "args": { + "External id": 5128, + "cbid": 210, "correlation": 40342 + } + }, + { + "ph": "f", "id": 40342, "pid": 435, "tid": 435, "ts": 1776649790928518, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928529, "dur": 23, + "args": { + "External id": 5128, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40343, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40343, "pid": 0, "tid": 7, "ts": 1776649790928529, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928519, "dur": 6, + "args": { + "External id": 5128, + "cbid": 211, "correlation": 40343 + } + }, + { + "ph": "s", "id": 40343, "pid": 435, "tid": 435, "ts": 1776649790928519, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928547, "dur": 0, + "args": { + "External id": 5130, + "cbid": 210, "correlation": 40370 + } + }, + { + "ph": "f", "id": 40370, "pid": 435, "tid": 435, "ts": 1776649790928547, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928559, "dur": 23, + "args": { + "External id": 5130, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40371, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40371, "pid": 0, "tid": 7, "ts": 1776649790928559, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928549, "dur": 6, + "args": { + "External id": 5130, + "cbid": 211, "correlation": 40371 + } + }, + { + "ph": "s", "id": 40371, "pid": 435, "tid": 435, "ts": 1776649790928549, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928577, "dur": 0, + "args": { + "External id": 5132, + "cbid": 210, "correlation": 40398 + } + }, + { + "ph": "f", "id": 40398, "pid": 435, "tid": 435, "ts": 1776649790928577, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928589, "dur": 23, + "args": { + "External id": 5132, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40399, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40399, "pid": 0, "tid": 7, "ts": 1776649790928589, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928578, "dur": 6, + "args": { + "External id": 5132, + "cbid": 211, "correlation": 40399 + } + }, + { + "ph": "s", "id": 40399, "pid": 435, "tid": 435, "ts": 1776649790928578, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928607, "dur": 0, + "args": { + "External id": 5134, + "cbid": 210, "correlation": 40426 + } + }, + { + "ph": "f", "id": 40426, "pid": 435, "tid": 435, "ts": 1776649790928607, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928618, "dur": 23, + "args": { + "External id": 5134, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40427, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40427, "pid": 0, "tid": 7, "ts": 1776649790928618, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928608, "dur": 5, + "args": { + "External id": 5134, + "cbid": 211, "correlation": 40427 + } + }, + { + "ph": "s", "id": 40427, "pid": 435, "tid": 435, "ts": 1776649790928608, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928637, "dur": 0, + "args": { + "External id": 5136, + "cbid": 210, "correlation": 40454 + } + }, + { + "ph": "f", "id": 40454, "pid": 435, "tid": 435, "ts": 1776649790928637, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928648, "dur": 23, + "args": { + "External id": 5136, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40455, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40455, "pid": 0, "tid": 7, "ts": 1776649790928648, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928638, "dur": 6, + "args": { + "External id": 5136, + "cbid": 211, "correlation": 40455 + } + }, + { + "ph": "s", "id": 40455, "pid": 435, "tid": 435, "ts": 1776649790928638, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928667, "dur": 0, + "args": { + "External id": 5138, + "cbid": 210, "correlation": 40482 + } + }, + { + "ph": "f", "id": 40482, "pid": 435, "tid": 435, "ts": 1776649790928667, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928678, "dur": 23, + "args": { + "External id": 5138, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40483, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40483, "pid": 0, "tid": 7, "ts": 1776649790928678, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928668, "dur": 5, + "args": { + "External id": 5138, + "cbid": 211, "correlation": 40483 + } + }, + { + "ph": "s", "id": 40483, "pid": 435, "tid": 435, "ts": 1776649790928668, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928696, "dur": 0, + "args": { + "External id": 5140, + "cbid": 210, "correlation": 40510 + } + }, + { + "ph": "f", "id": 40510, "pid": 435, "tid": 435, "ts": 1776649790928696, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928707, "dur": 23, + "args": { + "External id": 5140, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40511, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40511, "pid": 0, "tid": 7, "ts": 1776649790928707, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928697, "dur": 5, + "args": { + "External id": 5140, + "cbid": 211, "correlation": 40511 + } + }, + { + "ph": "s", "id": 40511, "pid": 435, "tid": 435, "ts": 1776649790928697, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928726, "dur": 0, + "args": { + "External id": 5142, + "cbid": 210, "correlation": 40538 + } + }, + { + "ph": "f", "id": 40538, "pid": 435, "tid": 435, "ts": 1776649790928726, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928737, "dur": 23, + "args": { + "External id": 5142, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40539, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40539, "pid": 0, "tid": 7, "ts": 1776649790928737, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928727, "dur": 6, + "args": { + "External id": 5142, + "cbid": 211, "correlation": 40539 + } + }, + { + "ph": "s", "id": 40539, "pid": 435, "tid": 435, "ts": 1776649790928727, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928755, "dur": 0, + "args": { + "External id": 5144, + "cbid": 210, "correlation": 40566 + } + }, + { + "ph": "f", "id": 40566, "pid": 435, "tid": 435, "ts": 1776649790928755, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928767, "dur": 23, + "args": { + "External id": 5144, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40567, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40567, "pid": 0, "tid": 7, "ts": 1776649790928767, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928756, "dur": 6, + "args": { + "External id": 5144, + "cbid": 211, "correlation": 40567 + } + }, + { + "ph": "s", "id": 40567, "pid": 435, "tid": 435, "ts": 1776649790928756, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928785, "dur": 0, + "args": { + "External id": 5146, + "cbid": 210, "correlation": 40594 + } + }, + { + "ph": "f", "id": 40594, "pid": 435, "tid": 435, "ts": 1776649790928785, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928797, "dur": 23, + "args": { + "External id": 5146, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40595, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40595, "pid": 0, "tid": 7, "ts": 1776649790928797, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928786, "dur": 5, + "args": { + "External id": 5146, + "cbid": 211, "correlation": 40595 + } + }, + { + "ph": "s", "id": 40595, "pid": 435, "tid": 435, "ts": 1776649790928786, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928815, "dur": 0, + "args": { + "External id": 5148, + "cbid": 210, "correlation": 40622 + } + }, + { + "ph": "f", "id": 40622, "pid": 435, "tid": 435, "ts": 1776649790928815, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928826, "dur": 23, + "args": { + "External id": 5148, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40623, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40623, "pid": 0, "tid": 7, "ts": 1776649790928826, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928816, "dur": 5, + "args": { + "External id": 5148, + "cbid": 211, "correlation": 40623 + } + }, + { + "ph": "s", "id": 40623, "pid": 435, "tid": 435, "ts": 1776649790928816, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928845, "dur": 0, + "args": { + "External id": 5150, + "cbid": 210, "correlation": 40650 + } + }, + { + "ph": "f", "id": 40650, "pid": 435, "tid": 435, "ts": 1776649790928845, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928857, "dur": 23, + "args": { + "External id": 5150, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40651, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40651, "pid": 0, "tid": 7, "ts": 1776649790928857, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928846, "dur": 6, + "args": { + "External id": 5150, + "cbid": 211, "correlation": 40651 + } + }, + { + "ph": "s", "id": 40651, "pid": 435, "tid": 435, "ts": 1776649790928846, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928875, "dur": 0, + "args": { + "External id": 5152, + "cbid": 210, "correlation": 40678 + } + }, + { + "ph": "f", "id": 40678, "pid": 435, "tid": 435, "ts": 1776649790928875, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928887, "dur": 23, + "args": { + "External id": 5152, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40679, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40679, "pid": 0, "tid": 7, "ts": 1776649790928887, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928876, "dur": 5, + "args": { + "External id": 5152, + "cbid": 211, "correlation": 40679 + } + }, + { + "ph": "s", "id": 40679, "pid": 435, "tid": 435, "ts": 1776649790928876, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928905, "dur": 0, + "args": { + "External id": 5154, + "cbid": 210, "correlation": 40706 + } + }, + { + "ph": "f", "id": 40706, "pid": 435, "tid": 435, "ts": 1776649790928905, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928916, "dur": 23, + "args": { + "External id": 5154, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40707, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40707, "pid": 0, "tid": 7, "ts": 1776649790928916, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928906, "dur": 6, + "args": { + "External id": 5154, + "cbid": 211, "correlation": 40707 + } + }, + { + "ph": "s", "id": 40707, "pid": 435, "tid": 435, "ts": 1776649790928906, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928935, "dur": 0, + "args": { + "External id": 5156, + "cbid": 210, "correlation": 40734 + } + }, + { + "ph": "f", "id": 40734, "pid": 435, "tid": 435, "ts": 1776649790928935, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928946, "dur": 23, + "args": { + "External id": 5156, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40735, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40735, "pid": 0, "tid": 7, "ts": 1776649790928946, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928936, "dur": 5, + "args": { + "External id": 5156, + "cbid": 211, "correlation": 40735 + } + }, + { + "ph": "s", "id": 40735, "pid": 435, "tid": 435, "ts": 1776649790928936, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928965, "dur": 0, + "args": { + "External id": 5158, + "cbid": 210, "correlation": 40762 + } + }, + { + "ph": "f", "id": 40762, "pid": 435, "tid": 435, "ts": 1776649790928965, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790928976, "dur": 23, + "args": { + "External id": 5158, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40763, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40763, "pid": 0, "tid": 7, "ts": 1776649790928976, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928966, "dur": 6, + "args": { + "External id": 5158, + "cbid": 211, "correlation": 40763 + } + }, + { + "ph": "s", "id": 40763, "pid": 435, "tid": 435, "ts": 1776649790928966, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790928994, "dur": 0, + "args": { + "External id": 5160, + "cbid": 210, "correlation": 40790 + } + }, + { + "ph": "f", "id": 40790, "pid": 435, "tid": 435, "ts": 1776649790928994, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929006, "dur": 23, + "args": { + "External id": 5160, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40791, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40791, "pid": 0, "tid": 7, "ts": 1776649790929006, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790928995, "dur": 6, + "args": { + "External id": 5160, + "cbid": 211, "correlation": 40791 + } + }, + { + "ph": "s", "id": 40791, "pid": 435, "tid": 435, "ts": 1776649790928995, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929024, "dur": 0, + "args": { + "External id": 5162, + "cbid": 210, "correlation": 40818 + } + }, + { + "ph": "f", "id": 40818, "pid": 435, "tid": 435, "ts": 1776649790929024, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929035, "dur": 23, + "args": { + "External id": 5162, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40819, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40819, "pid": 0, "tid": 7, "ts": 1776649790929035, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929025, "dur": 6, + "args": { + "External id": 5162, + "cbid": 211, "correlation": 40819 + } + }, + { + "ph": "s", "id": 40819, "pid": 435, "tid": 435, "ts": 1776649790929025, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929053, "dur": 0, + "args": { + "External id": 5164, + "cbid": 210, "correlation": 40846 + } + }, + { + "ph": "f", "id": 40846, "pid": 435, "tid": 435, "ts": 1776649790929053, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929065, "dur": 23, + "args": { + "External id": 5164, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40847, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40847, "pid": 0, "tid": 7, "ts": 1776649790929065, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929055, "dur": 5, + "args": { + "External id": 5164, + "cbid": 211, "correlation": 40847 + } + }, + { + "ph": "s", "id": 40847, "pid": 435, "tid": 435, "ts": 1776649790929055, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929083, "dur": 0, + "args": { + "External id": 5166, + "cbid": 210, "correlation": 40874 + } + }, + { + "ph": "f", "id": 40874, "pid": 435, "tid": 435, "ts": 1776649790929083, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929095, "dur": 23, + "args": { + "External id": 5166, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40875, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40875, "pid": 0, "tid": 7, "ts": 1776649790929095, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929084, "dur": 6, + "args": { + "External id": 5166, + "cbid": 211, "correlation": 40875 + } + }, + { + "ph": "s", "id": 40875, "pid": 435, "tid": 435, "ts": 1776649790929084, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929113, "dur": 0, + "args": { + "External id": 5168, + "cbid": 210, "correlation": 40902 + } + }, + { + "ph": "f", "id": 40902, "pid": 435, "tid": 435, "ts": 1776649790929113, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929124, "dur": 23, + "args": { + "External id": 5168, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40903, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40903, "pid": 0, "tid": 7, "ts": 1776649790929124, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929114, "dur": 5, + "args": { + "External id": 5168, + "cbid": 211, "correlation": 40903 + } + }, + { + "ph": "s", "id": 40903, "pid": 435, "tid": 435, "ts": 1776649790929114, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929142, "dur": 0, + "args": { + "External id": 5170, + "cbid": 210, "correlation": 40930 + } + }, + { + "ph": "f", "id": 40930, "pid": 435, "tid": 435, "ts": 1776649790929142, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929154, "dur": 23, + "args": { + "External id": 5170, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40931, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40931, "pid": 0, "tid": 7, "ts": 1776649790929154, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929144, "dur": 5, + "args": { + "External id": 5170, + "cbid": 211, "correlation": 40931 + } + }, + { + "ph": "s", "id": 40931, "pid": 435, "tid": 435, "ts": 1776649790929144, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929172, "dur": 0, + "args": { + "External id": 5172, + "cbid": 210, "correlation": 40958 + } + }, + { + "ph": "f", "id": 40958, "pid": 435, "tid": 435, "ts": 1776649790929172, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929184, "dur": 23, + "args": { + "External id": 5172, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40959, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40959, "pid": 0, "tid": 7, "ts": 1776649790929184, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929174, "dur": 5, + "args": { + "External id": 5172, + "cbid": 211, "correlation": 40959 + } + }, + { + "ph": "s", "id": 40959, "pid": 435, "tid": 435, "ts": 1776649790929174, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929202, "dur": 0, + "args": { + "External id": 5174, + "cbid": 210, "correlation": 40986 + } + }, + { + "ph": "f", "id": 40986, "pid": 435, "tid": 435, "ts": 1776649790929202, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929213, "dur": 23, + "args": { + "External id": 5174, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 40987, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 40987, "pid": 0, "tid": 7, "ts": 1776649790929213, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929203, "dur": 6, + "args": { + "External id": 5174, + "cbid": 211, "correlation": 40987 + } + }, + { + "ph": "s", "id": 40987, "pid": 435, "tid": 435, "ts": 1776649790929203, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929232, "dur": 0, + "args": { + "External id": 5176, + "cbid": 210, "correlation": 41014 + } + }, + { + "ph": "f", "id": 41014, "pid": 435, "tid": 435, "ts": 1776649790929232, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929243, "dur": 23, + "args": { + "External id": 5176, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41015, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41015, "pid": 0, "tid": 7, "ts": 1776649790929243, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929233, "dur": 5, + "args": { + "External id": 5176, + "cbid": 211, "correlation": 41015 + } + }, + { + "ph": "s", "id": 41015, "pid": 435, "tid": 435, "ts": 1776649790929233, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929261, "dur": 0, + "args": { + "External id": 5178, + "cbid": 210, "correlation": 41042 + } + }, + { + "ph": "f", "id": 41042, "pid": 435, "tid": 435, "ts": 1776649790929261, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929273, "dur": 23, + "args": { + "External id": 5178, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41043, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41043, "pid": 0, "tid": 7, "ts": 1776649790929273, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929262, "dur": 5, + "args": { + "External id": 5178, + "cbid": 211, "correlation": 41043 + } + }, + { + "ph": "s", "id": 41043, "pid": 435, "tid": 435, "ts": 1776649790929262, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929291, "dur": 0, + "args": { + "External id": 5180, + "cbid": 210, "correlation": 41070 + } + }, + { + "ph": "f", "id": 41070, "pid": 435, "tid": 435, "ts": 1776649790929291, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929302, "dur": 23, + "args": { + "External id": 5180, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41071, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41071, "pid": 0, "tid": 7, "ts": 1776649790929302, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929292, "dur": 5, + "args": { + "External id": 5180, + "cbid": 211, "correlation": 41071 + } + }, + { + "ph": "s", "id": 41071, "pid": 435, "tid": 435, "ts": 1776649790929292, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929327, "dur": 0, + "args": { + "External id": 5182, + "cbid": 210, "correlation": 41098 + } + }, + { + "ph": "f", "id": 41098, "pid": 435, "tid": 435, "ts": 1776649790929327, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929337, "dur": 23, + "args": { + "External id": 5182, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41099, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41099, "pid": 0, "tid": 7, "ts": 1776649790929337, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929328, "dur": 6, + "args": { + "External id": 5182, + "cbid": 211, "correlation": 41099 + } + }, + { + "ph": "s", "id": 41099, "pid": 435, "tid": 435, "ts": 1776649790929328, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929357, "dur": 0, + "args": { + "External id": 5184, + "cbid": 210, "correlation": 41126 + } + }, + { + "ph": "f", "id": 41126, "pid": 435, "tid": 435, "ts": 1776649790929357, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929368, "dur": 23, + "args": { + "External id": 5184, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41127, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41127, "pid": 0, "tid": 7, "ts": 1776649790929368, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929358, "dur": 6, + "args": { + "External id": 5184, + "cbid": 211, "correlation": 41127 + } + }, + { + "ph": "s", "id": 41127, "pid": 435, "tid": 435, "ts": 1776649790929358, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929387, "dur": 0, + "args": { + "External id": 5186, + "cbid": 210, "correlation": 41154 + } + }, + { + "ph": "f", "id": 41154, "pid": 435, "tid": 435, "ts": 1776649790929387, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929398, "dur": 23, + "args": { + "External id": 5186, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41155, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41155, "pid": 0, "tid": 7, "ts": 1776649790929398, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929388, "dur": 5, + "args": { + "External id": 5186, + "cbid": 211, "correlation": 41155 + } + }, + { + "ph": "s", "id": 41155, "pid": 435, "tid": 435, "ts": 1776649790929388, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929416, "dur": 0, + "args": { + "External id": 5188, + "cbid": 210, "correlation": 41182 + } + }, + { + "ph": "f", "id": 41182, "pid": 435, "tid": 435, "ts": 1776649790929416, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929428, "dur": 22, + "args": { + "External id": 5188, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41183, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41183, "pid": 0, "tid": 7, "ts": 1776649790929428, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929418, "dur": 5, + "args": { + "External id": 5188, + "cbid": 211, "correlation": 41183 + } + }, + { + "ph": "s", "id": 41183, "pid": 435, "tid": 435, "ts": 1776649790929418, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929446, "dur": 0, + "args": { + "External id": 5190, + "cbid": 210, "correlation": 41210 + } + }, + { + "ph": "f", "id": 41210, "pid": 435, "tid": 435, "ts": 1776649790929446, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929457, "dur": 23, + "args": { + "External id": 5190, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41211, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41211, "pid": 0, "tid": 7, "ts": 1776649790929457, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929447, "dur": 5, + "args": { + "External id": 5190, + "cbid": 211, "correlation": 41211 + } + }, + { + "ph": "s", "id": 41211, "pid": 435, "tid": 435, "ts": 1776649790929447, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929475, "dur": 0, + "args": { + "External id": 5192, + "cbid": 210, "correlation": 41238 + } + }, + { + "ph": "f", "id": 41238, "pid": 435, "tid": 435, "ts": 1776649790929475, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929487, "dur": 23, + "args": { + "External id": 5192, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41239, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41239, "pid": 0, "tid": 7, "ts": 1776649790929487, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929477, "dur": 6, + "args": { + "External id": 5192, + "cbid": 211, "correlation": 41239 + } + }, + { + "ph": "s", "id": 41239, "pid": 435, "tid": 435, "ts": 1776649790929477, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929505, "dur": 0, + "args": { + "External id": 5194, + "cbid": 210, "correlation": 41266 + } + }, + { + "ph": "f", "id": 41266, "pid": 435, "tid": 435, "ts": 1776649790929505, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929517, "dur": 23, + "args": { + "External id": 5194, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41267, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41267, "pid": 0, "tid": 7, "ts": 1776649790929517, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929507, "dur": 5, + "args": { + "External id": 5194, + "cbid": 211, "correlation": 41267 + } + }, + { + "ph": "s", "id": 41267, "pid": 435, "tid": 435, "ts": 1776649790929507, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929535, "dur": 0, + "args": { + "External id": 5196, + "cbid": 210, "correlation": 41294 + } + }, + { + "ph": "f", "id": 41294, "pid": 435, "tid": 435, "ts": 1776649790929535, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929546, "dur": 23, + "args": { + "External id": 5196, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41295, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41295, "pid": 0, "tid": 7, "ts": 1776649790929546, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929536, "dur": 5, + "args": { + "External id": 5196, + "cbid": 211, "correlation": 41295 + } + }, + { + "ph": "s", "id": 41295, "pid": 435, "tid": 435, "ts": 1776649790929536, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929564, "dur": 0, + "args": { + "External id": 5198, + "cbid": 210, "correlation": 41322 + } + }, + { + "ph": "f", "id": 41322, "pid": 435, "tid": 435, "ts": 1776649790929564, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929576, "dur": 23, + "args": { + "External id": 5198, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41323, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41323, "pid": 0, "tid": 7, "ts": 1776649790929576, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929566, "dur": 5, + "args": { + "External id": 5198, + "cbid": 211, "correlation": 41323 + } + }, + { + "ph": "s", "id": 41323, "pid": 435, "tid": 435, "ts": 1776649790929566, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929594, "dur": 0, + "args": { + "External id": 5200, + "cbid": 210, "correlation": 41350 + } + }, + { + "ph": "f", "id": 41350, "pid": 435, "tid": 435, "ts": 1776649790929594, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929606, "dur": 23, + "args": { + "External id": 5200, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41351, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41351, "pid": 0, "tid": 7, "ts": 1776649790929606, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929595, "dur": 5, + "args": { + "External id": 5200, + "cbid": 211, "correlation": 41351 + } + }, + { + "ph": "s", "id": 41351, "pid": 435, "tid": 435, "ts": 1776649790929595, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790929629, "dur": 4, + "args": { + "External id": 41362, + "cbid": 135, "correlation": 41362 + } + }, + { + "ph": "f", "id": 41362, "pid": 435, "tid": 435, "ts": 1776649790929629, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790929645, "dur": 5, + "args": { + "External id": 41367, + "cbid": 165, "correlation": 41367 + } + }, + { + "ph": "s", "id": 41367, "pid": 435, "tid": 435, "ts": 1776649790929645, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649790929653, "dur": 4, + "args": { + "External id": 41370, + "cbid": 139, "correlation": 41370 + } + }, + { + "ph": "f", "id": 41370, "pid": 435, "tid": 435, "ts": 1776649790929653, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929691, "dur": 1, + "args": { + "External id": 5202, + "cbid": 210, "correlation": 41404 + } + }, + { + "ph": "f", "id": 41404, "pid": 435, "tid": 435, "ts": 1776649790929691, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929703, "dur": 23, + "args": { + "External id": 5202, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41405, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41405, "pid": 0, "tid": 7, "ts": 1776649790929703, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929692, "dur": 8, + "args": { + "External id": 5202, + "cbid": 211, "correlation": 41405 + } + }, + { + "ph": "s", "id": 41405, "pid": 435, "tid": 435, "ts": 1776649790929692, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929726, "dur": 0, + "args": { + "External id": 5204, + "cbid": 210, "correlation": 41432 + } + }, + { + "ph": "f", "id": 41432, "pid": 435, "tid": 435, "ts": 1776649790929726, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929736, "dur": 23, + "args": { + "External id": 5204, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41433, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41433, "pid": 0, "tid": 7, "ts": 1776649790929736, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929727, "dur": 6, + "args": { + "External id": 5204, + "cbid": 211, "correlation": 41433 + } + }, + { + "ph": "s", "id": 41433, "pid": 435, "tid": 435, "ts": 1776649790929727, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929757, "dur": 0, + "args": { + "External id": 5206, + "cbid": 210, "correlation": 41460 + } + }, + { + "ph": "f", "id": 41460, "pid": 435, "tid": 435, "ts": 1776649790929757, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929767, "dur": 23, + "args": { + "External id": 5206, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41461, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41461, "pid": 0, "tid": 7, "ts": 1776649790929767, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929758, "dur": 6, + "args": { + "External id": 5206, + "cbid": 211, "correlation": 41461 + } + }, + { + "ph": "s", "id": 41461, "pid": 435, "tid": 435, "ts": 1776649790929758, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929787, "dur": 0, + "args": { + "External id": 5208, + "cbid": 210, "correlation": 41488 + } + }, + { + "ph": "f", "id": 41488, "pid": 435, "tid": 435, "ts": 1776649790929787, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929797, "dur": 23, + "args": { + "External id": 5208, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41489, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41489, "pid": 0, "tid": 7, "ts": 1776649790929797, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929788, "dur": 6, + "args": { + "External id": 5208, + "cbid": 211, "correlation": 41489 + } + }, + { + "ph": "s", "id": 41489, "pid": 435, "tid": 435, "ts": 1776649790929788, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929817, "dur": 0, + "args": { + "External id": 5210, + "cbid": 210, "correlation": 41516 + } + }, + { + "ph": "f", "id": 41516, "pid": 435, "tid": 435, "ts": 1776649790929817, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929827, "dur": 23, + "args": { + "External id": 5210, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41517, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41517, "pid": 0, "tid": 7, "ts": 1776649790929827, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929818, "dur": 6, + "args": { + "External id": 5210, + "cbid": 211, "correlation": 41517 + } + }, + { + "ph": "s", "id": 41517, "pid": 435, "tid": 435, "ts": 1776649790929818, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929847, "dur": 0, + "args": { + "External id": 5212, + "cbid": 210, "correlation": 41544 + } + }, + { + "ph": "f", "id": 41544, "pid": 435, "tid": 435, "ts": 1776649790929847, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929857, "dur": 23, + "args": { + "External id": 5212, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41545, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41545, "pid": 0, "tid": 7, "ts": 1776649790929857, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929848, "dur": 6, + "args": { + "External id": 5212, + "cbid": 211, "correlation": 41545 + } + }, + { + "ph": "s", "id": 41545, "pid": 435, "tid": 435, "ts": 1776649790929848, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929877, "dur": 0, + "args": { + "External id": 5214, + "cbid": 210, "correlation": 41572 + } + }, + { + "ph": "f", "id": 41572, "pid": 435, "tid": 435, "ts": 1776649790929877, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929887, "dur": 23, + "args": { + "External id": 5214, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41573, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41573, "pid": 0, "tid": 7, "ts": 1776649790929887, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929878, "dur": 5, + "args": { + "External id": 5214, + "cbid": 211, "correlation": 41573 + } + }, + { + "ph": "s", "id": 41573, "pid": 435, "tid": 435, "ts": 1776649790929878, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929907, "dur": 0, + "args": { + "External id": 5216, + "cbid": 210, "correlation": 41600 + } + }, + { + "ph": "f", "id": 41600, "pid": 435, "tid": 435, "ts": 1776649790929907, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929917, "dur": 23, + "args": { + "External id": 5216, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41601, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41601, "pid": 0, "tid": 7, "ts": 1776649790929917, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929908, "dur": 6, + "args": { + "External id": 5216, + "cbid": 211, "correlation": 41601 + } + }, + { + "ph": "s", "id": 41601, "pid": 435, "tid": 435, "ts": 1776649790929908, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929937, "dur": 0, + "args": { + "External id": 5218, + "cbid": 210, "correlation": 41628 + } + }, + { + "ph": "f", "id": 41628, "pid": 435, "tid": 435, "ts": 1776649790929937, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929947, "dur": 23, + "args": { + "External id": 5218, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41629, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41629, "pid": 0, "tid": 7, "ts": 1776649790929947, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929938, "dur": 6, + "args": { + "External id": 5218, + "cbid": 211, "correlation": 41629 + } + }, + { + "ph": "s", "id": 41629, "pid": 435, "tid": 435, "ts": 1776649790929938, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790929967, "dur": 0, + "args": { + "External id": 5220, + "cbid": 210, "correlation": 41656 + } + }, + { + "ph": "f", "id": 41656, "pid": 435, "tid": 435, "ts": 1776649790929967, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790929977, "dur": 23, + "args": { + "External id": 5220, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41657, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41657, "pid": 0, "tid": 7, "ts": 1776649790929977, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790929968, "dur": 5, + "args": { + "External id": 5220, + "cbid": 211, "correlation": 41657 + } + }, + { + "ph": "s", "id": 41657, "pid": 435, "tid": 435, "ts": 1776649790929968, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790929993, "dur": 11, + "args": { + "External id": 41663, + "cbid": 165, "correlation": 41663 + } + }, + { + "ph": "s", "id": 41663, "pid": 435, "tid": 435, "ts": 1776649790929993, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790930025, "dur": 3, + "args": { + "External id": 41672, + "cbid": 135, "correlation": 41672 + } + }, + { + "ph": "f", "id": 41672, "pid": 435, "tid": 435, "ts": 1776649790930025, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930055, "dur": 1, + "args": { + "External id": 5222, + "cbid": 210, "correlation": 41698 + } + }, + { + "ph": "f", "id": 41698, "pid": 435, "tid": 435, "ts": 1776649790930055, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930067, "dur": 23, + "args": { + "External id": 5222, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41699, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41699, "pid": 0, "tid": 7, "ts": 1776649790930067, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930057, "dur": 7, + "args": { + "External id": 5222, + "cbid": 211, "correlation": 41699 + } + }, + { + "ph": "s", "id": 41699, "pid": 435, "tid": 435, "ts": 1776649790930057, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930090, "dur": 0, + "args": { + "External id": 5224, + "cbid": 210, "correlation": 41726 + } + }, + { + "ph": "f", "id": 41726, "pid": 435, "tid": 435, "ts": 1776649790930090, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930100, "dur": 23, + "args": { + "External id": 5224, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41727, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41727, "pid": 0, "tid": 7, "ts": 1776649790930100, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930091, "dur": 6, + "args": { + "External id": 5224, + "cbid": 211, "correlation": 41727 + } + }, + { + "ph": "s", "id": 41727, "pid": 435, "tid": 435, "ts": 1776649790930091, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930120, "dur": 0, + "args": { + "External id": 5226, + "cbid": 210, "correlation": 41754 + } + }, + { + "ph": "f", "id": 41754, "pid": 435, "tid": 435, "ts": 1776649790930120, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930131, "dur": 23, + "args": { + "External id": 5226, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41755, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41755, "pid": 0, "tid": 7, "ts": 1776649790930131, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930122, "dur": 6, + "args": { + "External id": 5226, + "cbid": 211, "correlation": 41755 + } + }, + { + "ph": "s", "id": 41755, "pid": 435, "tid": 435, "ts": 1776649790930122, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930151, "dur": 0, + "args": { + "External id": 5228, + "cbid": 210, "correlation": 41782 + } + }, + { + "ph": "f", "id": 41782, "pid": 435, "tid": 435, "ts": 1776649790930151, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930161, "dur": 23, + "args": { + "External id": 5228, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41783, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41783, "pid": 0, "tid": 7, "ts": 1776649790930161, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930152, "dur": 6, + "args": { + "External id": 5228, + "cbid": 211, "correlation": 41783 + } + }, + { + "ph": "s", "id": 41783, "pid": 435, "tid": 435, "ts": 1776649790930152, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930181, "dur": 0, + "args": { + "External id": 5230, + "cbid": 210, "correlation": 41810 + } + }, + { + "ph": "f", "id": 41810, "pid": 435, "tid": 435, "ts": 1776649790930181, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930191, "dur": 23, + "args": { + "External id": 5230, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41811, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41811, "pid": 0, "tid": 7, "ts": 1776649790930191, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930182, "dur": 6, + "args": { + "External id": 5230, + "cbid": 211, "correlation": 41811 + } + }, + { + "ph": "s", "id": 41811, "pid": 435, "tid": 435, "ts": 1776649790930182, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930211, "dur": 0, + "args": { + "External id": 5232, + "cbid": 210, "correlation": 41838 + } + }, + { + "ph": "f", "id": 41838, "pid": 435, "tid": 435, "ts": 1776649790930211, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930222, "dur": 23, + "args": { + "External id": 5232, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41839, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41839, "pid": 0, "tid": 7, "ts": 1776649790930222, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930212, "dur": 6, + "args": { + "External id": 5232, + "cbid": 211, "correlation": 41839 + } + }, + { + "ph": "s", "id": 41839, "pid": 435, "tid": 435, "ts": 1776649790930212, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930241, "dur": 0, + "args": { + "External id": 5234, + "cbid": 210, "correlation": 41866 + } + }, + { + "ph": "f", "id": 41866, "pid": 435, "tid": 435, "ts": 1776649790930241, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930253, "dur": 23, + "args": { + "External id": 5234, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41867, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41867, "pid": 0, "tid": 7, "ts": 1776649790930253, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930242, "dur": 6, + "args": { + "External id": 5234, + "cbid": 211, "correlation": 41867 + } + }, + { + "ph": "s", "id": 41867, "pid": 435, "tid": 435, "ts": 1776649790930242, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930271, "dur": 0, + "args": { + "External id": 5236, + "cbid": 210, "correlation": 41894 + } + }, + { + "ph": "f", "id": 41894, "pid": 435, "tid": 435, "ts": 1776649790930271, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930282, "dur": 23, + "args": { + "External id": 5236, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41895, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41895, "pid": 0, "tid": 7, "ts": 1776649790930282, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930272, "dur": 6, + "args": { + "External id": 5236, + "cbid": 211, "correlation": 41895 + } + }, + { + "ph": "s", "id": 41895, "pid": 435, "tid": 435, "ts": 1776649790930272, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930301, "dur": 0, + "args": { + "External id": 5238, + "cbid": 210, "correlation": 41922 + } + }, + { + "ph": "f", "id": 41922, "pid": 435, "tid": 435, "ts": 1776649790930301, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930312, "dur": 23, + "args": { + "External id": 5238, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41923, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41923, "pid": 0, "tid": 7, "ts": 1776649790930312, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930302, "dur": 6, + "args": { + "External id": 5238, + "cbid": 211, "correlation": 41923 + } + }, + { + "ph": "s", "id": 41923, "pid": 435, "tid": 435, "ts": 1776649790930302, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930331, "dur": 0, + "args": { + "External id": 5240, + "cbid": 210, "correlation": 41950 + } + }, + { + "ph": "f", "id": 41950, "pid": 435, "tid": 435, "ts": 1776649790930331, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930342, "dur": 23, + "args": { + "External id": 5240, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41951, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41951, "pid": 0, "tid": 7, "ts": 1776649790930342, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930332, "dur": 5, + "args": { + "External id": 5240, + "cbid": 211, "correlation": 41951 + } + }, + { + "ph": "s", "id": 41951, "pid": 435, "tid": 435, "ts": 1776649790930332, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930361, "dur": 0, + "args": { + "External id": 5242, + "cbid": 210, "correlation": 41978 + } + }, + { + "ph": "f", "id": 41978, "pid": 435, "tid": 435, "ts": 1776649790930361, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930372, "dur": 23, + "args": { + "External id": 5242, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 41979, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 41979, "pid": 0, "tid": 7, "ts": 1776649790930372, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930362, "dur": 5, + "args": { + "External id": 5242, + "cbid": 211, "correlation": 41979 + } + }, + { + "ph": "s", "id": 41979, "pid": 435, "tid": 435, "ts": 1776649790930362, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930390, "dur": 0, + "args": { + "External id": 5244, + "cbid": 210, "correlation": 42006 + } + }, + { + "ph": "f", "id": 42006, "pid": 435, "tid": 435, "ts": 1776649790930390, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930401, "dur": 23, + "args": { + "External id": 5244, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42007, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42007, "pid": 0, "tid": 7, "ts": 1776649790930401, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930391, "dur": 5, + "args": { + "External id": 5244, + "cbid": 211, "correlation": 42007 + } + }, + { + "ph": "s", "id": 42007, "pid": 435, "tid": 435, "ts": 1776649790930391, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930420, "dur": 0, + "args": { + "External id": 5246, + "cbid": 210, "correlation": 42034 + } + }, + { + "ph": "f", "id": 42034, "pid": 435, "tid": 435, "ts": 1776649790930420, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930431, "dur": 23, + "args": { + "External id": 5246, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42035, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42035, "pid": 0, "tid": 7, "ts": 1776649790930431, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930421, "dur": 6, + "args": { + "External id": 5246, + "cbid": 211, "correlation": 42035 + } + }, + { + "ph": "s", "id": 42035, "pid": 435, "tid": 435, "ts": 1776649790930421, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930449, "dur": 0, + "args": { + "External id": 5248, + "cbid": 210, "correlation": 42062 + } + }, + { + "ph": "f", "id": 42062, "pid": 435, "tid": 435, "ts": 1776649790930449, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930461, "dur": 23, + "args": { + "External id": 5248, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42063, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42063, "pid": 0, "tid": 7, "ts": 1776649790930461, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930451, "dur": 6, + "args": { + "External id": 5248, + "cbid": 211, "correlation": 42063 + } + }, + { + "ph": "s", "id": 42063, "pid": 435, "tid": 435, "ts": 1776649790930451, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930479, "dur": 0, + "args": { + "External id": 5250, + "cbid": 210, "correlation": 42090 + } + }, + { + "ph": "f", "id": 42090, "pid": 435, "tid": 435, "ts": 1776649790930479, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930491, "dur": 23, + "args": { + "External id": 5250, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42091, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42091, "pid": 0, "tid": 7, "ts": 1776649790930491, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930481, "dur": 5, + "args": { + "External id": 5250, + "cbid": 211, "correlation": 42091 + } + }, + { + "ph": "s", "id": 42091, "pid": 435, "tid": 435, "ts": 1776649790930481, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930509, "dur": 0, + "args": { + "External id": 5252, + "cbid": 210, "correlation": 42118 + } + }, + { + "ph": "f", "id": 42118, "pid": 435, "tid": 435, "ts": 1776649790930509, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930521, "dur": 23, + "args": { + "External id": 5252, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42119, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42119, "pid": 0, "tid": 7, "ts": 1776649790930521, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930510, "dur": 5, + "args": { + "External id": 5252, + "cbid": 211, "correlation": 42119 + } + }, + { + "ph": "s", "id": 42119, "pid": 435, "tid": 435, "ts": 1776649790930510, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930539, "dur": 0, + "args": { + "External id": 5254, + "cbid": 210, "correlation": 42146 + } + }, + { + "ph": "f", "id": 42146, "pid": 435, "tid": 435, "ts": 1776649790930539, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930550, "dur": 23, + "args": { + "External id": 5254, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42147, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42147, "pid": 0, "tid": 7, "ts": 1776649790930550, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930540, "dur": 6, + "args": { + "External id": 5254, + "cbid": 211, "correlation": 42147 + } + }, + { + "ph": "s", "id": 42147, "pid": 435, "tid": 435, "ts": 1776649790930540, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930568, "dur": 0, + "args": { + "External id": 5256, + "cbid": 210, "correlation": 42174 + } + }, + { + "ph": "f", "id": 42174, "pid": 435, "tid": 435, "ts": 1776649790930568, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930580, "dur": 23, + "args": { + "External id": 5256, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42175, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42175, "pid": 0, "tid": 7, "ts": 1776649790930580, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930570, "dur": 6, + "args": { + "External id": 5256, + "cbid": 211, "correlation": 42175 + } + }, + { + "ph": "s", "id": 42175, "pid": 435, "tid": 435, "ts": 1776649790930570, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930598, "dur": 0, + "args": { + "External id": 5258, + "cbid": 210, "correlation": 42202 + } + }, + { + "ph": "f", "id": 42202, "pid": 435, "tid": 435, "ts": 1776649790930598, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930610, "dur": 23, + "args": { + "External id": 5258, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42203, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42203, "pid": 0, "tid": 7, "ts": 1776649790930610, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930600, "dur": 5, + "args": { + "External id": 5258, + "cbid": 211, "correlation": 42203 + } + }, + { + "ph": "s", "id": 42203, "pid": 435, "tid": 435, "ts": 1776649790930600, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930628, "dur": 0, + "args": { + "External id": 5260, + "cbid": 210, "correlation": 42230 + } + }, + { + "ph": "f", "id": 42230, "pid": 435, "tid": 435, "ts": 1776649790930628, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930639, "dur": 23, + "args": { + "External id": 5260, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42231, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42231, "pid": 0, "tid": 7, "ts": 1776649790930639, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930629, "dur": 5, + "args": { + "External id": 5260, + "cbid": 211, "correlation": 42231 + } + }, + { + "ph": "s", "id": 42231, "pid": 435, "tid": 435, "ts": 1776649790930629, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930658, "dur": 0, + "args": { + "External id": 5262, + "cbid": 210, "correlation": 42258 + } + }, + { + "ph": "f", "id": 42258, "pid": 435, "tid": 435, "ts": 1776649790930658, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930669, "dur": 23, + "args": { + "External id": 5262, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42259, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42259, "pid": 0, "tid": 7, "ts": 1776649790930669, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930659, "dur": 5, + "args": { + "External id": 5262, + "cbid": 211, "correlation": 42259 + } + }, + { + "ph": "s", "id": 42259, "pid": 435, "tid": 435, "ts": 1776649790930659, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930688, "dur": 0, + "args": { + "External id": 5264, + "cbid": 210, "correlation": 42286 + } + }, + { + "ph": "f", "id": 42286, "pid": 435, "tid": 435, "ts": 1776649790930688, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930699, "dur": 23, + "args": { + "External id": 5264, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42287, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42287, "pid": 0, "tid": 7, "ts": 1776649790930699, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930689, "dur": 5, + "args": { + "External id": 5264, + "cbid": 211, "correlation": 42287 + } + }, + { + "ph": "s", "id": 42287, "pid": 435, "tid": 435, "ts": 1776649790930689, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930721, "dur": 0, + "args": { + "External id": 5266, + "cbid": 210, "correlation": 42314 + } + }, + { + "ph": "f", "id": 42314, "pid": 435, "tid": 435, "ts": 1776649790930721, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930731, "dur": 23, + "args": { + "External id": 5266, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42315, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42315, "pid": 0, "tid": 7, "ts": 1776649790930731, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930722, "dur": 5, + "args": { + "External id": 5266, + "cbid": 211, "correlation": 42315 + } + }, + { + "ph": "s", "id": 42315, "pid": 435, "tid": 435, "ts": 1776649790930722, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930751, "dur": 0, + "args": { + "External id": 5268, + "cbid": 210, "correlation": 42342 + } + }, + { + "ph": "f", "id": 42342, "pid": 435, "tid": 435, "ts": 1776649790930751, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930761, "dur": 23, + "args": { + "External id": 5268, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42343, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42343, "pid": 0, "tid": 7, "ts": 1776649790930761, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930752, "dur": 6, + "args": { + "External id": 5268, + "cbid": 211, "correlation": 42343 + } + }, + { + "ph": "s", "id": 42343, "pid": 435, "tid": 435, "ts": 1776649790930752, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930781, "dur": 0, + "args": { + "External id": 5270, + "cbid": 210, "correlation": 42370 + } + }, + { + "ph": "f", "id": 42370, "pid": 435, "tid": 435, "ts": 1776649790930781, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930791, "dur": 23, + "args": { + "External id": 5270, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42371, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42371, "pid": 0, "tid": 7, "ts": 1776649790930791, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930782, "dur": 6, + "args": { + "External id": 5270, + "cbid": 211, "correlation": 42371 + } + }, + { + "ph": "s", "id": 42371, "pid": 435, "tid": 435, "ts": 1776649790930782, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930811, "dur": 0, + "args": { + "External id": 5272, + "cbid": 210, "correlation": 42398 + } + }, + { + "ph": "f", "id": 42398, "pid": 435, "tid": 435, "ts": 1776649790930811, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930821, "dur": 23, + "args": { + "External id": 5272, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42399, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42399, "pid": 0, "tid": 7, "ts": 1776649790930821, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930812, "dur": 6, + "args": { + "External id": 5272, + "cbid": 211, "correlation": 42399 + } + }, + { + "ph": "s", "id": 42399, "pid": 435, "tid": 435, "ts": 1776649790930812, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930841, "dur": 0, + "args": { + "External id": 5274, + "cbid": 210, "correlation": 42426 + } + }, + { + "ph": "f", "id": 42426, "pid": 435, "tid": 435, "ts": 1776649790930841, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930852, "dur": 23, + "args": { + "External id": 5274, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42427, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42427, "pid": 0, "tid": 7, "ts": 1776649790930852, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930843, "dur": 6, + "args": { + "External id": 5274, + "cbid": 211, "correlation": 42427 + } + }, + { + "ph": "s", "id": 42427, "pid": 435, "tid": 435, "ts": 1776649790930843, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930871, "dur": 0, + "args": { + "External id": 5276, + "cbid": 210, "correlation": 42454 + } + }, + { + "ph": "f", "id": 42454, "pid": 435, "tid": 435, "ts": 1776649790930871, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930881, "dur": 23, + "args": { + "External id": 5276, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42455, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42455, "pid": 0, "tid": 7, "ts": 1776649790930881, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930872, "dur": 6, + "args": { + "External id": 5276, + "cbid": 211, "correlation": 42455 + } + }, + { + "ph": "s", "id": 42455, "pid": 435, "tid": 435, "ts": 1776649790930872, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930901, "dur": 0, + "args": { + "External id": 5278, + "cbid": 210, "correlation": 42482 + } + }, + { + "ph": "f", "id": 42482, "pid": 435, "tid": 435, "ts": 1776649790930901, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930912, "dur": 23, + "args": { + "External id": 5278, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42483, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42483, "pid": 0, "tid": 7, "ts": 1776649790930912, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930903, "dur": 6, + "args": { + "External id": 5278, + "cbid": 211, "correlation": 42483 + } + }, + { + "ph": "s", "id": 42483, "pid": 435, "tid": 435, "ts": 1776649790930903, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930931, "dur": 0, + "args": { + "External id": 5280, + "cbid": 210, "correlation": 42510 + } + }, + { + "ph": "f", "id": 42510, "pid": 435, "tid": 435, "ts": 1776649790930931, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930941, "dur": 23, + "args": { + "External id": 5280, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42511, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42511, "pid": 0, "tid": 7, "ts": 1776649790930941, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930933, "dur": 6, + "args": { + "External id": 5280, + "cbid": 211, "correlation": 42511 + } + }, + { + "ph": "s", "id": 42511, "pid": 435, "tid": 435, "ts": 1776649790930933, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930961, "dur": 0, + "args": { + "External id": 5282, + "cbid": 210, "correlation": 42538 + } + }, + { + "ph": "f", "id": 42538, "pid": 435, "tid": 435, "ts": 1776649790930961, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790930971, "dur": 23, + "args": { + "External id": 5282, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42539, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42539, "pid": 0, "tid": 7, "ts": 1776649790930971, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930962, "dur": 5, + "args": { + "External id": 5282, + "cbid": 211, "correlation": 42539 + } + }, + { + "ph": "s", "id": 42539, "pid": 435, "tid": 435, "ts": 1776649790930962, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790930991, "dur": 0, + "args": { + "External id": 5284, + "cbid": 210, "correlation": 42566 + } + }, + { + "ph": "f", "id": 42566, "pid": 435, "tid": 435, "ts": 1776649790930991, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931002, "dur": 23, + "args": { + "External id": 5284, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42567, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42567, "pid": 0, "tid": 7, "ts": 1776649790931002, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790930992, "dur": 5, + "args": { + "External id": 5284, + "cbid": 211, "correlation": 42567 + } + }, + { + "ph": "s", "id": 42567, "pid": 435, "tid": 435, "ts": 1776649790930992, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931021, "dur": 0, + "args": { + "External id": 5286, + "cbid": 210, "correlation": 42594 + } + }, + { + "ph": "f", "id": 42594, "pid": 435, "tid": 435, "ts": 1776649790931021, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931033, "dur": 23, + "args": { + "External id": 5286, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42595, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42595, "pid": 0, "tid": 7, "ts": 1776649790931033, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931023, "dur": 6, + "args": { + "External id": 5286, + "cbid": 211, "correlation": 42595 + } + }, + { + "ph": "s", "id": 42595, "pid": 435, "tid": 435, "ts": 1776649790931023, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931052, "dur": 0, + "args": { + "External id": 5288, + "cbid": 210, "correlation": 42622 + } + }, + { + "ph": "f", "id": 42622, "pid": 435, "tid": 435, "ts": 1776649790931052, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931063, "dur": 23, + "args": { + "External id": 5288, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42623, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42623, "pid": 0, "tid": 7, "ts": 1776649790931063, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931053, "dur": 5, + "args": { + "External id": 5288, + "cbid": 211, "correlation": 42623 + } + }, + { + "ph": "s", "id": 42623, "pid": 435, "tid": 435, "ts": 1776649790931053, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931081, "dur": 0, + "args": { + "External id": 5290, + "cbid": 210, "correlation": 42650 + } + }, + { + "ph": "f", "id": 42650, "pid": 435, "tid": 435, "ts": 1776649790931081, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931093, "dur": 23, + "args": { + "External id": 5290, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42651, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42651, "pid": 0, "tid": 7, "ts": 1776649790931093, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931083, "dur": 6, + "args": { + "External id": 5290, + "cbid": 211, "correlation": 42651 + } + }, + { + "ph": "s", "id": 42651, "pid": 435, "tid": 435, "ts": 1776649790931083, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931111, "dur": 0, + "args": { + "External id": 5292, + "cbid": 210, "correlation": 42678 + } + }, + { + "ph": "f", "id": 42678, "pid": 435, "tid": 435, "ts": 1776649790931111, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931122, "dur": 23, + "args": { + "External id": 5292, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42679, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42679, "pid": 0, "tid": 7, "ts": 1776649790931122, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931113, "dur": 5, + "args": { + "External id": 5292, + "cbid": 211, "correlation": 42679 + } + }, + { + "ph": "s", "id": 42679, "pid": 435, "tid": 435, "ts": 1776649790931113, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931142, "dur": 0, + "args": { + "External id": 5294, + "cbid": 210, "correlation": 42706 + } + }, + { + "ph": "f", "id": 42706, "pid": 435, "tid": 435, "ts": 1776649790931142, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931154, "dur": 23, + "args": { + "External id": 5294, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42707, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42707, "pid": 0, "tid": 7, "ts": 1776649790931154, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931143, "dur": 6, + "args": { + "External id": 5294, + "cbid": 211, "correlation": 42707 + } + }, + { + "ph": "s", "id": 42707, "pid": 435, "tid": 435, "ts": 1776649790931143, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931172, "dur": 0, + "args": { + "External id": 5296, + "cbid": 210, "correlation": 42734 + } + }, + { + "ph": "f", "id": 42734, "pid": 435, "tid": 435, "ts": 1776649790931172, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931183, "dur": 22, + "args": { + "External id": 5296, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42735, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42735, "pid": 0, "tid": 7, "ts": 1776649790931183, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931173, "dur": 5, + "args": { + "External id": 5296, + "cbid": 211, "correlation": 42735 + } + }, + { + "ph": "s", "id": 42735, "pid": 435, "tid": 435, "ts": 1776649790931173, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931202, "dur": 0, + "args": { + "External id": 5298, + "cbid": 210, "correlation": 42762 + } + }, + { + "ph": "f", "id": 42762, "pid": 435, "tid": 435, "ts": 1776649790931202, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931214, "dur": 23, + "args": { + "External id": 5298, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42763, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42763, "pid": 0, "tid": 7, "ts": 1776649790931214, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931203, "dur": 6, + "args": { + "External id": 5298, + "cbid": 211, "correlation": 42763 + } + }, + { + "ph": "s", "id": 42763, "pid": 435, "tid": 435, "ts": 1776649790931203, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931232, "dur": 0, + "args": { + "External id": 5300, + "cbid": 210, "correlation": 42790 + } + }, + { + "ph": "f", "id": 42790, "pid": 435, "tid": 435, "ts": 1776649790931232, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931243, "dur": 23, + "args": { + "External id": 5300, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42791, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42791, "pid": 0, "tid": 7, "ts": 1776649790931243, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931233, "dur": 5, + "args": { + "External id": 5300, + "cbid": 211, "correlation": 42791 + } + }, + { + "ph": "s", "id": 42791, "pid": 435, "tid": 435, "ts": 1776649790931233, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931262, "dur": 0, + "args": { + "External id": 5302, + "cbid": 210, "correlation": 42818 + } + }, + { + "ph": "f", "id": 42818, "pid": 435, "tid": 435, "ts": 1776649790931262, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931273, "dur": 23, + "args": { + "External id": 5302, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42819, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42819, "pid": 0, "tid": 7, "ts": 1776649790931273, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931263, "dur": 6, + "args": { + "External id": 5302, + "cbid": 211, "correlation": 42819 + } + }, + { + "ph": "s", "id": 42819, "pid": 435, "tid": 435, "ts": 1776649790931263, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931292, "dur": 0, + "args": { + "External id": 5304, + "cbid": 210, "correlation": 42846 + } + }, + { + "ph": "f", "id": 42846, "pid": 435, "tid": 435, "ts": 1776649790931292, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931303, "dur": 23, + "args": { + "External id": 5304, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42847, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42847, "pid": 0, "tid": 7, "ts": 1776649790931303, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931293, "dur": 5, + "args": { + "External id": 5304, + "cbid": 211, "correlation": 42847 + } + }, + { + "ph": "s", "id": 42847, "pid": 435, "tid": 435, "ts": 1776649790931293, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931322, "dur": 0, + "args": { + "External id": 5306, + "cbid": 210, "correlation": 42874 + } + }, + { + "ph": "f", "id": 42874, "pid": 435, "tid": 435, "ts": 1776649790931322, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931333, "dur": 23, + "args": { + "External id": 5306, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42875, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42875, "pid": 0, "tid": 7, "ts": 1776649790931333, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931323, "dur": 5, + "args": { + "External id": 5306, + "cbid": 211, "correlation": 42875 + } + }, + { + "ph": "s", "id": 42875, "pid": 435, "tid": 435, "ts": 1776649790931323, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931351, "dur": 0, + "args": { + "External id": 5308, + "cbid": 210, "correlation": 42902 + } + }, + { + "ph": "f", "id": 42902, "pid": 435, "tid": 435, "ts": 1776649790931351, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931363, "dur": 23, + "args": { + "External id": 5308, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42903, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42903, "pid": 0, "tid": 7, "ts": 1776649790931363, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931353, "dur": 5, + "args": { + "External id": 5308, + "cbid": 211, "correlation": 42903 + } + }, + { + "ph": "s", "id": 42903, "pid": 435, "tid": 435, "ts": 1776649790931353, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931381, "dur": 0, + "args": { + "External id": 5310, + "cbid": 210, "correlation": 42930 + } + }, + { + "ph": "f", "id": 42930, "pid": 435, "tid": 435, "ts": 1776649790931381, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931393, "dur": 23, + "args": { + "External id": 5310, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42931, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42931, "pid": 0, "tid": 7, "ts": 1776649790931393, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931383, "dur": 6, + "args": { + "External id": 5310, + "cbid": 211, "correlation": 42931 + } + }, + { + "ph": "s", "id": 42931, "pid": 435, "tid": 435, "ts": 1776649790931383, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931412, "dur": 0, + "args": { + "External id": 5312, + "cbid": 210, "correlation": 42958 + } + }, + { + "ph": "f", "id": 42958, "pid": 435, "tid": 435, "ts": 1776649790931412, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931424, "dur": 23, + "args": { + "External id": 5312, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42959, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42959, "pid": 0, "tid": 7, "ts": 1776649790931424, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931413, "dur": 5, + "args": { + "External id": 5312, + "cbid": 211, "correlation": 42959 + } + }, + { + "ph": "s", "id": 42959, "pid": 435, "tid": 435, "ts": 1776649790931413, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931442, "dur": 0, + "args": { + "External id": 5314, + "cbid": 210, "correlation": 42986 + } + }, + { + "ph": "f", "id": 42986, "pid": 435, "tid": 435, "ts": 1776649790931442, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931453, "dur": 23, + "args": { + "External id": 5314, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 42987, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 42987, "pid": 0, "tid": 7, "ts": 1776649790931453, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931443, "dur": 5, + "args": { + "External id": 5314, + "cbid": 211, "correlation": 42987 + } + }, + { + "ph": "s", "id": 42987, "pid": 435, "tid": 435, "ts": 1776649790931443, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931472, "dur": 0, + "args": { + "External id": 5316, + "cbid": 210, "correlation": 43014 + } + }, + { + "ph": "f", "id": 43014, "pid": 435, "tid": 435, "ts": 1776649790931472, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931483, "dur": 23, + "args": { + "External id": 5316, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43015, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43015, "pid": 0, "tid": 7, "ts": 1776649790931483, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931473, "dur": 6, + "args": { + "External id": 5316, + "cbid": 211, "correlation": 43015 + } + }, + { + "ph": "s", "id": 43015, "pid": 435, "tid": 435, "ts": 1776649790931473, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931502, "dur": 0, + "args": { + "External id": 5318, + "cbid": 210, "correlation": 43042 + } + }, + { + "ph": "f", "id": 43042, "pid": 435, "tid": 435, "ts": 1776649790931502, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931513, "dur": 23, + "args": { + "External id": 5318, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43043, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43043, "pid": 0, "tid": 7, "ts": 1776649790931513, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931503, "dur": 5, + "args": { + "External id": 5318, + "cbid": 211, "correlation": 43043 + } + }, + { + "ph": "s", "id": 43043, "pid": 435, "tid": 435, "ts": 1776649790931503, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931531, "dur": 0, + "args": { + "External id": 5320, + "cbid": 210, "correlation": 43070 + } + }, + { + "ph": "f", "id": 43070, "pid": 435, "tid": 435, "ts": 1776649790931531, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931542, "dur": 23, + "args": { + "External id": 5320, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43071, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43071, "pid": 0, "tid": 7, "ts": 1776649790931542, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931533, "dur": 6, + "args": { + "External id": 5320, + "cbid": 211, "correlation": 43071 + } + }, + { + "ph": "s", "id": 43071, "pid": 435, "tid": 435, "ts": 1776649790931533, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790931565, "dur": 4, + "args": { + "External id": 43082, + "cbid": 135, "correlation": 43082 + } + }, + { + "ph": "f", "id": 43082, "pid": 435, "tid": 435, "ts": 1776649790931565, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790931580, "dur": 5, + "args": { + "External id": 43087, + "cbid": 165, "correlation": 43087 + } + }, + { + "ph": "s", "id": 43087, "pid": 435, "tid": 435, "ts": 1776649790931580, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649790931588, "dur": 4, + "args": { + "External id": 43090, + "cbid": 139, "correlation": 43090 + } + }, + { + "ph": "f", "id": 43090, "pid": 435, "tid": 435, "ts": 1776649790931588, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931624, "dur": 1, + "args": { + "External id": 5322, + "cbid": 210, "correlation": 43124 + } + }, + { + "ph": "f", "id": 43124, "pid": 435, "tid": 435, "ts": 1776649790931624, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931637, "dur": 23, + "args": { + "External id": 5322, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43125, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43125, "pid": 0, "tid": 7, "ts": 1776649790931637, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931626, "dur": 8, + "args": { + "External id": 5322, + "cbid": 211, "correlation": 43125 + } + }, + { + "ph": "s", "id": 43125, "pid": 435, "tid": 435, "ts": 1776649790931626, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931660, "dur": 0, + "args": { + "External id": 5324, + "cbid": 210, "correlation": 43152 + } + }, + { + "ph": "f", "id": 43152, "pid": 435, "tid": 435, "ts": 1776649790931660, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931670, "dur": 23, + "args": { + "External id": 5324, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43153, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43153, "pid": 0, "tid": 7, "ts": 1776649790931670, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931661, "dur": 6, + "args": { + "External id": 5324, + "cbid": 211, "correlation": 43153 + } + }, + { + "ph": "s", "id": 43153, "pid": 435, "tid": 435, "ts": 1776649790931661, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931690, "dur": 0, + "args": { + "External id": 5326, + "cbid": 210, "correlation": 43180 + } + }, + { + "ph": "f", "id": 43180, "pid": 435, "tid": 435, "ts": 1776649790931690, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931701, "dur": 23, + "args": { + "External id": 5326, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43181, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43181, "pid": 0, "tid": 7, "ts": 1776649790931701, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931691, "dur": 6, + "args": { + "External id": 5326, + "cbid": 211, "correlation": 43181 + } + }, + { + "ph": "s", "id": 43181, "pid": 435, "tid": 435, "ts": 1776649790931691, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931721, "dur": 0, + "args": { + "External id": 5328, + "cbid": 210, "correlation": 43208 + } + }, + { + "ph": "f", "id": 43208, "pid": 435, "tid": 435, "ts": 1776649790931721, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931731, "dur": 23, + "args": { + "External id": 5328, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43209, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43209, "pid": 0, "tid": 7, "ts": 1776649790931731, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931722, "dur": 6, + "args": { + "External id": 5328, + "cbid": 211, "correlation": 43209 + } + }, + { + "ph": "s", "id": 43209, "pid": 435, "tid": 435, "ts": 1776649790931722, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931751, "dur": 0, + "args": { + "External id": 5330, + "cbid": 210, "correlation": 43236 + } + }, + { + "ph": "f", "id": 43236, "pid": 435, "tid": 435, "ts": 1776649790931751, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931762, "dur": 23, + "args": { + "External id": 5330, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43237, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43237, "pid": 0, "tid": 7, "ts": 1776649790931762, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931753, "dur": 6, + "args": { + "External id": 5330, + "cbid": 211, "correlation": 43237 + } + }, + { + "ph": "s", "id": 43237, "pid": 435, "tid": 435, "ts": 1776649790931753, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931782, "dur": 0, + "args": { + "External id": 5332, + "cbid": 210, "correlation": 43264 + } + }, + { + "ph": "f", "id": 43264, "pid": 435, "tid": 435, "ts": 1776649790931782, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931792, "dur": 23, + "args": { + "External id": 5332, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43265, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43265, "pid": 0, "tid": 7, "ts": 1776649790931792, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931783, "dur": 6, + "args": { + "External id": 5332, + "cbid": 211, "correlation": 43265 + } + }, + { + "ph": "s", "id": 43265, "pid": 435, "tid": 435, "ts": 1776649790931783, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931812, "dur": 0, + "args": { + "External id": 5334, + "cbid": 210, "correlation": 43292 + } + }, + { + "ph": "f", "id": 43292, "pid": 435, "tid": 435, "ts": 1776649790931812, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931823, "dur": 23, + "args": { + "External id": 5334, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43293, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43293, "pid": 0, "tid": 7, "ts": 1776649790931823, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931814, "dur": 6, + "args": { + "External id": 5334, + "cbid": 211, "correlation": 43293 + } + }, + { + "ph": "s", "id": 43293, "pid": 435, "tid": 435, "ts": 1776649790931814, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931842, "dur": 0, + "args": { + "External id": 5336, + "cbid": 210, "correlation": 43320 + } + }, + { + "ph": "f", "id": 43320, "pid": 435, "tid": 435, "ts": 1776649790931842, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931853, "dur": 23, + "args": { + "External id": 5336, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43321, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43321, "pid": 0, "tid": 7, "ts": 1776649790931853, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931843, "dur": 5, + "args": { + "External id": 5336, + "cbid": 211, "correlation": 43321 + } + }, + { + "ph": "s", "id": 43321, "pid": 435, "tid": 435, "ts": 1776649790931843, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931872, "dur": 0, + "args": { + "External id": 5338, + "cbid": 210, "correlation": 43348 + } + }, + { + "ph": "f", "id": 43348, "pid": 435, "tid": 435, "ts": 1776649790931872, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931883, "dur": 23, + "args": { + "External id": 5338, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43349, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43349, "pid": 0, "tid": 7, "ts": 1776649790931883, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931873, "dur": 6, + "args": { + "External id": 5338, + "cbid": 211, "correlation": 43349 + } + }, + { + "ph": "s", "id": 43349, "pid": 435, "tid": 435, "ts": 1776649790931873, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931901, "dur": 0, + "args": { + "External id": 5340, + "cbid": 210, "correlation": 43376 + } + }, + { + "ph": "f", "id": 43376, "pid": 435, "tid": 435, "ts": 1776649790931901, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790931913, "dur": 23, + "args": { + "External id": 5340, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43377, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43377, "pid": 0, "tid": 7, "ts": 1776649790931913, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931903, "dur": 6, + "args": { + "External id": 5340, + "cbid": 211, "correlation": 43377 + } + }, + { + "ph": "s", "id": 43377, "pid": 435, "tid": 435, "ts": 1776649790931903, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790931927, "dur": 13, + "args": { + "External id": 43383, + "cbid": 165, "correlation": 43383 + } + }, + { + "ph": "s", "id": 43383, "pid": 435, "tid": 435, "ts": 1776649790931927, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790931960, "dur": 3, + "args": { + "External id": 43392, + "cbid": 135, "correlation": 43392 + } + }, + { + "ph": "f", "id": 43392, "pid": 435, "tid": 435, "ts": 1776649790931960, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790931988, "dur": 1, + "args": { + "External id": 5342, + "cbid": 210, "correlation": 43418 + } + }, + { + "ph": "f", "id": 43418, "pid": 435, "tid": 435, "ts": 1776649790931988, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932000, "dur": 23, + "args": { + "External id": 5342, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43419, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43419, "pid": 0, "tid": 7, "ts": 1776649790932000, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790931990, "dur": 7, + "args": { + "External id": 5342, + "cbid": 211, "correlation": 43419 + } + }, + { + "ph": "s", "id": 43419, "pid": 435, "tid": 435, "ts": 1776649790931990, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932022, "dur": 0, + "args": { + "External id": 5344, + "cbid": 210, "correlation": 43446 + } + }, + { + "ph": "f", "id": 43446, "pid": 435, "tid": 435, "ts": 1776649790932022, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932033, "dur": 23, + "args": { + "External id": 5344, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43447, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43447, "pid": 0, "tid": 7, "ts": 1776649790932033, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932023, "dur": 6, + "args": { + "External id": 5344, + "cbid": 211, "correlation": 43447 + } + }, + { + "ph": "s", "id": 43447, "pid": 435, "tid": 435, "ts": 1776649790932023, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932052, "dur": 0, + "args": { + "External id": 5346, + "cbid": 210, "correlation": 43474 + } + }, + { + "ph": "f", "id": 43474, "pid": 435, "tid": 435, "ts": 1776649790932052, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932062, "dur": 23, + "args": { + "External id": 5346, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43475, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43475, "pid": 0, "tid": 7, "ts": 1776649790932062, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932053, "dur": 6, + "args": { + "External id": 5346, + "cbid": 211, "correlation": 43475 + } + }, + { + "ph": "s", "id": 43475, "pid": 435, "tid": 435, "ts": 1776649790932053, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932082, "dur": 0, + "args": { + "External id": 5348, + "cbid": 210, "correlation": 43502 + } + }, + { + "ph": "f", "id": 43502, "pid": 435, "tid": 435, "ts": 1776649790932082, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932093, "dur": 23, + "args": { + "External id": 5348, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43503, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43503, "pid": 0, "tid": 7, "ts": 1776649790932093, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932083, "dur": 6, + "args": { + "External id": 5348, + "cbid": 211, "correlation": 43503 + } + }, + { + "ph": "s", "id": 43503, "pid": 435, "tid": 435, "ts": 1776649790932083, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932111, "dur": 0, + "args": { + "External id": 5350, + "cbid": 210, "correlation": 43530 + } + }, + { + "ph": "f", "id": 43530, "pid": 435, "tid": 435, "ts": 1776649790932111, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932123, "dur": 23, + "args": { + "External id": 5350, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43531, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43531, "pid": 0, "tid": 7, "ts": 1776649790932123, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932113, "dur": 6, + "args": { + "External id": 5350, + "cbid": 211, "correlation": 43531 + } + }, + { + "ph": "s", "id": 43531, "pid": 435, "tid": 435, "ts": 1776649790932113, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932141, "dur": 0, + "args": { + "External id": 5352, + "cbid": 210, "correlation": 43558 + } + }, + { + "ph": "f", "id": 43558, "pid": 435, "tid": 435, "ts": 1776649790932141, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932152, "dur": 23, + "args": { + "External id": 5352, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43559, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43559, "pid": 0, "tid": 7, "ts": 1776649790932152, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932142, "dur": 6, + "args": { + "External id": 5352, + "cbid": 211, "correlation": 43559 + } + }, + { + "ph": "s", "id": 43559, "pid": 435, "tid": 435, "ts": 1776649790932142, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932170, "dur": 0, + "args": { + "External id": 5354, + "cbid": 210, "correlation": 43586 + } + }, + { + "ph": "f", "id": 43586, "pid": 435, "tid": 435, "ts": 1776649790932170, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932182, "dur": 23, + "args": { + "External id": 5354, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43587, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43587, "pid": 0, "tid": 7, "ts": 1776649790932182, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932171, "dur": 5, + "args": { + "External id": 5354, + "cbid": 211, "correlation": 43587 + } + }, + { + "ph": "s", "id": 43587, "pid": 435, "tid": 435, "ts": 1776649790932171, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932199, "dur": 0, + "args": { + "External id": 5356, + "cbid": 210, "correlation": 43614 + } + }, + { + "ph": "f", "id": 43614, "pid": 435, "tid": 435, "ts": 1776649790932199, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932211, "dur": 23, + "args": { + "External id": 5356, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43615, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43615, "pid": 0, "tid": 7, "ts": 1776649790932211, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932201, "dur": 5, + "args": { + "External id": 5356, + "cbid": 211, "correlation": 43615 + } + }, + { + "ph": "s", "id": 43615, "pid": 435, "tid": 435, "ts": 1776649790932201, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932252, "dur": 0, + "args": { + "External id": 5358, + "cbid": 210, "correlation": 43642 + } + }, + { + "ph": "f", "id": 43642, "pid": 435, "tid": 435, "ts": 1776649790932252, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932263, "dur": 23, + "args": { + "External id": 5358, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43643, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43643, "pid": 0, "tid": 7, "ts": 1776649790932263, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932254, "dur": 6, + "args": { + "External id": 5358, + "cbid": 211, "correlation": 43643 + } + }, + { + "ph": "s", "id": 43643, "pid": 435, "tid": 435, "ts": 1776649790932254, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932285, "dur": 0, + "args": { + "External id": 5360, + "cbid": 210, "correlation": 43670 + } + }, + { + "ph": "f", "id": 43670, "pid": 435, "tid": 435, "ts": 1776649790932285, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932295, "dur": 23, + "args": { + "External id": 5360, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43671, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43671, "pid": 0, "tid": 7, "ts": 1776649790932295, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932286, "dur": 6, + "args": { + "External id": 5360, + "cbid": 211, "correlation": 43671 + } + }, + { + "ph": "s", "id": 43671, "pid": 435, "tid": 435, "ts": 1776649790932286, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932314, "dur": 0, + "args": { + "External id": 5362, + "cbid": 210, "correlation": 43698 + } + }, + { + "ph": "f", "id": 43698, "pid": 435, "tid": 435, "ts": 1776649790932314, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932325, "dur": 23, + "args": { + "External id": 5362, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43699, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43699, "pid": 0, "tid": 7, "ts": 1776649790932325, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932315, "dur": 5, + "args": { + "External id": 5362, + "cbid": 211, "correlation": 43699 + } + }, + { + "ph": "s", "id": 43699, "pid": 435, "tid": 435, "ts": 1776649790932315, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932344, "dur": 0, + "args": { + "External id": 5364, + "cbid": 210, "correlation": 43726 + } + }, + { + "ph": "f", "id": 43726, "pid": 435, "tid": 435, "ts": 1776649790932344, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932355, "dur": 23, + "args": { + "External id": 5364, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43727, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43727, "pid": 0, "tid": 7, "ts": 1776649790932355, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932345, "dur": 5, + "args": { + "External id": 5364, + "cbid": 211, "correlation": 43727 + } + }, + { + "ph": "s", "id": 43727, "pid": 435, "tid": 435, "ts": 1776649790932345, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932373, "dur": 0, + "args": { + "External id": 5366, + "cbid": 210, "correlation": 43754 + } + }, + { + "ph": "f", "id": 43754, "pid": 435, "tid": 435, "ts": 1776649790932373, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932385, "dur": 23, + "args": { + "External id": 5366, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43755, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43755, "pid": 0, "tid": 7, "ts": 1776649790932385, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932374, "dur": 5, + "args": { + "External id": 5366, + "cbid": 211, "correlation": 43755 + } + }, + { + "ph": "s", "id": 43755, "pid": 435, "tid": 435, "ts": 1776649790932374, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932403, "dur": 0, + "args": { + "External id": 5368, + "cbid": 210, "correlation": 43782 + } + }, + { + "ph": "f", "id": 43782, "pid": 435, "tid": 435, "ts": 1776649790932403, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932414, "dur": 23, + "args": { + "External id": 5368, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43783, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43783, "pid": 0, "tid": 7, "ts": 1776649790932414, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932404, "dur": 5, + "args": { + "External id": 5368, + "cbid": 211, "correlation": 43783 + } + }, + { + "ph": "s", "id": 43783, "pid": 435, "tid": 435, "ts": 1776649790932404, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932432, "dur": 0, + "args": { + "External id": 5370, + "cbid": 210, "correlation": 43810 + } + }, + { + "ph": "f", "id": 43810, "pid": 435, "tid": 435, "ts": 1776649790932432, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932443, "dur": 23, + "args": { + "External id": 5370, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43811, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43811, "pid": 0, "tid": 7, "ts": 1776649790932443, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932433, "dur": 5, + "args": { + "External id": 5370, + "cbid": 211, "correlation": 43811 + } + }, + { + "ph": "s", "id": 43811, "pid": 435, "tid": 435, "ts": 1776649790932433, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932461, "dur": 0, + "args": { + "External id": 5372, + "cbid": 210, "correlation": 43838 + } + }, + { + "ph": "f", "id": 43838, "pid": 435, "tid": 435, "ts": 1776649790932461, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932473, "dur": 23, + "args": { + "External id": 5372, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43839, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43839, "pid": 0, "tid": 7, "ts": 1776649790932473, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932462, "dur": 5, + "args": { + "External id": 5372, + "cbid": 211, "correlation": 43839 + } + }, + { + "ph": "s", "id": 43839, "pid": 435, "tid": 435, "ts": 1776649790932462, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932490, "dur": 0, + "args": { + "External id": 5374, + "cbid": 210, "correlation": 43866 + } + }, + { + "ph": "f", "id": 43866, "pid": 435, "tid": 435, "ts": 1776649790932490, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932502, "dur": 23, + "args": { + "External id": 5374, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43867, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43867, "pid": 0, "tid": 7, "ts": 1776649790932502, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932491, "dur": 5, + "args": { + "External id": 5374, + "cbid": 211, "correlation": 43867 + } + }, + { + "ph": "s", "id": 43867, "pid": 435, "tid": 435, "ts": 1776649790932491, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932519, "dur": 0, + "args": { + "External id": 5376, + "cbid": 210, "correlation": 43894 + } + }, + { + "ph": "f", "id": 43894, "pid": 435, "tid": 435, "ts": 1776649790932519, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932531, "dur": 23, + "args": { + "External id": 5376, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43895, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43895, "pid": 0, "tid": 7, "ts": 1776649790932531, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932520, "dur": 6, + "args": { + "External id": 5376, + "cbid": 211, "correlation": 43895 + } + }, + { + "ph": "s", "id": 43895, "pid": 435, "tid": 435, "ts": 1776649790932520, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932548, "dur": 0, + "args": { + "External id": 5378, + "cbid": 210, "correlation": 43922 + } + }, + { + "ph": "f", "id": 43922, "pid": 435, "tid": 435, "ts": 1776649790932548, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932560, "dur": 23, + "args": { + "External id": 5378, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43923, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43923, "pid": 0, "tid": 7, "ts": 1776649790932560, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932549, "dur": 5, + "args": { + "External id": 5378, + "cbid": 211, "correlation": 43923 + } + }, + { + "ph": "s", "id": 43923, "pid": 435, "tid": 435, "ts": 1776649790932549, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932577, "dur": 0, + "args": { + "External id": 5380, + "cbid": 210, "correlation": 43950 + } + }, + { + "ph": "f", "id": 43950, "pid": 435, "tid": 435, "ts": 1776649790932577, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932588, "dur": 23, + "args": { + "External id": 5380, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43951, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43951, "pid": 0, "tid": 7, "ts": 1776649790932588, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932578, "dur": 6, + "args": { + "External id": 5380, + "cbid": 211, "correlation": 43951 + } + }, + { + "ph": "s", "id": 43951, "pid": 435, "tid": 435, "ts": 1776649790932578, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932606, "dur": 0, + "args": { + "External id": 5382, + "cbid": 210, "correlation": 43978 + } + }, + { + "ph": "f", "id": 43978, "pid": 435, "tid": 435, "ts": 1776649790932606, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932618, "dur": 23, + "args": { + "External id": 5382, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 43979, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 43979, "pid": 0, "tid": 7, "ts": 1776649790932618, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932607, "dur": 5, + "args": { + "External id": 5382, + "cbid": 211, "correlation": 43979 + } + }, + { + "ph": "s", "id": 43979, "pid": 435, "tid": 435, "ts": 1776649790932607, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932635, "dur": 0, + "args": { + "External id": 5384, + "cbid": 210, "correlation": 44006 + } + }, + { + "ph": "f", "id": 44006, "pid": 435, "tid": 435, "ts": 1776649790932635, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932646, "dur": 23, + "args": { + "External id": 5384, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44007, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44007, "pid": 0, "tid": 7, "ts": 1776649790932646, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932636, "dur": 5, + "args": { + "External id": 5384, + "cbid": 211, "correlation": 44007 + } + }, + { + "ph": "s", "id": 44007, "pid": 435, "tid": 435, "ts": 1776649790932636, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932664, "dur": 0, + "args": { + "External id": 5386, + "cbid": 210, "correlation": 44034 + } + }, + { + "ph": "f", "id": 44034, "pid": 435, "tid": 435, "ts": 1776649790932664, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932675, "dur": 23, + "args": { + "External id": 5386, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44035, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44035, "pid": 0, "tid": 7, "ts": 1776649790932675, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932665, "dur": 5, + "args": { + "External id": 5386, + "cbid": 211, "correlation": 44035 + } + }, + { + "ph": "s", "id": 44035, "pid": 435, "tid": 435, "ts": 1776649790932665, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932693, "dur": 0, + "args": { + "External id": 5388, + "cbid": 210, "correlation": 44062 + } + }, + { + "ph": "f", "id": 44062, "pid": 435, "tid": 435, "ts": 1776649790932693, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932704, "dur": 23, + "args": { + "External id": 5388, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44063, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44063, "pid": 0, "tid": 7, "ts": 1776649790932704, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932694, "dur": 5, + "args": { + "External id": 5388, + "cbid": 211, "correlation": 44063 + } + }, + { + "ph": "s", "id": 44063, "pid": 435, "tid": 435, "ts": 1776649790932694, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932722, "dur": 0, + "args": { + "External id": 5390, + "cbid": 210, "correlation": 44090 + } + }, + { + "ph": "f", "id": 44090, "pid": 435, "tid": 435, "ts": 1776649790932722, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932733, "dur": 23, + "args": { + "External id": 5390, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44091, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44091, "pid": 0, "tid": 7, "ts": 1776649790932733, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932723, "dur": 5, + "args": { + "External id": 5390, + "cbid": 211, "correlation": 44091 + } + }, + { + "ph": "s", "id": 44091, "pid": 435, "tid": 435, "ts": 1776649790932723, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932751, "dur": 0, + "args": { + "External id": 5392, + "cbid": 210, "correlation": 44118 + } + }, + { + "ph": "f", "id": 44118, "pid": 435, "tid": 435, "ts": 1776649790932751, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932762, "dur": 23, + "args": { + "External id": 5392, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44119, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44119, "pid": 0, "tid": 7, "ts": 1776649790932762, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932752, "dur": 5, + "args": { + "External id": 5392, + "cbid": 211, "correlation": 44119 + } + }, + { + "ph": "s", "id": 44119, "pid": 435, "tid": 435, "ts": 1776649790932752, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932780, "dur": 0, + "args": { + "External id": 5394, + "cbid": 210, "correlation": 44146 + } + }, + { + "ph": "f", "id": 44146, "pid": 435, "tid": 435, "ts": 1776649790932780, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932791, "dur": 23, + "args": { + "External id": 5394, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44147, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44147, "pid": 0, "tid": 7, "ts": 1776649790932791, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932781, "dur": 5, + "args": { + "External id": 5394, + "cbid": 211, "correlation": 44147 + } + }, + { + "ph": "s", "id": 44147, "pid": 435, "tid": 435, "ts": 1776649790932781, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932809, "dur": 0, + "args": { + "External id": 5396, + "cbid": 210, "correlation": 44174 + } + }, + { + "ph": "f", "id": 44174, "pid": 435, "tid": 435, "ts": 1776649790932809, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932820, "dur": 23, + "args": { + "External id": 5396, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44175, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44175, "pid": 0, "tid": 7, "ts": 1776649790932820, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932810, "dur": 5, + "args": { + "External id": 5396, + "cbid": 211, "correlation": 44175 + } + }, + { + "ph": "s", "id": 44175, "pid": 435, "tid": 435, "ts": 1776649790932810, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932838, "dur": 0, + "args": { + "External id": 5398, + "cbid": 210, "correlation": 44202 + } + }, + { + "ph": "f", "id": 44202, "pid": 435, "tid": 435, "ts": 1776649790932838, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932849, "dur": 23, + "args": { + "External id": 5398, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44203, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44203, "pid": 0, "tid": 7, "ts": 1776649790932849, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932839, "dur": 5, + "args": { + "External id": 5398, + "cbid": 211, "correlation": 44203 + } + }, + { + "ph": "s", "id": 44203, "pid": 435, "tid": 435, "ts": 1776649790932839, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932867, "dur": 0, + "args": { + "External id": 5400, + "cbid": 210, "correlation": 44230 + } + }, + { + "ph": "f", "id": 44230, "pid": 435, "tid": 435, "ts": 1776649790932867, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932878, "dur": 23, + "args": { + "External id": 5400, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44231, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44231, "pid": 0, "tid": 7, "ts": 1776649790932878, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932868, "dur": 5, + "args": { + "External id": 5400, + "cbid": 211, "correlation": 44231 + } + }, + { + "ph": "s", "id": 44231, "pid": 435, "tid": 435, "ts": 1776649790932868, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932896, "dur": 0, + "args": { + "External id": 5402, + "cbid": 210, "correlation": 44258 + } + }, + { + "ph": "f", "id": 44258, "pid": 435, "tid": 435, "ts": 1776649790932896, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932907, "dur": 22, + "args": { + "External id": 5402, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44259, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44259, "pid": 0, "tid": 7, "ts": 1776649790932907, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932897, "dur": 5, + "args": { + "External id": 5402, + "cbid": 211, "correlation": 44259 + } + }, + { + "ph": "s", "id": 44259, "pid": 435, "tid": 435, "ts": 1776649790932897, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932925, "dur": 0, + "args": { + "External id": 5404, + "cbid": 210, "correlation": 44286 + } + }, + { + "ph": "f", "id": 44286, "pid": 435, "tid": 435, "ts": 1776649790932925, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932936, "dur": 23, + "args": { + "External id": 5404, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44287, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44287, "pid": 0, "tid": 7, "ts": 1776649790932936, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932926, "dur": 5, + "args": { + "External id": 5404, + "cbid": 211, "correlation": 44287 + } + }, + { + "ph": "s", "id": 44287, "pid": 435, "tid": 435, "ts": 1776649790932926, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932953, "dur": 0, + "args": { + "External id": 5406, + "cbid": 210, "correlation": 44314 + } + }, + { + "ph": "f", "id": 44314, "pid": 435, "tid": 435, "ts": 1776649790932953, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932965, "dur": 22, + "args": { + "External id": 5406, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44315, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44315, "pid": 0, "tid": 7, "ts": 1776649790932965, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932955, "dur": 5, + "args": { + "External id": 5406, + "cbid": 211, "correlation": 44315 + } + }, + { + "ph": "s", "id": 44315, "pid": 435, "tid": 435, "ts": 1776649790932955, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790932983, "dur": 0, + "args": { + "External id": 5408, + "cbid": 210, "correlation": 44342 + } + }, + { + "ph": "f", "id": 44342, "pid": 435, "tid": 435, "ts": 1776649790932983, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790932994, "dur": 22, + "args": { + "External id": 5408, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44343, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44343, "pid": 0, "tid": 7, "ts": 1776649790932994, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790932984, "dur": 5, + "args": { + "External id": 5408, + "cbid": 211, "correlation": 44343 + } + }, + { + "ph": "s", "id": 44343, "pid": 435, "tid": 435, "ts": 1776649790932984, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790933012, "dur": 0, + "args": { + "External id": 5410, + "cbid": 210, "correlation": 44370 + } + }, + { + "ph": "f", "id": 44370, "pid": 435, "tid": 435, "ts": 1776649790933012, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790933023, "dur": 23, + "args": { + "External id": 5410, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44371, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44371, "pid": 0, "tid": 7, "ts": 1776649790933023, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790933013, "dur": 5, + "args": { + "External id": 5410, + "cbid": 211, "correlation": 44371 + } + }, + { + "ph": "s", "id": 44371, "pid": 435, "tid": 435, "ts": 1776649790933013, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790933041, "dur": 0, + "args": { + "External id": 5412, + "cbid": 210, "correlation": 44398 + } + }, + { + "ph": "f", "id": 44398, "pid": 435, "tid": 435, "ts": 1776649790933041, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790933052, "dur": 23, + "args": { + "External id": 5412, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44399, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44399, "pid": 0, "tid": 7, "ts": 1776649790933052, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790933042, "dur": 5, + "args": { + "External id": 5412, + "cbid": 211, "correlation": 44399 + } + }, + { + "ph": "s", "id": 44399, "pid": 435, "tid": 435, "ts": 1776649790933042, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790933071, "dur": 0, + "args": { + "External id": 5414, + "cbid": 210, "correlation": 44426 + } + }, + { + "ph": "f", "id": 44426, "pid": 435, "tid": 435, "ts": 1776649790933071, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790933082, "dur": 23, + "args": { + "External id": 5414, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44427, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44427, "pid": 0, "tid": 7, "ts": 1776649790933082, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790933072, "dur": 5, + "args": { + "External id": 5414, + "cbid": 211, "correlation": 44427 + } + }, + { + "ph": "s", "id": 44427, "pid": 435, "tid": 435, "ts": 1776649790933072, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790933100, "dur": 0, + "args": { + "External id": 5416, + "cbid": 210, "correlation": 44454 + } + }, + { + "ph": "f", "id": 44454, "pid": 435, "tid": 435, "ts": 1776649790933100, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790933111, "dur": 23, + "args": { + "External id": 5416, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44455, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44455, "pid": 0, "tid": 7, "ts": 1776649790933111, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790933101, "dur": 5, + "args": { + "External id": 5416, + "cbid": 211, "correlation": 44455 + } + }, + { + "ph": "s", "id": 44455, "pid": 435, "tid": 435, "ts": 1776649790933101, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790933129, "dur": 0, + "args": { + "External id": 5418, + "cbid": 210, "correlation": 44482 + } + }, + { + "ph": "f", "id": 44482, "pid": 435, "tid": 435, "ts": 1776649790933129, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790933140, "dur": 23, + "args": { + "External id": 5418, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44483, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44483, "pid": 0, "tid": 7, "ts": 1776649790933140, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790933130, "dur": 5, + "args": { + "External id": 5418, + "cbid": 211, "correlation": 44483 + } + }, + { + "ph": "s", "id": 44483, "pid": 435, "tid": 435, "ts": 1776649790933130, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790933158, "dur": 0, + "args": { + "External id": 5420, + "cbid": 210, "correlation": 44510 + } + }, + { + "ph": "f", "id": 44510, "pid": 435, "tid": 435, "ts": 1776649790933158, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790933169, "dur": 23, + "args": { + "External id": 5420, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44511, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44511, "pid": 0, "tid": 7, "ts": 1776649790933169, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790933159, "dur": 5, + "args": { + "External id": 5420, + "cbid": 211, "correlation": 44511 + } + }, + { + "ph": "s", "id": 44511, "pid": 435, "tid": 435, "ts": 1776649790933159, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790933187, "dur": 0, + "args": { + "External id": 5422, + "cbid": 210, "correlation": 44538 + } + }, + { + "ph": "f", "id": 44538, "pid": 435, "tid": 435, "ts": 1776649790933187, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790933198, "dur": 23, + "args": { + "External id": 5422, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44539, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44539, "pid": 0, "tid": 7, "ts": 1776649790933198, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790933188, "dur": 5, + "args": { + "External id": 5422, + "cbid": 211, "correlation": 44539 + } + }, + { + "ph": "s", "id": 44539, "pid": 435, "tid": 435, "ts": 1776649790933188, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790933216, "dur": 0, + "args": { + "External id": 5424, + "cbid": 210, "correlation": 44566 + } + }, + { + "ph": "f", "id": 44566, "pid": 435, "tid": 435, "ts": 1776649790933216, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790933227, "dur": 23, + "args": { + "External id": 5424, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44567, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44567, "pid": 0, "tid": 7, "ts": 1776649790933227, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790933217, "dur": 5, + "args": { + "External id": 5424, + "cbid": 211, "correlation": 44567 + } + }, + { + "ph": "s", "id": 44567, "pid": 435, "tid": 435, "ts": 1776649790933217, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790933245, "dur": 0, + "args": { + "External id": 5426, + "cbid": 210, "correlation": 44594 + } + }, + { + "ph": "f", "id": 44594, "pid": 435, "tid": 435, "ts": 1776649790933245, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790933257, "dur": 23, + "args": { + "External id": 5426, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44595, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44595, "pid": 0, "tid": 7, "ts": 1776649790933257, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790933246, "dur": 6, + "args": { + "External id": 5426, + "cbid": 211, "correlation": 44595 + } + }, + { + "ph": "s", "id": 44595, "pid": 435, "tid": 435, "ts": 1776649790933246, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790933277, "dur": 0, + "args": { + "External id": 5428, + "cbid": 210, "correlation": 44622 + } + }, + { + "ph": "f", "id": 44622, "pid": 435, "tid": 435, "ts": 1776649790933277, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790933287, "dur": 23, + "args": { + "External id": 5428, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44623, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44623, "pid": 0, "tid": 7, "ts": 1776649790933287, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790933278, "dur": 6, + "args": { + "External id": 5428, + "cbid": 211, "correlation": 44623 + } + }, + { + "ph": "s", "id": 44623, "pid": 435, "tid": 435, "ts": 1776649790933278, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790933308, "dur": 0, + "args": { + "External id": 5430, + "cbid": 210, "correlation": 44650 + } + }, + { + "ph": "f", "id": 44650, "pid": 435, "tid": 435, "ts": 1776649790933308, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790933318, "dur": 23, + "args": { + "External id": 5430, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44651, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44651, "pid": 0, "tid": 7, "ts": 1776649790933318, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790933309, "dur": 5, + "args": { + "External id": 5430, + "cbid": 211, "correlation": 44651 + } + }, + { + "ph": "s", "id": 44651, "pid": 435, "tid": 435, "ts": 1776649790933309, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790933337, "dur": 0, + "args": { + "External id": 5432, + "cbid": 210, "correlation": 44678 + } + }, + { + "ph": "f", "id": 44678, "pid": 435, "tid": 435, "ts": 1776649790933337, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790933348, "dur": 23, + "args": { + "External id": 5432, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44679, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44679, "pid": 0, "tid": 7, "ts": 1776649790933348, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790933339, "dur": 5, + "args": { + "External id": 5432, + "cbid": 211, "correlation": 44679 + } + }, + { + "ph": "s", "id": 44679, "pid": 435, "tid": 435, "ts": 1776649790933339, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790933366, "dur": 0, + "args": { + "External id": 5434, + "cbid": 210, "correlation": 44706 + } + }, + { + "ph": "f", "id": 44706, "pid": 435, "tid": 435, "ts": 1776649790933366, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790933378, "dur": 23, + "args": { + "External id": 5434, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44707, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44707, "pid": 0, "tid": 7, "ts": 1776649790933378, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790933367, "dur": 6, + "args": { + "External id": 5434, + "cbid": 211, "correlation": 44707 + } + }, + { + "ph": "s", "id": 44707, "pid": 435, "tid": 435, "ts": 1776649790933367, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790933396, "dur": 0, + "args": { + "External id": 5436, + "cbid": 210, "correlation": 44734 + } + }, + { + "ph": "f", "id": 44734, "pid": 435, "tid": 435, "ts": 1776649790933396, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790933407, "dur": 23, + "args": { + "External id": 5436, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44735, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44735, "pid": 0, "tid": 7, "ts": 1776649790933407, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790933397, "dur": 6, + "args": { + "External id": 5436, + "cbid": 211, "correlation": 44735 + } + }, + { + "ph": "s", "id": 44735, "pid": 435, "tid": 435, "ts": 1776649790933397, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790933425, "dur": 0, + "args": { + "External id": 5438, + "cbid": 210, "correlation": 44762 + } + }, + { + "ph": "f", "id": 44762, "pid": 435, "tid": 435, "ts": 1776649790933425, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790933436, "dur": 23, + "args": { + "External id": 5438, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44763, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44763, "pid": 0, "tid": 7, "ts": 1776649790933436, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790933426, "dur": 5, + "args": { + "External id": 5438, + "cbid": 211, "correlation": 44763 + } + }, + { + "ph": "s", "id": 44763, "pid": 435, "tid": 435, "ts": 1776649790933426, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649790933454, "dur": 0, + "args": { + "External id": 5440, + "cbid": 210, "correlation": 44790 + } + }, + { + "ph": "f", "id": 44790, "pid": 435, "tid": 435, "ts": 1776649790933454, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_64x64_sliced1x2_ldg8_f2f_stages_64x6_nn", "pid": 0, "tid": 7, + "ts": 1776649790933465, "dur": 23, + "args": { + "External id": 5440, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44791, + "registers per thread": 146, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 44791, "pid": 0, "tid": 7, "ts": 1776649790933465, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790933455, "dur": 5, + "args": { + "External id": 5440, + "cbid": 211, "correlation": 44791 + } + }, + { + "ph": "s", "id": 44791, "pid": 435, "tid": 435, "ts": 1776649790933455, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649790933488, "dur": 4, + "args": { + "External id": 44802, + "cbid": 135, "correlation": 44802 + } + }, + { + "ph": "f", "id": 44802, "pid": 435, "tid": 435, "ts": 1776649790933488, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649790933503, "dur": 5, + "args": { + "External id": 44807, + "cbid": 165, "correlation": 44807 + } + }, + { + "ph": "s", "id": 44807, "pid": 435, "tid": 435, "ts": 1776649790933503, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649790933511, "dur": 4, + "args": { + "External id": 44810, + "cbid": 139, "correlation": 44810 + } + }, + { + "ph": "f", "id": 44810, "pid": 435, "tid": 435, "ts": 1776649790933511, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor >, at::detail::Array >(int, at::native::AUnaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790933606, "dur": 18, + "args": { + "External id": 5443, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44827, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 44827, "pid": 0, "tid": 7, "ts": 1776649790933606, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790933594, "dur": 10, + "args": { + "External id": 5443, + "cbid": 211, "correlation": 44827 + } + }, + { + "ph": "s", "id": 44827, "pid": 435, "tid": 435, "ts": 1776649790933594, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649790933648, "dur": 29, + "args": { + "External id": 5447, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44840, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 44840, "pid": 0, "tid": 7, "ts": 1776649790933648, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790933638, "dur": 8, + "args": { + "External id": 5447, + "cbid": 211, "correlation": 44840 + } + }, + { + "ph": "s", "id": 44840, "pid": 435, "tid": 435, "ts": 1776649790933638, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BUnaryFunctor, at::detail::Array >(int, at::native::BUnaryFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790933685, "dur": 17, + "args": { + "External id": 5448, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44854, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 44854, "pid": 0, "tid": 7, "ts": 1776649790933685, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790933674, "dur": 7, + "args": { + "External id": 5448, + "cbid": 211, "correlation": 44854 + } + }, + { + "ph": "s", "id": 44854, "pid": 435, "tid": 435, "ts": 1776649790933674, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649790933710, "dur": 29, + "args": { + "External id": 5453, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44867, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 151.703705, + "warps per SM": 606.814819, + "grid": [16384, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 44867, "pid": 0, "tid": 7, "ts": 1776649790933710, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790933701, "dur": 6, + "args": { + "External id": 5453, + "cbid": 211, "correlation": 44867 + } + }, + { + "ph": "s", "id": 44867, "pid": 435, "tid": 435, "ts": 1776649790933701, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy_contig, unsigned int, 3, 128, 1>(at::native::(anonymous namespace)::OpaqueType<1u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata, unsigned int, 128, 1>, at::native::(anonymous namespace)::TensorSizeStride, int, unsigned int)", "pid": 0, "tid": 7, + "ts": 1776649790933803, "dur": 63, + "args": { + "External id": 5459, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44879, + "registers per thread": 21, + "shared memory": 0, + "blocks per SM": 4.000000, + "warps per SM": 64.000000, + "grid": [216, 2, 1], + "block": [512, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 44879, "pid": 0, "tid": 7, "ts": 1776649790933803, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790933792, "dur": 9, + "args": { + "External id": 5459, + "cbid": 211, "correlation": 44879 + } + }, + { + "ph": "s", "id": 44879, "pid": 435, "tid": 435, "ts": 1776649790933792, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649790933876, "dur": 72, + "args": { + "External id": 5470, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44892, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 44892, "pid": 0, "tid": 7, "ts": 1776649790933876, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790933861, "dur": 10, + "args": { + "External id": 5470, + "cbid": 211, "correlation": 44892 + } + }, + { + "ph": "s", "id": 44892, "pid": 435, "tid": 435, "ts": 1776649790933861, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, at::detail::Array >(int, at::native::CUDAFunctorOnSelf_add, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790933949, "dur": 33, + "args": { + "External id": 5473, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44903, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 44903, "pid": 0, "tid": 7, "ts": 1776649790933949, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790933902, "dur": 8, + "args": { + "External id": 5473, + "cbid": 211, "correlation": 44903 + } + }, + { + "ph": "s", "id": 44903, "pid": 435, "tid": 435, "ts": 1776649790933902, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649790933983, "dur": 91, + "args": { + "External id": 5477, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44916, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 44916, "pid": 0, "tid": 7, "ts": 1776649790933983, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790933933, "dur": 7, + "args": { + "External id": 5477, + "cbid": 211, "correlation": 44916 + } + }, + { + "ph": "s", "id": 44916, "pid": 435, "tid": 435, "ts": 1776649790933933, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor >, at::detail::Array >(int, at::native::BinaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649790934076, "dur": 60, + "args": { + "External id": 5478, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44927, + "registers per thread": 22, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 44927, "pid": 0, "tid": 7, "ts": 1776649790934076, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790933960, "dur": 6, + "args": { + "External id": 5478, + "cbid": 211, "correlation": 44927 + } + }, + { + "ph": "s", "id": 44927, "pid": 435, "tid": 435, "ts": 1776649790933960, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649790934138, "dur": 221, + "args": { + "External id": 5486, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 44941, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 44941, "pid": 0, "tid": 7, "ts": 1776649790934138, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649790934017, "dur": 7, + "args": { + "External id": 5486, + "cbid": 211, "correlation": 44941 + } + }, + { + "ph": "s", "id": 44941, "pid": 435, "tid": 435, "ts": 1776649790934017, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649791072911, "dur": 3622, + "args": { + "External id": 5502, + "device": 0, "context": 1, + "stream": 7, "correlation": 44957, + "bytes": 33554432, "memory bandwidth (GB/s)": 9.263898153097239 + } + }, + { + "ph": "f", "id": 44957, "pid": 0, "tid": 7, "ts": 1776649791072911, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649791072766, "dur": 3733, + "args": { + "External id": 5502, + "cbid": 41, "correlation": 44957 + } + }, + { + "ph": "s", "id": 44957, "pid": 435, "tid": 435, "ts": 1776649791072766, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791076507, "dur": 37, + "args": { + "External id": 5502, + "cbid": 131, "correlation": 44958 + } + }, + { + "ph": "s", "id": 44958, "pid": 435, "tid": 435, "ts": 1776649791076507, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7, + "ts": 1776649791177512, "dur": 40, + "args": { + "External id": 5505, + "device": 0, "context": 1, + "stream": 7, "correlation": 44963, + "bytes": 33554432, "memory bandwidth (GB/s)": 836.8523543495611 + } + }, + { + "ph": "f", "id": 44963, "pid": 0, "tid": 7, "ts": 1776649791177512, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649791177440, "dur": 72, + "args": { + "External id": 5505, + "cbid": 41, "correlation": 44963 + } + }, + { + "ph": "s", "id": 44963, "pid": 435, "tid": 435, "ts": 1776649791177440, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649791178099, "dur": 2102, + "args": { + "External id": 5514, + "device": 0, "context": 1, + "stream": 7, "correlation": 44978, + "bytes": 8388608, "memory bandwidth (GB/s)": 3.9893036582611074 + } + }, + { + "ph": "f", "id": 44978, "pid": 0, "tid": 7, "ts": 1776649791178099, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649791177804, "dur": 2340, + "args": { + "External id": 5514, + "cbid": 41, "correlation": 44978 + } + }, + { + "ph": "s", "id": 44978, "pid": 435, "tid": 435, "ts": 1776649791177804, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791180146, "dur": 65, + "args": { + "External id": 5514, + "cbid": 131, "correlation": 44979 + } + }, + { + "ph": "s", "id": 44979, "pid": 435, "tid": 435, "ts": 1776649791180146, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649791180297, "dur": 17, + "args": { + "External id": 5518, + "device": 0, "context": 1, + "stream": 7, "correlation": 44991, + "bytes": 262144, "memory bandwidth (GB/s)": 15.086556169429098 + } + }, + { + "ph": "f", "id": 44991, "pid": 0, "tid": 7, "ts": 1776649791180297, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649791180266, "dur": 29, + "args": { + "External id": 5518, + "cbid": 41, "correlation": 44991 + } + }, + { + "ph": "s", "id": 44991, "pid": 435, "tid": 435, "ts": 1776649791180266, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791180295, "dur": 26, + "args": { + "External id": 5518, + "cbid": 131, "correlation": 44992 + } + }, + { + "ph": "s", "id": 44992, "pid": 435, "tid": 435, "ts": 1776649791180295, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649791180354, "dur": 1, + "args": { + "External id": 5522, + "device": 0, "context": 1, + "stream": 7, "correlation": 45004, + "bytes": 2048, "memory bandwidth (GB/s)": 1.4883720930232558 + } + }, + { + "ph": "f", "id": 45004, "pid": 0, "tid": 7, "ts": 1776649791180354, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649791180344, "dur": 6, + "args": { + "External id": 5522, + "cbid": 41, "correlation": 45004 + } + }, + { + "ph": "s", "id": 45004, "pid": 435, "tid": 435, "ts": 1776649791180344, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791180350, "dur": 10, + "args": { + "External id": 5522, + "cbid": 131, "correlation": 45005 + } + }, + { + "ph": "s", "id": 45005, "pid": 435, "tid": 435, "ts": 1776649791180350, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649791180521, "dur": 229, + "args": { + "External id": 5537, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45017, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 45017, "pid": 0, "tid": 7, "ts": 1776649791180521, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791180501, "dur": 21, + "args": { + "External id": 5537, + "cbid": 211, "correlation": 45017 + } + }, + { + "ph": "s", "id": 45017, "pid": 435, "tid": 435, "ts": 1776649791180501, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649791180752, "dur": 6, + "args": { + "External id": 5543, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45031, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 2.370370, + "warps per SM": 9.481482, + "grid": [256, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 15 + } + }, + { + "ph": "f", "id": 45031, "pid": 0, "tid": 7, "ts": 1776649791180752, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791180560, "dur": 8, + "args": { + "External id": 5543, + "cbid": 211, "correlation": 45031 + } + }, + { + "ph": "s", "id": 45031, "pid": 435, "tid": 435, "ts": 1776649791180560, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649791180759, "dur": 126, + "args": { + "External id": 5545, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45042, + "registers per thread": 20, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 45042, "pid": 0, "tid": 7, "ts": 1776649791180759, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791180607, "dur": 8, + "args": { + "External id": 5545, + "cbid": 211, "correlation": 45042 + } + }, + { + "ph": "s", "id": 45042, "pid": 435, "tid": 435, "ts": 1776649791180607, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791180887, "dur": 39, + "args": { + "External id": 5546, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45052, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 45052, "pid": 0, "tid": 7, "ts": 1776649791180887, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791180639, "dur": 7, + "args": { + "External id": 5546, + "cbid": 211, "correlation": 45052 + } + }, + { + "ph": "s", "id": 45052, "pid": 435, "tid": 435, "ts": 1776649791180639, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649791180928, "dur": 98, + "args": { + "External id": 5550, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45065, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 45065, "pid": 0, "tid": 7, "ts": 1776649791180928, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791180673, "dur": 9, + "args": { + "External id": 5550, + "cbid": 211, "correlation": 45065 + } + }, + { + "ph": "s", "id": 45065, "pid": 435, "tid": 435, "ts": 1776649791180673, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, at::detail::Array >(int, at::native::CUDAFunctorOnSelf_add, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791181028, "dur": 75, + "args": { + "External id": 5551, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45072, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 45072, "pid": 0, "tid": 7, "ts": 1776649791181028, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791180718, "dur": 9, + "args": { + "External id": 5551, + "cbid": 211, "correlation": 45072 + } + }, + { + "ph": "s", "id": 45072, "pid": 435, "tid": 435, "ts": 1776649791180718, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1}, at::detail::Array >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791181105, "dur": 78, + "args": { + "External id": 5552, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45082, + "registers per thread": 22, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 45082, "pid": 0, "tid": 7, "ts": 1776649791181105, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791180760, "dur": 7, + "args": { + "External id": 5552, + "cbid": 211, "correlation": 45082 + } + }, + { + "ph": "s", "id": 45082, "pid": 435, "tid": 435, "ts": 1776649791180760, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649791181184, "dur": 102, + "args": { + "External id": 5562, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45095, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 45095, "pid": 0, "tid": 7, "ts": 1776649791181184, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791180817, "dur": 8, + "args": { + "External id": 5562, + "cbid": 211, "correlation": 45095 + } + }, + { + "ph": "s", "id": 45095, "pid": 435, "tid": 435, "ts": 1776649791180817, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649791181290, "dur": 1, + "args": { + "External id": 5572, + "device": 0, "context": 1, + "stream": 7, "correlation": 45110, + "bytes": 512, "memory bandwidth (GB/s)": 0.4 + } + }, + { + "ph": "f", "id": 45110, "pid": 0, "tid": 7, "ts": 1776649791181290, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649791180909, "dur": 10, + "args": { + "External id": 5572, + "cbid": 41, "correlation": 45110 + } + }, + { + "ph": "s", "id": 45110, "pid": 435, "tid": 435, "ts": 1776649791180909, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791180920, "dur": 377, + "args": { + "External id": 5572, + "cbid": 131, "correlation": 45111 + } + }, + { + "ph": "s", "id": 45111, "pid": 435, "tid": 435, "ts": 1776649791180920, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649791181361, "dur": 9, + "args": { + "External id": 5574, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45120, + "registers per thread": 40, + "shared memory": 0, + "blocks per SM": 2.370370, + "warps per SM": 9.481482, + "grid": [256, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 15 + } + }, + { + "ph": "f", "id": 45120, "pid": 0, "tid": 7, "ts": 1776649791181361, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791181347, "dur": 12, + "args": { + "External id": 5574, + "cbid": 211, "correlation": 45120 + } + }, + { + "ph": "s", "id": 45120, "pid": 435, "tid": 435, "ts": 1776649791181347, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#3}::operator()() const::{lambda(int)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649791181428, "dur": 111, + "args": { + "External id": 5588, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45135, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 45135, "pid": 0, "tid": 7, "ts": 1776649791181428, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791181417, "dur": 8, + "args": { + "External id": 5588, + "cbid": 211, "correlation": 45135 + } + }, + { + "ph": "s", "id": 45135, "pid": 435, "tid": 435, "ts": 1776649791181417, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649791181542, "dur": 1, + "args": { + "External id": 5598, + "device": 0, "context": 1, + "stream": 7, "correlation": 45150, + "bytes": 8192, "memory bandwidth (GB/s)": 4.491228070175438 + } + }, + { + "ph": "f", "id": 45150, "pid": 0, "tid": 7, "ts": 1776649791181542, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649791181492, "dur": 10, + "args": { + "External id": 5598, + "cbid": 41, "correlation": 45150 + } + }, + { + "ph": "s", "id": 45150, "pid": 435, "tid": 435, "ts": 1776649791181492, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791181502, "dur": 47, + "args": { + "External id": 5598, + "cbid": 131, "correlation": 45151 + } + }, + { + "ph": "s", "id": 45151, "pid": 435, "tid": 435, "ts": 1776649791181502, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::index_elementwise_kernel<128, 4, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1}>(long, at::native::gpu_index_kernel >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1}>(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef, at::native::index_kernel_impl >(at::TensorIteratorBase&, c10::ArrayRef, c10::ArrayRef)::{lambda(char*, char const*, long)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649791181581, "dur": 127, + "args": { + "External id": 5594, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45160, + "registers per thread": 40, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 75 + } + }, + { + "ph": "f", "id": 45160, "pid": 0, "tid": 7, "ts": 1776649791181581, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791181569, "dur": 11, + "args": { + "External id": 5594, + "cbid": 211, "correlation": 45160 + } + }, + { + "ph": "s", "id": 45160, "pid": 435, "tid": 435, "ts": 1776649791181569, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoH (Device -> Pageable)", "pid": 0, "tid": 7, + "ts": 1776649791195727, "dur": 47705, + "args": { + "External id": 5607, + "device": 0, "context": 1, + "stream": 7, "correlation": 45168, + "bytes": 67108864, "memory bandwidth (GB/s)": 1.4067392103047414 + } + }, + { + "ph": "f", "id": 45168, "pid": 0, "tid": 7, "ts": 1776649791195727, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649791195704, "dur": 48985, + "args": { + "External id": 5607, + "cbid": 41, "correlation": 45168 + } + }, + { + "ph": "s", "id": 45168, "pid": 435, "tid": 435, "ts": 1776649791195704, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791244692, "dur": 16, + "args": { + "External id": 5607, + "cbid": 131, "correlation": 45169 + } + }, + { + "ph": "s", "id": 45169, "pid": 435, "tid": 435, "ts": 1776649791244692, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy HtoD (Pageable -> Device)", "pid": 0, "tid": 7, + "ts": 1776649791305150, "dur": 515, + "args": { + "External id": 5617, + "device": 0, "context": 1, + "stream": 7, "correlation": 45181, + "bytes": 8388608, "memory bandwidth (GB/s)": 16.267143449982353 + } + }, + { + "ph": "f", "id": 45181, "pid": 0, "tid": 7, "ts": 1776649791305150, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649791305002, "dur": 607, + "args": { + "External id": 5617, + "cbid": 41, "correlation": 45181 + } + }, + { + "ph": "s", "id": 45181, "pid": 435, "tid": 435, "ts": 1776649791305002, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791305611, "dur": 64, + "args": { + "External id": 5617, + "cbid": 131, "correlation": 45182 + } + }, + { + "ph": "s", "id": 45182, "pid": 435, "tid": 435, "ts": 1776649791305611, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7, + "ts": 1776649791305876, "dur": 9, + "args": { + "External id": 5623, + "device": 0, "context": 1, + "stream": 7, "correlation": 45187, + "bytes": 8388608, "memory bandwidth (GB/s)": 870.9102990033223 + } + }, + { + "ph": "f", "id": 45187, "pid": 0, "tid": 7, "ts": 1776649791305876, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649791305833, "dur": 40, + "args": { + "External id": 5623, + "cbid": 41, "correlation": 45187 + } + }, + { + "ph": "s", "id": 45187, "pid": 435, "tid": 435, "ts": 1776649791305833, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 0, "tid": 7, + "ts": 1776649791305928, "dur": 3, + "args": { + "External id": 5629, + "device": 0, "context": 1, + "stream": 7, "correlation": 45195, + "bytes": 262144, "memory bandwidth (GB/s)": 79.53398058252426 + } + }, + { + "ph": "f", "id": 45195, "pid": 0, "tid": 7, "ts": 1776649791305928, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMemcpyAsync", "pid": 435, "tid": 435, + "ts": 1776649791305911, "dur": 14, + "args": { + "External id": 5629, + "cbid": 41, "correlation": 45195 + } + }, + { + "ph": "s", "id": 45195, "pid": 435, "tid": 435, "ts": 1776649791305911, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791306682, "dur": 1, + "args": { + "External id": 5634, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45218, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45218, "pid": 0, "tid": 7, "ts": 1776649791306682, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791306658, "dur": 25, + "args": { + "External id": 5634, + "cbid": 211, "correlation": 45218 + } + }, + { + "ph": "s", "id": 45218, "pid": 435, "tid": 435, "ts": 1776649791306658, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791306702, "dur": 2, + "args": { + "External id": 45224, + "cbid": 200, "correlation": 45224 + } + }, + { + "ph": "f", "id": 45224, "pid": 435, "tid": 435, "ts": 1776649791306702, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791306705, "dur": 3, + "args": { + "External id": 45225, + "cbid": 273, "correlation": 45225 + } + }, + { + "ph": "f", "id": 45225, "pid": 435, "tid": 435, "ts": 1776649791306705, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791306719, "dur": 16, + "args": { + "External id": 45226, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45226, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45226, "pid": 0, "tid": 7, "ts": 1776649791306719, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791306710, "dur": 6, + "args": { + "External id": 45226, + "cbid": 211, "correlation": 45226 + } + }, + { + "ph": "s", "id": 45226, "pid": 435, "tid": 435, "ts": 1776649791306710, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791306765, "dur": 1, + "args": { + "External id": 5639, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45244, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45244, "pid": 0, "tid": 7, "ts": 1776649791306765, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791306755, "dur": 7, + "args": { + "External id": 5639, + "cbid": 211, "correlation": 45244 + } + }, + { + "ph": "s", "id": 45244, "pid": 435, "tid": 435, "ts": 1776649791306755, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791306769, "dur": 0, + "args": { + "External id": 45250, + "cbid": 200, "correlation": 45250 + } + }, + { + "ph": "f", "id": 45250, "pid": 435, "tid": 435, "ts": 1776649791306769, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791306769, "dur": 0, + "args": { + "External id": 45251, + "cbid": 273, "correlation": 45251 + } + }, + { + "ph": "f", "id": 45251, "pid": 435, "tid": 435, "ts": 1776649791306769, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791306779, "dur": 15, + "args": { + "External id": 45252, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45252, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45252, "pid": 0, "tid": 7, "ts": 1776649791306779, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791306771, "dur": 5, + "args": { + "External id": 45252, + "cbid": 211, "correlation": 45252 + } + }, + { + "ph": "s", "id": 45252, "pid": 435, "tid": 435, "ts": 1776649791306771, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791306816, "dur": 1, + "args": { + "External id": 5644, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45270, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45270, "pid": 0, "tid": 7, "ts": 1776649791306816, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791306806, "dur": 7, + "args": { + "External id": 5644, + "cbid": 211, "correlation": 45270 + } + }, + { + "ph": "s", "id": 45270, "pid": 435, "tid": 435, "ts": 1776649791306806, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791306819, "dur": 0, + "args": { + "External id": 45276, + "cbid": 200, "correlation": 45276 + } + }, + { + "ph": "f", "id": 45276, "pid": 435, "tid": 435, "ts": 1776649791306819, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791306820, "dur": 0, + "args": { + "External id": 45277, + "cbid": 273, "correlation": 45277 + } + }, + { + "ph": "f", "id": 45277, "pid": 435, "tid": 435, "ts": 1776649791306820, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791306829, "dur": 15, + "args": { + "External id": 45278, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45278, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45278, "pid": 0, "tid": 7, "ts": 1776649791306829, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791306821, "dur": 4, + "args": { + "External id": 45278, + "cbid": 211, "correlation": 45278 + } + }, + { + "ph": "s", "id": 45278, "pid": 435, "tid": 435, "ts": 1776649791306821, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791306865, "dur": 1, + "args": { + "External id": 5649, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45296, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45296, "pid": 0, "tid": 7, "ts": 1776649791306865, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791306856, "dur": 6, + "args": { + "External id": 5649, + "cbid": 211, "correlation": 45296 + } + }, + { + "ph": "s", "id": 45296, "pid": 435, "tid": 435, "ts": 1776649791306856, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791306869, "dur": 0, + "args": { + "External id": 45302, + "cbid": 200, "correlation": 45302 + } + }, + { + "ph": "f", "id": 45302, "pid": 435, "tid": 435, "ts": 1776649791306869, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791306869, "dur": 0, + "args": { + "External id": 45303, + "cbid": 273, "correlation": 45303 + } + }, + { + "ph": "f", "id": 45303, "pid": 435, "tid": 435, "ts": 1776649791306869, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791306878, "dur": 17, + "args": { + "External id": 45304, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45304, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45304, "pid": 0, "tid": 7, "ts": 1776649791306878, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791306870, "dur": 5, + "args": { + "External id": 45304, + "cbid": 211, "correlation": 45304 + } + }, + { + "ph": "s", "id": 45304, "pid": 435, "tid": 435, "ts": 1776649791306870, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791306914, "dur": 1, + "args": { + "External id": 5654, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45322, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45322, "pid": 0, "tid": 7, "ts": 1776649791306914, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791306905, "dur": 6, + "args": { + "External id": 5654, + "cbid": 211, "correlation": 45322 + } + }, + { + "ph": "s", "id": 45322, "pid": 435, "tid": 435, "ts": 1776649791306905, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791306917, "dur": 0, + "args": { + "External id": 45328, + "cbid": 200, "correlation": 45328 + } + }, + { + "ph": "f", "id": 45328, "pid": 435, "tid": 435, "ts": 1776649791306917, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791306918, "dur": 0, + "args": { + "External id": 45329, + "cbid": 273, "correlation": 45329 + } + }, + { + "ph": "f", "id": 45329, "pid": 435, "tid": 435, "ts": 1776649791306918, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791306927, "dur": 16, + "args": { + "External id": 45330, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45330, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45330, "pid": 0, "tid": 7, "ts": 1776649791306927, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791306919, "dur": 4, + "args": { + "External id": 45330, + "cbid": 211, "correlation": 45330 + } + }, + { + "ph": "s", "id": 45330, "pid": 435, "tid": 435, "ts": 1776649791306919, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791306964, "dur": 1, + "args": { + "External id": 5659, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45348, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45348, "pid": 0, "tid": 7, "ts": 1776649791306964, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791306954, "dur": 7, + "args": { + "External id": 5659, + "cbid": 211, "correlation": 45348 + } + }, + { + "ph": "s", "id": 45348, "pid": 435, "tid": 435, "ts": 1776649791306954, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791306967, "dur": 0, + "args": { + "External id": 45354, + "cbid": 200, "correlation": 45354 + } + }, + { + "ph": "f", "id": 45354, "pid": 435, "tid": 435, "ts": 1776649791306967, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791306968, "dur": 0, + "args": { + "External id": 45355, + "cbid": 273, "correlation": 45355 + } + }, + { + "ph": "f", "id": 45355, "pid": 435, "tid": 435, "ts": 1776649791306968, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791306977, "dur": 15, + "args": { + "External id": 45356, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45356, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45356, "pid": 0, "tid": 7, "ts": 1776649791306977, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791306969, "dur": 4, + "args": { + "External id": 45356, + "cbid": 211, "correlation": 45356 + } + }, + { + "ph": "s", "id": 45356, "pid": 435, "tid": 435, "ts": 1776649791306969, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791307012, "dur": 1, + "args": { + "External id": 5664, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45374, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45374, "pid": 0, "tid": 7, "ts": 1776649791307012, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307003, "dur": 6, + "args": { + "External id": 5664, + "cbid": 211, "correlation": 45374 + } + }, + { + "ph": "s", "id": 45374, "pid": 435, "tid": 435, "ts": 1776649791307003, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307015, "dur": 0, + "args": { + "External id": 45380, + "cbid": 200, "correlation": 45380 + } + }, + { + "ph": "f", "id": 45380, "pid": 435, "tid": 435, "ts": 1776649791307015, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307016, "dur": 0, + "args": { + "External id": 45381, + "cbid": 273, "correlation": 45381 + } + }, + { + "ph": "f", "id": 45381, "pid": 435, "tid": 435, "ts": 1776649791307016, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791307024, "dur": 15, + "args": { + "External id": 45382, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45382, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45382, "pid": 0, "tid": 7, "ts": 1776649791307024, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307017, "dur": 4, + "args": { + "External id": 45382, + "cbid": 211, "correlation": 45382 + } + }, + { + "ph": "s", "id": 45382, "pid": 435, "tid": 435, "ts": 1776649791307017, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791307060, "dur": 1, + "args": { + "External id": 5669, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45400, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45400, "pid": 0, "tid": 7, "ts": 1776649791307060, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307051, "dur": 6, + "args": { + "External id": 5669, + "cbid": 211, "correlation": 45400 + } + }, + { + "ph": "s", "id": 45400, "pid": 435, "tid": 435, "ts": 1776649791307051, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307063, "dur": 0, + "args": { + "External id": 45406, + "cbid": 200, "correlation": 45406 + } + }, + { + "ph": "f", "id": 45406, "pid": 435, "tid": 435, "ts": 1776649791307063, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307064, "dur": 0, + "args": { + "External id": 45407, + "cbid": 273, "correlation": 45407 + } + }, + { + "ph": "f", "id": 45407, "pid": 435, "tid": 435, "ts": 1776649791307064, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791307073, "dur": 15, + "args": { + "External id": 45408, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45408, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45408, "pid": 0, "tid": 7, "ts": 1776649791307073, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307065, "dur": 5, + "args": { + "External id": 45408, + "cbid": 211, "correlation": 45408 + } + }, + { + "ph": "s", "id": 45408, "pid": 435, "tid": 435, "ts": 1776649791307065, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791307108, "dur": 1, + "args": { + "External id": 5674, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45426, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45426, "pid": 0, "tid": 7, "ts": 1776649791307108, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307099, "dur": 6, + "args": { + "External id": 5674, + "cbid": 211, "correlation": 45426 + } + }, + { + "ph": "s", "id": 45426, "pid": 435, "tid": 435, "ts": 1776649791307099, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307112, "dur": 0, + "args": { + "External id": 45432, + "cbid": 200, "correlation": 45432 + } + }, + { + "ph": "f", "id": 45432, "pid": 435, "tid": 435, "ts": 1776649791307112, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307112, "dur": 0, + "args": { + "External id": 45433, + "cbid": 273, "correlation": 45433 + } + }, + { + "ph": "f", "id": 45433, "pid": 435, "tid": 435, "ts": 1776649791307112, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791307121, "dur": 15, + "args": { + "External id": 45434, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45434, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45434, "pid": 0, "tid": 7, "ts": 1776649791307121, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307113, "dur": 4, + "args": { + "External id": 45434, + "cbid": 211, "correlation": 45434 + } + }, + { + "ph": "s", "id": 45434, "pid": 435, "tid": 435, "ts": 1776649791307113, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791307156, "dur": 1, + "args": { + "External id": 5679, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45452, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45452, "pid": 0, "tid": 7, "ts": 1776649791307156, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307147, "dur": 6, + "args": { + "External id": 5679, + "cbid": 211, "correlation": 45452 + } + }, + { + "ph": "s", "id": 45452, "pid": 435, "tid": 435, "ts": 1776649791307147, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307159, "dur": 0, + "args": { + "External id": 45458, + "cbid": 200, "correlation": 45458 + } + }, + { + "ph": "f", "id": 45458, "pid": 435, "tid": 435, "ts": 1776649791307159, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307160, "dur": 0, + "args": { + "External id": 45459, + "cbid": 273, "correlation": 45459 + } + }, + { + "ph": "f", "id": 45459, "pid": 435, "tid": 435, "ts": 1776649791307160, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791307169, "dur": 15, + "args": { + "External id": 45460, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45460, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45460, "pid": 0, "tid": 7, "ts": 1776649791307169, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307161, "dur": 4, + "args": { + "External id": 45460, + "cbid": 211, "correlation": 45460 + } + }, + { + "ph": "s", "id": 45460, "pid": 435, "tid": 435, "ts": 1776649791307161, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791307239, "dur": 13, + "args": { + "External id": 45463, + "cbid": 165, "correlation": 45463 + } + }, + { + "ph": "s", "id": 45463, "pid": 435, "tid": 435, "ts": 1776649791307239, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791307300, "dur": 8, + "args": { + "External id": 45472, + "cbid": 135, "correlation": 45472 + } + }, + { + "ph": "f", "id": 45472, "pid": 435, "tid": 435, "ts": 1776649791307300, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791307358, "dur": 1, + "args": { + "External id": 5684, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45492, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45492, "pid": 0, "tid": 7, "ts": 1776649791307358, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307347, "dur": 9, + "args": { + "External id": 5684, + "cbid": 211, "correlation": 45492 + } + }, + { + "ph": "s", "id": 45492, "pid": 435, "tid": 435, "ts": 1776649791307347, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307363, "dur": 0, + "args": { + "External id": 45498, + "cbid": 200, "correlation": 45498 + } + }, + { + "ph": "f", "id": 45498, "pid": 435, "tid": 435, "ts": 1776649791307363, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307364, "dur": 0, + "args": { + "External id": 45499, + "cbid": 273, "correlation": 45499 + } + }, + { + "ph": "f", "id": 45499, "pid": 435, "tid": 435, "ts": 1776649791307364, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791307373, "dur": 15, + "args": { + "External id": 45500, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45500, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45500, "pid": 0, "tid": 7, "ts": 1776649791307373, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307365, "dur": 5, + "args": { + "External id": 45500, + "cbid": 211, "correlation": 45500 + } + }, + { + "ph": "s", "id": 45500, "pid": 435, "tid": 435, "ts": 1776649791307365, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791307412, "dur": 1, + "args": { + "External id": 5689, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45518, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45518, "pid": 0, "tid": 7, "ts": 1776649791307412, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307403, "dur": 6, + "args": { + "External id": 5689, + "cbid": 211, "correlation": 45518 + } + }, + { + "ph": "s", "id": 45518, "pid": 435, "tid": 435, "ts": 1776649791307403, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307416, "dur": 0, + "args": { + "External id": 45524, + "cbid": 200, "correlation": 45524 + } + }, + { + "ph": "f", "id": 45524, "pid": 435, "tid": 435, "ts": 1776649791307416, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307416, "dur": 0, + "args": { + "External id": 45525, + "cbid": 273, "correlation": 45525 + } + }, + { + "ph": "f", "id": 45525, "pid": 435, "tid": 435, "ts": 1776649791307416, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791307425, "dur": 15, + "args": { + "External id": 45526, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45526, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45526, "pid": 0, "tid": 7, "ts": 1776649791307425, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307417, "dur": 5, + "args": { + "External id": 45526, + "cbid": 211, "correlation": 45526 + } + }, + { + "ph": "s", "id": 45526, "pid": 435, "tid": 435, "ts": 1776649791307417, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791307462, "dur": 1, + "args": { + "External id": 5694, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45544, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45544, "pid": 0, "tid": 7, "ts": 1776649791307462, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307452, "dur": 6, + "args": { + "External id": 5694, + "cbid": 211, "correlation": 45544 + } + }, + { + "ph": "s", "id": 45544, "pid": 435, "tid": 435, "ts": 1776649791307452, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307465, "dur": 0, + "args": { + "External id": 45550, + "cbid": 200, "correlation": 45550 + } + }, + { + "ph": "f", "id": 45550, "pid": 435, "tid": 435, "ts": 1776649791307465, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307466, "dur": 0, + "args": { + "External id": 45551, + "cbid": 273, "correlation": 45551 + } + }, + { + "ph": "f", "id": 45551, "pid": 435, "tid": 435, "ts": 1776649791307466, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791307475, "dur": 16, + "args": { + "External id": 45552, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45552, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45552, "pid": 0, "tid": 7, "ts": 1776649791307475, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307467, "dur": 5, + "args": { + "External id": 45552, + "cbid": 211, "correlation": 45552 + } + }, + { + "ph": "s", "id": 45552, "pid": 435, "tid": 435, "ts": 1776649791307467, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791307510, "dur": 1, + "args": { + "External id": 5699, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45570, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45570, "pid": 0, "tid": 7, "ts": 1776649791307510, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307501, "dur": 6, + "args": { + "External id": 5699, + "cbid": 211, "correlation": 45570 + } + }, + { + "ph": "s", "id": 45570, "pid": 435, "tid": 435, "ts": 1776649791307501, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307513, "dur": 0, + "args": { + "External id": 45576, + "cbid": 200, "correlation": 45576 + } + }, + { + "ph": "f", "id": 45576, "pid": 435, "tid": 435, "ts": 1776649791307513, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307514, "dur": 0, + "args": { + "External id": 45577, + "cbid": 273, "correlation": 45577 + } + }, + { + "ph": "f", "id": 45577, "pid": 435, "tid": 435, "ts": 1776649791307514, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791307523, "dur": 15, + "args": { + "External id": 45578, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45578, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45578, "pid": 0, "tid": 7, "ts": 1776649791307523, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307515, "dur": 5, + "args": { + "External id": 45578, + "cbid": 211, "correlation": 45578 + } + }, + { + "ph": "s", "id": 45578, "pid": 435, "tid": 435, "ts": 1776649791307515, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791307559, "dur": 1, + "args": { + "External id": 5704, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45596, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45596, "pid": 0, "tid": 7, "ts": 1776649791307559, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307550, "dur": 6, + "args": { + "External id": 5704, + "cbid": 211, "correlation": 45596 + } + }, + { + "ph": "s", "id": 45596, "pid": 435, "tid": 435, "ts": 1776649791307550, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307562, "dur": 0, + "args": { + "External id": 45602, + "cbid": 200, "correlation": 45602 + } + }, + { + "ph": "f", "id": 45602, "pid": 435, "tid": 435, "ts": 1776649791307562, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307563, "dur": 0, + "args": { + "External id": 45603, + "cbid": 273, "correlation": 45603 + } + }, + { + "ph": "f", "id": 45603, "pid": 435, "tid": 435, "ts": 1776649791307563, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791307572, "dur": 15, + "args": { + "External id": 45604, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45604, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45604, "pid": 0, "tid": 7, "ts": 1776649791307572, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307564, "dur": 5, + "args": { + "External id": 45604, + "cbid": 211, "correlation": 45604 + } + }, + { + "ph": "s", "id": 45604, "pid": 435, "tid": 435, "ts": 1776649791307564, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791307608, "dur": 1, + "args": { + "External id": 5709, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45622, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45622, "pid": 0, "tid": 7, "ts": 1776649791307608, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307599, "dur": 6, + "args": { + "External id": 5709, + "cbid": 211, "correlation": 45622 + } + }, + { + "ph": "s", "id": 45622, "pid": 435, "tid": 435, "ts": 1776649791307599, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307611, "dur": 0, + "args": { + "External id": 45628, + "cbid": 200, "correlation": 45628 + } + }, + { + "ph": "f", "id": 45628, "pid": 435, "tid": 435, "ts": 1776649791307611, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307612, "dur": 0, + "args": { + "External id": 45629, + "cbid": 273, "correlation": 45629 + } + }, + { + "ph": "f", "id": 45629, "pid": 435, "tid": 435, "ts": 1776649791307612, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791307621, "dur": 15, + "args": { + "External id": 45630, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45630, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45630, "pid": 0, "tid": 7, "ts": 1776649791307621, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307613, "dur": 4, + "args": { + "External id": 45630, + "cbid": 211, "correlation": 45630 + } + }, + { + "ph": "s", "id": 45630, "pid": 435, "tid": 435, "ts": 1776649791307613, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791307656, "dur": 1, + "args": { + "External id": 5714, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45648, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45648, "pid": 0, "tid": 7, "ts": 1776649791307656, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307647, "dur": 6, + "args": { + "External id": 5714, + "cbid": 211, "correlation": 45648 + } + }, + { + "ph": "s", "id": 45648, "pid": 435, "tid": 435, "ts": 1776649791307647, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307659, "dur": 0, + "args": { + "External id": 45654, + "cbid": 200, "correlation": 45654 + } + }, + { + "ph": "f", "id": 45654, "pid": 435, "tid": 435, "ts": 1776649791307659, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307660, "dur": 0, + "args": { + "External id": 45655, + "cbid": 273, "correlation": 45655 + } + }, + { + "ph": "f", "id": 45655, "pid": 435, "tid": 435, "ts": 1776649791307660, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791307669, "dur": 15, + "args": { + "External id": 45656, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45656, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45656, "pid": 0, "tid": 7, "ts": 1776649791307669, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307661, "dur": 4, + "args": { + "External id": 45656, + "cbid": 211, "correlation": 45656 + } + }, + { + "ph": "s", "id": 45656, "pid": 435, "tid": 435, "ts": 1776649791307661, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791307704, "dur": 1, + "args": { + "External id": 5719, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45674, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45674, "pid": 0, "tid": 7, "ts": 1776649791307704, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307695, "dur": 6, + "args": { + "External id": 5719, + "cbid": 211, "correlation": 45674 + } + }, + { + "ph": "s", "id": 45674, "pid": 435, "tid": 435, "ts": 1776649791307695, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307707, "dur": 0, + "args": { + "External id": 45680, + "cbid": 200, "correlation": 45680 + } + }, + { + "ph": "f", "id": 45680, "pid": 435, "tid": 435, "ts": 1776649791307707, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307708, "dur": 0, + "args": { + "External id": 45681, + "cbid": 273, "correlation": 45681 + } + }, + { + "ph": "f", "id": 45681, "pid": 435, "tid": 435, "ts": 1776649791307708, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791307717, "dur": 15, + "args": { + "External id": 45682, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45682, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45682, "pid": 0, "tid": 7, "ts": 1776649791307717, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307709, "dur": 4, + "args": { + "External id": 45682, + "cbid": 211, "correlation": 45682 + } + }, + { + "ph": "s", "id": 45682, "pid": 435, "tid": 435, "ts": 1776649791307709, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791307752, "dur": 1, + "args": { + "External id": 5724, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45700, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45700, "pid": 0, "tid": 7, "ts": 1776649791307752, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307743, "dur": 6, + "args": { + "External id": 5724, + "cbid": 211, "correlation": 45700 + } + }, + { + "ph": "s", "id": 45700, "pid": 435, "tid": 435, "ts": 1776649791307743, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307756, "dur": 0, + "args": { + "External id": 45706, + "cbid": 200, "correlation": 45706 + } + }, + { + "ph": "f", "id": 45706, "pid": 435, "tid": 435, "ts": 1776649791307756, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307756, "dur": 0, + "args": { + "External id": 45707, + "cbid": 273, "correlation": 45707 + } + }, + { + "ph": "f", "id": 45707, "pid": 435, "tid": 435, "ts": 1776649791307756, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791307765, "dur": 15, + "args": { + "External id": 45708, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45708, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45708, "pid": 0, "tid": 7, "ts": 1776649791307765, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307757, "dur": 4, + "args": { + "External id": 45708, + "cbid": 211, "correlation": 45708 + } + }, + { + "ph": "s", "id": 45708, "pid": 435, "tid": 435, "ts": 1776649791307757, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791307800, "dur": 1, + "args": { + "External id": 5729, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45726, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45726, "pid": 0, "tid": 7, "ts": 1776649791307800, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307791, "dur": 6, + "args": { + "External id": 5729, + "cbid": 211, "correlation": 45726 + } + }, + { + "ph": "s", "id": 45726, "pid": 435, "tid": 435, "ts": 1776649791307791, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307803, "dur": 0, + "args": { + "External id": 45732, + "cbid": 200, "correlation": 45732 + } + }, + { + "ph": "f", "id": 45732, "pid": 435, "tid": 435, "ts": 1776649791307803, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307804, "dur": 0, + "args": { + "External id": 45733, + "cbid": 273, "correlation": 45733 + } + }, + { + "ph": "f", "id": 45733, "pid": 435, "tid": 435, "ts": 1776649791307804, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791307813, "dur": 15, + "args": { + "External id": 45734, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45734, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45734, "pid": 0, "tid": 7, "ts": 1776649791307813, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307805, "dur": 4, + "args": { + "External id": 45734, + "cbid": 211, "correlation": 45734 + } + }, + { + "ph": "s", "id": 45734, "pid": 435, "tid": 435, "ts": 1776649791307805, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791307849, "dur": 1, + "args": { + "External id": 5734, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45752, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45752, "pid": 0, "tid": 7, "ts": 1776649791307849, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307839, "dur": 7, + "args": { + "External id": 5734, + "cbid": 211, "correlation": 45752 + } + }, + { + "ph": "s", "id": 45752, "pid": 435, "tid": 435, "ts": 1776649791307839, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307852, "dur": 0, + "args": { + "External id": 45758, + "cbid": 200, "correlation": 45758 + } + }, + { + "ph": "f", "id": 45758, "pid": 435, "tid": 435, "ts": 1776649791307852, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307853, "dur": 0, + "args": { + "External id": 45759, + "cbid": 273, "correlation": 45759 + } + }, + { + "ph": "f", "id": 45759, "pid": 435, "tid": 435, "ts": 1776649791307853, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791307862, "dur": 16, + "args": { + "External id": 45760, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45760, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45760, "pid": 0, "tid": 7, "ts": 1776649791307862, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307854, "dur": 4, + "args": { + "External id": 45760, + "cbid": 211, "correlation": 45760 + } + }, + { + "ph": "s", "id": 45760, "pid": 435, "tid": 435, "ts": 1776649791307854, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791307897, "dur": 1, + "args": { + "External id": 5739, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45778, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45778, "pid": 0, "tid": 7, "ts": 1776649791307897, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307888, "dur": 6, + "args": { + "External id": 5739, + "cbid": 211, "correlation": 45778 + } + }, + { + "ph": "s", "id": 45778, "pid": 435, "tid": 435, "ts": 1776649791307888, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307900, "dur": 0, + "args": { + "External id": 45784, + "cbid": 200, "correlation": 45784 + } + }, + { + "ph": "f", "id": 45784, "pid": 435, "tid": 435, "ts": 1776649791307900, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307901, "dur": 0, + "args": { + "External id": 45785, + "cbid": 273, "correlation": 45785 + } + }, + { + "ph": "f", "id": 45785, "pid": 435, "tid": 435, "ts": 1776649791307901, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791307910, "dur": 16, + "args": { + "External id": 45786, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45786, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45786, "pid": 0, "tid": 7, "ts": 1776649791307910, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307902, "dur": 4, + "args": { + "External id": 45786, + "cbid": 211, "correlation": 45786 + } + }, + { + "ph": "s", "id": 45786, "pid": 435, "tid": 435, "ts": 1776649791307902, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791307945, "dur": 1, + "args": { + "External id": 5744, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45804, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45804, "pid": 0, "tid": 7, "ts": 1776649791307945, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307936, "dur": 6, + "args": { + "External id": 5744, + "cbid": 211, "correlation": 45804 + } + }, + { + "ph": "s", "id": 45804, "pid": 435, "tid": 435, "ts": 1776649791307936, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307948, "dur": 0, + "args": { + "External id": 45810, + "cbid": 200, "correlation": 45810 + } + }, + { + "ph": "f", "id": 45810, "pid": 435, "tid": 435, "ts": 1776649791307948, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307949, "dur": 0, + "args": { + "External id": 45811, + "cbid": 273, "correlation": 45811 + } + }, + { + "ph": "f", "id": 45811, "pid": 435, "tid": 435, "ts": 1776649791307949, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791307958, "dur": 15, + "args": { + "External id": 45812, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45812, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45812, "pid": 0, "tid": 7, "ts": 1776649791307958, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307950, "dur": 4, + "args": { + "External id": 45812, + "cbid": 211, "correlation": 45812 + } + }, + { + "ph": "s", "id": 45812, "pid": 435, "tid": 435, "ts": 1776649791307950, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791307993, "dur": 1, + "args": { + "External id": 5749, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45830, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45830, "pid": 0, "tid": 7, "ts": 1776649791307993, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307983, "dur": 6, + "args": { + "External id": 5749, + "cbid": 211, "correlation": 45830 + } + }, + { + "ph": "s", "id": 45830, "pid": 435, "tid": 435, "ts": 1776649791307983, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307996, "dur": 0, + "args": { + "External id": 45836, + "cbid": 200, "correlation": 45836 + } + }, + { + "ph": "f", "id": 45836, "pid": 435, "tid": 435, "ts": 1776649791307996, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791307996, "dur": 0, + "args": { + "External id": 45837, + "cbid": 273, "correlation": 45837 + } + }, + { + "ph": "f", "id": 45837, "pid": 435, "tid": 435, "ts": 1776649791307996, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791308006, "dur": 15, + "args": { + "External id": 45838, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45838, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45838, "pid": 0, "tid": 7, "ts": 1776649791308006, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791307997, "dur": 5, + "args": { + "External id": 45838, + "cbid": 211, "correlation": 45838 + } + }, + { + "ph": "s", "id": 45838, "pid": 435, "tid": 435, "ts": 1776649791307997, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791308040, "dur": 1, + "args": { + "External id": 5754, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45856, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45856, "pid": 0, "tid": 7, "ts": 1776649791308040, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308031, "dur": 6, + "args": { + "External id": 5754, + "cbid": 211, "correlation": 45856 + } + }, + { + "ph": "s", "id": 45856, "pid": 435, "tid": 435, "ts": 1776649791308031, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308044, "dur": 0, + "args": { + "External id": 45862, + "cbid": 200, "correlation": 45862 + } + }, + { + "ph": "f", "id": 45862, "pid": 435, "tid": 435, "ts": 1776649791308044, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308044, "dur": 0, + "args": { + "External id": 45863, + "cbid": 273, "correlation": 45863 + } + }, + { + "ph": "f", "id": 45863, "pid": 435, "tid": 435, "ts": 1776649791308044, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791308053, "dur": 15, + "args": { + "External id": 45864, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45864, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45864, "pid": 0, "tid": 7, "ts": 1776649791308053, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308045, "dur": 4, + "args": { + "External id": 45864, + "cbid": 211, "correlation": 45864 + } + }, + { + "ph": "s", "id": 45864, "pid": 435, "tid": 435, "ts": 1776649791308045, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791308088, "dur": 1, + "args": { + "External id": 5759, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45882, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45882, "pid": 0, "tid": 7, "ts": 1776649791308088, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308079, "dur": 6, + "args": { + "External id": 5759, + "cbid": 211, "correlation": 45882 + } + }, + { + "ph": "s", "id": 45882, "pid": 435, "tid": 435, "ts": 1776649791308079, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308092, "dur": 0, + "args": { + "External id": 45888, + "cbid": 200, "correlation": 45888 + } + }, + { + "ph": "f", "id": 45888, "pid": 435, "tid": 435, "ts": 1776649791308092, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308092, "dur": 0, + "args": { + "External id": 45889, + "cbid": 273, "correlation": 45889 + } + }, + { + "ph": "f", "id": 45889, "pid": 435, "tid": 435, "ts": 1776649791308092, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791308101, "dur": 15, + "args": { + "External id": 45890, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45890, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45890, "pid": 0, "tid": 7, "ts": 1776649791308101, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308093, "dur": 4, + "args": { + "External id": 45890, + "cbid": 211, "correlation": 45890 + } + }, + { + "ph": "s", "id": 45890, "pid": 435, "tid": 435, "ts": 1776649791308093, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791308136, "dur": 1, + "args": { + "External id": 5764, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45908, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45908, "pid": 0, "tid": 7, "ts": 1776649791308136, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308127, "dur": 6, + "args": { + "External id": 5764, + "cbid": 211, "correlation": 45908 + } + }, + { + "ph": "s", "id": 45908, "pid": 435, "tid": 435, "ts": 1776649791308127, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308140, "dur": 0, + "args": { + "External id": 45914, + "cbid": 200, "correlation": 45914 + } + }, + { + "ph": "f", "id": 45914, "pid": 435, "tid": 435, "ts": 1776649791308140, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308140, "dur": 0, + "args": { + "External id": 45915, + "cbid": 273, "correlation": 45915 + } + }, + { + "ph": "f", "id": 45915, "pid": 435, "tid": 435, "ts": 1776649791308140, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791308149, "dur": 15, + "args": { + "External id": 45916, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45916, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45916, "pid": 0, "tid": 7, "ts": 1776649791308149, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308141, "dur": 4, + "args": { + "External id": 45916, + "cbid": 211, "correlation": 45916 + } + }, + { + "ph": "s", "id": 45916, "pid": 435, "tid": 435, "ts": 1776649791308141, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791308185, "dur": 1, + "args": { + "External id": 5769, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45934, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45934, "pid": 0, "tid": 7, "ts": 1776649791308185, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308175, "dur": 6, + "args": { + "External id": 5769, + "cbid": 211, "correlation": 45934 + } + }, + { + "ph": "s", "id": 45934, "pid": 435, "tid": 435, "ts": 1776649791308175, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308188, "dur": 0, + "args": { + "External id": 45940, + "cbid": 200, "correlation": 45940 + } + }, + { + "ph": "f", "id": 45940, "pid": 435, "tid": 435, "ts": 1776649791308188, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308188, "dur": 0, + "args": { + "External id": 45941, + "cbid": 273, "correlation": 45941 + } + }, + { + "ph": "f", "id": 45941, "pid": 435, "tid": 435, "ts": 1776649791308188, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791308197, "dur": 15, + "args": { + "External id": 45942, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45942, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45942, "pid": 0, "tid": 7, "ts": 1776649791308197, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308189, "dur": 4, + "args": { + "External id": 45942, + "cbid": 211, "correlation": 45942 + } + }, + { + "ph": "s", "id": 45942, "pid": 435, "tid": 435, "ts": 1776649791308189, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791308274, "dur": 1, + "args": { + "External id": 5774, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45960, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45960, "pid": 0, "tid": 7, "ts": 1776649791308274, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308263, "dur": 8, + "args": { + "External id": 5774, + "cbid": 211, "correlation": 45960 + } + }, + { + "ph": "s", "id": 45960, "pid": 435, "tid": 435, "ts": 1776649791308263, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308279, "dur": 0, + "args": { + "External id": 45966, + "cbid": 200, "correlation": 45966 + } + }, + { + "ph": "f", "id": 45966, "pid": 435, "tid": 435, "ts": 1776649791308279, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308279, "dur": 0, + "args": { + "External id": 45967, + "cbid": 273, "correlation": 45967 + } + }, + { + "ph": "f", "id": 45967, "pid": 435, "tid": 435, "ts": 1776649791308279, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791308289, "dur": 15, + "args": { + "External id": 45968, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45968, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45968, "pid": 0, "tid": 7, "ts": 1776649791308289, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308281, "dur": 5, + "args": { + "External id": 45968, + "cbid": 211, "correlation": 45968 + } + }, + { + "ph": "s", "id": 45968, "pid": 435, "tid": 435, "ts": 1776649791308281, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791308328, "dur": 1, + "args": { + "External id": 5779, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45986, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45986, "pid": 0, "tid": 7, "ts": 1776649791308328, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308319, "dur": 6, + "args": { + "External id": 5779, + "cbid": 211, "correlation": 45986 + } + }, + { + "ph": "s", "id": 45986, "pid": 435, "tid": 435, "ts": 1776649791308319, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308331, "dur": 0, + "args": { + "External id": 45992, + "cbid": 200, "correlation": 45992 + } + }, + { + "ph": "f", "id": 45992, "pid": 435, "tid": 435, "ts": 1776649791308331, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308332, "dur": 0, + "args": { + "External id": 45993, + "cbid": 273, "correlation": 45993 + } + }, + { + "ph": "f", "id": 45993, "pid": 435, "tid": 435, "ts": 1776649791308332, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791308341, "dur": 16, + "args": { + "External id": 45994, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 45994, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 45994, "pid": 0, "tid": 7, "ts": 1776649791308341, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308333, "dur": 5, + "args": { + "External id": 45994, + "cbid": 211, "correlation": 45994 + } + }, + { + "ph": "s", "id": 45994, "pid": 435, "tid": 435, "ts": 1776649791308333, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791308377, "dur": 1, + "args": { + "External id": 5784, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46012, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46012, "pid": 0, "tid": 7, "ts": 1776649791308377, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308368, "dur": 6, + "args": { + "External id": 5784, + "cbid": 211, "correlation": 46012 + } + }, + { + "ph": "s", "id": 46012, "pid": 435, "tid": 435, "ts": 1776649791308368, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308380, "dur": 0, + "args": { + "External id": 46018, + "cbid": 200, "correlation": 46018 + } + }, + { + "ph": "f", "id": 46018, "pid": 435, "tid": 435, "ts": 1776649791308380, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308381, "dur": 0, + "args": { + "External id": 46019, + "cbid": 273, "correlation": 46019 + } + }, + { + "ph": "f", "id": 46019, "pid": 435, "tid": 435, "ts": 1776649791308381, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791308390, "dur": 16, + "args": { + "External id": 46020, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46020, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46020, "pid": 0, "tid": 7, "ts": 1776649791308390, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308382, "dur": 5, + "args": { + "External id": 46020, + "cbid": 211, "correlation": 46020 + } + }, + { + "ph": "s", "id": 46020, "pid": 435, "tid": 435, "ts": 1776649791308382, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791308425, "dur": 1, + "args": { + "External id": 5789, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46038, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46038, "pid": 0, "tid": 7, "ts": 1776649791308425, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308416, "dur": 6, + "args": { + "External id": 5789, + "cbid": 211, "correlation": 46038 + } + }, + { + "ph": "s", "id": 46038, "pid": 435, "tid": 435, "ts": 1776649791308416, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308428, "dur": 0, + "args": { + "External id": 46044, + "cbid": 200, "correlation": 46044 + } + }, + { + "ph": "f", "id": 46044, "pid": 435, "tid": 435, "ts": 1776649791308428, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308429, "dur": 0, + "args": { + "External id": 46045, + "cbid": 273, "correlation": 46045 + } + }, + { + "ph": "f", "id": 46045, "pid": 435, "tid": 435, "ts": 1776649791308429, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791308438, "dur": 15, + "args": { + "External id": 46046, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46046, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46046, "pid": 0, "tid": 7, "ts": 1776649791308438, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308430, "dur": 4, + "args": { + "External id": 46046, + "cbid": 211, "correlation": 46046 + } + }, + { + "ph": "s", "id": 46046, "pid": 435, "tid": 435, "ts": 1776649791308430, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791308473, "dur": 1, + "args": { + "External id": 5794, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46064, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46064, "pid": 0, "tid": 7, "ts": 1776649791308473, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308464, "dur": 6, + "args": { + "External id": 5794, + "cbid": 211, "correlation": 46064 + } + }, + { + "ph": "s", "id": 46064, "pid": 435, "tid": 435, "ts": 1776649791308464, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308476, "dur": 0, + "args": { + "External id": 46070, + "cbid": 200, "correlation": 46070 + } + }, + { + "ph": "f", "id": 46070, "pid": 435, "tid": 435, "ts": 1776649791308476, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308477, "dur": 0, + "args": { + "External id": 46071, + "cbid": 273, "correlation": 46071 + } + }, + { + "ph": "f", "id": 46071, "pid": 435, "tid": 435, "ts": 1776649791308477, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791308486, "dur": 15, + "args": { + "External id": 46072, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46072, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46072, "pid": 0, "tid": 7, "ts": 1776649791308486, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308478, "dur": 4, + "args": { + "External id": 46072, + "cbid": 211, "correlation": 46072 + } + }, + { + "ph": "s", "id": 46072, "pid": 435, "tid": 435, "ts": 1776649791308478, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791308521, "dur": 1, + "args": { + "External id": 5799, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46090, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46090, "pid": 0, "tid": 7, "ts": 1776649791308521, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308512, "dur": 6, + "args": { + "External id": 5799, + "cbid": 211, "correlation": 46090 + } + }, + { + "ph": "s", "id": 46090, "pid": 435, "tid": 435, "ts": 1776649791308512, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308524, "dur": 0, + "args": { + "External id": 46096, + "cbid": 200, "correlation": 46096 + } + }, + { + "ph": "f", "id": 46096, "pid": 435, "tid": 435, "ts": 1776649791308524, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308525, "dur": 0, + "args": { + "External id": 46097, + "cbid": 273, "correlation": 46097 + } + }, + { + "ph": "f", "id": 46097, "pid": 435, "tid": 435, "ts": 1776649791308525, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791308534, "dur": 15, + "args": { + "External id": 46098, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46098, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46098, "pid": 0, "tid": 7, "ts": 1776649791308534, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308526, "dur": 5, + "args": { + "External id": 46098, + "cbid": 211, "correlation": 46098 + } + }, + { + "ph": "s", "id": 46098, "pid": 435, "tid": 435, "ts": 1776649791308526, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791308569, "dur": 1, + "args": { + "External id": 5804, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46116, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46116, "pid": 0, "tid": 7, "ts": 1776649791308569, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308560, "dur": 6, + "args": { + "External id": 5804, + "cbid": 211, "correlation": 46116 + } + }, + { + "ph": "s", "id": 46116, "pid": 435, "tid": 435, "ts": 1776649791308560, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308572, "dur": 0, + "args": { + "External id": 46122, + "cbid": 200, "correlation": 46122 + } + }, + { + "ph": "f", "id": 46122, "pid": 435, "tid": 435, "ts": 1776649791308572, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308573, "dur": 0, + "args": { + "External id": 46123, + "cbid": 273, "correlation": 46123 + } + }, + { + "ph": "f", "id": 46123, "pid": 435, "tid": 435, "ts": 1776649791308573, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791308582, "dur": 15, + "args": { + "External id": 46124, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46124, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46124, "pid": 0, "tid": 7, "ts": 1776649791308582, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308574, "dur": 5, + "args": { + "External id": 46124, + "cbid": 211, "correlation": 46124 + } + }, + { + "ph": "s", "id": 46124, "pid": 435, "tid": 435, "ts": 1776649791308574, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791308617, "dur": 1, + "args": { + "External id": 5809, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46142, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46142, "pid": 0, "tid": 7, "ts": 1776649791308617, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308608, "dur": 6, + "args": { + "External id": 5809, + "cbid": 211, "correlation": 46142 + } + }, + { + "ph": "s", "id": 46142, "pid": 435, "tid": 435, "ts": 1776649791308608, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308621, "dur": 0, + "args": { + "External id": 46148, + "cbid": 200, "correlation": 46148 + } + }, + { + "ph": "f", "id": 46148, "pid": 435, "tid": 435, "ts": 1776649791308621, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308621, "dur": 0, + "args": { + "External id": 46149, + "cbid": 273, "correlation": 46149 + } + }, + { + "ph": "f", "id": 46149, "pid": 435, "tid": 435, "ts": 1776649791308621, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791308630, "dur": 15, + "args": { + "External id": 46150, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46150, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46150, "pid": 0, "tid": 7, "ts": 1776649791308630, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308622, "dur": 5, + "args": { + "External id": 46150, + "cbid": 211, "correlation": 46150 + } + }, + { + "ph": "s", "id": 46150, "pid": 435, "tid": 435, "ts": 1776649791308622, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791308665, "dur": 1, + "args": { + "External id": 5814, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46168, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46168, "pid": 0, "tid": 7, "ts": 1776649791308665, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308656, "dur": 6, + "args": { + "External id": 5814, + "cbid": 211, "correlation": 46168 + } + }, + { + "ph": "s", "id": 46168, "pid": 435, "tid": 435, "ts": 1776649791308656, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308668, "dur": 0, + "args": { + "External id": 46174, + "cbid": 200, "correlation": 46174 + } + }, + { + "ph": "f", "id": 46174, "pid": 435, "tid": 435, "ts": 1776649791308668, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308669, "dur": 0, + "args": { + "External id": 46175, + "cbid": 273, "correlation": 46175 + } + }, + { + "ph": "f", "id": 46175, "pid": 435, "tid": 435, "ts": 1776649791308669, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791308678, "dur": 15, + "args": { + "External id": 46176, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46176, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46176, "pid": 0, "tid": 7, "ts": 1776649791308678, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308670, "dur": 4, + "args": { + "External id": 46176, + "cbid": 211, "correlation": 46176 + } + }, + { + "ph": "s", "id": 46176, "pid": 435, "tid": 435, "ts": 1776649791308670, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791308713, "dur": 1, + "args": { + "External id": 5819, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46194, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46194, "pid": 0, "tid": 7, "ts": 1776649791308713, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308704, "dur": 6, + "args": { + "External id": 5819, + "cbid": 211, "correlation": 46194 + } + }, + { + "ph": "s", "id": 46194, "pid": 435, "tid": 435, "ts": 1776649791308704, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308716, "dur": 0, + "args": { + "External id": 46200, + "cbid": 200, "correlation": 46200 + } + }, + { + "ph": "f", "id": 46200, "pid": 435, "tid": 435, "ts": 1776649791308716, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308717, "dur": 0, + "args": { + "External id": 46201, + "cbid": 273, "correlation": 46201 + } + }, + { + "ph": "f", "id": 46201, "pid": 435, "tid": 435, "ts": 1776649791308717, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791308726, "dur": 15, + "args": { + "External id": 46202, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46202, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46202, "pid": 0, "tid": 7, "ts": 1776649791308726, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308718, "dur": 4, + "args": { + "External id": 46202, + "cbid": 211, "correlation": 46202 + } + }, + { + "ph": "s", "id": 46202, "pid": 435, "tid": 435, "ts": 1776649791308718, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791308761, "dur": 1, + "args": { + "External id": 5824, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46220, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46220, "pid": 0, "tid": 7, "ts": 1776649791308761, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308752, "dur": 6, + "args": { + "External id": 5824, + "cbid": 211, "correlation": 46220 + } + }, + { + "ph": "s", "id": 46220, "pid": 435, "tid": 435, "ts": 1776649791308752, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308765, "dur": 0, + "args": { + "External id": 46226, + "cbid": 200, "correlation": 46226 + } + }, + { + "ph": "f", "id": 46226, "pid": 435, "tid": 435, "ts": 1776649791308765, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308765, "dur": 0, + "args": { + "External id": 46227, + "cbid": 273, "correlation": 46227 + } + }, + { + "ph": "f", "id": 46227, "pid": 435, "tid": 435, "ts": 1776649791308765, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791308774, "dur": 17, + "args": { + "External id": 46228, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46228, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46228, "pid": 0, "tid": 7, "ts": 1776649791308774, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308766, "dur": 5, + "args": { + "External id": 46228, + "cbid": 211, "correlation": 46228 + } + }, + { + "ph": "s", "id": 46228, "pid": 435, "tid": 435, "ts": 1776649791308766, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791308810, "dur": 1, + "args": { + "External id": 5829, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46246, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46246, "pid": 0, "tid": 7, "ts": 1776649791308810, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308800, "dur": 6, + "args": { + "External id": 5829, + "cbid": 211, "correlation": 46246 + } + }, + { + "ph": "s", "id": 46246, "pid": 435, "tid": 435, "ts": 1776649791308800, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308813, "dur": 0, + "args": { + "External id": 46252, + "cbid": 200, "correlation": 46252 + } + }, + { + "ph": "f", "id": 46252, "pid": 435, "tid": 435, "ts": 1776649791308813, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308813, "dur": 0, + "args": { + "External id": 46253, + "cbid": 273, "correlation": 46253 + } + }, + { + "ph": "f", "id": 46253, "pid": 435, "tid": 435, "ts": 1776649791308813, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791308822, "dur": 15, + "args": { + "External id": 46254, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46254, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46254, "pid": 0, "tid": 7, "ts": 1776649791308822, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308814, "dur": 5, + "args": { + "External id": 46254, + "cbid": 211, "correlation": 46254 + } + }, + { + "ph": "s", "id": 46254, "pid": 435, "tid": 435, "ts": 1776649791308814, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791308857, "dur": 1, + "args": { + "External id": 5834, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46272, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46272, "pid": 0, "tid": 7, "ts": 1776649791308857, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308848, "dur": 6, + "args": { + "External id": 5834, + "cbid": 211, "correlation": 46272 + } + }, + { + "ph": "s", "id": 46272, "pid": 435, "tid": 435, "ts": 1776649791308848, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308861, "dur": 0, + "args": { + "External id": 46278, + "cbid": 200, "correlation": 46278 + } + }, + { + "ph": "f", "id": 46278, "pid": 435, "tid": 435, "ts": 1776649791308861, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308861, "dur": 0, + "args": { + "External id": 46279, + "cbid": 273, "correlation": 46279 + } + }, + { + "ph": "f", "id": 46279, "pid": 435, "tid": 435, "ts": 1776649791308861, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791308870, "dur": 15, + "args": { + "External id": 46280, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46280, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46280, "pid": 0, "tid": 7, "ts": 1776649791308870, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308862, "dur": 4, + "args": { + "External id": 46280, + "cbid": 211, "correlation": 46280 + } + }, + { + "ph": "s", "id": 46280, "pid": 435, "tid": 435, "ts": 1776649791308862, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791308906, "dur": 1, + "args": { + "External id": 5839, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46298, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46298, "pid": 0, "tid": 7, "ts": 1776649791308906, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308896, "dur": 6, + "args": { + "External id": 5839, + "cbid": 211, "correlation": 46298 + } + }, + { + "ph": "s", "id": 46298, "pid": 435, "tid": 435, "ts": 1776649791308896, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308909, "dur": 0, + "args": { + "External id": 46304, + "cbid": 200, "correlation": 46304 + } + }, + { + "ph": "f", "id": 46304, "pid": 435, "tid": 435, "ts": 1776649791308909, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308909, "dur": 0, + "args": { + "External id": 46305, + "cbid": 273, "correlation": 46305 + } + }, + { + "ph": "f", "id": 46305, "pid": 435, "tid": 435, "ts": 1776649791308909, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791308918, "dur": 16, + "args": { + "External id": 46306, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46306, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46306, "pid": 0, "tid": 7, "ts": 1776649791308918, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308910, "dur": 4, + "args": { + "External id": 46306, + "cbid": 211, "correlation": 46306 + } + }, + { + "ph": "s", "id": 46306, "pid": 435, "tid": 435, "ts": 1776649791308910, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791308954, "dur": 1, + "args": { + "External id": 5844, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46324, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46324, "pid": 0, "tid": 7, "ts": 1776649791308954, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308944, "dur": 6, + "args": { + "External id": 5844, + "cbid": 211, "correlation": 46324 + } + }, + { + "ph": "s", "id": 46324, "pid": 435, "tid": 435, "ts": 1776649791308944, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308956, "dur": 0, + "args": { + "External id": 46330, + "cbid": 200, "correlation": 46330 + } + }, + { + "ph": "f", "id": 46330, "pid": 435, "tid": 435, "ts": 1776649791308956, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791308957, "dur": 0, + "args": { + "External id": 46331, + "cbid": 273, "correlation": 46331 + } + }, + { + "ph": "f", "id": 46331, "pid": 435, "tid": 435, "ts": 1776649791308957, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791308966, "dur": 15, + "args": { + "External id": 46332, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46332, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46332, "pid": 0, "tid": 7, "ts": 1776649791308966, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308958, "dur": 4, + "args": { + "External id": 46332, + "cbid": 211, "correlation": 46332 + } + }, + { + "ph": "s", "id": 46332, "pid": 435, "tid": 435, "ts": 1776649791308958, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791309001, "dur": 1, + "args": { + "External id": 5849, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46350, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46350, "pid": 0, "tid": 7, "ts": 1776649791309001, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791308992, "dur": 6, + "args": { + "External id": 5849, + "cbid": 211, "correlation": 46350 + } + }, + { + "ph": "s", "id": 46350, "pid": 435, "tid": 435, "ts": 1776649791308992, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309004, "dur": 0, + "args": { + "External id": 46356, + "cbid": 200, "correlation": 46356 + } + }, + { + "ph": "f", "id": 46356, "pid": 435, "tid": 435, "ts": 1776649791309004, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309005, "dur": 0, + "args": { + "External id": 46357, + "cbid": 273, "correlation": 46357 + } + }, + { + "ph": "f", "id": 46357, "pid": 435, "tid": 435, "ts": 1776649791309005, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791309014, "dur": 15, + "args": { + "External id": 46358, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46358, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46358, "pid": 0, "tid": 7, "ts": 1776649791309014, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309006, "dur": 5, + "args": { + "External id": 46358, + "cbid": 211, "correlation": 46358 + } + }, + { + "ph": "s", "id": 46358, "pid": 435, "tid": 435, "ts": 1776649791309006, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791309049, "dur": 1, + "args": { + "External id": 5854, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46376, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46376, "pid": 0, "tid": 7, "ts": 1776649791309049, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309040, "dur": 6, + "args": { + "External id": 5854, + "cbid": 211, "correlation": 46376 + } + }, + { + "ph": "s", "id": 46376, "pid": 435, "tid": 435, "ts": 1776649791309040, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309052, "dur": 0, + "args": { + "External id": 46382, + "cbid": 200, "correlation": 46382 + } + }, + { + "ph": "f", "id": 46382, "pid": 435, "tid": 435, "ts": 1776649791309052, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309053, "dur": 0, + "args": { + "External id": 46383, + "cbid": 273, "correlation": 46383 + } + }, + { + "ph": "f", "id": 46383, "pid": 435, "tid": 435, "ts": 1776649791309053, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791309062, "dur": 15, + "args": { + "External id": 46384, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46384, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46384, "pid": 0, "tid": 7, "ts": 1776649791309062, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309054, "dur": 4, + "args": { + "External id": 46384, + "cbid": 211, "correlation": 46384 + } + }, + { + "ph": "s", "id": 46384, "pid": 435, "tid": 435, "ts": 1776649791309054, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791309097, "dur": 1, + "args": { + "External id": 5859, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46402, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46402, "pid": 0, "tid": 7, "ts": 1776649791309097, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309088, "dur": 6, + "args": { + "External id": 5859, + "cbid": 211, "correlation": 46402 + } + }, + { + "ph": "s", "id": 46402, "pid": 435, "tid": 435, "ts": 1776649791309088, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309100, "dur": 0, + "args": { + "External id": 46408, + "cbid": 200, "correlation": 46408 + } + }, + { + "ph": "f", "id": 46408, "pid": 435, "tid": 435, "ts": 1776649791309100, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309101, "dur": 0, + "args": { + "External id": 46409, + "cbid": 273, "correlation": 46409 + } + }, + { + "ph": "f", "id": 46409, "pid": 435, "tid": 435, "ts": 1776649791309101, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791309110, "dur": 15, + "args": { + "External id": 46410, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46410, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46410, "pid": 0, "tid": 7, "ts": 1776649791309110, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309102, "dur": 5, + "args": { + "External id": 46410, + "cbid": 211, "correlation": 46410 + } + }, + { + "ph": "s", "id": 46410, "pid": 435, "tid": 435, "ts": 1776649791309102, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791309146, "dur": 1, + "args": { + "External id": 5864, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46428, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46428, "pid": 0, "tid": 7, "ts": 1776649791309146, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309137, "dur": 6, + "args": { + "External id": 5864, + "cbid": 211, "correlation": 46428 + } + }, + { + "ph": "s", "id": 46428, "pid": 435, "tid": 435, "ts": 1776649791309137, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309149, "dur": 0, + "args": { + "External id": 46434, + "cbid": 200, "correlation": 46434 + } + }, + { + "ph": "f", "id": 46434, "pid": 435, "tid": 435, "ts": 1776649791309149, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309150, "dur": 0, + "args": { + "External id": 46435, + "cbid": 273, "correlation": 46435 + } + }, + { + "ph": "f", "id": 46435, "pid": 435, "tid": 435, "ts": 1776649791309150, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791309159, "dur": 15, + "args": { + "External id": 46436, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46436, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46436, "pid": 0, "tid": 7, "ts": 1776649791309159, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309151, "dur": 5, + "args": { + "External id": 46436, + "cbid": 211, "correlation": 46436 + } + }, + { + "ph": "s", "id": 46436, "pid": 435, "tid": 435, "ts": 1776649791309151, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791309194, "dur": 1, + "args": { + "External id": 5869, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46454, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46454, "pid": 0, "tid": 7, "ts": 1776649791309194, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309185, "dur": 6, + "args": { + "External id": 5869, + "cbid": 211, "correlation": 46454 + } + }, + { + "ph": "s", "id": 46454, "pid": 435, "tid": 435, "ts": 1776649791309185, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309197, "dur": 0, + "args": { + "External id": 46460, + "cbid": 200, "correlation": 46460 + } + }, + { + "ph": "f", "id": 46460, "pid": 435, "tid": 435, "ts": 1776649791309197, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309198, "dur": 0, + "args": { + "External id": 46461, + "cbid": 273, "correlation": 46461 + } + }, + { + "ph": "f", "id": 46461, "pid": 435, "tid": 435, "ts": 1776649791309198, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791309207, "dur": 16, + "args": { + "External id": 46462, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46462, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46462, "pid": 0, "tid": 7, "ts": 1776649791309207, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309199, "dur": 5, + "args": { + "External id": 46462, + "cbid": 211, "correlation": 46462 + } + }, + { + "ph": "s", "id": 46462, "pid": 435, "tid": 435, "ts": 1776649791309199, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791309243, "dur": 1, + "args": { + "External id": 5874, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46480, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46480, "pid": 0, "tid": 7, "ts": 1776649791309243, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309233, "dur": 6, + "args": { + "External id": 5874, + "cbid": 211, "correlation": 46480 + } + }, + { + "ph": "s", "id": 46480, "pid": 435, "tid": 435, "ts": 1776649791309233, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309246, "dur": 0, + "args": { + "External id": 46486, + "cbid": 200, "correlation": 46486 + } + }, + { + "ph": "f", "id": 46486, "pid": 435, "tid": 435, "ts": 1776649791309246, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309247, "dur": 0, + "args": { + "External id": 46487, + "cbid": 273, "correlation": 46487 + } + }, + { + "ph": "f", "id": 46487, "pid": 435, "tid": 435, "ts": 1776649791309247, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791309256, "dur": 15, + "args": { + "External id": 46488, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46488, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46488, "pid": 0, "tid": 7, "ts": 1776649791309256, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309248, "dur": 5, + "args": { + "External id": 46488, + "cbid": 211, "correlation": 46488 + } + }, + { + "ph": "s", "id": 46488, "pid": 435, "tid": 435, "ts": 1776649791309248, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791309291, "dur": 1, + "args": { + "External id": 5879, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46506, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46506, "pid": 0, "tid": 7, "ts": 1776649791309291, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309282, "dur": 6, + "args": { + "External id": 5879, + "cbid": 211, "correlation": 46506 + } + }, + { + "ph": "s", "id": 46506, "pid": 435, "tid": 435, "ts": 1776649791309282, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309294, "dur": 0, + "args": { + "External id": 46512, + "cbid": 200, "correlation": 46512 + } + }, + { + "ph": "f", "id": 46512, "pid": 435, "tid": 435, "ts": 1776649791309294, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309295, "dur": 0, + "args": { + "External id": 46513, + "cbid": 273, "correlation": 46513 + } + }, + { + "ph": "f", "id": 46513, "pid": 435, "tid": 435, "ts": 1776649791309295, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791309304, "dur": 15, + "args": { + "External id": 46514, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46514, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46514, "pid": 0, "tid": 7, "ts": 1776649791309304, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309296, "dur": 4, + "args": { + "External id": 46514, + "cbid": 211, "correlation": 46514 + } + }, + { + "ph": "s", "id": 46514, "pid": 435, "tid": 435, "ts": 1776649791309296, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791309339, "dur": 1, + "args": { + "External id": 5884, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46532, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46532, "pid": 0, "tid": 7, "ts": 1776649791309339, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309330, "dur": 6, + "args": { + "External id": 5884, + "cbid": 211, "correlation": 46532 + } + }, + { + "ph": "s", "id": 46532, "pid": 435, "tid": 435, "ts": 1776649791309330, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309342, "dur": 0, + "args": { + "External id": 46538, + "cbid": 200, "correlation": 46538 + } + }, + { + "ph": "f", "id": 46538, "pid": 435, "tid": 435, "ts": 1776649791309342, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309343, "dur": 0, + "args": { + "External id": 46539, + "cbid": 273, "correlation": 46539 + } + }, + { + "ph": "f", "id": 46539, "pid": 435, "tid": 435, "ts": 1776649791309343, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791309352, "dur": 16, + "args": { + "External id": 46540, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46540, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46540, "pid": 0, "tid": 7, "ts": 1776649791309352, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309344, "dur": 4, + "args": { + "External id": 46540, + "cbid": 211, "correlation": 46540 + } + }, + { + "ph": "s", "id": 46540, "pid": 435, "tid": 435, "ts": 1776649791309344, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791309387, "dur": 1, + "args": { + "External id": 5889, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46558, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46558, "pid": 0, "tid": 7, "ts": 1776649791309387, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309378, "dur": 6, + "args": { + "External id": 5889, + "cbid": 211, "correlation": 46558 + } + }, + { + "ph": "s", "id": 46558, "pid": 435, "tid": 435, "ts": 1776649791309378, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309390, "dur": 0, + "args": { + "External id": 46564, + "cbid": 200, "correlation": 46564 + } + }, + { + "ph": "f", "id": 46564, "pid": 435, "tid": 435, "ts": 1776649791309390, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309391, "dur": 0, + "args": { + "External id": 46565, + "cbid": 273, "correlation": 46565 + } + }, + { + "ph": "f", "id": 46565, "pid": 435, "tid": 435, "ts": 1776649791309391, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791309400, "dur": 15, + "args": { + "External id": 46566, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46566, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46566, "pid": 0, "tid": 7, "ts": 1776649791309400, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309392, "dur": 4, + "args": { + "External id": 46566, + "cbid": 211, "correlation": 46566 + } + }, + { + "ph": "s", "id": 46566, "pid": 435, "tid": 435, "ts": 1776649791309392, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791309435, "dur": 1, + "args": { + "External id": 5894, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46584, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46584, "pid": 0, "tid": 7, "ts": 1776649791309435, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309426, "dur": 6, + "args": { + "External id": 5894, + "cbid": 211, "correlation": 46584 + } + }, + { + "ph": "s", "id": 46584, "pid": 435, "tid": 435, "ts": 1776649791309426, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309438, "dur": 0, + "args": { + "External id": 46590, + "cbid": 200, "correlation": 46590 + } + }, + { + "ph": "f", "id": 46590, "pid": 435, "tid": 435, "ts": 1776649791309438, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309438, "dur": 0, + "args": { + "External id": 46591, + "cbid": 273, "correlation": 46591 + } + }, + { + "ph": "f", "id": 46591, "pid": 435, "tid": 435, "ts": 1776649791309438, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791309447, "dur": 15, + "args": { + "External id": 46592, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46592, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46592, "pid": 0, "tid": 7, "ts": 1776649791309447, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309439, "dur": 4, + "args": { + "External id": 46592, + "cbid": 211, "correlation": 46592 + } + }, + { + "ph": "s", "id": 46592, "pid": 435, "tid": 435, "ts": 1776649791309439, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791309482, "dur": 1, + "args": { + "External id": 5899, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46610, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46610, "pid": 0, "tid": 7, "ts": 1776649791309482, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309473, "dur": 6, + "args": { + "External id": 5899, + "cbid": 211, "correlation": 46610 + } + }, + { + "ph": "s", "id": 46610, "pid": 435, "tid": 435, "ts": 1776649791309473, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309486, "dur": 0, + "args": { + "External id": 46616, + "cbid": 200, "correlation": 46616 + } + }, + { + "ph": "f", "id": 46616, "pid": 435, "tid": 435, "ts": 1776649791309486, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309486, "dur": 0, + "args": { + "External id": 46617, + "cbid": 273, "correlation": 46617 + } + }, + { + "ph": "f", "id": 46617, "pid": 435, "tid": 435, "ts": 1776649791309486, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791309495, "dur": 15, + "args": { + "External id": 46618, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46618, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46618, "pid": 0, "tid": 7, "ts": 1776649791309495, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309487, "dur": 4, + "args": { + "External id": 46618, + "cbid": 211, "correlation": 46618 + } + }, + { + "ph": "s", "id": 46618, "pid": 435, "tid": 435, "ts": 1776649791309487, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791309530, "dur": 1, + "args": { + "External id": 5904, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46636, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46636, "pid": 0, "tid": 7, "ts": 1776649791309530, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309521, "dur": 6, + "args": { + "External id": 5904, + "cbid": 211, "correlation": 46636 + } + }, + { + "ph": "s", "id": 46636, "pid": 435, "tid": 435, "ts": 1776649791309521, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309533, "dur": 0, + "args": { + "External id": 46642, + "cbid": 200, "correlation": 46642 + } + }, + { + "ph": "f", "id": 46642, "pid": 435, "tid": 435, "ts": 1776649791309533, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309534, "dur": 0, + "args": { + "External id": 46643, + "cbid": 273, "correlation": 46643 + } + }, + { + "ph": "f", "id": 46643, "pid": 435, "tid": 435, "ts": 1776649791309534, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791309543, "dur": 15, + "args": { + "External id": 46644, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46644, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46644, "pid": 0, "tid": 7, "ts": 1776649791309543, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309535, "dur": 4, + "args": { + "External id": 46644, + "cbid": 211, "correlation": 46644 + } + }, + { + "ph": "s", "id": 46644, "pid": 435, "tid": 435, "ts": 1776649791309535, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791309578, "dur": 1, + "args": { + "External id": 5909, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46662, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46662, "pid": 0, "tid": 7, "ts": 1776649791309578, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309569, "dur": 6, + "args": { + "External id": 5909, + "cbid": 211, "correlation": 46662 + } + }, + { + "ph": "s", "id": 46662, "pid": 435, "tid": 435, "ts": 1776649791309569, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309581, "dur": 0, + "args": { + "External id": 46668, + "cbid": 200, "correlation": 46668 + } + }, + { + "ph": "f", "id": 46668, "pid": 435, "tid": 435, "ts": 1776649791309581, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309582, "dur": 0, + "args": { + "External id": 46669, + "cbid": 273, "correlation": 46669 + } + }, + { + "ph": "f", "id": 46669, "pid": 435, "tid": 435, "ts": 1776649791309582, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791309591, "dur": 16, + "args": { + "External id": 46670, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46670, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46670, "pid": 0, "tid": 7, "ts": 1776649791309591, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309583, "dur": 4, + "args": { + "External id": 46670, + "cbid": 211, "correlation": 46670 + } + }, + { + "ph": "s", "id": 46670, "pid": 435, "tid": 435, "ts": 1776649791309583, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791309628, "dur": 1, + "args": { + "External id": 5914, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46688, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46688, "pid": 0, "tid": 7, "ts": 1776649791309628, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309619, "dur": 6, + "args": { + "External id": 5914, + "cbid": 211, "correlation": 46688 + } + }, + { + "ph": "s", "id": 46688, "pid": 435, "tid": 435, "ts": 1776649791309619, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309631, "dur": 0, + "args": { + "External id": 46694, + "cbid": 200, "correlation": 46694 + } + }, + { + "ph": "f", "id": 46694, "pid": 435, "tid": 435, "ts": 1776649791309631, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309632, "dur": 0, + "args": { + "External id": 46695, + "cbid": 273, "correlation": 46695 + } + }, + { + "ph": "f", "id": 46695, "pid": 435, "tid": 435, "ts": 1776649791309632, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791309641, "dur": 16, + "args": { + "External id": 46696, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46696, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46696, "pid": 0, "tid": 7, "ts": 1776649791309641, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309633, "dur": 4, + "args": { + "External id": 46696, + "cbid": 211, "correlation": 46696 + } + }, + { + "ph": "s", "id": 46696, "pid": 435, "tid": 435, "ts": 1776649791309633, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791309676, "dur": 1, + "args": { + "External id": 5919, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46714, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46714, "pid": 0, "tid": 7, "ts": 1776649791309676, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309667, "dur": 6, + "args": { + "External id": 5919, + "cbid": 211, "correlation": 46714 + } + }, + { + "ph": "s", "id": 46714, "pid": 435, "tid": 435, "ts": 1776649791309667, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309679, "dur": 0, + "args": { + "External id": 46720, + "cbid": 200, "correlation": 46720 + } + }, + { + "ph": "f", "id": 46720, "pid": 435, "tid": 435, "ts": 1776649791309679, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309680, "dur": 0, + "args": { + "External id": 46721, + "cbid": 273, "correlation": 46721 + } + }, + { + "ph": "f", "id": 46721, "pid": 435, "tid": 435, "ts": 1776649791309680, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791309689, "dur": 15, + "args": { + "External id": 46722, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46722, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46722, "pid": 0, "tid": 7, "ts": 1776649791309689, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309681, "dur": 5, + "args": { + "External id": 46722, + "cbid": 211, "correlation": 46722 + } + }, + { + "ph": "s", "id": 46722, "pid": 435, "tid": 435, "ts": 1776649791309681, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791309724, "dur": 1, + "args": { + "External id": 5924, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46740, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46740, "pid": 0, "tid": 7, "ts": 1776649791309724, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309715, "dur": 6, + "args": { + "External id": 5924, + "cbid": 211, "correlation": 46740 + } + }, + { + "ph": "s", "id": 46740, "pid": 435, "tid": 435, "ts": 1776649791309715, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309727, "dur": 0, + "args": { + "External id": 46746, + "cbid": 200, "correlation": 46746 + } + }, + { + "ph": "f", "id": 46746, "pid": 435, "tid": 435, "ts": 1776649791309727, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309728, "dur": 0, + "args": { + "External id": 46747, + "cbid": 273, "correlation": 46747 + } + }, + { + "ph": "f", "id": 46747, "pid": 435, "tid": 435, "ts": 1776649791309728, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791309737, "dur": 16, + "args": { + "External id": 46748, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46748, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46748, "pid": 0, "tid": 7, "ts": 1776649791309737, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309729, "dur": 4, + "args": { + "External id": 46748, + "cbid": 211, "correlation": 46748 + } + }, + { + "ph": "s", "id": 46748, "pid": 435, "tid": 435, "ts": 1776649791309729, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791309772, "dur": 1, + "args": { + "External id": 5929, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46766, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46766, "pid": 0, "tid": 7, "ts": 1776649791309772, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309763, "dur": 6, + "args": { + "External id": 5929, + "cbid": 211, "correlation": 46766 + } + }, + { + "ph": "s", "id": 46766, "pid": 435, "tid": 435, "ts": 1776649791309763, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309775, "dur": 0, + "args": { + "External id": 46772, + "cbid": 200, "correlation": 46772 + } + }, + { + "ph": "f", "id": 46772, "pid": 435, "tid": 435, "ts": 1776649791309775, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309776, "dur": 0, + "args": { + "External id": 46773, + "cbid": 273, "correlation": 46773 + } + }, + { + "ph": "f", "id": 46773, "pid": 435, "tid": 435, "ts": 1776649791309776, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791309785, "dur": 15, + "args": { + "External id": 46774, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46774, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46774, "pid": 0, "tid": 7, "ts": 1776649791309785, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309777, "dur": 5, + "args": { + "External id": 46774, + "cbid": 211, "correlation": 46774 + } + }, + { + "ph": "s", "id": 46774, "pid": 435, "tid": 435, "ts": 1776649791309777, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791309809, "dur": 3, + "args": { + "External id": 46782, + "cbid": 135, "correlation": 46782 + } + }, + { + "ph": "f", "id": 46782, "pid": 435, "tid": 435, "ts": 1776649791309809, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791309824, "dur": 5, + "args": { + "External id": 46787, + "cbid": 165, "correlation": 46787 + } + }, + { + "ph": "s", "id": 46787, "pid": 435, "tid": 435, "ts": 1776649791309824, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649791309834, "dur": 7, + "args": { + "External id": 46790, + "cbid": 139, "correlation": 46790 + } + }, + { + "ph": "f", "id": 46790, "pid": 435, "tid": 435, "ts": 1776649791309834, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791309897, "dur": 1, + "args": { + "External id": 5934, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46818, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46818, "pid": 0, "tid": 7, "ts": 1776649791309897, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309886, "dur": 9, + "args": { + "External id": 5934, + "cbid": 211, "correlation": 46818 + } + }, + { + "ph": "s", "id": 46818, "pid": 435, "tid": 435, "ts": 1776649791309886, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309901, "dur": 0, + "args": { + "External id": 46824, + "cbid": 200, "correlation": 46824 + } + }, + { + "ph": "f", "id": 46824, "pid": 435, "tid": 435, "ts": 1776649791309901, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309902, "dur": 1, + "args": { + "External id": 46825, + "cbid": 273, "correlation": 46825 + } + }, + { + "ph": "f", "id": 46825, "pid": 435, "tid": 435, "ts": 1776649791309902, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791309912, "dur": 15, + "args": { + "External id": 46826, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46826, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46826, "pid": 0, "tid": 7, "ts": 1776649791309912, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309903, "dur": 5, + "args": { + "External id": 46826, + "cbid": 211, "correlation": 46826 + } + }, + { + "ph": "s", "id": 46826, "pid": 435, "tid": 435, "ts": 1776649791309903, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791309950, "dur": 1, + "args": { + "External id": 5939, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46844, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46844, "pid": 0, "tid": 7, "ts": 1776649791309950, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309940, "dur": 7, + "args": { + "External id": 5939, + "cbid": 211, "correlation": 46844 + } + }, + { + "ph": "s", "id": 46844, "pid": 435, "tid": 435, "ts": 1776649791309940, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309953, "dur": 0, + "args": { + "External id": 46850, + "cbid": 200, "correlation": 46850 + } + }, + { + "ph": "f", "id": 46850, "pid": 435, "tid": 435, "ts": 1776649791309953, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791309954, "dur": 0, + "args": { + "External id": 46851, + "cbid": 273, "correlation": 46851 + } + }, + { + "ph": "f", "id": 46851, "pid": 435, "tid": 435, "ts": 1776649791309954, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791309963, "dur": 15, + "args": { + "External id": 46852, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46852, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46852, "pid": 0, "tid": 7, "ts": 1776649791309963, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309955, "dur": 4, + "args": { + "External id": 46852, + "cbid": 211, "correlation": 46852 + } + }, + { + "ph": "s", "id": 46852, "pid": 435, "tid": 435, "ts": 1776649791309955, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791309999, "dur": 1, + "args": { + "External id": 5944, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46870, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46870, "pid": 0, "tid": 7, "ts": 1776649791309999, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791309989, "dur": 6, + "args": { + "External id": 5944, + "cbid": 211, "correlation": 46870 + } + }, + { + "ph": "s", "id": 46870, "pid": 435, "tid": 435, "ts": 1776649791309989, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310002, "dur": 0, + "args": { + "External id": 46876, + "cbid": 200, "correlation": 46876 + } + }, + { + "ph": "f", "id": 46876, "pid": 435, "tid": 435, "ts": 1776649791310002, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310003, "dur": 0, + "args": { + "External id": 46877, + "cbid": 273, "correlation": 46877 + } + }, + { + "ph": "f", "id": 46877, "pid": 435, "tid": 435, "ts": 1776649791310003, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791310012, "dur": 15, + "args": { + "External id": 46878, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46878, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46878, "pid": 0, "tid": 7, "ts": 1776649791310012, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310004, "dur": 4, + "args": { + "External id": 46878, + "cbid": 211, "correlation": 46878 + } + }, + { + "ph": "s", "id": 46878, "pid": 435, "tid": 435, "ts": 1776649791310004, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791310048, "dur": 1, + "args": { + "External id": 5949, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46896, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46896, "pid": 0, "tid": 7, "ts": 1776649791310048, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310038, "dur": 6, + "args": { + "External id": 5949, + "cbid": 211, "correlation": 46896 + } + }, + { + "ph": "s", "id": 46896, "pid": 435, "tid": 435, "ts": 1776649791310038, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310051, "dur": 0, + "args": { + "External id": 46902, + "cbid": 200, "correlation": 46902 + } + }, + { + "ph": "f", "id": 46902, "pid": 435, "tid": 435, "ts": 1776649791310051, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310051, "dur": 0, + "args": { + "External id": 46903, + "cbid": 273, "correlation": 46903 + } + }, + { + "ph": "f", "id": 46903, "pid": 435, "tid": 435, "ts": 1776649791310051, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791310060, "dur": 15, + "args": { + "External id": 46904, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46904, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46904, "pid": 0, "tid": 7, "ts": 1776649791310060, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310052, "dur": 4, + "args": { + "External id": 46904, + "cbid": 211, "correlation": 46904 + } + }, + { + "ph": "s", "id": 46904, "pid": 435, "tid": 435, "ts": 1776649791310052, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791310096, "dur": 1, + "args": { + "External id": 5954, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46922, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46922, "pid": 0, "tid": 7, "ts": 1776649791310096, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310086, "dur": 6, + "args": { + "External id": 5954, + "cbid": 211, "correlation": 46922 + } + }, + { + "ph": "s", "id": 46922, "pid": 435, "tid": 435, "ts": 1776649791310086, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310099, "dur": 0, + "args": { + "External id": 46928, + "cbid": 200, "correlation": 46928 + } + }, + { + "ph": "f", "id": 46928, "pid": 435, "tid": 435, "ts": 1776649791310099, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310100, "dur": 0, + "args": { + "External id": 46929, + "cbid": 273, "correlation": 46929 + } + }, + { + "ph": "f", "id": 46929, "pid": 435, "tid": 435, "ts": 1776649791310100, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791310109, "dur": 16, + "args": { + "External id": 46930, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46930, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46930, "pid": 0, "tid": 7, "ts": 1776649791310109, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310101, "dur": 4, + "args": { + "External id": 46930, + "cbid": 211, "correlation": 46930 + } + }, + { + "ph": "s", "id": 46930, "pid": 435, "tid": 435, "ts": 1776649791310101, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791310144, "dur": 1, + "args": { + "External id": 5959, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46948, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46948, "pid": 0, "tid": 7, "ts": 1776649791310144, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310135, "dur": 7, + "args": { + "External id": 5959, + "cbid": 211, "correlation": 46948 + } + }, + { + "ph": "s", "id": 46948, "pid": 435, "tid": 435, "ts": 1776649791310135, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310147, "dur": 0, + "args": { + "External id": 46954, + "cbid": 200, "correlation": 46954 + } + }, + { + "ph": "f", "id": 46954, "pid": 435, "tid": 435, "ts": 1776649791310147, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310148, "dur": 0, + "args": { + "External id": 46955, + "cbid": 273, "correlation": 46955 + } + }, + { + "ph": "f", "id": 46955, "pid": 435, "tid": 435, "ts": 1776649791310148, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791310157, "dur": 15, + "args": { + "External id": 46956, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46956, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46956, "pid": 0, "tid": 7, "ts": 1776649791310157, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310149, "dur": 4, + "args": { + "External id": 46956, + "cbid": 211, "correlation": 46956 + } + }, + { + "ph": "s", "id": 46956, "pid": 435, "tid": 435, "ts": 1776649791310149, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791310192, "dur": 1, + "args": { + "External id": 5964, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46974, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46974, "pid": 0, "tid": 7, "ts": 1776649791310192, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310183, "dur": 6, + "args": { + "External id": 5964, + "cbid": 211, "correlation": 46974 + } + }, + { + "ph": "s", "id": 46974, "pid": 435, "tid": 435, "ts": 1776649791310183, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310196, "dur": 0, + "args": { + "External id": 46980, + "cbid": 200, "correlation": 46980 + } + }, + { + "ph": "f", "id": 46980, "pid": 435, "tid": 435, "ts": 1776649791310196, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310196, "dur": 0, + "args": { + "External id": 46981, + "cbid": 273, "correlation": 46981 + } + }, + { + "ph": "f", "id": 46981, "pid": 435, "tid": 435, "ts": 1776649791310196, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791310205, "dur": 15, + "args": { + "External id": 46982, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 46982, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 46982, "pid": 0, "tid": 7, "ts": 1776649791310205, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310197, "dur": 4, + "args": { + "External id": 46982, + "cbid": 211, "correlation": 46982 + } + }, + { + "ph": "s", "id": 46982, "pid": 435, "tid": 435, "ts": 1776649791310197, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791310241, "dur": 1, + "args": { + "External id": 5969, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47000, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47000, "pid": 0, "tid": 7, "ts": 1776649791310241, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310232, "dur": 6, + "args": { + "External id": 5969, + "cbid": 211, "correlation": 47000 + } + }, + { + "ph": "s", "id": 47000, "pid": 435, "tid": 435, "ts": 1776649791310232, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310244, "dur": 0, + "args": { + "External id": 47006, + "cbid": 200, "correlation": 47006 + } + }, + { + "ph": "f", "id": 47006, "pid": 435, "tid": 435, "ts": 1776649791310244, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310245, "dur": 0, + "args": { + "External id": 47007, + "cbid": 273, "correlation": 47007 + } + }, + { + "ph": "f", "id": 47007, "pid": 435, "tid": 435, "ts": 1776649791310245, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791310254, "dur": 15, + "args": { + "External id": 47008, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47008, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47008, "pid": 0, "tid": 7, "ts": 1776649791310254, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310246, "dur": 4, + "args": { + "External id": 47008, + "cbid": 211, "correlation": 47008 + } + }, + { + "ph": "s", "id": 47008, "pid": 435, "tid": 435, "ts": 1776649791310246, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791310288, "dur": 1, + "args": { + "External id": 5974, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47026, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47026, "pid": 0, "tid": 7, "ts": 1776649791310288, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310279, "dur": 6, + "args": { + "External id": 5974, + "cbid": 211, "correlation": 47026 + } + }, + { + "ph": "s", "id": 47026, "pid": 435, "tid": 435, "ts": 1776649791310279, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310292, "dur": 0, + "args": { + "External id": 47032, + "cbid": 200, "correlation": 47032 + } + }, + { + "ph": "f", "id": 47032, "pid": 435, "tid": 435, "ts": 1776649791310292, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310292, "dur": 0, + "args": { + "External id": 47033, + "cbid": 273, "correlation": 47033 + } + }, + { + "ph": "f", "id": 47033, "pid": 435, "tid": 435, "ts": 1776649791310292, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791310302, "dur": 15, + "args": { + "External id": 47034, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47034, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47034, "pid": 0, "tid": 7, "ts": 1776649791310302, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310294, "dur": 4, + "args": { + "External id": 47034, + "cbid": 211, "correlation": 47034 + } + }, + { + "ph": "s", "id": 47034, "pid": 435, "tid": 435, "ts": 1776649791310294, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791310338, "dur": 1, + "args": { + "External id": 5979, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47052, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47052, "pid": 0, "tid": 7, "ts": 1776649791310338, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310328, "dur": 6, + "args": { + "External id": 5979, + "cbid": 211, "correlation": 47052 + } + }, + { + "ph": "s", "id": 47052, "pid": 435, "tid": 435, "ts": 1776649791310328, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310340, "dur": 0, + "args": { + "External id": 47058, + "cbid": 200, "correlation": 47058 + } + }, + { + "ph": "f", "id": 47058, "pid": 435, "tid": 435, "ts": 1776649791310340, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310341, "dur": 0, + "args": { + "External id": 47059, + "cbid": 273, "correlation": 47059 + } + }, + { + "ph": "f", "id": 47059, "pid": 435, "tid": 435, "ts": 1776649791310341, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791310350, "dur": 16, + "args": { + "External id": 47060, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47060, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47060, "pid": 0, "tid": 7, "ts": 1776649791310350, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310342, "dur": 5, + "args": { + "External id": 47060, + "cbid": 211, "correlation": 47060 + } + }, + { + "ph": "s", "id": 47060, "pid": 435, "tid": 435, "ts": 1776649791310342, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791310364, "dur": 5, + "args": { + "External id": 47063, + "cbid": 165, "correlation": 47063 + } + }, + { + "ph": "s", "id": 47063, "pid": 435, "tid": 435, "ts": 1776649791310364, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791310392, "dur": 3, + "args": { + "External id": 47072, + "cbid": 135, "correlation": 47072 + } + }, + { + "ph": "f", "id": 47072, "pid": 435, "tid": 435, "ts": 1776649791310392, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791310440, "dur": 1, + "args": { + "External id": 5984, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47092, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47092, "pid": 0, "tid": 7, "ts": 1776649791310440, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310429, "dur": 8, + "args": { + "External id": 5984, + "cbid": 211, "correlation": 47092 + } + }, + { + "ph": "s", "id": 47092, "pid": 435, "tid": 435, "ts": 1776649791310429, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310444, "dur": 0, + "args": { + "External id": 47098, + "cbid": 200, "correlation": 47098 + } + }, + { + "ph": "f", "id": 47098, "pid": 435, "tid": 435, "ts": 1776649791310444, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310445, "dur": 1, + "args": { + "External id": 47099, + "cbid": 273, "correlation": 47099 + } + }, + { + "ph": "f", "id": 47099, "pid": 435, "tid": 435, "ts": 1776649791310445, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791310456, "dur": 15, + "args": { + "External id": 47100, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47100, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47100, "pid": 0, "tid": 7, "ts": 1776649791310456, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310447, "dur": 5, + "args": { + "External id": 47100, + "cbid": 211, "correlation": 47100 + } + }, + { + "ph": "s", "id": 47100, "pid": 435, "tid": 435, "ts": 1776649791310447, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791310492, "dur": 1, + "args": { + "External id": 5989, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47118, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47118, "pid": 0, "tid": 7, "ts": 1776649791310492, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310483, "dur": 6, + "args": { + "External id": 5989, + "cbid": 211, "correlation": 47118 + } + }, + { + "ph": "s", "id": 47118, "pid": 435, "tid": 435, "ts": 1776649791310483, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310496, "dur": 0, + "args": { + "External id": 47124, + "cbid": 200, "correlation": 47124 + } + }, + { + "ph": "f", "id": 47124, "pid": 435, "tid": 435, "ts": 1776649791310496, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310496, "dur": 0, + "args": { + "External id": 47125, + "cbid": 273, "correlation": 47125 + } + }, + { + "ph": "f", "id": 47125, "pid": 435, "tid": 435, "ts": 1776649791310496, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791310506, "dur": 15, + "args": { + "External id": 47126, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47126, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47126, "pid": 0, "tid": 7, "ts": 1776649791310506, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310498, "dur": 4, + "args": { + "External id": 47126, + "cbid": 211, "correlation": 47126 + } + }, + { + "ph": "s", "id": 47126, "pid": 435, "tid": 435, "ts": 1776649791310498, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791310541, "dur": 1, + "args": { + "External id": 5994, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47144, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47144, "pid": 0, "tid": 7, "ts": 1776649791310541, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310532, "dur": 6, + "args": { + "External id": 5994, + "cbid": 211, "correlation": 47144 + } + }, + { + "ph": "s", "id": 47144, "pid": 435, "tid": 435, "ts": 1776649791310532, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310545, "dur": 0, + "args": { + "External id": 47150, + "cbid": 200, "correlation": 47150 + } + }, + { + "ph": "f", "id": 47150, "pid": 435, "tid": 435, "ts": 1776649791310545, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310545, "dur": 0, + "args": { + "External id": 47151, + "cbid": 273, "correlation": 47151 + } + }, + { + "ph": "f", "id": 47151, "pid": 435, "tid": 435, "ts": 1776649791310545, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791310554, "dur": 16, + "args": { + "External id": 47152, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47152, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47152, "pid": 0, "tid": 7, "ts": 1776649791310554, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310546, "dur": 5, + "args": { + "External id": 47152, + "cbid": 211, "correlation": 47152 + } + }, + { + "ph": "s", "id": 47152, "pid": 435, "tid": 435, "ts": 1776649791310546, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791310590, "dur": 1, + "args": { + "External id": 5999, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47170, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47170, "pid": 0, "tid": 7, "ts": 1776649791310590, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310581, "dur": 6, + "args": { + "External id": 5999, + "cbid": 211, "correlation": 47170 + } + }, + { + "ph": "s", "id": 47170, "pid": 435, "tid": 435, "ts": 1776649791310581, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310593, "dur": 0, + "args": { + "External id": 47176, + "cbid": 200, "correlation": 47176 + } + }, + { + "ph": "f", "id": 47176, "pid": 435, "tid": 435, "ts": 1776649791310593, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310594, "dur": 0, + "args": { + "External id": 47177, + "cbid": 273, "correlation": 47177 + } + }, + { + "ph": "f", "id": 47177, "pid": 435, "tid": 435, "ts": 1776649791310594, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791310603, "dur": 16, + "args": { + "External id": 47178, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47178, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47178, "pid": 0, "tid": 7, "ts": 1776649791310603, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310595, "dur": 4, + "args": { + "External id": 47178, + "cbid": 211, "correlation": 47178 + } + }, + { + "ph": "s", "id": 47178, "pid": 435, "tid": 435, "ts": 1776649791310595, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791310639, "dur": 1, + "args": { + "External id": 6004, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47196, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47196, "pid": 0, "tid": 7, "ts": 1776649791310639, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310630, "dur": 6, + "args": { + "External id": 6004, + "cbid": 211, "correlation": 47196 + } + }, + { + "ph": "s", "id": 47196, "pid": 435, "tid": 435, "ts": 1776649791310630, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310642, "dur": 0, + "args": { + "External id": 47202, + "cbid": 200, "correlation": 47202 + } + }, + { + "ph": "f", "id": 47202, "pid": 435, "tid": 435, "ts": 1776649791310642, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310643, "dur": 0, + "args": { + "External id": 47203, + "cbid": 273, "correlation": 47203 + } + }, + { + "ph": "f", "id": 47203, "pid": 435, "tid": 435, "ts": 1776649791310643, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791310652, "dur": 15, + "args": { + "External id": 47204, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47204, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47204, "pid": 0, "tid": 7, "ts": 1776649791310652, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310644, "dur": 5, + "args": { + "External id": 47204, + "cbid": 211, "correlation": 47204 + } + }, + { + "ph": "s", "id": 47204, "pid": 435, "tid": 435, "ts": 1776649791310644, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791310687, "dur": 1, + "args": { + "External id": 6009, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47222, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47222, "pid": 0, "tid": 7, "ts": 1776649791310687, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310678, "dur": 6, + "args": { + "External id": 6009, + "cbid": 211, "correlation": 47222 + } + }, + { + "ph": "s", "id": 47222, "pid": 435, "tid": 435, "ts": 1776649791310678, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310690, "dur": 0, + "args": { + "External id": 47228, + "cbid": 200, "correlation": 47228 + } + }, + { + "ph": "f", "id": 47228, "pid": 435, "tid": 435, "ts": 1776649791310690, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310691, "dur": 0, + "args": { + "External id": 47229, + "cbid": 273, "correlation": 47229 + } + }, + { + "ph": "f", "id": 47229, "pid": 435, "tid": 435, "ts": 1776649791310691, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791310699, "dur": 15, + "args": { + "External id": 47230, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47230, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47230, "pid": 0, "tid": 7, "ts": 1776649791310699, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310692, "dur": 5, + "args": { + "External id": 47230, + "cbid": 211, "correlation": 47230 + } + }, + { + "ph": "s", "id": 47230, "pid": 435, "tid": 435, "ts": 1776649791310692, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791310735, "dur": 1, + "args": { + "External id": 6014, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47248, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47248, "pid": 0, "tid": 7, "ts": 1776649791310735, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310726, "dur": 6, + "args": { + "External id": 6014, + "cbid": 211, "correlation": 47248 + } + }, + { + "ph": "s", "id": 47248, "pid": 435, "tid": 435, "ts": 1776649791310726, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310738, "dur": 0, + "args": { + "External id": 47254, + "cbid": 200, "correlation": 47254 + } + }, + { + "ph": "f", "id": 47254, "pid": 435, "tid": 435, "ts": 1776649791310738, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310739, "dur": 0, + "args": { + "External id": 47255, + "cbid": 273, "correlation": 47255 + } + }, + { + "ph": "f", "id": 47255, "pid": 435, "tid": 435, "ts": 1776649791310739, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791310748, "dur": 15, + "args": { + "External id": 47256, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47256, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47256, "pid": 0, "tid": 7, "ts": 1776649791310748, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310740, "dur": 4, + "args": { + "External id": 47256, + "cbid": 211, "correlation": 47256 + } + }, + { + "ph": "s", "id": 47256, "pid": 435, "tid": 435, "ts": 1776649791310740, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791310783, "dur": 1, + "args": { + "External id": 6019, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47274, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47274, "pid": 0, "tid": 7, "ts": 1776649791310783, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310774, "dur": 6, + "args": { + "External id": 6019, + "cbid": 211, "correlation": 47274 + } + }, + { + "ph": "s", "id": 47274, "pid": 435, "tid": 435, "ts": 1776649791310774, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310786, "dur": 0, + "args": { + "External id": 47280, + "cbid": 200, "correlation": 47280 + } + }, + { + "ph": "f", "id": 47280, "pid": 435, "tid": 435, "ts": 1776649791310786, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310787, "dur": 0, + "args": { + "External id": 47281, + "cbid": 273, "correlation": 47281 + } + }, + { + "ph": "f", "id": 47281, "pid": 435, "tid": 435, "ts": 1776649791310787, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791310796, "dur": 15, + "args": { + "External id": 47282, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47282, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47282, "pid": 0, "tid": 7, "ts": 1776649791310796, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310788, "dur": 4, + "args": { + "External id": 47282, + "cbid": 211, "correlation": 47282 + } + }, + { + "ph": "s", "id": 47282, "pid": 435, "tid": 435, "ts": 1776649791310788, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791310831, "dur": 1, + "args": { + "External id": 6024, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47300, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47300, "pid": 0, "tid": 7, "ts": 1776649791310831, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310822, "dur": 6, + "args": { + "External id": 6024, + "cbid": 211, "correlation": 47300 + } + }, + { + "ph": "s", "id": 47300, "pid": 435, "tid": 435, "ts": 1776649791310822, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310835, "dur": 0, + "args": { + "External id": 47306, + "cbid": 200, "correlation": 47306 + } + }, + { + "ph": "f", "id": 47306, "pid": 435, "tid": 435, "ts": 1776649791310835, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310835, "dur": 0, + "args": { + "External id": 47307, + "cbid": 273, "correlation": 47307 + } + }, + { + "ph": "f", "id": 47307, "pid": 435, "tid": 435, "ts": 1776649791310835, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791310844, "dur": 15, + "args": { + "External id": 47308, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47308, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47308, "pid": 0, "tid": 7, "ts": 1776649791310844, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310836, "dur": 4, + "args": { + "External id": 47308, + "cbid": 211, "correlation": 47308 + } + }, + { + "ph": "s", "id": 47308, "pid": 435, "tid": 435, "ts": 1776649791310836, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791310880, "dur": 1, + "args": { + "External id": 6029, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47326, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47326, "pid": 0, "tid": 7, "ts": 1776649791310880, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310870, "dur": 6, + "args": { + "External id": 6029, + "cbid": 211, "correlation": 47326 + } + }, + { + "ph": "s", "id": 47326, "pid": 435, "tid": 435, "ts": 1776649791310870, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310883, "dur": 0, + "args": { + "External id": 47332, + "cbid": 200, "correlation": 47332 + } + }, + { + "ph": "f", "id": 47332, "pid": 435, "tid": 435, "ts": 1776649791310883, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310883, "dur": 0, + "args": { + "External id": 47333, + "cbid": 273, "correlation": 47333 + } + }, + { + "ph": "f", "id": 47333, "pid": 435, "tid": 435, "ts": 1776649791310883, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791310892, "dur": 15, + "args": { + "External id": 47334, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47334, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47334, "pid": 0, "tid": 7, "ts": 1776649791310892, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310885, "dur": 4, + "args": { + "External id": 47334, + "cbid": 211, "correlation": 47334 + } + }, + { + "ph": "s", "id": 47334, "pid": 435, "tid": 435, "ts": 1776649791310885, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791310928, "dur": 1, + "args": { + "External id": 6034, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47352, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47352, "pid": 0, "tid": 7, "ts": 1776649791310928, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310918, "dur": 6, + "args": { + "External id": 6034, + "cbid": 211, "correlation": 47352 + } + }, + { + "ph": "s", "id": 47352, "pid": 435, "tid": 435, "ts": 1776649791310918, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310931, "dur": 0, + "args": { + "External id": 47358, + "cbid": 200, "correlation": 47358 + } + }, + { + "ph": "f", "id": 47358, "pid": 435, "tid": 435, "ts": 1776649791310931, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310931, "dur": 0, + "args": { + "External id": 47359, + "cbid": 273, "correlation": 47359 + } + }, + { + "ph": "f", "id": 47359, "pid": 435, "tid": 435, "ts": 1776649791310931, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791310941, "dur": 15, + "args": { + "External id": 47360, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47360, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47360, "pid": 0, "tid": 7, "ts": 1776649791310941, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310932, "dur": 4, + "args": { + "External id": 47360, + "cbid": 211, "correlation": 47360 + } + }, + { + "ph": "s", "id": 47360, "pid": 435, "tid": 435, "ts": 1776649791310932, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791310976, "dur": 1, + "args": { + "External id": 6039, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47378, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47378, "pid": 0, "tid": 7, "ts": 1776649791310976, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310966, "dur": 6, + "args": { + "External id": 6039, + "cbid": 211, "correlation": 47378 + } + }, + { + "ph": "s", "id": 47378, "pid": 435, "tid": 435, "ts": 1776649791310966, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310979, "dur": 0, + "args": { + "External id": 47384, + "cbid": 200, "correlation": 47384 + } + }, + { + "ph": "f", "id": 47384, "pid": 435, "tid": 435, "ts": 1776649791310979, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791310979, "dur": 0, + "args": { + "External id": 47385, + "cbid": 273, "correlation": 47385 + } + }, + { + "ph": "f", "id": 47385, "pid": 435, "tid": 435, "ts": 1776649791310979, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791310988, "dur": 16, + "args": { + "External id": 47386, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47386, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47386, "pid": 0, "tid": 7, "ts": 1776649791310988, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791310980, "dur": 4, + "args": { + "External id": 47386, + "cbid": 211, "correlation": 47386 + } + }, + { + "ph": "s", "id": 47386, "pid": 435, "tid": 435, "ts": 1776649791310980, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791311023, "dur": 1, + "args": { + "External id": 6044, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47404, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47404, "pid": 0, "tid": 7, "ts": 1776649791311023, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311014, "dur": 6, + "args": { + "External id": 6044, + "cbid": 211, "correlation": 47404 + } + }, + { + "ph": "s", "id": 47404, "pid": 435, "tid": 435, "ts": 1776649791311014, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311026, "dur": 0, + "args": { + "External id": 47410, + "cbid": 200, "correlation": 47410 + } + }, + { + "ph": "f", "id": 47410, "pid": 435, "tid": 435, "ts": 1776649791311026, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311027, "dur": 0, + "args": { + "External id": 47411, + "cbid": 273, "correlation": 47411 + } + }, + { + "ph": "f", "id": 47411, "pid": 435, "tid": 435, "ts": 1776649791311027, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791311036, "dur": 16, + "args": { + "External id": 47412, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47412, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47412, "pid": 0, "tid": 7, "ts": 1776649791311036, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311028, "dur": 4, + "args": { + "External id": 47412, + "cbid": 211, "correlation": 47412 + } + }, + { + "ph": "s", "id": 47412, "pid": 435, "tid": 435, "ts": 1776649791311028, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791311071, "dur": 1, + "args": { + "External id": 6049, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47430, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47430, "pid": 0, "tid": 7, "ts": 1776649791311071, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311062, "dur": 6, + "args": { + "External id": 6049, + "cbid": 211, "correlation": 47430 + } + }, + { + "ph": "s", "id": 47430, "pid": 435, "tid": 435, "ts": 1776649791311062, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311074, "dur": 0, + "args": { + "External id": 47436, + "cbid": 200, "correlation": 47436 + } + }, + { + "ph": "f", "id": 47436, "pid": 435, "tid": 435, "ts": 1776649791311074, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311075, "dur": 0, + "args": { + "External id": 47437, + "cbid": 273, "correlation": 47437 + } + }, + { + "ph": "f", "id": 47437, "pid": 435, "tid": 435, "ts": 1776649791311075, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791311084, "dur": 15, + "args": { + "External id": 47438, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47438, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47438, "pid": 0, "tid": 7, "ts": 1776649791311084, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311076, "dur": 4, + "args": { + "External id": 47438, + "cbid": 211, "correlation": 47438 + } + }, + { + "ph": "s", "id": 47438, "pid": 435, "tid": 435, "ts": 1776649791311076, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791311119, "dur": 1, + "args": { + "External id": 6054, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47456, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47456, "pid": 0, "tid": 7, "ts": 1776649791311119, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311110, "dur": 6, + "args": { + "External id": 6054, + "cbid": 211, "correlation": 47456 + } + }, + { + "ph": "s", "id": 47456, "pid": 435, "tid": 435, "ts": 1776649791311110, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311122, "dur": 0, + "args": { + "External id": 47462, + "cbid": 200, "correlation": 47462 + } + }, + { + "ph": "f", "id": 47462, "pid": 435, "tid": 435, "ts": 1776649791311122, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311123, "dur": 0, + "args": { + "External id": 47463, + "cbid": 273, "correlation": 47463 + } + }, + { + "ph": "f", "id": 47463, "pid": 435, "tid": 435, "ts": 1776649791311123, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791311132, "dur": 15, + "args": { + "External id": 47464, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47464, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47464, "pid": 0, "tid": 7, "ts": 1776649791311132, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311124, "dur": 5, + "args": { + "External id": 47464, + "cbid": 211, "correlation": 47464 + } + }, + { + "ph": "s", "id": 47464, "pid": 435, "tid": 435, "ts": 1776649791311124, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791311167, "dur": 1, + "args": { + "External id": 6059, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47482, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47482, "pid": 0, "tid": 7, "ts": 1776649791311167, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311158, "dur": 6, + "args": { + "External id": 6059, + "cbid": 211, "correlation": 47482 + } + }, + { + "ph": "s", "id": 47482, "pid": 435, "tid": 435, "ts": 1776649791311158, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311171, "dur": 0, + "args": { + "External id": 47488, + "cbid": 200, "correlation": 47488 + } + }, + { + "ph": "f", "id": 47488, "pid": 435, "tid": 435, "ts": 1776649791311171, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311171, "dur": 0, + "args": { + "External id": 47489, + "cbid": 273, "correlation": 47489 + } + }, + { + "ph": "f", "id": 47489, "pid": 435, "tid": 435, "ts": 1776649791311171, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791311180, "dur": 15, + "args": { + "External id": 47490, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47490, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47490, "pid": 0, "tid": 7, "ts": 1776649791311180, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311172, "dur": 4, + "args": { + "External id": 47490, + "cbid": 211, "correlation": 47490 + } + }, + { + "ph": "s", "id": 47490, "pid": 435, "tid": 435, "ts": 1776649791311172, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791311216, "dur": 1, + "args": { + "External id": 6064, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47508, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47508, "pid": 0, "tid": 7, "ts": 1776649791311216, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311206, "dur": 6, + "args": { + "External id": 6064, + "cbid": 211, "correlation": 47508 + } + }, + { + "ph": "s", "id": 47508, "pid": 435, "tid": 435, "ts": 1776649791311206, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311219, "dur": 0, + "args": { + "External id": 47514, + "cbid": 200, "correlation": 47514 + } + }, + { + "ph": "f", "id": 47514, "pid": 435, "tid": 435, "ts": 1776649791311219, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311219, "dur": 0, + "args": { + "External id": 47515, + "cbid": 273, "correlation": 47515 + } + }, + { + "ph": "f", "id": 47515, "pid": 435, "tid": 435, "ts": 1776649791311219, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791311228, "dur": 15, + "args": { + "External id": 47516, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47516, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47516, "pid": 0, "tid": 7, "ts": 1776649791311228, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311221, "dur": 5, + "args": { + "External id": 47516, + "cbid": 211, "correlation": 47516 + } + }, + { + "ph": "s", "id": 47516, "pid": 435, "tid": 435, "ts": 1776649791311221, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791311265, "dur": 1, + "args": { + "External id": 6069, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47534, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47534, "pid": 0, "tid": 7, "ts": 1776649791311265, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311255, "dur": 6, + "args": { + "External id": 6069, + "cbid": 211, "correlation": 47534 + } + }, + { + "ph": "s", "id": 47534, "pid": 435, "tid": 435, "ts": 1776649791311255, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311268, "dur": 0, + "args": { + "External id": 47540, + "cbid": 200, "correlation": 47540 + } + }, + { + "ph": "f", "id": 47540, "pid": 435, "tid": 435, "ts": 1776649791311268, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311269, "dur": 0, + "args": { + "External id": 47541, + "cbid": 273, "correlation": 47541 + } + }, + { + "ph": "f", "id": 47541, "pid": 435, "tid": 435, "ts": 1776649791311269, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791311278, "dur": 16, + "args": { + "External id": 47542, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47542, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47542, "pid": 0, "tid": 7, "ts": 1776649791311278, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311270, "dur": 4, + "args": { + "External id": 47542, + "cbid": 211, "correlation": 47542 + } + }, + { + "ph": "s", "id": 47542, "pid": 435, "tid": 435, "ts": 1776649791311270, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791311313, "dur": 1, + "args": { + "External id": 6074, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47560, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47560, "pid": 0, "tid": 7, "ts": 1776649791311313, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311304, "dur": 6, + "args": { + "External id": 6074, + "cbid": 211, "correlation": 47560 + } + }, + { + "ph": "s", "id": 47560, "pid": 435, "tid": 435, "ts": 1776649791311304, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311316, "dur": 0, + "args": { + "External id": 47566, + "cbid": 200, "correlation": 47566 + } + }, + { + "ph": "f", "id": 47566, "pid": 435, "tid": 435, "ts": 1776649791311316, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311317, "dur": 0, + "args": { + "External id": 47567, + "cbid": 273, "correlation": 47567 + } + }, + { + "ph": "f", "id": 47567, "pid": 435, "tid": 435, "ts": 1776649791311317, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791311326, "dur": 15, + "args": { + "External id": 47568, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47568, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47568, "pid": 0, "tid": 7, "ts": 1776649791311326, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311318, "dur": 4, + "args": { + "External id": 47568, + "cbid": 211, "correlation": 47568 + } + }, + { + "ph": "s", "id": 47568, "pid": 435, "tid": 435, "ts": 1776649791311318, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791311362, "dur": 1, + "args": { + "External id": 6079, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47586, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47586, "pid": 0, "tid": 7, "ts": 1776649791311362, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311353, "dur": 6, + "args": { + "External id": 6079, + "cbid": 211, "correlation": 47586 + } + }, + { + "ph": "s", "id": 47586, "pid": 435, "tid": 435, "ts": 1776649791311353, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311365, "dur": 0, + "args": { + "External id": 47592, + "cbid": 200, "correlation": 47592 + } + }, + { + "ph": "f", "id": 47592, "pid": 435, "tid": 435, "ts": 1776649791311365, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311366, "dur": 0, + "args": { + "External id": 47593, + "cbid": 273, "correlation": 47593 + } + }, + { + "ph": "f", "id": 47593, "pid": 435, "tid": 435, "ts": 1776649791311366, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791311374, "dur": 15, + "args": { + "External id": 47594, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47594, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47594, "pid": 0, "tid": 7, "ts": 1776649791311374, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311367, "dur": 4, + "args": { + "External id": 47594, + "cbid": 211, "correlation": 47594 + } + }, + { + "ph": "s", "id": 47594, "pid": 435, "tid": 435, "ts": 1776649791311367, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791311410, "dur": 1, + "args": { + "External id": 6084, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47612, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47612, "pid": 0, "tid": 7, "ts": 1776649791311410, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311401, "dur": 6, + "args": { + "External id": 6084, + "cbid": 211, "correlation": 47612 + } + }, + { + "ph": "s", "id": 47612, "pid": 435, "tid": 435, "ts": 1776649791311401, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311413, "dur": 0, + "args": { + "External id": 47618, + "cbid": 200, "correlation": 47618 + } + }, + { + "ph": "f", "id": 47618, "pid": 435, "tid": 435, "ts": 1776649791311413, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311414, "dur": 0, + "args": { + "External id": 47619, + "cbid": 273, "correlation": 47619 + } + }, + { + "ph": "f", "id": 47619, "pid": 435, "tid": 435, "ts": 1776649791311414, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791311423, "dur": 17, + "args": { + "External id": 47620, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47620, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47620, "pid": 0, "tid": 7, "ts": 1776649791311423, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311415, "dur": 4, + "args": { + "External id": 47620, + "cbid": 211, "correlation": 47620 + } + }, + { + "ph": "s", "id": 47620, "pid": 435, "tid": 435, "ts": 1776649791311415, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791311458, "dur": 1, + "args": { + "External id": 6089, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47638, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47638, "pid": 0, "tid": 7, "ts": 1776649791311458, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311449, "dur": 6, + "args": { + "External id": 6089, + "cbid": 211, "correlation": 47638 + } + }, + { + "ph": "s", "id": 47638, "pid": 435, "tid": 435, "ts": 1776649791311449, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311461, "dur": 0, + "args": { + "External id": 47644, + "cbid": 200, "correlation": 47644 + } + }, + { + "ph": "f", "id": 47644, "pid": 435, "tid": 435, "ts": 1776649791311461, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311462, "dur": 0, + "args": { + "External id": 47645, + "cbid": 273, "correlation": 47645 + } + }, + { + "ph": "f", "id": 47645, "pid": 435, "tid": 435, "ts": 1776649791311462, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791311471, "dur": 15, + "args": { + "External id": 47646, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47646, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47646, "pid": 0, "tid": 7, "ts": 1776649791311471, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311463, "dur": 4, + "args": { + "External id": 47646, + "cbid": 211, "correlation": 47646 + } + }, + { + "ph": "s", "id": 47646, "pid": 435, "tid": 435, "ts": 1776649791311463, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791311506, "dur": 1, + "args": { + "External id": 6094, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47664, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47664, "pid": 0, "tid": 7, "ts": 1776649791311506, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311497, "dur": 6, + "args": { + "External id": 6094, + "cbid": 211, "correlation": 47664 + } + }, + { + "ph": "s", "id": 47664, "pid": 435, "tid": 435, "ts": 1776649791311497, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311510, "dur": 0, + "args": { + "External id": 47670, + "cbid": 200, "correlation": 47670 + } + }, + { + "ph": "f", "id": 47670, "pid": 435, "tid": 435, "ts": 1776649791311510, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311510, "dur": 0, + "args": { + "External id": 47671, + "cbid": 273, "correlation": 47671 + } + }, + { + "ph": "f", "id": 47671, "pid": 435, "tid": 435, "ts": 1776649791311510, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791311519, "dur": 15, + "args": { + "External id": 47672, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47672, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47672, "pid": 0, "tid": 7, "ts": 1776649791311519, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311511, "dur": 4, + "args": { + "External id": 47672, + "cbid": 211, "correlation": 47672 + } + }, + { + "ph": "s", "id": 47672, "pid": 435, "tid": 435, "ts": 1776649791311511, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791311554, "dur": 1, + "args": { + "External id": 6099, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47690, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47690, "pid": 0, "tid": 7, "ts": 1776649791311554, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311545, "dur": 6, + "args": { + "External id": 6099, + "cbid": 211, "correlation": 47690 + } + }, + { + "ph": "s", "id": 47690, "pid": 435, "tid": 435, "ts": 1776649791311545, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311557, "dur": 0, + "args": { + "External id": 47696, + "cbid": 200, "correlation": 47696 + } + }, + { + "ph": "f", "id": 47696, "pid": 435, "tid": 435, "ts": 1776649791311557, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311558, "dur": 0, + "args": { + "External id": 47697, + "cbid": 273, "correlation": 47697 + } + }, + { + "ph": "f", "id": 47697, "pid": 435, "tid": 435, "ts": 1776649791311558, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791311567, "dur": 15, + "args": { + "External id": 47698, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47698, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47698, "pid": 0, "tid": 7, "ts": 1776649791311567, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311559, "dur": 4, + "args": { + "External id": 47698, + "cbid": 211, "correlation": 47698 + } + }, + { + "ph": "s", "id": 47698, "pid": 435, "tid": 435, "ts": 1776649791311559, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791311603, "dur": 1, + "args": { + "External id": 6104, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47716, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47716, "pid": 0, "tid": 7, "ts": 1776649791311603, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311594, "dur": 6, + "args": { + "External id": 6104, + "cbid": 211, "correlation": 47716 + } + }, + { + "ph": "s", "id": 47716, "pid": 435, "tid": 435, "ts": 1776649791311594, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311606, "dur": 0, + "args": { + "External id": 47722, + "cbid": 200, "correlation": 47722 + } + }, + { + "ph": "f", "id": 47722, "pid": 435, "tid": 435, "ts": 1776649791311606, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311606, "dur": 0, + "args": { + "External id": 47723, + "cbid": 273, "correlation": 47723 + } + }, + { + "ph": "f", "id": 47723, "pid": 435, "tid": 435, "ts": 1776649791311606, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791311616, "dur": 15, + "args": { + "External id": 47724, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47724, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47724, "pid": 0, "tid": 7, "ts": 1776649791311616, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311608, "dur": 5, + "args": { + "External id": 47724, + "cbid": 211, "correlation": 47724 + } + }, + { + "ph": "s", "id": 47724, "pid": 435, "tid": 435, "ts": 1776649791311608, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791311651, "dur": 1, + "args": { + "External id": 6109, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47742, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47742, "pid": 0, "tid": 7, "ts": 1776649791311651, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311642, "dur": 6, + "args": { + "External id": 6109, + "cbid": 211, "correlation": 47742 + } + }, + { + "ph": "s", "id": 47742, "pid": 435, "tid": 435, "ts": 1776649791311642, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311654, "dur": 0, + "args": { + "External id": 47748, + "cbid": 200, "correlation": 47748 + } + }, + { + "ph": "f", "id": 47748, "pid": 435, "tid": 435, "ts": 1776649791311654, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311654, "dur": 0, + "args": { + "External id": 47749, + "cbid": 273, "correlation": 47749 + } + }, + { + "ph": "f", "id": 47749, "pid": 435, "tid": 435, "ts": 1776649791311654, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791311663, "dur": 15, + "args": { + "External id": 47750, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47750, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47750, "pid": 0, "tid": 7, "ts": 1776649791311663, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311655, "dur": 4, + "args": { + "External id": 47750, + "cbid": 211, "correlation": 47750 + } + }, + { + "ph": "s", "id": 47750, "pid": 435, "tid": 435, "ts": 1776649791311655, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791311699, "dur": 1, + "args": { + "External id": 6114, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47768, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47768, "pid": 0, "tid": 7, "ts": 1776649791311699, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311690, "dur": 6, + "args": { + "External id": 6114, + "cbid": 211, "correlation": 47768 + } + }, + { + "ph": "s", "id": 47768, "pid": 435, "tid": 435, "ts": 1776649791311690, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311702, "dur": 0, + "args": { + "External id": 47774, + "cbid": 200, "correlation": 47774 + } + }, + { + "ph": "f", "id": 47774, "pid": 435, "tid": 435, "ts": 1776649791311702, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311703, "dur": 0, + "args": { + "External id": 47775, + "cbid": 273, "correlation": 47775 + } + }, + { + "ph": "f", "id": 47775, "pid": 435, "tid": 435, "ts": 1776649791311703, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791311712, "dur": 16, + "args": { + "External id": 47776, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47776, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47776, "pid": 0, "tid": 7, "ts": 1776649791311712, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311704, "dur": 4, + "args": { + "External id": 47776, + "cbid": 211, "correlation": 47776 + } + }, + { + "ph": "s", "id": 47776, "pid": 435, "tid": 435, "ts": 1776649791311704, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791311747, "dur": 1, + "args": { + "External id": 6119, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47794, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47794, "pid": 0, "tid": 7, "ts": 1776649791311747, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311738, "dur": 6, + "args": { + "External id": 6119, + "cbid": 211, "correlation": 47794 + } + }, + { + "ph": "s", "id": 47794, "pid": 435, "tid": 435, "ts": 1776649791311738, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311750, "dur": 0, + "args": { + "External id": 47800, + "cbid": 200, "correlation": 47800 + } + }, + { + "ph": "f", "id": 47800, "pid": 435, "tid": 435, "ts": 1776649791311750, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311751, "dur": 0, + "args": { + "External id": 47801, + "cbid": 273, "correlation": 47801 + } + }, + { + "ph": "f", "id": 47801, "pid": 435, "tid": 435, "ts": 1776649791311751, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791311760, "dur": 15, + "args": { + "External id": 47802, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47802, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47802, "pid": 0, "tid": 7, "ts": 1776649791311760, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311752, "dur": 4, + "args": { + "External id": 47802, + "cbid": 211, "correlation": 47802 + } + }, + { + "ph": "s", "id": 47802, "pid": 435, "tid": 435, "ts": 1776649791311752, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791311796, "dur": 1, + "args": { + "External id": 6124, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47820, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47820, "pid": 0, "tid": 7, "ts": 1776649791311796, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311786, "dur": 6, + "args": { + "External id": 6124, + "cbid": 211, "correlation": 47820 + } + }, + { + "ph": "s", "id": 47820, "pid": 435, "tid": 435, "ts": 1776649791311786, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311798, "dur": 0, + "args": { + "External id": 47826, + "cbid": 200, "correlation": 47826 + } + }, + { + "ph": "f", "id": 47826, "pid": 435, "tid": 435, "ts": 1776649791311798, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311799, "dur": 0, + "args": { + "External id": 47827, + "cbid": 273, "correlation": 47827 + } + }, + { + "ph": "f", "id": 47827, "pid": 435, "tid": 435, "ts": 1776649791311799, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791311808, "dur": 15, + "args": { + "External id": 47828, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47828, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47828, "pid": 0, "tid": 7, "ts": 1776649791311808, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311800, "dur": 4, + "args": { + "External id": 47828, + "cbid": 211, "correlation": 47828 + } + }, + { + "ph": "s", "id": 47828, "pid": 435, "tid": 435, "ts": 1776649791311800, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791311844, "dur": 1, + "args": { + "External id": 6129, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47846, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47846, "pid": 0, "tid": 7, "ts": 1776649791311844, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311835, "dur": 6, + "args": { + "External id": 6129, + "cbid": 211, "correlation": 47846 + } + }, + { + "ph": "s", "id": 47846, "pid": 435, "tid": 435, "ts": 1776649791311835, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311847, "dur": 0, + "args": { + "External id": 47852, + "cbid": 200, "correlation": 47852 + } + }, + { + "ph": "f", "id": 47852, "pid": 435, "tid": 435, "ts": 1776649791311847, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311847, "dur": 0, + "args": { + "External id": 47853, + "cbid": 273, "correlation": 47853 + } + }, + { + "ph": "f", "id": 47853, "pid": 435, "tid": 435, "ts": 1776649791311847, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791311856, "dur": 16, + "args": { + "External id": 47854, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47854, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47854, "pid": 0, "tid": 7, "ts": 1776649791311856, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311848, "dur": 4, + "args": { + "External id": 47854, + "cbid": 211, "correlation": 47854 + } + }, + { + "ph": "s", "id": 47854, "pid": 435, "tid": 435, "ts": 1776649791311848, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791311891, "dur": 1, + "args": { + "External id": 6134, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47872, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47872, "pid": 0, "tid": 7, "ts": 1776649791311891, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311882, "dur": 6, + "args": { + "External id": 6134, + "cbid": 211, "correlation": 47872 + } + }, + { + "ph": "s", "id": 47872, "pid": 435, "tid": 435, "ts": 1776649791311882, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311894, "dur": 0, + "args": { + "External id": 47878, + "cbid": 200, "correlation": 47878 + } + }, + { + "ph": "f", "id": 47878, "pid": 435, "tid": 435, "ts": 1776649791311894, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311895, "dur": 0, + "args": { + "External id": 47879, + "cbid": 273, "correlation": 47879 + } + }, + { + "ph": "f", "id": 47879, "pid": 435, "tid": 435, "ts": 1776649791311895, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791311904, "dur": 15, + "args": { + "External id": 47880, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47880, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47880, "pid": 0, "tid": 7, "ts": 1776649791311904, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311896, "dur": 4, + "args": { + "External id": 47880, + "cbid": 211, "correlation": 47880 + } + }, + { + "ph": "s", "id": 47880, "pid": 435, "tid": 435, "ts": 1776649791311896, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791311939, "dur": 1, + "args": { + "External id": 6139, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47898, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47898, "pid": 0, "tid": 7, "ts": 1776649791311939, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311930, "dur": 6, + "args": { + "External id": 6139, + "cbid": 211, "correlation": 47898 + } + }, + { + "ph": "s", "id": 47898, "pid": 435, "tid": 435, "ts": 1776649791311930, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311943, "dur": 0, + "args": { + "External id": 47904, + "cbid": 200, "correlation": 47904 + } + }, + { + "ph": "f", "id": 47904, "pid": 435, "tid": 435, "ts": 1776649791311943, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311943, "dur": 0, + "args": { + "External id": 47905, + "cbid": 273, "correlation": 47905 + } + }, + { + "ph": "f", "id": 47905, "pid": 435, "tid": 435, "ts": 1776649791311943, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791311952, "dur": 15, + "args": { + "External id": 47906, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47906, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47906, "pid": 0, "tid": 7, "ts": 1776649791311952, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311944, "dur": 4, + "args": { + "External id": 47906, + "cbid": 211, "correlation": 47906 + } + }, + { + "ph": "s", "id": 47906, "pid": 435, "tid": 435, "ts": 1776649791311944, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791311988, "dur": 1, + "args": { + "External id": 6144, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47924, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47924, "pid": 0, "tid": 7, "ts": 1776649791311988, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311978, "dur": 6, + "args": { + "External id": 6144, + "cbid": 211, "correlation": 47924 + } + }, + { + "ph": "s", "id": 47924, "pid": 435, "tid": 435, "ts": 1776649791311978, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311991, "dur": 0, + "args": { + "External id": 47930, + "cbid": 200, "correlation": 47930 + } + }, + { + "ph": "f", "id": 47930, "pid": 435, "tid": 435, "ts": 1776649791311991, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791311992, "dur": 0, + "args": { + "External id": 47931, + "cbid": 273, "correlation": 47931 + } + }, + { + "ph": "f", "id": 47931, "pid": 435, "tid": 435, "ts": 1776649791311992, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791312001, "dur": 15, + "args": { + "External id": 47932, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47932, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47932, "pid": 0, "tid": 7, "ts": 1776649791312001, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791311993, "dur": 4, + "args": { + "External id": 47932, + "cbid": 211, "correlation": 47932 + } + }, + { + "ph": "s", "id": 47932, "pid": 435, "tid": 435, "ts": 1776649791311993, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791312046, "dur": 1, + "args": { + "External id": 6149, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47950, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47950, "pid": 0, "tid": 7, "ts": 1776649791312046, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312037, "dur": 6, + "args": { + "External id": 6149, + "cbid": 211, "correlation": 47950 + } + }, + { + "ph": "s", "id": 47950, "pid": 435, "tid": 435, "ts": 1776649791312037, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312049, "dur": 0, + "args": { + "External id": 47956, + "cbid": 200, "correlation": 47956 + } + }, + { + "ph": "f", "id": 47956, "pid": 435, "tid": 435, "ts": 1776649791312049, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312050, "dur": 0, + "args": { + "External id": 47957, + "cbid": 273, "correlation": 47957 + } + }, + { + "ph": "f", "id": 47957, "pid": 435, "tid": 435, "ts": 1776649791312050, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791312059, "dur": 16, + "args": { + "External id": 47958, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47958, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47958, "pid": 0, "tid": 7, "ts": 1776649791312059, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312051, "dur": 4, + "args": { + "External id": 47958, + "cbid": 211, "correlation": 47958 + } + }, + { + "ph": "s", "id": 47958, "pid": 435, "tid": 435, "ts": 1776649791312051, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791312095, "dur": 1, + "args": { + "External id": 6154, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47976, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47976, "pid": 0, "tid": 7, "ts": 1776649791312095, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312085, "dur": 6, + "args": { + "External id": 6154, + "cbid": 211, "correlation": 47976 + } + }, + { + "ph": "s", "id": 47976, "pid": 435, "tid": 435, "ts": 1776649791312085, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312098, "dur": 0, + "args": { + "External id": 47982, + "cbid": 200, "correlation": 47982 + } + }, + { + "ph": "f", "id": 47982, "pid": 435, "tid": 435, "ts": 1776649791312098, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312098, "dur": 0, + "args": { + "External id": 47983, + "cbid": 273, "correlation": 47983 + } + }, + { + "ph": "f", "id": 47983, "pid": 435, "tid": 435, "ts": 1776649791312098, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791312107, "dur": 15, + "args": { + "External id": 47984, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 47984, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 47984, "pid": 0, "tid": 7, "ts": 1776649791312107, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312099, "dur": 4, + "args": { + "External id": 47984, + "cbid": 211, "correlation": 47984 + } + }, + { + "ph": "s", "id": 47984, "pid": 435, "tid": 435, "ts": 1776649791312099, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791312143, "dur": 1, + "args": { + "External id": 6159, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48002, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48002, "pid": 0, "tid": 7, "ts": 1776649791312143, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312134, "dur": 6, + "args": { + "External id": 6159, + "cbid": 211, "correlation": 48002 + } + }, + { + "ph": "s", "id": 48002, "pid": 435, "tid": 435, "ts": 1776649791312134, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312145, "dur": 0, + "args": { + "External id": 48008, + "cbid": 200, "correlation": 48008 + } + }, + { + "ph": "f", "id": 48008, "pid": 435, "tid": 435, "ts": 1776649791312145, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312146, "dur": 0, + "args": { + "External id": 48009, + "cbid": 273, "correlation": 48009 + } + }, + { + "ph": "f", "id": 48009, "pid": 435, "tid": 435, "ts": 1776649791312146, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791312155, "dur": 15, + "args": { + "External id": 48010, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48010, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48010, "pid": 0, "tid": 7, "ts": 1776649791312155, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312147, "dur": 4, + "args": { + "External id": 48010, + "cbid": 211, "correlation": 48010 + } + }, + { + "ph": "s", "id": 48010, "pid": 435, "tid": 435, "ts": 1776649791312147, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791312190, "dur": 1, + "args": { + "External id": 6164, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48028, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48028, "pid": 0, "tid": 7, "ts": 1776649791312190, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312181, "dur": 6, + "args": { + "External id": 6164, + "cbid": 211, "correlation": 48028 + } + }, + { + "ph": "s", "id": 48028, "pid": 435, "tid": 435, "ts": 1776649791312181, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312194, "dur": 0, + "args": { + "External id": 48034, + "cbid": 200, "correlation": 48034 + } + }, + { + "ph": "f", "id": 48034, "pid": 435, "tid": 435, "ts": 1776649791312194, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312194, "dur": 0, + "args": { + "External id": 48035, + "cbid": 273, "correlation": 48035 + } + }, + { + "ph": "f", "id": 48035, "pid": 435, "tid": 435, "ts": 1776649791312194, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791312204, "dur": 15, + "args": { + "External id": 48036, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48036, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48036, "pid": 0, "tid": 7, "ts": 1776649791312204, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312196, "dur": 4, + "args": { + "External id": 48036, + "cbid": 211, "correlation": 48036 + } + }, + { + "ph": "s", "id": 48036, "pid": 435, "tid": 435, "ts": 1776649791312196, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791312260, "dur": 1, + "args": { + "External id": 6169, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48054, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48054, "pid": 0, "tid": 7, "ts": 1776649791312260, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312250, "dur": 8, + "args": { + "External id": 6169, + "cbid": 211, "correlation": 48054 + } + }, + { + "ph": "s", "id": 48054, "pid": 435, "tid": 435, "ts": 1776649791312250, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312264, "dur": 0, + "args": { + "External id": 48060, + "cbid": 200, "correlation": 48060 + } + }, + { + "ph": "f", "id": 48060, "pid": 435, "tid": 435, "ts": 1776649791312264, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312265, "dur": 0, + "args": { + "External id": 48061, + "cbid": 273, "correlation": 48061 + } + }, + { + "ph": "f", "id": 48061, "pid": 435, "tid": 435, "ts": 1776649791312265, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791312274, "dur": 16, + "args": { + "External id": 48062, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48062, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48062, "pid": 0, "tid": 7, "ts": 1776649791312274, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312266, "dur": 4, + "args": { + "External id": 48062, + "cbid": 211, "correlation": 48062 + } + }, + { + "ph": "s", "id": 48062, "pid": 435, "tid": 435, "ts": 1776649791312266, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791312312, "dur": 1, + "args": { + "External id": 6174, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48080, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48080, "pid": 0, "tid": 7, "ts": 1776649791312312, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312302, "dur": 6, + "args": { + "External id": 6174, + "cbid": 211, "correlation": 48080 + } + }, + { + "ph": "s", "id": 48080, "pid": 435, "tid": 435, "ts": 1776649791312302, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312315, "dur": 0, + "args": { + "External id": 48086, + "cbid": 200, "correlation": 48086 + } + }, + { + "ph": "f", "id": 48086, "pid": 435, "tid": 435, "ts": 1776649791312315, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312316, "dur": 0, + "args": { + "External id": 48087, + "cbid": 273, "correlation": 48087 + } + }, + { + "ph": "f", "id": 48087, "pid": 435, "tid": 435, "ts": 1776649791312316, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791312325, "dur": 16, + "args": { + "External id": 48088, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48088, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48088, "pid": 0, "tid": 7, "ts": 1776649791312325, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312317, "dur": 5, + "args": { + "External id": 48088, + "cbid": 211, "correlation": 48088 + } + }, + { + "ph": "s", "id": 48088, "pid": 435, "tid": 435, "ts": 1776649791312317, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791312361, "dur": 1, + "args": { + "External id": 6179, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48106, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48106, "pid": 0, "tid": 7, "ts": 1776649791312361, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312351, "dur": 7, + "args": { + "External id": 6179, + "cbid": 211, "correlation": 48106 + } + }, + { + "ph": "s", "id": 48106, "pid": 435, "tid": 435, "ts": 1776649791312351, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312364, "dur": 0, + "args": { + "External id": 48112, + "cbid": 200, "correlation": 48112 + } + }, + { + "ph": "f", "id": 48112, "pid": 435, "tid": 435, "ts": 1776649791312364, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312365, "dur": 0, + "args": { + "External id": 48113, + "cbid": 273, "correlation": 48113 + } + }, + { + "ph": "f", "id": 48113, "pid": 435, "tid": 435, "ts": 1776649791312365, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791312374, "dur": 15, + "args": { + "External id": 48114, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48114, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48114, "pid": 0, "tid": 7, "ts": 1776649791312374, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312366, "dur": 5, + "args": { + "External id": 48114, + "cbid": 211, "correlation": 48114 + } + }, + { + "ph": "s", "id": 48114, "pid": 435, "tid": 435, "ts": 1776649791312366, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791312409, "dur": 1, + "args": { + "External id": 6184, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48132, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48132, "pid": 0, "tid": 7, "ts": 1776649791312409, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312400, "dur": 6, + "args": { + "External id": 6184, + "cbid": 211, "correlation": 48132 + } + }, + { + "ph": "s", "id": 48132, "pid": 435, "tid": 435, "ts": 1776649791312400, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312412, "dur": 0, + "args": { + "External id": 48138, + "cbid": 200, "correlation": 48138 + } + }, + { + "ph": "f", "id": 48138, "pid": 435, "tid": 435, "ts": 1776649791312412, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312413, "dur": 0, + "args": { + "External id": 48139, + "cbid": 273, "correlation": 48139 + } + }, + { + "ph": "f", "id": 48139, "pid": 435, "tid": 435, "ts": 1776649791312413, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791312422, "dur": 15, + "args": { + "External id": 48140, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48140, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48140, "pid": 0, "tid": 7, "ts": 1776649791312422, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312414, "dur": 4, + "args": { + "External id": 48140, + "cbid": 211, "correlation": 48140 + } + }, + { + "ph": "s", "id": 48140, "pid": 435, "tid": 435, "ts": 1776649791312414, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791312457, "dur": 1, + "args": { + "External id": 6189, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48158, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48158, "pid": 0, "tid": 7, "ts": 1776649791312457, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312448, "dur": 6, + "args": { + "External id": 6189, + "cbid": 211, "correlation": 48158 + } + }, + { + "ph": "s", "id": 48158, "pid": 435, "tid": 435, "ts": 1776649791312448, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312460, "dur": 0, + "args": { + "External id": 48164, + "cbid": 200, "correlation": 48164 + } + }, + { + "ph": "f", "id": 48164, "pid": 435, "tid": 435, "ts": 1776649791312460, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312461, "dur": 0, + "args": { + "External id": 48165, + "cbid": 273, "correlation": 48165 + } + }, + { + "ph": "f", "id": 48165, "pid": 435, "tid": 435, "ts": 1776649791312461, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791312470, "dur": 15, + "args": { + "External id": 48166, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48166, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48166, "pid": 0, "tid": 7, "ts": 1776649791312470, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312462, "dur": 4, + "args": { + "External id": 48166, + "cbid": 211, "correlation": 48166 + } + }, + { + "ph": "s", "id": 48166, "pid": 435, "tid": 435, "ts": 1776649791312462, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791312506, "dur": 1, + "args": { + "External id": 6194, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48184, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48184, "pid": 0, "tid": 7, "ts": 1776649791312506, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312496, "dur": 6, + "args": { + "External id": 6194, + "cbid": 211, "correlation": 48184 + } + }, + { + "ph": "s", "id": 48184, "pid": 435, "tid": 435, "ts": 1776649791312496, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312509, "dur": 0, + "args": { + "External id": 48190, + "cbid": 200, "correlation": 48190 + } + }, + { + "ph": "f", "id": 48190, "pid": 435, "tid": 435, "ts": 1776649791312509, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312510, "dur": 0, + "args": { + "External id": 48191, + "cbid": 273, "correlation": 48191 + } + }, + { + "ph": "f", "id": 48191, "pid": 435, "tid": 435, "ts": 1776649791312510, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791312519, "dur": 15, + "args": { + "External id": 48192, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48192, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48192, "pid": 0, "tid": 7, "ts": 1776649791312519, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312511, "dur": 5, + "args": { + "External id": 48192, + "cbid": 211, "correlation": 48192 + } + }, + { + "ph": "s", "id": 48192, "pid": 435, "tid": 435, "ts": 1776649791312511, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791312554, "dur": 1, + "args": { + "External id": 6199, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48210, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48210, "pid": 0, "tid": 7, "ts": 1776649791312554, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312545, "dur": 6, + "args": { + "External id": 6199, + "cbid": 211, "correlation": 48210 + } + }, + { + "ph": "s", "id": 48210, "pid": 435, "tid": 435, "ts": 1776649791312545, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312557, "dur": 0, + "args": { + "External id": 48216, + "cbid": 200, "correlation": 48216 + } + }, + { + "ph": "f", "id": 48216, "pid": 435, "tid": 435, "ts": 1776649791312557, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312558, "dur": 0, + "args": { + "External id": 48217, + "cbid": 273, "correlation": 48217 + } + }, + { + "ph": "f", "id": 48217, "pid": 435, "tid": 435, "ts": 1776649791312558, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791312567, "dur": 15, + "args": { + "External id": 48218, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48218, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48218, "pid": 0, "tid": 7, "ts": 1776649791312567, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312559, "dur": 4, + "args": { + "External id": 48218, + "cbid": 211, "correlation": 48218 + } + }, + { + "ph": "s", "id": 48218, "pid": 435, "tid": 435, "ts": 1776649791312559, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791312603, "dur": 1, + "args": { + "External id": 6204, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48236, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48236, "pid": 0, "tid": 7, "ts": 1776649791312603, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312594, "dur": 6, + "args": { + "External id": 6204, + "cbid": 211, "correlation": 48236 + } + }, + { + "ph": "s", "id": 48236, "pid": 435, "tid": 435, "ts": 1776649791312594, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312606, "dur": 0, + "args": { + "External id": 48242, + "cbid": 200, "correlation": 48242 + } + }, + { + "ph": "f", "id": 48242, "pid": 435, "tid": 435, "ts": 1776649791312606, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312606, "dur": 0, + "args": { + "External id": 48243, + "cbid": 273, "correlation": 48243 + } + }, + { + "ph": "f", "id": 48243, "pid": 435, "tid": 435, "ts": 1776649791312606, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791312616, "dur": 15, + "args": { + "External id": 48244, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48244, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48244, "pid": 0, "tid": 7, "ts": 1776649791312616, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312608, "dur": 4, + "args": { + "External id": 48244, + "cbid": 211, "correlation": 48244 + } + }, + { + "ph": "s", "id": 48244, "pid": 435, "tid": 435, "ts": 1776649791312608, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791312651, "dur": 1, + "args": { + "External id": 6209, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48262, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48262, "pid": 0, "tid": 7, "ts": 1776649791312651, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312642, "dur": 6, + "args": { + "External id": 6209, + "cbid": 211, "correlation": 48262 + } + }, + { + "ph": "s", "id": 48262, "pid": 435, "tid": 435, "ts": 1776649791312642, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312654, "dur": 0, + "args": { + "External id": 48268, + "cbid": 200, "correlation": 48268 + } + }, + { + "ph": "f", "id": 48268, "pid": 435, "tid": 435, "ts": 1776649791312654, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312655, "dur": 0, + "args": { + "External id": 48269, + "cbid": 273, "correlation": 48269 + } + }, + { + "ph": "f", "id": 48269, "pid": 435, "tid": 435, "ts": 1776649791312655, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791312664, "dur": 15, + "args": { + "External id": 48270, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48270, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48270, "pid": 0, "tid": 7, "ts": 1776649791312664, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312656, "dur": 4, + "args": { + "External id": 48270, + "cbid": 211, "correlation": 48270 + } + }, + { + "ph": "s", "id": 48270, "pid": 435, "tid": 435, "ts": 1776649791312656, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791312699, "dur": 1, + "args": { + "External id": 6214, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48288, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48288, "pid": 0, "tid": 7, "ts": 1776649791312699, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312690, "dur": 6, + "args": { + "External id": 6214, + "cbid": 211, "correlation": 48288 + } + }, + { + "ph": "s", "id": 48288, "pid": 435, "tid": 435, "ts": 1776649791312690, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312702, "dur": 0, + "args": { + "External id": 48294, + "cbid": 200, "correlation": 48294 + } + }, + { + "ph": "f", "id": 48294, "pid": 435, "tid": 435, "ts": 1776649791312702, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312703, "dur": 0, + "args": { + "External id": 48295, + "cbid": 273, "correlation": 48295 + } + }, + { + "ph": "f", "id": 48295, "pid": 435, "tid": 435, "ts": 1776649791312703, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791312712, "dur": 17, + "args": { + "External id": 48296, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48296, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48296, "pid": 0, "tid": 7, "ts": 1776649791312712, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312704, "dur": 4, + "args": { + "External id": 48296, + "cbid": 211, "correlation": 48296 + } + }, + { + "ph": "s", "id": 48296, "pid": 435, "tid": 435, "ts": 1776649791312704, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791312747, "dur": 1, + "args": { + "External id": 6219, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48314, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48314, "pid": 0, "tid": 7, "ts": 1776649791312747, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312738, "dur": 6, + "args": { + "External id": 6219, + "cbid": 211, "correlation": 48314 + } + }, + { + "ph": "s", "id": 48314, "pid": 435, "tid": 435, "ts": 1776649791312738, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312750, "dur": 0, + "args": { + "External id": 48320, + "cbid": 200, "correlation": 48320 + } + }, + { + "ph": "f", "id": 48320, "pid": 435, "tid": 435, "ts": 1776649791312750, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312751, "dur": 0, + "args": { + "External id": 48321, + "cbid": 273, "correlation": 48321 + } + }, + { + "ph": "f", "id": 48321, "pid": 435, "tid": 435, "ts": 1776649791312751, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791312760, "dur": 16, + "args": { + "External id": 48322, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48322, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48322, "pid": 0, "tid": 7, "ts": 1776649791312760, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312752, "dur": 5, + "args": { + "External id": 48322, + "cbid": 211, "correlation": 48322 + } + }, + { + "ph": "s", "id": 48322, "pid": 435, "tid": 435, "ts": 1776649791312752, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791312795, "dur": 1, + "args": { + "External id": 6224, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48340, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48340, "pid": 0, "tid": 7, "ts": 1776649791312795, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312786, "dur": 6, + "args": { + "External id": 6224, + "cbid": 211, "correlation": 48340 + } + }, + { + "ph": "s", "id": 48340, "pid": 435, "tid": 435, "ts": 1776649791312786, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312798, "dur": 0, + "args": { + "External id": 48346, + "cbid": 200, "correlation": 48346 + } + }, + { + "ph": "f", "id": 48346, "pid": 435, "tid": 435, "ts": 1776649791312798, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312799, "dur": 0, + "args": { + "External id": 48347, + "cbid": 273, "correlation": 48347 + } + }, + { + "ph": "f", "id": 48347, "pid": 435, "tid": 435, "ts": 1776649791312799, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791312808, "dur": 15, + "args": { + "External id": 48348, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48348, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48348, "pid": 0, "tid": 7, "ts": 1776649791312808, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312800, "dur": 5, + "args": { + "External id": 48348, + "cbid": 211, "correlation": 48348 + } + }, + { + "ph": "s", "id": 48348, "pid": 435, "tid": 435, "ts": 1776649791312800, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791312843, "dur": 1, + "args": { + "External id": 6229, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48366, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48366, "pid": 0, "tid": 7, "ts": 1776649791312843, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312834, "dur": 6, + "args": { + "External id": 6229, + "cbid": 211, "correlation": 48366 + } + }, + { + "ph": "s", "id": 48366, "pid": 435, "tid": 435, "ts": 1776649791312834, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312847, "dur": 0, + "args": { + "External id": 48372, + "cbid": 200, "correlation": 48372 + } + }, + { + "ph": "f", "id": 48372, "pid": 435, "tid": 435, "ts": 1776649791312847, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312847, "dur": 0, + "args": { + "External id": 48373, + "cbid": 273, "correlation": 48373 + } + }, + { + "ph": "f", "id": 48373, "pid": 435, "tid": 435, "ts": 1776649791312847, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791312856, "dur": 15, + "args": { + "External id": 48374, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48374, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48374, "pid": 0, "tid": 7, "ts": 1776649791312856, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312848, "dur": 5, + "args": { + "External id": 48374, + "cbid": 211, "correlation": 48374 + } + }, + { + "ph": "s", "id": 48374, "pid": 435, "tid": 435, "ts": 1776649791312848, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791312879, "dur": 3, + "args": { + "External id": 48382, + "cbid": 135, "correlation": 48382 + } + }, + { + "ph": "f", "id": 48382, "pid": 435, "tid": 435, "ts": 1776649791312879, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791312894, "dur": 5, + "args": { + "External id": 48387, + "cbid": 165, "correlation": 48387 + } + }, + { + "ph": "s", "id": 48387, "pid": 435, "tid": 435, "ts": 1776649791312894, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649791312903, "dur": 4, + "args": { + "External id": 48390, + "cbid": 139, "correlation": 48390 + } + }, + { + "ph": "f", "id": 48390, "pid": 435, "tid": 435, "ts": 1776649791312903, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791312960, "dur": 1, + "args": { + "External id": 6234, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48418, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48418, "pid": 0, "tid": 7, "ts": 1776649791312960, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312949, "dur": 8, + "args": { + "External id": 6234, + "cbid": 211, "correlation": 48418 + } + }, + { + "ph": "s", "id": 48418, "pid": 435, "tid": 435, "ts": 1776649791312949, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312964, "dur": 0, + "args": { + "External id": 48424, + "cbid": 200, "correlation": 48424 + } + }, + { + "ph": "f", "id": 48424, "pid": 435, "tid": 435, "ts": 1776649791312964, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791312965, "dur": 0, + "args": { + "External id": 48425, + "cbid": 273, "correlation": 48425 + } + }, + { + "ph": "f", "id": 48425, "pid": 435, "tid": 435, "ts": 1776649791312965, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791312975, "dur": 15, + "args": { + "External id": 48426, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48426, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48426, "pid": 0, "tid": 7, "ts": 1776649791312975, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791312967, "dur": 5, + "args": { + "External id": 48426, + "cbid": 211, "correlation": 48426 + } + }, + { + "ph": "s", "id": 48426, "pid": 435, "tid": 435, "ts": 1776649791312967, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791313012, "dur": 1, + "args": { + "External id": 6239, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48444, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48444, "pid": 0, "tid": 7, "ts": 1776649791313012, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313003, "dur": 6, + "args": { + "External id": 6239, + "cbid": 211, "correlation": 48444 + } + }, + { + "ph": "s", "id": 48444, "pid": 435, "tid": 435, "ts": 1776649791313003, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313016, "dur": 0, + "args": { + "External id": 48450, + "cbid": 200, "correlation": 48450 + } + }, + { + "ph": "f", "id": 48450, "pid": 435, "tid": 435, "ts": 1776649791313016, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313016, "dur": 0, + "args": { + "External id": 48451, + "cbid": 273, "correlation": 48451 + } + }, + { + "ph": "f", "id": 48451, "pid": 435, "tid": 435, "ts": 1776649791313016, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791313025, "dur": 15, + "args": { + "External id": 48452, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48452, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48452, "pid": 0, "tid": 7, "ts": 1776649791313025, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313018, "dur": 4, + "args": { + "External id": 48452, + "cbid": 211, "correlation": 48452 + } + }, + { + "ph": "s", "id": 48452, "pid": 435, "tid": 435, "ts": 1776649791313018, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791313062, "dur": 1, + "args": { + "External id": 6244, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48470, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48470, "pid": 0, "tid": 7, "ts": 1776649791313062, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313052, "dur": 6, + "args": { + "External id": 6244, + "cbid": 211, "correlation": 48470 + } + }, + { + "ph": "s", "id": 48470, "pid": 435, "tid": 435, "ts": 1776649791313052, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313065, "dur": 0, + "args": { + "External id": 48476, + "cbid": 200, "correlation": 48476 + } + }, + { + "ph": "f", "id": 48476, "pid": 435, "tid": 435, "ts": 1776649791313065, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313065, "dur": 0, + "args": { + "External id": 48477, + "cbid": 273, "correlation": 48477 + } + }, + { + "ph": "f", "id": 48477, "pid": 435, "tid": 435, "ts": 1776649791313065, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791313074, "dur": 15, + "args": { + "External id": 48478, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48478, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48478, "pid": 0, "tid": 7, "ts": 1776649791313074, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313066, "dur": 4, + "args": { + "External id": 48478, + "cbid": 211, "correlation": 48478 + } + }, + { + "ph": "s", "id": 48478, "pid": 435, "tid": 435, "ts": 1776649791313066, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791313110, "dur": 1, + "args": { + "External id": 6249, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48496, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48496, "pid": 0, "tid": 7, "ts": 1776649791313110, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313101, "dur": 6, + "args": { + "External id": 6249, + "cbid": 211, "correlation": 48496 + } + }, + { + "ph": "s", "id": 48496, "pid": 435, "tid": 435, "ts": 1776649791313101, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313113, "dur": 0, + "args": { + "External id": 48502, + "cbid": 200, "correlation": 48502 + } + }, + { + "ph": "f", "id": 48502, "pid": 435, "tid": 435, "ts": 1776649791313113, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313114, "dur": 0, + "args": { + "External id": 48503, + "cbid": 273, "correlation": 48503 + } + }, + { + "ph": "f", "id": 48503, "pid": 435, "tid": 435, "ts": 1776649791313114, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791313123, "dur": 15, + "args": { + "External id": 48504, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48504, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48504, "pid": 0, "tid": 7, "ts": 1776649791313123, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313115, "dur": 5, + "args": { + "External id": 48504, + "cbid": 211, "correlation": 48504 + } + }, + { + "ph": "s", "id": 48504, "pid": 435, "tid": 435, "ts": 1776649791313115, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791313159, "dur": 1, + "args": { + "External id": 6254, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48522, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48522, "pid": 0, "tid": 7, "ts": 1776649791313159, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313149, "dur": 6, + "args": { + "External id": 6254, + "cbid": 211, "correlation": 48522 + } + }, + { + "ph": "s", "id": 48522, "pid": 435, "tid": 435, "ts": 1776649791313149, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313162, "dur": 0, + "args": { + "External id": 48528, + "cbid": 200, "correlation": 48528 + } + }, + { + "ph": "f", "id": 48528, "pid": 435, "tid": 435, "ts": 1776649791313162, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313162, "dur": 0, + "args": { + "External id": 48529, + "cbid": 273, "correlation": 48529 + } + }, + { + "ph": "f", "id": 48529, "pid": 435, "tid": 435, "ts": 1776649791313162, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791313171, "dur": 15, + "args": { + "External id": 48530, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48530, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48530, "pid": 0, "tid": 7, "ts": 1776649791313171, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313163, "dur": 4, + "args": { + "External id": 48530, + "cbid": 211, "correlation": 48530 + } + }, + { + "ph": "s", "id": 48530, "pid": 435, "tid": 435, "ts": 1776649791313163, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791313206, "dur": 1, + "args": { + "External id": 6259, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48548, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48548, "pid": 0, "tid": 7, "ts": 1776649791313206, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313197, "dur": 6, + "args": { + "External id": 6259, + "cbid": 211, "correlation": 48548 + } + }, + { + "ph": "s", "id": 48548, "pid": 435, "tid": 435, "ts": 1776649791313197, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313209, "dur": 0, + "args": { + "External id": 48554, + "cbid": 200, "correlation": 48554 + } + }, + { + "ph": "f", "id": 48554, "pid": 435, "tid": 435, "ts": 1776649791313209, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313210, "dur": 0, + "args": { + "External id": 48555, + "cbid": 273, "correlation": 48555 + } + }, + { + "ph": "f", "id": 48555, "pid": 435, "tid": 435, "ts": 1776649791313210, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791313219, "dur": 16, + "args": { + "External id": 48556, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48556, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48556, "pid": 0, "tid": 7, "ts": 1776649791313219, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313211, "dur": 4, + "args": { + "External id": 48556, + "cbid": 211, "correlation": 48556 + } + }, + { + "ph": "s", "id": 48556, "pid": 435, "tid": 435, "ts": 1776649791313211, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791313255, "dur": 1, + "args": { + "External id": 6264, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48574, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48574, "pid": 0, "tid": 7, "ts": 1776649791313255, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313246, "dur": 6, + "args": { + "External id": 6264, + "cbid": 211, "correlation": 48574 + } + }, + { + "ph": "s", "id": 48574, "pid": 435, "tid": 435, "ts": 1776649791313246, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313258, "dur": 0, + "args": { + "External id": 48580, + "cbid": 200, "correlation": 48580 + } + }, + { + "ph": "f", "id": 48580, "pid": 435, "tid": 435, "ts": 1776649791313258, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313259, "dur": 0, + "args": { + "External id": 48581, + "cbid": 273, "correlation": 48581 + } + }, + { + "ph": "f", "id": 48581, "pid": 435, "tid": 435, "ts": 1776649791313259, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791313268, "dur": 15, + "args": { + "External id": 48582, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48582, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48582, "pid": 0, "tid": 7, "ts": 1776649791313268, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313260, "dur": 5, + "args": { + "External id": 48582, + "cbid": 211, "correlation": 48582 + } + }, + { + "ph": "s", "id": 48582, "pid": 435, "tid": 435, "ts": 1776649791313260, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791313304, "dur": 1, + "args": { + "External id": 6269, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48600, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48600, "pid": 0, "tid": 7, "ts": 1776649791313304, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313294, "dur": 6, + "args": { + "External id": 6269, + "cbid": 211, "correlation": 48600 + } + }, + { + "ph": "s", "id": 48600, "pid": 435, "tid": 435, "ts": 1776649791313294, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313307, "dur": 0, + "args": { + "External id": 48606, + "cbid": 200, "correlation": 48606 + } + }, + { + "ph": "f", "id": 48606, "pid": 435, "tid": 435, "ts": 1776649791313307, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313308, "dur": 0, + "args": { + "External id": 48607, + "cbid": 273, "correlation": 48607 + } + }, + { + "ph": "f", "id": 48607, "pid": 435, "tid": 435, "ts": 1776649791313308, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791313317, "dur": 15, + "args": { + "External id": 48608, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48608, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48608, "pid": 0, "tid": 7, "ts": 1776649791313317, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313309, "dur": 5, + "args": { + "External id": 48608, + "cbid": 211, "correlation": 48608 + } + }, + { + "ph": "s", "id": 48608, "pid": 435, "tid": 435, "ts": 1776649791313309, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791313353, "dur": 1, + "args": { + "External id": 6274, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48626, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48626, "pid": 0, "tid": 7, "ts": 1776649791313353, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313343, "dur": 7, + "args": { + "External id": 6274, + "cbid": 211, "correlation": 48626 + } + }, + { + "ph": "s", "id": 48626, "pid": 435, "tid": 435, "ts": 1776649791313343, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313356, "dur": 0, + "args": { + "External id": 48632, + "cbid": 200, "correlation": 48632 + } + }, + { + "ph": "f", "id": 48632, "pid": 435, "tid": 435, "ts": 1776649791313356, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313356, "dur": 0, + "args": { + "External id": 48633, + "cbid": 273, "correlation": 48633 + } + }, + { + "ph": "f", "id": 48633, "pid": 435, "tid": 435, "ts": 1776649791313356, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791313365, "dur": 15, + "args": { + "External id": 48634, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48634, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48634, "pid": 0, "tid": 7, "ts": 1776649791313365, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313358, "dur": 4, + "args": { + "External id": 48634, + "cbid": 211, "correlation": 48634 + } + }, + { + "ph": "s", "id": 48634, "pid": 435, "tid": 435, "ts": 1776649791313358, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791313400, "dur": 1, + "args": { + "External id": 6279, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48652, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48652, "pid": 0, "tid": 7, "ts": 1776649791313400, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313391, "dur": 6, + "args": { + "External id": 6279, + "cbid": 211, "correlation": 48652 + } + }, + { + "ph": "s", "id": 48652, "pid": 435, "tid": 435, "ts": 1776649791313391, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313404, "dur": 0, + "args": { + "External id": 48658, + "cbid": 200, "correlation": 48658 + } + }, + { + "ph": "f", "id": 48658, "pid": 435, "tid": 435, "ts": 1776649791313404, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313404, "dur": 0, + "args": { + "External id": 48659, + "cbid": 273, "correlation": 48659 + } + }, + { + "ph": "f", "id": 48659, "pid": 435, "tid": 435, "ts": 1776649791313404, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791313413, "dur": 15, + "args": { + "External id": 48660, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48660, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48660, "pid": 0, "tid": 7, "ts": 1776649791313413, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313405, "dur": 5, + "args": { + "External id": 48660, + "cbid": 211, "correlation": 48660 + } + }, + { + "ph": "s", "id": 48660, "pid": 435, "tid": 435, "ts": 1776649791313405, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791313428, "dur": 6, + "args": { + "External id": 48663, + "cbid": 165, "correlation": 48663 + } + }, + { + "ph": "s", "id": 48663, "pid": 435, "tid": 435, "ts": 1776649791313428, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791313454, "dur": 3, + "args": { + "External id": 48672, + "cbid": 135, "correlation": 48672 + } + }, + { + "ph": "f", "id": 48672, "pid": 435, "tid": 435, "ts": 1776649791313454, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791313503, "dur": 1, + "args": { + "External id": 6284, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48692, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48692, "pid": 0, "tid": 7, "ts": 1776649791313503, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313492, "dur": 8, + "args": { + "External id": 6284, + "cbid": 211, "correlation": 48692 + } + }, + { + "ph": "s", "id": 48692, "pid": 435, "tid": 435, "ts": 1776649791313492, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313507, "dur": 0, + "args": { + "External id": 48698, + "cbid": 200, "correlation": 48698 + } + }, + { + "ph": "f", "id": 48698, "pid": 435, "tid": 435, "ts": 1776649791313507, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313508, "dur": 0, + "args": { + "External id": 48699, + "cbid": 273, "correlation": 48699 + } + }, + { + "ph": "f", "id": 48699, "pid": 435, "tid": 435, "ts": 1776649791313508, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791313518, "dur": 15, + "args": { + "External id": 48700, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48700, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48700, "pid": 0, "tid": 7, "ts": 1776649791313518, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313509, "dur": 5, + "args": { + "External id": 48700, + "cbid": 211, "correlation": 48700 + } + }, + { + "ph": "s", "id": 48700, "pid": 435, "tid": 435, "ts": 1776649791313509, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791313555, "dur": 1, + "args": { + "External id": 6289, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48718, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48718, "pid": 0, "tid": 7, "ts": 1776649791313555, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313546, "dur": 7, + "args": { + "External id": 6289, + "cbid": 211, "correlation": 48718 + } + }, + { + "ph": "s", "id": 48718, "pid": 435, "tid": 435, "ts": 1776649791313546, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313559, "dur": 0, + "args": { + "External id": 48724, + "cbid": 200, "correlation": 48724 + } + }, + { + "ph": "f", "id": 48724, "pid": 435, "tid": 435, "ts": 1776649791313559, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313560, "dur": 0, + "args": { + "External id": 48725, + "cbid": 273, "correlation": 48725 + } + }, + { + "ph": "f", "id": 48725, "pid": 435, "tid": 435, "ts": 1776649791313560, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791313569, "dur": 15, + "args": { + "External id": 48726, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48726, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48726, "pid": 0, "tid": 7, "ts": 1776649791313569, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313561, "dur": 5, + "args": { + "External id": 48726, + "cbid": 211, "correlation": 48726 + } + }, + { + "ph": "s", "id": 48726, "pid": 435, "tid": 435, "ts": 1776649791313561, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791313605, "dur": 1, + "args": { + "External id": 6294, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48744, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48744, "pid": 0, "tid": 7, "ts": 1776649791313605, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313596, "dur": 6, + "args": { + "External id": 6294, + "cbid": 211, "correlation": 48744 + } + }, + { + "ph": "s", "id": 48744, "pid": 435, "tid": 435, "ts": 1776649791313596, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313608, "dur": 0, + "args": { + "External id": 48750, + "cbid": 200, "correlation": 48750 + } + }, + { + "ph": "f", "id": 48750, "pid": 435, "tid": 435, "ts": 1776649791313608, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313609, "dur": 0, + "args": { + "External id": 48751, + "cbid": 273, "correlation": 48751 + } + }, + { + "ph": "f", "id": 48751, "pid": 435, "tid": 435, "ts": 1776649791313609, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791313618, "dur": 15, + "args": { + "External id": 48752, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48752, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48752, "pid": 0, "tid": 7, "ts": 1776649791313618, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313610, "dur": 4, + "args": { + "External id": 48752, + "cbid": 211, "correlation": 48752 + } + }, + { + "ph": "s", "id": 48752, "pid": 435, "tid": 435, "ts": 1776649791313610, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791313653, "dur": 1, + "args": { + "External id": 6299, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48770, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48770, "pid": 0, "tid": 7, "ts": 1776649791313653, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313644, "dur": 6, + "args": { + "External id": 6299, + "cbid": 211, "correlation": 48770 + } + }, + { + "ph": "s", "id": 48770, "pid": 435, "tid": 435, "ts": 1776649791313644, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313657, "dur": 0, + "args": { + "External id": 48776, + "cbid": 200, "correlation": 48776 + } + }, + { + "ph": "f", "id": 48776, "pid": 435, "tid": 435, "ts": 1776649791313657, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313657, "dur": 0, + "args": { + "External id": 48777, + "cbid": 273, "correlation": 48777 + } + }, + { + "ph": "f", "id": 48777, "pid": 435, "tid": 435, "ts": 1776649791313657, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791313666, "dur": 17, + "args": { + "External id": 48778, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48778, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48778, "pid": 0, "tid": 7, "ts": 1776649791313666, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313658, "dur": 5, + "args": { + "External id": 48778, + "cbid": 211, "correlation": 48778 + } + }, + { + "ph": "s", "id": 48778, "pid": 435, "tid": 435, "ts": 1776649791313658, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791313702, "dur": 1, + "args": { + "External id": 6304, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48796, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48796, "pid": 0, "tid": 7, "ts": 1776649791313702, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313692, "dur": 7, + "args": { + "External id": 6304, + "cbid": 211, "correlation": 48796 + } + }, + { + "ph": "s", "id": 48796, "pid": 435, "tid": 435, "ts": 1776649791313692, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313705, "dur": 0, + "args": { + "External id": 48802, + "cbid": 200, "correlation": 48802 + } + }, + { + "ph": "f", "id": 48802, "pid": 435, "tid": 435, "ts": 1776649791313705, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313706, "dur": 0, + "args": { + "External id": 48803, + "cbid": 273, "correlation": 48803 + } + }, + { + "ph": "f", "id": 48803, "pid": 435, "tid": 435, "ts": 1776649791313706, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791313715, "dur": 16, + "args": { + "External id": 48804, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48804, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48804, "pid": 0, "tid": 7, "ts": 1776649791313715, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313707, "dur": 5, + "args": { + "External id": 48804, + "cbid": 211, "correlation": 48804 + } + }, + { + "ph": "s", "id": 48804, "pid": 435, "tid": 435, "ts": 1776649791313707, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791313751, "dur": 1, + "args": { + "External id": 6309, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48822, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48822, "pid": 0, "tid": 7, "ts": 1776649791313751, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313741, "dur": 6, + "args": { + "External id": 6309, + "cbid": 211, "correlation": 48822 + } + }, + { + "ph": "s", "id": 48822, "pid": 435, "tid": 435, "ts": 1776649791313741, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313754, "dur": 0, + "args": { + "External id": 48828, + "cbid": 200, "correlation": 48828 + } + }, + { + "ph": "f", "id": 48828, "pid": 435, "tid": 435, "ts": 1776649791313754, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313755, "dur": 0, + "args": { + "External id": 48829, + "cbid": 273, "correlation": 48829 + } + }, + { + "ph": "f", "id": 48829, "pid": 435, "tid": 435, "ts": 1776649791313755, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791313764, "dur": 15, + "args": { + "External id": 48830, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48830, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48830, "pid": 0, "tid": 7, "ts": 1776649791313764, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313756, "dur": 5, + "args": { + "External id": 48830, + "cbid": 211, "correlation": 48830 + } + }, + { + "ph": "s", "id": 48830, "pid": 435, "tid": 435, "ts": 1776649791313756, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791313799, "dur": 1, + "args": { + "External id": 6314, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48848, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48848, "pid": 0, "tid": 7, "ts": 1776649791313799, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313790, "dur": 7, + "args": { + "External id": 6314, + "cbid": 211, "correlation": 48848 + } + }, + { + "ph": "s", "id": 48848, "pid": 435, "tid": 435, "ts": 1776649791313790, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313803, "dur": 0, + "args": { + "External id": 48854, + "cbid": 200, "correlation": 48854 + } + }, + { + "ph": "f", "id": 48854, "pid": 435, "tid": 435, "ts": 1776649791313803, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313803, "dur": 0, + "args": { + "External id": 48855, + "cbid": 273, "correlation": 48855 + } + }, + { + "ph": "f", "id": 48855, "pid": 435, "tid": 435, "ts": 1776649791313803, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791313812, "dur": 15, + "args": { + "External id": 48856, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48856, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48856, "pid": 0, "tid": 7, "ts": 1776649791313812, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313804, "dur": 4, + "args": { + "External id": 48856, + "cbid": 211, "correlation": 48856 + } + }, + { + "ph": "s", "id": 48856, "pid": 435, "tid": 435, "ts": 1776649791313804, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791313848, "dur": 1, + "args": { + "External id": 6319, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48874, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48874, "pid": 0, "tid": 7, "ts": 1776649791313848, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313838, "dur": 7, + "args": { + "External id": 6319, + "cbid": 211, "correlation": 48874 + } + }, + { + "ph": "s", "id": 48874, "pid": 435, "tid": 435, "ts": 1776649791313838, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313851, "dur": 0, + "args": { + "External id": 48880, + "cbid": 200, "correlation": 48880 + } + }, + { + "ph": "f", "id": 48880, "pid": 435, "tid": 435, "ts": 1776649791313851, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313851, "dur": 0, + "args": { + "External id": 48881, + "cbid": 273, "correlation": 48881 + } + }, + { + "ph": "f", "id": 48881, "pid": 435, "tid": 435, "ts": 1776649791313851, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791313861, "dur": 15, + "args": { + "External id": 48882, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48882, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48882, "pid": 0, "tid": 7, "ts": 1776649791313861, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313853, "dur": 5, + "args": { + "External id": 48882, + "cbid": 211, "correlation": 48882 + } + }, + { + "ph": "s", "id": 48882, "pid": 435, "tid": 435, "ts": 1776649791313853, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791313896, "dur": 1, + "args": { + "External id": 6324, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48900, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48900, "pid": 0, "tid": 7, "ts": 1776649791313896, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313887, "dur": 6, + "args": { + "External id": 6324, + "cbid": 211, "correlation": 48900 + } + }, + { + "ph": "s", "id": 48900, "pid": 435, "tid": 435, "ts": 1776649791313887, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313899, "dur": 0, + "args": { + "External id": 48906, + "cbid": 200, "correlation": 48906 + } + }, + { + "ph": "f", "id": 48906, "pid": 435, "tid": 435, "ts": 1776649791313899, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313900, "dur": 0, + "args": { + "External id": 48907, + "cbid": 273, "correlation": 48907 + } + }, + { + "ph": "f", "id": 48907, "pid": 435, "tid": 435, "ts": 1776649791313900, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791313909, "dur": 15, + "args": { + "External id": 48908, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48908, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48908, "pid": 0, "tid": 7, "ts": 1776649791313909, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313901, "dur": 4, + "args": { + "External id": 48908, + "cbid": 211, "correlation": 48908 + } + }, + { + "ph": "s", "id": 48908, "pid": 435, "tid": 435, "ts": 1776649791313901, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791313945, "dur": 1, + "args": { + "External id": 6329, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48926, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48926, "pid": 0, "tid": 7, "ts": 1776649791313945, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313935, "dur": 6, + "args": { + "External id": 6329, + "cbid": 211, "correlation": 48926 + } + }, + { + "ph": "s", "id": 48926, "pid": 435, "tid": 435, "ts": 1776649791313935, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313948, "dur": 0, + "args": { + "External id": 48932, + "cbid": 200, "correlation": 48932 + } + }, + { + "ph": "f", "id": 48932, "pid": 435, "tid": 435, "ts": 1776649791313948, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313948, "dur": 0, + "args": { + "External id": 48933, + "cbid": 273, "correlation": 48933 + } + }, + { + "ph": "f", "id": 48933, "pid": 435, "tid": 435, "ts": 1776649791313948, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791313957, "dur": 15, + "args": { + "External id": 48934, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48934, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48934, "pid": 0, "tid": 7, "ts": 1776649791313957, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313949, "dur": 5, + "args": { + "External id": 48934, + "cbid": 211, "correlation": 48934 + } + }, + { + "ph": "s", "id": 48934, "pid": 435, "tid": 435, "ts": 1776649791313949, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791313993, "dur": 1, + "args": { + "External id": 6334, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48952, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48952, "pid": 0, "tid": 7, "ts": 1776649791313993, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313984, "dur": 6, + "args": { + "External id": 6334, + "cbid": 211, "correlation": 48952 + } + }, + { + "ph": "s", "id": 48952, "pid": 435, "tid": 435, "ts": 1776649791313984, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313996, "dur": 0, + "args": { + "External id": 48958, + "cbid": 200, "correlation": 48958 + } + }, + { + "ph": "f", "id": 48958, "pid": 435, "tid": 435, "ts": 1776649791313996, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791313997, "dur": 0, + "args": { + "External id": 48959, + "cbid": 273, "correlation": 48959 + } + }, + { + "ph": "f", "id": 48959, "pid": 435, "tid": 435, "ts": 1776649791313997, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791314006, "dur": 15, + "args": { + "External id": 48960, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48960, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48960, "pid": 0, "tid": 7, "ts": 1776649791314006, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791313998, "dur": 5, + "args": { + "External id": 48960, + "cbid": 211, "correlation": 48960 + } + }, + { + "ph": "s", "id": 48960, "pid": 435, "tid": 435, "ts": 1776649791313998, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791314041, "dur": 1, + "args": { + "External id": 6339, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48978, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48978, "pid": 0, "tid": 7, "ts": 1776649791314041, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314032, "dur": 6, + "args": { + "External id": 6339, + "cbid": 211, "correlation": 48978 + } + }, + { + "ph": "s", "id": 48978, "pid": 435, "tid": 435, "ts": 1776649791314032, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314044, "dur": 0, + "args": { + "External id": 48984, + "cbid": 200, "correlation": 48984 + } + }, + { + "ph": "f", "id": 48984, "pid": 435, "tid": 435, "ts": 1776649791314044, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314045, "dur": 0, + "args": { + "External id": 48985, + "cbid": 273, "correlation": 48985 + } + }, + { + "ph": "f", "id": 48985, "pid": 435, "tid": 435, "ts": 1776649791314045, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791314054, "dur": 15, + "args": { + "External id": 48986, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 48986, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 48986, "pid": 0, "tid": 7, "ts": 1776649791314054, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314046, "dur": 5, + "args": { + "External id": 48986, + "cbid": 211, "correlation": 48986 + } + }, + { + "ph": "s", "id": 48986, "pid": 435, "tid": 435, "ts": 1776649791314046, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791314089, "dur": 1, + "args": { + "External id": 6344, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49004, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49004, "pid": 0, "tid": 7, "ts": 1776649791314089, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314080, "dur": 6, + "args": { + "External id": 6344, + "cbid": 211, "correlation": 49004 + } + }, + { + "ph": "s", "id": 49004, "pid": 435, "tid": 435, "ts": 1776649791314080, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314092, "dur": 0, + "args": { + "External id": 49010, + "cbid": 200, "correlation": 49010 + } + }, + { + "ph": "f", "id": 49010, "pid": 435, "tid": 435, "ts": 1776649791314092, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314093, "dur": 0, + "args": { + "External id": 49011, + "cbid": 273, "correlation": 49011 + } + }, + { + "ph": "f", "id": 49011, "pid": 435, "tid": 435, "ts": 1776649791314093, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791314102, "dur": 16, + "args": { + "External id": 49012, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49012, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49012, "pid": 0, "tid": 7, "ts": 1776649791314102, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314094, "dur": 5, + "args": { + "External id": 49012, + "cbid": 211, "correlation": 49012 + } + }, + { + "ph": "s", "id": 49012, "pid": 435, "tid": 435, "ts": 1776649791314094, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791314138, "dur": 1, + "args": { + "External id": 6349, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49030, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49030, "pid": 0, "tid": 7, "ts": 1776649791314138, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314128, "dur": 6, + "args": { + "External id": 6349, + "cbid": 211, "correlation": 49030 + } + }, + { + "ph": "s", "id": 49030, "pid": 435, "tid": 435, "ts": 1776649791314128, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314141, "dur": 0, + "args": { + "External id": 49036, + "cbid": 200, "correlation": 49036 + } + }, + { + "ph": "f", "id": 49036, "pid": 435, "tid": 435, "ts": 1776649791314141, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314141, "dur": 0, + "args": { + "External id": 49037, + "cbid": 273, "correlation": 49037 + } + }, + { + "ph": "f", "id": 49037, "pid": 435, "tid": 435, "ts": 1776649791314141, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791314150, "dur": 15, + "args": { + "External id": 49038, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49038, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49038, "pid": 0, "tid": 7, "ts": 1776649791314150, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314142, "dur": 5, + "args": { + "External id": 49038, + "cbid": 211, "correlation": 49038 + } + }, + { + "ph": "s", "id": 49038, "pid": 435, "tid": 435, "ts": 1776649791314142, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791314186, "dur": 1, + "args": { + "External id": 6354, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49056, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49056, "pid": 0, "tid": 7, "ts": 1776649791314186, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314177, "dur": 6, + "args": { + "External id": 6354, + "cbid": 211, "correlation": 49056 + } + }, + { + "ph": "s", "id": 49056, "pid": 435, "tid": 435, "ts": 1776649791314177, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314189, "dur": 0, + "args": { + "External id": 49062, + "cbid": 200, "correlation": 49062 + } + }, + { + "ph": "f", "id": 49062, "pid": 435, "tid": 435, "ts": 1776649791314189, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314190, "dur": 0, + "args": { + "External id": 49063, + "cbid": 273, "correlation": 49063 + } + }, + { + "ph": "f", "id": 49063, "pid": 435, "tid": 435, "ts": 1776649791314190, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791314199, "dur": 15, + "args": { + "External id": 49064, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49064, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49064, "pid": 0, "tid": 7, "ts": 1776649791314199, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314191, "dur": 4, + "args": { + "External id": 49064, + "cbid": 211, "correlation": 49064 + } + }, + { + "ph": "s", "id": 49064, "pid": 435, "tid": 435, "ts": 1776649791314191, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791314234, "dur": 1, + "args": { + "External id": 6359, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49082, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49082, "pid": 0, "tid": 7, "ts": 1776649791314234, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314225, "dur": 7, + "args": { + "External id": 6359, + "cbid": 211, "correlation": 49082 + } + }, + { + "ph": "s", "id": 49082, "pid": 435, "tid": 435, "ts": 1776649791314225, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314238, "dur": 0, + "args": { + "External id": 49088, + "cbid": 200, "correlation": 49088 + } + }, + { + "ph": "f", "id": 49088, "pid": 435, "tid": 435, "ts": 1776649791314238, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314238, "dur": 0, + "args": { + "External id": 49089, + "cbid": 273, "correlation": 49089 + } + }, + { + "ph": "f", "id": 49089, "pid": 435, "tid": 435, "ts": 1776649791314238, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791314248, "dur": 15, + "args": { + "External id": 49090, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49090, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49090, "pid": 0, "tid": 7, "ts": 1776649791314248, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314240, "dur": 5, + "args": { + "External id": 49090, + "cbid": 211, "correlation": 49090 + } + }, + { + "ph": "s", "id": 49090, "pid": 435, "tid": 435, "ts": 1776649791314240, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791314283, "dur": 1, + "args": { + "External id": 6364, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49108, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49108, "pid": 0, "tid": 7, "ts": 1776649791314283, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314274, "dur": 7, + "args": { + "External id": 6364, + "cbid": 211, "correlation": 49108 + } + }, + { + "ph": "s", "id": 49108, "pid": 435, "tid": 435, "ts": 1776649791314274, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314286, "dur": 0, + "args": { + "External id": 49114, + "cbid": 200, "correlation": 49114 + } + }, + { + "ph": "f", "id": 49114, "pid": 435, "tid": 435, "ts": 1776649791314286, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314287, "dur": 0, + "args": { + "External id": 49115, + "cbid": 273, "correlation": 49115 + } + }, + { + "ph": "f", "id": 49115, "pid": 435, "tid": 435, "ts": 1776649791314287, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791314296, "dur": 15, + "args": { + "External id": 49116, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49116, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49116, "pid": 0, "tid": 7, "ts": 1776649791314296, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314288, "dur": 5, + "args": { + "External id": 49116, + "cbid": 211, "correlation": 49116 + } + }, + { + "ph": "s", "id": 49116, "pid": 435, "tid": 435, "ts": 1776649791314288, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791314331, "dur": 1, + "args": { + "External id": 6369, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49134, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49134, "pid": 0, "tid": 7, "ts": 1776649791314331, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314322, "dur": 6, + "args": { + "External id": 6369, + "cbid": 211, "correlation": 49134 + } + }, + { + "ph": "s", "id": 49134, "pid": 435, "tid": 435, "ts": 1776649791314322, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314334, "dur": 0, + "args": { + "External id": 49140, + "cbid": 200, "correlation": 49140 + } + }, + { + "ph": "f", "id": 49140, "pid": 435, "tid": 435, "ts": 1776649791314334, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314335, "dur": 0, + "args": { + "External id": 49141, + "cbid": 273, "correlation": 49141 + } + }, + { + "ph": "f", "id": 49141, "pid": 435, "tid": 435, "ts": 1776649791314335, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791314344, "dur": 15, + "args": { + "External id": 49142, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49142, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49142, "pid": 0, "tid": 7, "ts": 1776649791314344, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314336, "dur": 5, + "args": { + "External id": 49142, + "cbid": 211, "correlation": 49142 + } + }, + { + "ph": "s", "id": 49142, "pid": 435, "tid": 435, "ts": 1776649791314336, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791314380, "dur": 1, + "args": { + "External id": 6374, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49160, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49160, "pid": 0, "tid": 7, "ts": 1776649791314380, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314371, "dur": 6, + "args": { + "External id": 6374, + "cbid": 211, "correlation": 49160 + } + }, + { + "ph": "s", "id": 49160, "pid": 435, "tid": 435, "ts": 1776649791314371, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314383, "dur": 0, + "args": { + "External id": 49166, + "cbid": 200, "correlation": 49166 + } + }, + { + "ph": "f", "id": 49166, "pid": 435, "tid": 435, "ts": 1776649791314383, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314384, "dur": 0, + "args": { + "External id": 49167, + "cbid": 273, "correlation": 49167 + } + }, + { + "ph": "f", "id": 49167, "pid": 435, "tid": 435, "ts": 1776649791314384, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791314392, "dur": 15, + "args": { + "External id": 49168, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49168, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49168, "pid": 0, "tid": 7, "ts": 1776649791314392, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314385, "dur": 4, + "args": { + "External id": 49168, + "cbid": 211, "correlation": 49168 + } + }, + { + "ph": "s", "id": 49168, "pid": 435, "tid": 435, "ts": 1776649791314385, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791314427, "dur": 1, + "args": { + "External id": 6379, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49186, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49186, "pid": 0, "tid": 7, "ts": 1776649791314427, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314418, "dur": 6, + "args": { + "External id": 6379, + "cbid": 211, "correlation": 49186 + } + }, + { + "ph": "s", "id": 49186, "pid": 435, "tid": 435, "ts": 1776649791314418, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314431, "dur": 0, + "args": { + "External id": 49192, + "cbid": 200, "correlation": 49192 + } + }, + { + "ph": "f", "id": 49192, "pid": 435, "tid": 435, "ts": 1776649791314431, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314431, "dur": 0, + "args": { + "External id": 49193, + "cbid": 273, "correlation": 49193 + } + }, + { + "ph": "f", "id": 49193, "pid": 435, "tid": 435, "ts": 1776649791314431, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791314440, "dur": 15, + "args": { + "External id": 49194, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49194, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49194, "pid": 0, "tid": 7, "ts": 1776649791314440, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314432, "dur": 4, + "args": { + "External id": 49194, + "cbid": 211, "correlation": 49194 + } + }, + { + "ph": "s", "id": 49194, "pid": 435, "tid": 435, "ts": 1776649791314432, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791314476, "dur": 1, + "args": { + "External id": 6384, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49212, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49212, "pid": 0, "tid": 7, "ts": 1776649791314476, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314467, "dur": 7, + "args": { + "External id": 6384, + "cbid": 211, "correlation": 49212 + } + }, + { + "ph": "s", "id": 49212, "pid": 435, "tid": 435, "ts": 1776649791314467, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314479, "dur": 0, + "args": { + "External id": 49218, + "cbid": 200, "correlation": 49218 + } + }, + { + "ph": "f", "id": 49218, "pid": 435, "tid": 435, "ts": 1776649791314479, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314480, "dur": 0, + "args": { + "External id": 49219, + "cbid": 273, "correlation": 49219 + } + }, + { + "ph": "f", "id": 49219, "pid": 435, "tid": 435, "ts": 1776649791314480, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791314489, "dur": 15, + "args": { + "External id": 49220, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49220, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49220, "pid": 0, "tid": 7, "ts": 1776649791314489, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314481, "dur": 5, + "args": { + "External id": 49220, + "cbid": 211, "correlation": 49220 + } + }, + { + "ph": "s", "id": 49220, "pid": 435, "tid": 435, "ts": 1776649791314481, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791314524, "dur": 1, + "args": { + "External id": 6389, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49238, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49238, "pid": 0, "tid": 7, "ts": 1776649791314524, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314515, "dur": 6, + "args": { + "External id": 6389, + "cbid": 211, "correlation": 49238 + } + }, + { + "ph": "s", "id": 49238, "pid": 435, "tid": 435, "ts": 1776649791314515, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314528, "dur": 0, + "args": { + "External id": 49244, + "cbid": 200, "correlation": 49244 + } + }, + { + "ph": "f", "id": 49244, "pid": 435, "tid": 435, "ts": 1776649791314528, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314528, "dur": 0, + "args": { + "External id": 49245, + "cbid": 273, "correlation": 49245 + } + }, + { + "ph": "f", "id": 49245, "pid": 435, "tid": 435, "ts": 1776649791314528, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791314537, "dur": 16, + "args": { + "External id": 49246, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49246, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49246, "pid": 0, "tid": 7, "ts": 1776649791314537, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314529, "dur": 5, + "args": { + "External id": 49246, + "cbid": 211, "correlation": 49246 + } + }, + { + "ph": "s", "id": 49246, "pid": 435, "tid": 435, "ts": 1776649791314529, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791314572, "dur": 1, + "args": { + "External id": 6394, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49264, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49264, "pid": 0, "tid": 7, "ts": 1776649791314572, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314563, "dur": 6, + "args": { + "External id": 6394, + "cbid": 211, "correlation": 49264 + } + }, + { + "ph": "s", "id": 49264, "pid": 435, "tid": 435, "ts": 1776649791314563, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314575, "dur": 0, + "args": { + "External id": 49270, + "cbid": 200, "correlation": 49270 + } + }, + { + "ph": "f", "id": 49270, "pid": 435, "tid": 435, "ts": 1776649791314575, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314576, "dur": 0, + "args": { + "External id": 49271, + "cbid": 273, "correlation": 49271 + } + }, + { + "ph": "f", "id": 49271, "pid": 435, "tid": 435, "ts": 1776649791314576, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791314585, "dur": 15, + "args": { + "External id": 49272, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49272, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49272, "pid": 0, "tid": 7, "ts": 1776649791314585, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314577, "dur": 4, + "args": { + "External id": 49272, + "cbid": 211, "correlation": 49272 + } + }, + { + "ph": "s", "id": 49272, "pid": 435, "tid": 435, "ts": 1776649791314577, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791314620, "dur": 1, + "args": { + "External id": 6399, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49290, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49290, "pid": 0, "tid": 7, "ts": 1776649791314620, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314611, "dur": 6, + "args": { + "External id": 6399, + "cbid": 211, "correlation": 49290 + } + }, + { + "ph": "s", "id": 49290, "pid": 435, "tid": 435, "ts": 1776649791314611, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314624, "dur": 0, + "args": { + "External id": 49296, + "cbid": 200, "correlation": 49296 + } + }, + { + "ph": "f", "id": 49296, "pid": 435, "tid": 435, "ts": 1776649791314624, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314624, "dur": 0, + "args": { + "External id": 49297, + "cbid": 273, "correlation": 49297 + } + }, + { + "ph": "f", "id": 49297, "pid": 435, "tid": 435, "ts": 1776649791314624, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791314633, "dur": 15, + "args": { + "External id": 49298, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49298, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49298, "pid": 0, "tid": 7, "ts": 1776649791314633, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314625, "dur": 4, + "args": { + "External id": 49298, + "cbid": 211, "correlation": 49298 + } + }, + { + "ph": "s", "id": 49298, "pid": 435, "tid": 435, "ts": 1776649791314625, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791314669, "dur": 1, + "args": { + "External id": 6404, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49316, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49316, "pid": 0, "tid": 7, "ts": 1776649791314669, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314659, "dur": 6, + "args": { + "External id": 6404, + "cbid": 211, "correlation": 49316 + } + }, + { + "ph": "s", "id": 49316, "pid": 435, "tid": 435, "ts": 1776649791314659, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314672, "dur": 0, + "args": { + "External id": 49322, + "cbid": 200, "correlation": 49322 + } + }, + { + "ph": "f", "id": 49322, "pid": 435, "tid": 435, "ts": 1776649791314672, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314672, "dur": 0, + "args": { + "External id": 49323, + "cbid": 273, "correlation": 49323 + } + }, + { + "ph": "f", "id": 49323, "pid": 435, "tid": 435, "ts": 1776649791314672, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791314682, "dur": 15, + "args": { + "External id": 49324, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49324, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49324, "pid": 0, "tid": 7, "ts": 1776649791314682, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314674, "dur": 5, + "args": { + "External id": 49324, + "cbid": 211, "correlation": 49324 + } + }, + { + "ph": "s", "id": 49324, "pid": 435, "tid": 435, "ts": 1776649791314674, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791314717, "dur": 1, + "args": { + "External id": 6409, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49342, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49342, "pid": 0, "tid": 7, "ts": 1776649791314717, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314708, "dur": 6, + "args": { + "External id": 6409, + "cbid": 211, "correlation": 49342 + } + }, + { + "ph": "s", "id": 49342, "pid": 435, "tid": 435, "ts": 1776649791314708, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314720, "dur": 0, + "args": { + "External id": 49348, + "cbid": 200, "correlation": 49348 + } + }, + { + "ph": "f", "id": 49348, "pid": 435, "tid": 435, "ts": 1776649791314720, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314721, "dur": 0, + "args": { + "External id": 49349, + "cbid": 273, "correlation": 49349 + } + }, + { + "ph": "f", "id": 49349, "pid": 435, "tid": 435, "ts": 1776649791314721, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791314730, "dur": 15, + "args": { + "External id": 49350, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49350, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49350, "pid": 0, "tid": 7, "ts": 1776649791314730, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314722, "dur": 5, + "args": { + "External id": 49350, + "cbid": 211, "correlation": 49350 + } + }, + { + "ph": "s", "id": 49350, "pid": 435, "tid": 435, "ts": 1776649791314722, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791314765, "dur": 1, + "args": { + "External id": 6414, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49368, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49368, "pid": 0, "tid": 7, "ts": 1776649791314765, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314756, "dur": 6, + "args": { + "External id": 6414, + "cbid": 211, "correlation": 49368 + } + }, + { + "ph": "s", "id": 49368, "pid": 435, "tid": 435, "ts": 1776649791314756, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314768, "dur": 0, + "args": { + "External id": 49374, + "cbid": 200, "correlation": 49374 + } + }, + { + "ph": "f", "id": 49374, "pid": 435, "tid": 435, "ts": 1776649791314768, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314769, "dur": 0, + "args": { + "External id": 49375, + "cbid": 273, "correlation": 49375 + } + }, + { + "ph": "f", "id": 49375, "pid": 435, "tid": 435, "ts": 1776649791314769, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791314778, "dur": 15, + "args": { + "External id": 49376, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49376, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49376, "pid": 0, "tid": 7, "ts": 1776649791314778, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314770, "dur": 4, + "args": { + "External id": 49376, + "cbid": 211, "correlation": 49376 + } + }, + { + "ph": "s", "id": 49376, "pid": 435, "tid": 435, "ts": 1776649791314770, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791314813, "dur": 1, + "args": { + "External id": 6419, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49394, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49394, "pid": 0, "tid": 7, "ts": 1776649791314813, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314804, "dur": 6, + "args": { + "External id": 6419, + "cbid": 211, "correlation": 49394 + } + }, + { + "ph": "s", "id": 49394, "pid": 435, "tid": 435, "ts": 1776649791314804, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314816, "dur": 0, + "args": { + "External id": 49400, + "cbid": 200, "correlation": 49400 + } + }, + { + "ph": "f", "id": 49400, "pid": 435, "tid": 435, "ts": 1776649791314816, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314817, "dur": 0, + "args": { + "External id": 49401, + "cbid": 273, "correlation": 49401 + } + }, + { + "ph": "f", "id": 49401, "pid": 435, "tid": 435, "ts": 1776649791314817, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791314827, "dur": 15, + "args": { + "External id": 49402, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49402, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49402, "pid": 0, "tid": 7, "ts": 1776649791314827, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314818, "dur": 5, + "args": { + "External id": 49402, + "cbid": 211, "correlation": 49402 + } + }, + { + "ph": "s", "id": 49402, "pid": 435, "tid": 435, "ts": 1776649791314818, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791314864, "dur": 1, + "args": { + "External id": 6424, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49420, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49420, "pid": 0, "tid": 7, "ts": 1776649791314864, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314855, "dur": 6, + "args": { + "External id": 6424, + "cbid": 211, "correlation": 49420 + } + }, + { + "ph": "s", "id": 49420, "pid": 435, "tid": 435, "ts": 1776649791314855, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314873, "dur": 0, + "args": { + "External id": 49426, + "cbid": 200, "correlation": 49426 + } + }, + { + "ph": "f", "id": 49426, "pid": 435, "tid": 435, "ts": 1776649791314873, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314874, "dur": 0, + "args": { + "External id": 49427, + "cbid": 273, "correlation": 49427 + } + }, + { + "ph": "f", "id": 49427, "pid": 435, "tid": 435, "ts": 1776649791314874, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791314883, "dur": 15, + "args": { + "External id": 49428, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49428, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49428, "pid": 0, "tid": 7, "ts": 1776649791314883, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314875, "dur": 5, + "args": { + "External id": 49428, + "cbid": 211, "correlation": 49428 + } + }, + { + "ph": "s", "id": 49428, "pid": 435, "tid": 435, "ts": 1776649791314875, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791314921, "dur": 1, + "args": { + "External id": 6429, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49446, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49446, "pid": 0, "tid": 7, "ts": 1776649791314921, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314912, "dur": 6, + "args": { + "External id": 6429, + "cbid": 211, "correlation": 49446 + } + }, + { + "ph": "s", "id": 49446, "pid": 435, "tid": 435, "ts": 1776649791314912, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314925, "dur": 0, + "args": { + "External id": 49452, + "cbid": 200, "correlation": 49452 + } + }, + { + "ph": "f", "id": 49452, "pid": 435, "tid": 435, "ts": 1776649791314925, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314925, "dur": 0, + "args": { + "External id": 49453, + "cbid": 273, "correlation": 49453 + } + }, + { + "ph": "f", "id": 49453, "pid": 435, "tid": 435, "ts": 1776649791314925, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791314934, "dur": 16, + "args": { + "External id": 49454, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49454, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49454, "pid": 0, "tid": 7, "ts": 1776649791314934, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314926, "dur": 5, + "args": { + "External id": 49454, + "cbid": 211, "correlation": 49454 + } + }, + { + "ph": "s", "id": 49454, "pid": 435, "tid": 435, "ts": 1776649791314926, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791314970, "dur": 1, + "args": { + "External id": 6434, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49472, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49472, "pid": 0, "tid": 7, "ts": 1776649791314970, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314961, "dur": 6, + "args": { + "External id": 6434, + "cbid": 211, "correlation": 49472 + } + }, + { + "ph": "s", "id": 49472, "pid": 435, "tid": 435, "ts": 1776649791314961, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314973, "dur": 0, + "args": { + "External id": 49478, + "cbid": 200, "correlation": 49478 + } + }, + { + "ph": "f", "id": 49478, "pid": 435, "tid": 435, "ts": 1776649791314973, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791314974, "dur": 0, + "args": { + "External id": 49479, + "cbid": 273, "correlation": 49479 + } + }, + { + "ph": "f", "id": 49479, "pid": 435, "tid": 435, "ts": 1776649791314974, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791314982, "dur": 17, + "args": { + "External id": 49480, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49480, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49480, "pid": 0, "tid": 7, "ts": 1776649791314982, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791314975, "dur": 4, + "args": { + "External id": 49480, + "cbid": 211, "correlation": 49480 + } + }, + { + "ph": "s", "id": 49480, "pid": 435, "tid": 435, "ts": 1776649791314975, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791315018, "dur": 1, + "args": { + "External id": 6439, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49498, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49498, "pid": 0, "tid": 7, "ts": 1776649791315018, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315009, "dur": 6, + "args": { + "External id": 6439, + "cbid": 211, "correlation": 49498 + } + }, + { + "ph": "s", "id": 49498, "pid": 435, "tid": 435, "ts": 1776649791315009, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315022, "dur": 0, + "args": { + "External id": 49504, + "cbid": 200, "correlation": 49504 + } + }, + { + "ph": "f", "id": 49504, "pid": 435, "tid": 435, "ts": 1776649791315022, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315022, "dur": 0, + "args": { + "External id": 49505, + "cbid": 273, "correlation": 49505 + } + }, + { + "ph": "f", "id": 49505, "pid": 435, "tid": 435, "ts": 1776649791315022, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791315031, "dur": 15, + "args": { + "External id": 49506, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49506, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49506, "pid": 0, "tid": 7, "ts": 1776649791315031, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315023, "dur": 5, + "args": { + "External id": 49506, + "cbid": 211, "correlation": 49506 + } + }, + { + "ph": "s", "id": 49506, "pid": 435, "tid": 435, "ts": 1776649791315023, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791315067, "dur": 1, + "args": { + "External id": 6444, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49524, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49524, "pid": 0, "tid": 7, "ts": 1776649791315067, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315057, "dur": 6, + "args": { + "External id": 6444, + "cbid": 211, "correlation": 49524 + } + }, + { + "ph": "s", "id": 49524, "pid": 435, "tid": 435, "ts": 1776649791315057, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315070, "dur": 0, + "args": { + "External id": 49530, + "cbid": 200, "correlation": 49530 + } + }, + { + "ph": "f", "id": 49530, "pid": 435, "tid": 435, "ts": 1776649791315070, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315070, "dur": 0, + "args": { + "External id": 49531, + "cbid": 273, "correlation": 49531 + } + }, + { + "ph": "f", "id": 49531, "pid": 435, "tid": 435, "ts": 1776649791315070, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791315079, "dur": 15, + "args": { + "External id": 49532, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49532, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49532, "pid": 0, "tid": 7, "ts": 1776649791315079, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315071, "dur": 5, + "args": { + "External id": 49532, + "cbid": 211, "correlation": 49532 + } + }, + { + "ph": "s", "id": 49532, "pid": 435, "tid": 435, "ts": 1776649791315071, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791315114, "dur": 1, + "args": { + "External id": 6449, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49550, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49550, "pid": 0, "tid": 7, "ts": 1776649791315114, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315105, "dur": 6, + "args": { + "External id": 6449, + "cbid": 211, "correlation": 49550 + } + }, + { + "ph": "s", "id": 49550, "pid": 435, "tid": 435, "ts": 1776649791315105, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315118, "dur": 0, + "args": { + "External id": 49556, + "cbid": 200, "correlation": 49556 + } + }, + { + "ph": "f", "id": 49556, "pid": 435, "tid": 435, "ts": 1776649791315118, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315118, "dur": 0, + "args": { + "External id": 49557, + "cbid": 273, "correlation": 49557 + } + }, + { + "ph": "f", "id": 49557, "pid": 435, "tid": 435, "ts": 1776649791315118, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791315128, "dur": 15, + "args": { + "External id": 49558, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49558, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49558, "pid": 0, "tid": 7, "ts": 1776649791315128, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315119, "dur": 4, + "args": { + "External id": 49558, + "cbid": 211, "correlation": 49558 + } + }, + { + "ph": "s", "id": 49558, "pid": 435, "tid": 435, "ts": 1776649791315119, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791315162, "dur": 1, + "args": { + "External id": 6454, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49576, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49576, "pid": 0, "tid": 7, "ts": 1776649791315162, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315153, "dur": 6, + "args": { + "External id": 6454, + "cbid": 211, "correlation": 49576 + } + }, + { + "ph": "s", "id": 49576, "pid": 435, "tid": 435, "ts": 1776649791315153, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315165, "dur": 0, + "args": { + "External id": 49582, + "cbid": 200, "correlation": 49582 + } + }, + { + "ph": "f", "id": 49582, "pid": 435, "tid": 435, "ts": 1776649791315165, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315166, "dur": 0, + "args": { + "External id": 49583, + "cbid": 273, "correlation": 49583 + } + }, + { + "ph": "f", "id": 49583, "pid": 435, "tid": 435, "ts": 1776649791315166, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791315175, "dur": 15, + "args": { + "External id": 49584, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49584, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49584, "pid": 0, "tid": 7, "ts": 1776649791315175, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315167, "dur": 5, + "args": { + "External id": 49584, + "cbid": 211, "correlation": 49584 + } + }, + { + "ph": "s", "id": 49584, "pid": 435, "tid": 435, "ts": 1776649791315167, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791315210, "dur": 1, + "args": { + "External id": 6459, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49602, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49602, "pid": 0, "tid": 7, "ts": 1776649791315210, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315201, "dur": 6, + "args": { + "External id": 6459, + "cbid": 211, "correlation": 49602 + } + }, + { + "ph": "s", "id": 49602, "pid": 435, "tid": 435, "ts": 1776649791315201, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315213, "dur": 0, + "args": { + "External id": 49608, + "cbid": 200, "correlation": 49608 + } + }, + { + "ph": "f", "id": 49608, "pid": 435, "tid": 435, "ts": 1776649791315213, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315214, "dur": 0, + "args": { + "External id": 49609, + "cbid": 273, "correlation": 49609 + } + }, + { + "ph": "f", "id": 49609, "pid": 435, "tid": 435, "ts": 1776649791315214, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791315223, "dur": 15, + "args": { + "External id": 49610, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49610, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49610, "pid": 0, "tid": 7, "ts": 1776649791315223, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315215, "dur": 5, + "args": { + "External id": 49610, + "cbid": 211, "correlation": 49610 + } + }, + { + "ph": "s", "id": 49610, "pid": 435, "tid": 435, "ts": 1776649791315215, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791315258, "dur": 1, + "args": { + "External id": 6464, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49628, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49628, "pid": 0, "tid": 7, "ts": 1776649791315258, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315249, "dur": 6, + "args": { + "External id": 6464, + "cbid": 211, "correlation": 49628 + } + }, + { + "ph": "s", "id": 49628, "pid": 435, "tid": 435, "ts": 1776649791315249, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315261, "dur": 0, + "args": { + "External id": 49634, + "cbid": 200, "correlation": 49634 + } + }, + { + "ph": "f", "id": 49634, "pid": 435, "tid": 435, "ts": 1776649791315261, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315262, "dur": 0, + "args": { + "External id": 49635, + "cbid": 273, "correlation": 49635 + } + }, + { + "ph": "f", "id": 49635, "pid": 435, "tid": 435, "ts": 1776649791315262, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791315271, "dur": 15, + "args": { + "External id": 49636, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49636, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49636, "pid": 0, "tid": 7, "ts": 1776649791315271, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315263, "dur": 5, + "args": { + "External id": 49636, + "cbid": 211, "correlation": 49636 + } + }, + { + "ph": "s", "id": 49636, "pid": 435, "tid": 435, "ts": 1776649791315263, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791315307, "dur": 1, + "args": { + "External id": 6469, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49654, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49654, "pid": 0, "tid": 7, "ts": 1776649791315307, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315297, "dur": 6, + "args": { + "External id": 6469, + "cbid": 211, "correlation": 49654 + } + }, + { + "ph": "s", "id": 49654, "pid": 435, "tid": 435, "ts": 1776649791315297, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315310, "dur": 0, + "args": { + "External id": 49660, + "cbid": 200, "correlation": 49660 + } + }, + { + "ph": "f", "id": 49660, "pid": 435, "tid": 435, "ts": 1776649791315310, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315310, "dur": 0, + "args": { + "External id": 49661, + "cbid": 273, "correlation": 49661 + } + }, + { + "ph": "f", "id": 49661, "pid": 435, "tid": 435, "ts": 1776649791315310, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791315319, "dur": 15, + "args": { + "External id": 49662, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49662, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49662, "pid": 0, "tid": 7, "ts": 1776649791315319, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315311, "dur": 4, + "args": { + "External id": 49662, + "cbid": 211, "correlation": 49662 + } + }, + { + "ph": "s", "id": 49662, "pid": 435, "tid": 435, "ts": 1776649791315311, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791315356, "dur": 1, + "args": { + "External id": 6474, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49680, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49680, "pid": 0, "tid": 7, "ts": 1776649791315356, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315346, "dur": 7, + "args": { + "External id": 6474, + "cbid": 211, "correlation": 49680 + } + }, + { + "ph": "s", "id": 49680, "pid": 435, "tid": 435, "ts": 1776649791315346, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315359, "dur": 0, + "args": { + "External id": 49686, + "cbid": 200, "correlation": 49686 + } + }, + { + "ph": "f", "id": 49686, "pid": 435, "tid": 435, "ts": 1776649791315359, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315359, "dur": 0, + "args": { + "External id": 49687, + "cbid": 273, "correlation": 49687 + } + }, + { + "ph": "f", "id": 49687, "pid": 435, "tid": 435, "ts": 1776649791315359, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791315369, "dur": 16, + "args": { + "External id": 49688, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49688, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49688, "pid": 0, "tid": 7, "ts": 1776649791315369, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315361, "dur": 4, + "args": { + "External id": 49688, + "cbid": 211, "correlation": 49688 + } + }, + { + "ph": "s", "id": 49688, "pid": 435, "tid": 435, "ts": 1776649791315361, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791315404, "dur": 1, + "args": { + "External id": 6479, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49706, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49706, "pid": 0, "tid": 7, "ts": 1776649791315404, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315395, "dur": 6, + "args": { + "External id": 6479, + "cbid": 211, "correlation": 49706 + } + }, + { + "ph": "s", "id": 49706, "pid": 435, "tid": 435, "ts": 1776649791315395, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315407, "dur": 0, + "args": { + "External id": 49712, + "cbid": 200, "correlation": 49712 + } + }, + { + "ph": "f", "id": 49712, "pid": 435, "tid": 435, "ts": 1776649791315407, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315408, "dur": 0, + "args": { + "External id": 49713, + "cbid": 273, "correlation": 49713 + } + }, + { + "ph": "f", "id": 49713, "pid": 435, "tid": 435, "ts": 1776649791315408, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791315417, "dur": 16, + "args": { + "External id": 49714, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49714, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49714, "pid": 0, "tid": 7, "ts": 1776649791315417, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315409, "dur": 5, + "args": { + "External id": 49714, + "cbid": 211, "correlation": 49714 + } + }, + { + "ph": "s", "id": 49714, "pid": 435, "tid": 435, "ts": 1776649791315409, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791315452, "dur": 1, + "args": { + "External id": 6484, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49732, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49732, "pid": 0, "tid": 7, "ts": 1776649791315452, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315442, "dur": 6, + "args": { + "External id": 6484, + "cbid": 211, "correlation": 49732 + } + }, + { + "ph": "s", "id": 49732, "pid": 435, "tid": 435, "ts": 1776649791315442, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315455, "dur": 0, + "args": { + "External id": 49738, + "cbid": 200, "correlation": 49738 + } + }, + { + "ph": "f", "id": 49738, "pid": 435, "tid": 435, "ts": 1776649791315455, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315455, "dur": 0, + "args": { + "External id": 49739, + "cbid": 273, "correlation": 49739 + } + }, + { + "ph": "f", "id": 49739, "pid": 435, "tid": 435, "ts": 1776649791315455, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791315465, "dur": 15, + "args": { + "External id": 49740, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49740, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49740, "pid": 0, "tid": 7, "ts": 1776649791315465, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315457, "dur": 5, + "args": { + "External id": 49740, + "cbid": 211, "correlation": 49740 + } + }, + { + "ph": "s", "id": 49740, "pid": 435, "tid": 435, "ts": 1776649791315457, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791315500, "dur": 1, + "args": { + "External id": 6489, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49758, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49758, "pid": 0, "tid": 7, "ts": 1776649791315500, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315490, "dur": 6, + "args": { + "External id": 6489, + "cbid": 211, "correlation": 49758 + } + }, + { + "ph": "s", "id": 49758, "pid": 435, "tid": 435, "ts": 1776649791315490, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315503, "dur": 0, + "args": { + "External id": 49764, + "cbid": 200, "correlation": 49764 + } + }, + { + "ph": "f", "id": 49764, "pid": 435, "tid": 435, "ts": 1776649791315503, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315503, "dur": 0, + "args": { + "External id": 49765, + "cbid": 273, "correlation": 49765 + } + }, + { + "ph": "f", "id": 49765, "pid": 435, "tid": 435, "ts": 1776649791315503, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791315513, "dur": 15, + "args": { + "External id": 49766, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49766, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49766, "pid": 0, "tid": 7, "ts": 1776649791315513, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315505, "dur": 4, + "args": { + "External id": 49766, + "cbid": 211, "correlation": 49766 + } + }, + { + "ph": "s", "id": 49766, "pid": 435, "tid": 435, "ts": 1776649791315505, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791315548, "dur": 1, + "args": { + "External id": 6494, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49784, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49784, "pid": 0, "tid": 7, "ts": 1776649791315548, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315539, "dur": 6, + "args": { + "External id": 6494, + "cbid": 211, "correlation": 49784 + } + }, + { + "ph": "s", "id": 49784, "pid": 435, "tid": 435, "ts": 1776649791315539, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315551, "dur": 0, + "args": { + "External id": 49790, + "cbid": 200, "correlation": 49790 + } + }, + { + "ph": "f", "id": 49790, "pid": 435, "tid": 435, "ts": 1776649791315551, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315552, "dur": 0, + "args": { + "External id": 49791, + "cbid": 273, "correlation": 49791 + } + }, + { + "ph": "f", "id": 49791, "pid": 435, "tid": 435, "ts": 1776649791315552, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791315560, "dur": 15, + "args": { + "External id": 49792, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49792, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49792, "pid": 0, "tid": 7, "ts": 1776649791315560, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315553, "dur": 4, + "args": { + "External id": 49792, + "cbid": 211, "correlation": 49792 + } + }, + { + "ph": "s", "id": 49792, "pid": 435, "tid": 435, "ts": 1776649791315553, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791315596, "dur": 1, + "args": { + "External id": 6499, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49810, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49810, "pid": 0, "tid": 7, "ts": 1776649791315596, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315587, "dur": 6, + "args": { + "External id": 6499, + "cbid": 211, "correlation": 49810 + } + }, + { + "ph": "s", "id": 49810, "pid": 435, "tid": 435, "ts": 1776649791315587, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315599, "dur": 0, + "args": { + "External id": 49816, + "cbid": 200, "correlation": 49816 + } + }, + { + "ph": "f", "id": 49816, "pid": 435, "tid": 435, "ts": 1776649791315599, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315599, "dur": 0, + "args": { + "External id": 49817, + "cbid": 273, "correlation": 49817 + } + }, + { + "ph": "f", "id": 49817, "pid": 435, "tid": 435, "ts": 1776649791315599, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791315609, "dur": 15, + "args": { + "External id": 49818, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49818, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49818, "pid": 0, "tid": 7, "ts": 1776649791315609, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315601, "dur": 5, + "args": { + "External id": 49818, + "cbid": 211, "correlation": 49818 + } + }, + { + "ph": "s", "id": 49818, "pid": 435, "tid": 435, "ts": 1776649791315601, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791315644, "dur": 1, + "args": { + "External id": 6504, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49836, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49836, "pid": 0, "tid": 7, "ts": 1776649791315644, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315634, "dur": 6, + "args": { + "External id": 6504, + "cbid": 211, "correlation": 49836 + } + }, + { + "ph": "s", "id": 49836, "pid": 435, "tid": 435, "ts": 1776649791315634, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315647, "dur": 0, + "args": { + "External id": 49842, + "cbid": 200, "correlation": 49842 + } + }, + { + "ph": "f", "id": 49842, "pid": 435, "tid": 435, "ts": 1776649791315647, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315648, "dur": 0, + "args": { + "External id": 49843, + "cbid": 273, "correlation": 49843 + } + }, + { + "ph": "f", "id": 49843, "pid": 435, "tid": 435, "ts": 1776649791315648, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791315657, "dur": 15, + "args": { + "External id": 49844, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49844, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49844, "pid": 0, "tid": 7, "ts": 1776649791315657, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315649, "dur": 4, + "args": { + "External id": 49844, + "cbid": 211, "correlation": 49844 + } + }, + { + "ph": "s", "id": 49844, "pid": 435, "tid": 435, "ts": 1776649791315649, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791315692, "dur": 1, + "args": { + "External id": 6509, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49862, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49862, "pid": 0, "tid": 7, "ts": 1776649791315692, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315683, "dur": 6, + "args": { + "External id": 6509, + "cbid": 211, "correlation": 49862 + } + }, + { + "ph": "s", "id": 49862, "pid": 435, "tid": 435, "ts": 1776649791315683, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315695, "dur": 0, + "args": { + "External id": 49868, + "cbid": 200, "correlation": 49868 + } + }, + { + "ph": "f", "id": 49868, "pid": 435, "tid": 435, "ts": 1776649791315695, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315696, "dur": 0, + "args": { + "External id": 49869, + "cbid": 273, "correlation": 49869 + } + }, + { + "ph": "f", "id": 49869, "pid": 435, "tid": 435, "ts": 1776649791315696, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791315705, "dur": 15, + "args": { + "External id": 49870, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49870, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49870, "pid": 0, "tid": 7, "ts": 1776649791315705, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315697, "dur": 4, + "args": { + "External id": 49870, + "cbid": 211, "correlation": 49870 + } + }, + { + "ph": "s", "id": 49870, "pid": 435, "tid": 435, "ts": 1776649791315697, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791315740, "dur": 1, + "args": { + "External id": 6514, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49888, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49888, "pid": 0, "tid": 7, "ts": 1776649791315740, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315731, "dur": 6, + "args": { + "External id": 6514, + "cbid": 211, "correlation": 49888 + } + }, + { + "ph": "s", "id": 49888, "pid": 435, "tid": 435, "ts": 1776649791315731, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315743, "dur": 0, + "args": { + "External id": 49894, + "cbid": 200, "correlation": 49894 + } + }, + { + "ph": "f", "id": 49894, "pid": 435, "tid": 435, "ts": 1776649791315743, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315744, "dur": 0, + "args": { + "External id": 49895, + "cbid": 273, "correlation": 49895 + } + }, + { + "ph": "f", "id": 49895, "pid": 435, "tid": 435, "ts": 1776649791315744, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791315752, "dur": 16, + "args": { + "External id": 49896, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49896, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49896, "pid": 0, "tid": 7, "ts": 1776649791315752, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315745, "dur": 4, + "args": { + "External id": 49896, + "cbid": 211, "correlation": 49896 + } + }, + { + "ph": "s", "id": 49896, "pid": 435, "tid": 435, "ts": 1776649791315745, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791315787, "dur": 1, + "args": { + "External id": 6519, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49914, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49914, "pid": 0, "tid": 7, "ts": 1776649791315787, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315778, "dur": 6, + "args": { + "External id": 6519, + "cbid": 211, "correlation": 49914 + } + }, + { + "ph": "s", "id": 49914, "pid": 435, "tid": 435, "ts": 1776649791315778, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315791, "dur": 0, + "args": { + "External id": 49920, + "cbid": 200, "correlation": 49920 + } + }, + { + "ph": "f", "id": 49920, "pid": 435, "tid": 435, "ts": 1776649791315791, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315791, "dur": 0, + "args": { + "External id": 49921, + "cbid": 273, "correlation": 49921 + } + }, + { + "ph": "f", "id": 49921, "pid": 435, "tid": 435, "ts": 1776649791315791, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791315801, "dur": 16, + "args": { + "External id": 49922, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49922, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49922, "pid": 0, "tid": 7, "ts": 1776649791315801, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315792, "dur": 5, + "args": { + "External id": 49922, + "cbid": 211, "correlation": 49922 + } + }, + { + "ph": "s", "id": 49922, "pid": 435, "tid": 435, "ts": 1776649791315792, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791315836, "dur": 1, + "args": { + "External id": 6524, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49940, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49940, "pid": 0, "tid": 7, "ts": 1776649791315836, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315826, "dur": 6, + "args": { + "External id": 6524, + "cbid": 211, "correlation": 49940 + } + }, + { + "ph": "s", "id": 49940, "pid": 435, "tid": 435, "ts": 1776649791315826, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315839, "dur": 0, + "args": { + "External id": 49946, + "cbid": 200, "correlation": 49946 + } + }, + { + "ph": "f", "id": 49946, "pid": 435, "tid": 435, "ts": 1776649791315839, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315839, "dur": 0, + "args": { + "External id": 49947, + "cbid": 273, "correlation": 49947 + } + }, + { + "ph": "f", "id": 49947, "pid": 435, "tid": 435, "ts": 1776649791315839, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791315848, "dur": 15, + "args": { + "External id": 49948, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49948, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49948, "pid": 0, "tid": 7, "ts": 1776649791315848, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315840, "dur": 4, + "args": { + "External id": 49948, + "cbid": 211, "correlation": 49948 + } + }, + { + "ph": "s", "id": 49948, "pid": 435, "tid": 435, "ts": 1776649791315840, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791315883, "dur": 1, + "args": { + "External id": 6529, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49966, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49966, "pid": 0, "tid": 7, "ts": 1776649791315883, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315874, "dur": 6, + "args": { + "External id": 6529, + "cbid": 211, "correlation": 49966 + } + }, + { + "ph": "s", "id": 49966, "pid": 435, "tid": 435, "ts": 1776649791315874, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315886, "dur": 0, + "args": { + "External id": 49972, + "cbid": 200, "correlation": 49972 + } + }, + { + "ph": "f", "id": 49972, "pid": 435, "tid": 435, "ts": 1776649791315886, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791315887, "dur": 0, + "args": { + "External id": 49973, + "cbid": 273, "correlation": 49973 + } + }, + { + "ph": "f", "id": 49973, "pid": 435, "tid": 435, "ts": 1776649791315887, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791315896, "dur": 15, + "args": { + "External id": 49974, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 49974, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 49974, "pid": 0, "tid": 7, "ts": 1776649791315896, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315888, "dur": 4, + "args": { + "External id": 49974, + "cbid": 211, "correlation": 49974 + } + }, + { + "ph": "s", "id": 49974, "pid": 435, "tid": 435, "ts": 1776649791315888, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791315918, "dur": 3, + "args": { + "External id": 49982, + "cbid": 135, "correlation": 49982 + } + }, + { + "ph": "f", "id": 49982, "pid": 435, "tid": 435, "ts": 1776649791315918, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791315932, "dur": 5, + "args": { + "External id": 49987, + "cbid": 165, "correlation": 49987 + } + }, + { + "ph": "s", "id": 49987, "pid": 435, "tid": 435, "ts": 1776649791315932, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649791315941, "dur": 4, + "args": { + "External id": 49990, + "cbid": 139, "correlation": 49990 + } + }, + { + "ph": "f", "id": 49990, "pid": 435, "tid": 435, "ts": 1776649791315941, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791315997, "dur": 1, + "args": { + "External id": 6534, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50018, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50018, "pid": 0, "tid": 7, "ts": 1776649791315997, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791315986, "dur": 8, + "args": { + "External id": 6534, + "cbid": 211, "correlation": 50018 + } + }, + { + "ph": "s", "id": 50018, "pid": 435, "tid": 435, "ts": 1776649791315986, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316001, "dur": 0, + "args": { + "External id": 50024, + "cbid": 200, "correlation": 50024 + } + }, + { + "ph": "f", "id": 50024, "pid": 435, "tid": 435, "ts": 1776649791316001, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316002, "dur": 0, + "args": { + "External id": 50025, + "cbid": 273, "correlation": 50025 + } + }, + { + "ph": "f", "id": 50025, "pid": 435, "tid": 435, "ts": 1776649791316002, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791316012, "dur": 15, + "args": { + "External id": 50026, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50026, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50026, "pid": 0, "tid": 7, "ts": 1776649791316012, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316004, "dur": 5, + "args": { + "External id": 50026, + "cbid": 211, "correlation": 50026 + } + }, + { + "ph": "s", "id": 50026, "pid": 435, "tid": 435, "ts": 1776649791316004, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791316049, "dur": 1, + "args": { + "External id": 6539, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50044, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50044, "pid": 0, "tid": 7, "ts": 1776649791316049, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316040, "dur": 6, + "args": { + "External id": 6539, + "cbid": 211, "correlation": 50044 + } + }, + { + "ph": "s", "id": 50044, "pid": 435, "tid": 435, "ts": 1776649791316040, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316053, "dur": 0, + "args": { + "External id": 50050, + "cbid": 200, "correlation": 50050 + } + }, + { + "ph": "f", "id": 50050, "pid": 435, "tid": 435, "ts": 1776649791316053, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316053, "dur": 0, + "args": { + "External id": 50051, + "cbid": 273, "correlation": 50051 + } + }, + { + "ph": "f", "id": 50051, "pid": 435, "tid": 435, "ts": 1776649791316053, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791316062, "dur": 15, + "args": { + "External id": 50052, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50052, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50052, "pid": 0, "tid": 7, "ts": 1776649791316062, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316054, "dur": 4, + "args": { + "External id": 50052, + "cbid": 211, "correlation": 50052 + } + }, + { + "ph": "s", "id": 50052, "pid": 435, "tid": 435, "ts": 1776649791316054, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791316098, "dur": 1, + "args": { + "External id": 6544, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50070, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50070, "pid": 0, "tid": 7, "ts": 1776649791316098, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316089, "dur": 6, + "args": { + "External id": 6544, + "cbid": 211, "correlation": 50070 + } + }, + { + "ph": "s", "id": 50070, "pid": 435, "tid": 435, "ts": 1776649791316089, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316101, "dur": 0, + "args": { + "External id": 50076, + "cbid": 200, "correlation": 50076 + } + }, + { + "ph": "f", "id": 50076, "pid": 435, "tid": 435, "ts": 1776649791316101, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316102, "dur": 0, + "args": { + "External id": 50077, + "cbid": 273, "correlation": 50077 + } + }, + { + "ph": "f", "id": 50077, "pid": 435, "tid": 435, "ts": 1776649791316102, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791316111, "dur": 15, + "args": { + "External id": 50078, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50078, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50078, "pid": 0, "tid": 7, "ts": 1776649791316111, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316103, "dur": 4, + "args": { + "External id": 50078, + "cbid": 211, "correlation": 50078 + } + }, + { + "ph": "s", "id": 50078, "pid": 435, "tid": 435, "ts": 1776649791316103, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791316147, "dur": 1, + "args": { + "External id": 6549, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50096, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50096, "pid": 0, "tid": 7, "ts": 1776649791316147, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316138, "dur": 6, + "args": { + "External id": 6549, + "cbid": 211, "correlation": 50096 + } + }, + { + "ph": "s", "id": 50096, "pid": 435, "tid": 435, "ts": 1776649791316138, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316151, "dur": 0, + "args": { + "External id": 50102, + "cbid": 200, "correlation": 50102 + } + }, + { + "ph": "f", "id": 50102, "pid": 435, "tid": 435, "ts": 1776649791316151, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316151, "dur": 0, + "args": { + "External id": 50103, + "cbid": 273, "correlation": 50103 + } + }, + { + "ph": "f", "id": 50103, "pid": 435, "tid": 435, "ts": 1776649791316151, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791316161, "dur": 15, + "args": { + "External id": 50104, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50104, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50104, "pid": 0, "tid": 7, "ts": 1776649791316161, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316153, "dur": 5, + "args": { + "External id": 50104, + "cbid": 211, "correlation": 50104 + } + }, + { + "ph": "s", "id": 50104, "pid": 435, "tid": 435, "ts": 1776649791316153, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791316197, "dur": 1, + "args": { + "External id": 6554, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50122, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50122, "pid": 0, "tid": 7, "ts": 1776649791316197, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316188, "dur": 6, + "args": { + "External id": 6554, + "cbid": 211, "correlation": 50122 + } + }, + { + "ph": "s", "id": 50122, "pid": 435, "tid": 435, "ts": 1776649791316188, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316200, "dur": 0, + "args": { + "External id": 50128, + "cbid": 200, "correlation": 50128 + } + }, + { + "ph": "f", "id": 50128, "pid": 435, "tid": 435, "ts": 1776649791316200, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316201, "dur": 0, + "args": { + "External id": 50129, + "cbid": 273, "correlation": 50129 + } + }, + { + "ph": "f", "id": 50129, "pid": 435, "tid": 435, "ts": 1776649791316201, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791316210, "dur": 16, + "args": { + "External id": 50130, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50130, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50130, "pid": 0, "tid": 7, "ts": 1776649791316210, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316202, "dur": 4, + "args": { + "External id": 50130, + "cbid": 211, "correlation": 50130 + } + }, + { + "ph": "s", "id": 50130, "pid": 435, "tid": 435, "ts": 1776649791316202, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791316271, "dur": 1, + "args": { + "External id": 6559, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50148, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50148, "pid": 0, "tid": 7, "ts": 1776649791316271, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316260, "dur": 7, + "args": { + "External id": 6559, + "cbid": 211, "correlation": 50148 + } + }, + { + "ph": "s", "id": 50148, "pid": 435, "tid": 435, "ts": 1776649791316260, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316274, "dur": 0, + "args": { + "External id": 50154, + "cbid": 200, "correlation": 50154 + } + }, + { + "ph": "f", "id": 50154, "pid": 435, "tid": 435, "ts": 1776649791316274, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316275, "dur": 0, + "args": { + "External id": 50155, + "cbid": 273, "correlation": 50155 + } + }, + { + "ph": "f", "id": 50155, "pid": 435, "tid": 435, "ts": 1776649791316275, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791316284, "dur": 17, + "args": { + "External id": 50156, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50156, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50156, "pid": 0, "tid": 7, "ts": 1776649791316284, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316276, "dur": 5, + "args": { + "External id": 50156, + "cbid": 211, "correlation": 50156 + } + }, + { + "ph": "s", "id": 50156, "pid": 435, "tid": 435, "ts": 1776649791316276, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791316321, "dur": 1, + "args": { + "External id": 6564, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50174, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50174, "pid": 0, "tid": 7, "ts": 1776649791316321, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316312, "dur": 6, + "args": { + "External id": 6564, + "cbid": 211, "correlation": 50174 + } + }, + { + "ph": "s", "id": 50174, "pid": 435, "tid": 435, "ts": 1776649791316312, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316324, "dur": 0, + "args": { + "External id": 50180, + "cbid": 200, "correlation": 50180 + } + }, + { + "ph": "f", "id": 50180, "pid": 435, "tid": 435, "ts": 1776649791316324, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316325, "dur": 0, + "args": { + "External id": 50181, + "cbid": 273, "correlation": 50181 + } + }, + { + "ph": "f", "id": 50181, "pid": 435, "tid": 435, "ts": 1776649791316325, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791316334, "dur": 16, + "args": { + "External id": 50182, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50182, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50182, "pid": 0, "tid": 7, "ts": 1776649791316334, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316326, "dur": 5, + "args": { + "External id": 50182, + "cbid": 211, "correlation": 50182 + } + }, + { + "ph": "s", "id": 50182, "pid": 435, "tid": 435, "ts": 1776649791316326, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791316370, "dur": 1, + "args": { + "External id": 6569, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50200, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50200, "pid": 0, "tid": 7, "ts": 1776649791316370, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316361, "dur": 6, + "args": { + "External id": 6569, + "cbid": 211, "correlation": 50200 + } + }, + { + "ph": "s", "id": 50200, "pid": 435, "tid": 435, "ts": 1776649791316361, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316373, "dur": 0, + "args": { + "External id": 50206, + "cbid": 200, "correlation": 50206 + } + }, + { + "ph": "f", "id": 50206, "pid": 435, "tid": 435, "ts": 1776649791316373, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316374, "dur": 0, + "args": { + "External id": 50207, + "cbid": 273, "correlation": 50207 + } + }, + { + "ph": "f", "id": 50207, "pid": 435, "tid": 435, "ts": 1776649791316374, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791316383, "dur": 15, + "args": { + "External id": 50208, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50208, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50208, "pid": 0, "tid": 7, "ts": 1776649791316383, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316375, "dur": 4, + "args": { + "External id": 50208, + "cbid": 211, "correlation": 50208 + } + }, + { + "ph": "s", "id": 50208, "pid": 435, "tid": 435, "ts": 1776649791316375, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791316419, "dur": 1, + "args": { + "External id": 6574, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50226, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50226, "pid": 0, "tid": 7, "ts": 1776649791316419, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316410, "dur": 6, + "args": { + "External id": 6574, + "cbid": 211, "correlation": 50226 + } + }, + { + "ph": "s", "id": 50226, "pid": 435, "tid": 435, "ts": 1776649791316410, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316422, "dur": 0, + "args": { + "External id": 50232, + "cbid": 200, "correlation": 50232 + } + }, + { + "ph": "f", "id": 50232, "pid": 435, "tid": 435, "ts": 1776649791316422, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316423, "dur": 0, + "args": { + "External id": 50233, + "cbid": 273, "correlation": 50233 + } + }, + { + "ph": "f", "id": 50233, "pid": 435, "tid": 435, "ts": 1776649791316423, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791316431, "dur": 15, + "args": { + "External id": 50234, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50234, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50234, "pid": 0, "tid": 7, "ts": 1776649791316431, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316424, "dur": 4, + "args": { + "External id": 50234, + "cbid": 211, "correlation": 50234 + } + }, + { + "ph": "s", "id": 50234, "pid": 435, "tid": 435, "ts": 1776649791316424, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791316467, "dur": 1, + "args": { + "External id": 6579, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50252, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50252, "pid": 0, "tid": 7, "ts": 1776649791316467, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316458, "dur": 6, + "args": { + "External id": 6579, + "cbid": 211, "correlation": 50252 + } + }, + { + "ph": "s", "id": 50252, "pid": 435, "tid": 435, "ts": 1776649791316458, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316470, "dur": 0, + "args": { + "External id": 50258, + "cbid": 200, "correlation": 50258 + } + }, + { + "ph": "f", "id": 50258, "pid": 435, "tid": 435, "ts": 1776649791316470, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316471, "dur": 0, + "args": { + "External id": 50259, + "cbid": 273, "correlation": 50259 + } + }, + { + "ph": "f", "id": 50259, "pid": 435, "tid": 435, "ts": 1776649791316471, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791316480, "dur": 15, + "args": { + "External id": 50260, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50260, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50260, "pid": 0, "tid": 7, "ts": 1776649791316480, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316472, "dur": 4, + "args": { + "External id": 50260, + "cbid": 211, "correlation": 50260 + } + }, + { + "ph": "s", "id": 50260, "pid": 435, "tid": 435, "ts": 1776649791316472, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791316494, "dur": 6, + "args": { + "External id": 50263, + "cbid": 165, "correlation": 50263 + } + }, + { + "ph": "s", "id": 50263, "pid": 435, "tid": 435, "ts": 1776649791316494, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791316522, "dur": 3, + "args": { + "External id": 50272, + "cbid": 135, "correlation": 50272 + } + }, + { + "ph": "f", "id": 50272, "pid": 435, "tid": 435, "ts": 1776649791316522, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791316572, "dur": 1, + "args": { + "External id": 6584, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50292, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50292, "pid": 0, "tid": 7, "ts": 1776649791316572, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316561, "dur": 8, + "args": { + "External id": 6584, + "cbid": 211, "correlation": 50292 + } + }, + { + "ph": "s", "id": 50292, "pid": 435, "tid": 435, "ts": 1776649791316561, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316576, "dur": 0, + "args": { + "External id": 50298, + "cbid": 200, "correlation": 50298 + } + }, + { + "ph": "f", "id": 50298, "pid": 435, "tid": 435, "ts": 1776649791316576, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316576, "dur": 0, + "args": { + "External id": 50299, + "cbid": 273, "correlation": 50299 + } + }, + { + "ph": "f", "id": 50299, "pid": 435, "tid": 435, "ts": 1776649791316576, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791316586, "dur": 15, + "args": { + "External id": 50300, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50300, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50300, "pid": 0, "tid": 7, "ts": 1776649791316586, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316578, "dur": 5, + "args": { + "External id": 50300, + "cbid": 211, "correlation": 50300 + } + }, + { + "ph": "s", "id": 50300, "pid": 435, "tid": 435, "ts": 1776649791316578, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791316623, "dur": 1, + "args": { + "External id": 6589, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50318, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50318, "pid": 0, "tid": 7, "ts": 1776649791316623, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316613, "dur": 6, + "args": { + "External id": 6589, + "cbid": 211, "correlation": 50318 + } + }, + { + "ph": "s", "id": 50318, "pid": 435, "tid": 435, "ts": 1776649791316613, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316626, "dur": 0, + "args": { + "External id": 50324, + "cbid": 200, "correlation": 50324 + } + }, + { + "ph": "f", "id": 50324, "pid": 435, "tid": 435, "ts": 1776649791316626, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316627, "dur": 0, + "args": { + "External id": 50325, + "cbid": 273, "correlation": 50325 + } + }, + { + "ph": "f", "id": 50325, "pid": 435, "tid": 435, "ts": 1776649791316627, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791316636, "dur": 15, + "args": { + "External id": 50326, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50326, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50326, "pid": 0, "tid": 7, "ts": 1776649791316636, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316628, "dur": 4, + "args": { + "External id": 50326, + "cbid": 211, "correlation": 50326 + } + }, + { + "ph": "s", "id": 50326, "pid": 435, "tid": 435, "ts": 1776649791316628, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791316671, "dur": 1, + "args": { + "External id": 6594, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50344, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50344, "pid": 0, "tid": 7, "ts": 1776649791316671, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316662, "dur": 6, + "args": { + "External id": 6594, + "cbid": 211, "correlation": 50344 + } + }, + { + "ph": "s", "id": 50344, "pid": 435, "tid": 435, "ts": 1776649791316662, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316675, "dur": 0, + "args": { + "External id": 50350, + "cbid": 200, "correlation": 50350 + } + }, + { + "ph": "f", "id": 50350, "pid": 435, "tid": 435, "ts": 1776649791316675, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316675, "dur": 0, + "args": { + "External id": 50351, + "cbid": 273, "correlation": 50351 + } + }, + { + "ph": "f", "id": 50351, "pid": 435, "tid": 435, "ts": 1776649791316675, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791316685, "dur": 15, + "args": { + "External id": 50352, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50352, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50352, "pid": 0, "tid": 7, "ts": 1776649791316685, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316677, "dur": 5, + "args": { + "External id": 50352, + "cbid": 211, "correlation": 50352 + } + }, + { + "ph": "s", "id": 50352, "pid": 435, "tid": 435, "ts": 1776649791316677, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791316720, "dur": 1, + "args": { + "External id": 6599, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50370, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50370, "pid": 0, "tid": 7, "ts": 1776649791316720, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316711, "dur": 6, + "args": { + "External id": 6599, + "cbid": 211, "correlation": 50370 + } + }, + { + "ph": "s", "id": 50370, "pid": 435, "tid": 435, "ts": 1776649791316711, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316723, "dur": 0, + "args": { + "External id": 50376, + "cbid": 200, "correlation": 50376 + } + }, + { + "ph": "f", "id": 50376, "pid": 435, "tid": 435, "ts": 1776649791316723, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316724, "dur": 0, + "args": { + "External id": 50377, + "cbid": 273, "correlation": 50377 + } + }, + { + "ph": "f", "id": 50377, "pid": 435, "tid": 435, "ts": 1776649791316724, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791316733, "dur": 15, + "args": { + "External id": 50378, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50378, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50378, "pid": 0, "tid": 7, "ts": 1776649791316733, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316725, "dur": 4, + "args": { + "External id": 50378, + "cbid": 211, "correlation": 50378 + } + }, + { + "ph": "s", "id": 50378, "pid": 435, "tid": 435, "ts": 1776649791316725, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791316769, "dur": 1, + "args": { + "External id": 6604, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50396, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50396, "pid": 0, "tid": 7, "ts": 1776649791316769, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316759, "dur": 6, + "args": { + "External id": 6604, + "cbid": 211, "correlation": 50396 + } + }, + { + "ph": "s", "id": 50396, "pid": 435, "tid": 435, "ts": 1776649791316759, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316772, "dur": 0, + "args": { + "External id": 50402, + "cbid": 200, "correlation": 50402 + } + }, + { + "ph": "f", "id": 50402, "pid": 435, "tid": 435, "ts": 1776649791316772, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316772, "dur": 0, + "args": { + "External id": 50403, + "cbid": 273, "correlation": 50403 + } + }, + { + "ph": "f", "id": 50403, "pid": 435, "tid": 435, "ts": 1776649791316772, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791316781, "dur": 16, + "args": { + "External id": 50404, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50404, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50404, "pid": 0, "tid": 7, "ts": 1776649791316781, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316773, "dur": 5, + "args": { + "External id": 50404, + "cbid": 211, "correlation": 50404 + } + }, + { + "ph": "s", "id": 50404, "pid": 435, "tid": 435, "ts": 1776649791316773, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791316817, "dur": 1, + "args": { + "External id": 6609, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50422, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50422, "pid": 0, "tid": 7, "ts": 1776649791316817, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316808, "dur": 6, + "args": { + "External id": 6609, + "cbid": 211, "correlation": 50422 + } + }, + { + "ph": "s", "id": 50422, "pid": 435, "tid": 435, "ts": 1776649791316808, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316820, "dur": 0, + "args": { + "External id": 50428, + "cbid": 200, "correlation": 50428 + } + }, + { + "ph": "f", "id": 50428, "pid": 435, "tid": 435, "ts": 1776649791316820, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316821, "dur": 0, + "args": { + "External id": 50429, + "cbid": 273, "correlation": 50429 + } + }, + { + "ph": "f", "id": 50429, "pid": 435, "tid": 435, "ts": 1776649791316821, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791316830, "dur": 15, + "args": { + "External id": 50430, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50430, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50430, "pid": 0, "tid": 7, "ts": 1776649791316830, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316822, "dur": 5, + "args": { + "External id": 50430, + "cbid": 211, "correlation": 50430 + } + }, + { + "ph": "s", "id": 50430, "pid": 435, "tid": 435, "ts": 1776649791316822, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791316870, "dur": 1, + "args": { + "External id": 6614, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50448, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50448, "pid": 0, "tid": 7, "ts": 1776649791316870, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316861, "dur": 6, + "args": { + "External id": 6614, + "cbid": 211, "correlation": 50448 + } + }, + { + "ph": "s", "id": 50448, "pid": 435, "tid": 435, "ts": 1776649791316861, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316873, "dur": 0, + "args": { + "External id": 50454, + "cbid": 200, "correlation": 50454 + } + }, + { + "ph": "f", "id": 50454, "pid": 435, "tid": 435, "ts": 1776649791316873, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316874, "dur": 0, + "args": { + "External id": 50455, + "cbid": 273, "correlation": 50455 + } + }, + { + "ph": "f", "id": 50455, "pid": 435, "tid": 435, "ts": 1776649791316874, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791316883, "dur": 15, + "args": { + "External id": 50456, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50456, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50456, "pid": 0, "tid": 7, "ts": 1776649791316883, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316875, "dur": 4, + "args": { + "External id": 50456, + "cbid": 211, "correlation": 50456 + } + }, + { + "ph": "s", "id": 50456, "pid": 435, "tid": 435, "ts": 1776649791316875, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791316918, "dur": 1, + "args": { + "External id": 6619, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50474, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50474, "pid": 0, "tid": 7, "ts": 1776649791316918, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316909, "dur": 6, + "args": { + "External id": 6619, + "cbid": 211, "correlation": 50474 + } + }, + { + "ph": "s", "id": 50474, "pid": 435, "tid": 435, "ts": 1776649791316909, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316921, "dur": 0, + "args": { + "External id": 50480, + "cbid": 200, "correlation": 50480 + } + }, + { + "ph": "f", "id": 50480, "pid": 435, "tid": 435, "ts": 1776649791316921, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316922, "dur": 0, + "args": { + "External id": 50481, + "cbid": 273, "correlation": 50481 + } + }, + { + "ph": "f", "id": 50481, "pid": 435, "tid": 435, "ts": 1776649791316922, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791316931, "dur": 15, + "args": { + "External id": 50482, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50482, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50482, "pid": 0, "tid": 7, "ts": 1776649791316931, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316923, "dur": 5, + "args": { + "External id": 50482, + "cbid": 211, "correlation": 50482 + } + }, + { + "ph": "s", "id": 50482, "pid": 435, "tid": 435, "ts": 1776649791316923, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791316966, "dur": 1, + "args": { + "External id": 6624, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50500, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50500, "pid": 0, "tid": 7, "ts": 1776649791316966, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316957, "dur": 6, + "args": { + "External id": 6624, + "cbid": 211, "correlation": 50500 + } + }, + { + "ph": "s", "id": 50500, "pid": 435, "tid": 435, "ts": 1776649791316957, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316969, "dur": 0, + "args": { + "External id": 50506, + "cbid": 200, "correlation": 50506 + } + }, + { + "ph": "f", "id": 50506, "pid": 435, "tid": 435, "ts": 1776649791316969, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791316970, "dur": 0, + "args": { + "External id": 50507, + "cbid": 273, "correlation": 50507 + } + }, + { + "ph": "f", "id": 50507, "pid": 435, "tid": 435, "ts": 1776649791316970, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791316979, "dur": 15, + "args": { + "External id": 50508, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50508, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50508, "pid": 0, "tid": 7, "ts": 1776649791316979, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791316971, "dur": 5, + "args": { + "External id": 50508, + "cbid": 211, "correlation": 50508 + } + }, + { + "ph": "s", "id": 50508, "pid": 435, "tid": 435, "ts": 1776649791316971, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791317015, "dur": 1, + "args": { + "External id": 6629, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50526, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50526, "pid": 0, "tid": 7, "ts": 1776649791317015, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317005, "dur": 6, + "args": { + "External id": 6629, + "cbid": 211, "correlation": 50526 + } + }, + { + "ph": "s", "id": 50526, "pid": 435, "tid": 435, "ts": 1776649791317005, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317018, "dur": 0, + "args": { + "External id": 50532, + "cbid": 200, "correlation": 50532 + } + }, + { + "ph": "f", "id": 50532, "pid": 435, "tid": 435, "ts": 1776649791317018, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317019, "dur": 0, + "args": { + "External id": 50533, + "cbid": 273, "correlation": 50533 + } + }, + { + "ph": "f", "id": 50533, "pid": 435, "tid": 435, "ts": 1776649791317019, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791317028, "dur": 15, + "args": { + "External id": 50534, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50534, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50534, "pid": 0, "tid": 7, "ts": 1776649791317028, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317020, "dur": 4, + "args": { + "External id": 50534, + "cbid": 211, "correlation": 50534 + } + }, + { + "ph": "s", "id": 50534, "pid": 435, "tid": 435, "ts": 1776649791317020, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791317063, "dur": 1, + "args": { + "External id": 6634, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50552, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50552, "pid": 0, "tid": 7, "ts": 1776649791317063, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317054, "dur": 6, + "args": { + "External id": 6634, + "cbid": 211, "correlation": 50552 + } + }, + { + "ph": "s", "id": 50552, "pid": 435, "tid": 435, "ts": 1776649791317054, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317066, "dur": 0, + "args": { + "External id": 50558, + "cbid": 200, "correlation": 50558 + } + }, + { + "ph": "f", "id": 50558, "pid": 435, "tid": 435, "ts": 1776649791317066, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317067, "dur": 0, + "args": { + "External id": 50559, + "cbid": 273, "correlation": 50559 + } + }, + { + "ph": "f", "id": 50559, "pid": 435, "tid": 435, "ts": 1776649791317067, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791317076, "dur": 15, + "args": { + "External id": 50560, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50560, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50560, "pid": 0, "tid": 7, "ts": 1776649791317076, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317068, "dur": 5, + "args": { + "External id": 50560, + "cbid": 211, "correlation": 50560 + } + }, + { + "ph": "s", "id": 50560, "pid": 435, "tid": 435, "ts": 1776649791317068, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791317111, "dur": 1, + "args": { + "External id": 6639, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50578, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50578, "pid": 0, "tid": 7, "ts": 1776649791317111, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317102, "dur": 6, + "args": { + "External id": 6639, + "cbid": 211, "correlation": 50578 + } + }, + { + "ph": "s", "id": 50578, "pid": 435, "tid": 435, "ts": 1776649791317102, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317114, "dur": 0, + "args": { + "External id": 50584, + "cbid": 200, "correlation": 50584 + } + }, + { + "ph": "f", "id": 50584, "pid": 435, "tid": 435, "ts": 1776649791317114, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317115, "dur": 0, + "args": { + "External id": 50585, + "cbid": 273, "correlation": 50585 + } + }, + { + "ph": "f", "id": 50585, "pid": 435, "tid": 435, "ts": 1776649791317115, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791317124, "dur": 15, + "args": { + "External id": 50586, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50586, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50586, "pid": 0, "tid": 7, "ts": 1776649791317124, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317116, "dur": 4, + "args": { + "External id": 50586, + "cbid": 211, "correlation": 50586 + } + }, + { + "ph": "s", "id": 50586, "pid": 435, "tid": 435, "ts": 1776649791317116, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791317160, "dur": 1, + "args": { + "External id": 6644, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50604, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50604, "pid": 0, "tid": 7, "ts": 1776649791317160, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317150, "dur": 6, + "args": { + "External id": 6644, + "cbid": 211, "correlation": 50604 + } + }, + { + "ph": "s", "id": 50604, "pid": 435, "tid": 435, "ts": 1776649791317150, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317163, "dur": 0, + "args": { + "External id": 50610, + "cbid": 200, "correlation": 50610 + } + }, + { + "ph": "f", "id": 50610, "pid": 435, "tid": 435, "ts": 1776649791317163, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317163, "dur": 0, + "args": { + "External id": 50611, + "cbid": 273, "correlation": 50611 + } + }, + { + "ph": "f", "id": 50611, "pid": 435, "tid": 435, "ts": 1776649791317163, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791317172, "dur": 16, + "args": { + "External id": 50612, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50612, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50612, "pid": 0, "tid": 7, "ts": 1776649791317172, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317164, "dur": 4, + "args": { + "External id": 50612, + "cbid": 211, "correlation": 50612 + } + }, + { + "ph": "s", "id": 50612, "pid": 435, "tid": 435, "ts": 1776649791317164, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791317208, "dur": 1, + "args": { + "External id": 6649, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50630, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50630, "pid": 0, "tid": 7, "ts": 1776649791317208, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317199, "dur": 6, + "args": { + "External id": 6649, + "cbid": 211, "correlation": 50630 + } + }, + { + "ph": "s", "id": 50630, "pid": 435, "tid": 435, "ts": 1776649791317199, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317211, "dur": 0, + "args": { + "External id": 50636, + "cbid": 200, "correlation": 50636 + } + }, + { + "ph": "f", "id": 50636, "pid": 435, "tid": 435, "ts": 1776649791317211, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317211, "dur": 0, + "args": { + "External id": 50637, + "cbid": 273, "correlation": 50637 + } + }, + { + "ph": "f", "id": 50637, "pid": 435, "tid": 435, "ts": 1776649791317211, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791317220, "dur": 16, + "args": { + "External id": 50638, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50638, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50638, "pid": 0, "tid": 7, "ts": 1776649791317220, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317213, "dur": 4, + "args": { + "External id": 50638, + "cbid": 211, "correlation": 50638 + } + }, + { + "ph": "s", "id": 50638, "pid": 435, "tid": 435, "ts": 1776649791317213, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791317256, "dur": 1, + "args": { + "External id": 6654, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50656, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50656, "pid": 0, "tid": 7, "ts": 1776649791317256, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317247, "dur": 6, + "args": { + "External id": 6654, + "cbid": 211, "correlation": 50656 + } + }, + { + "ph": "s", "id": 50656, "pid": 435, "tid": 435, "ts": 1776649791317247, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317259, "dur": 0, + "args": { + "External id": 50662, + "cbid": 200, "correlation": 50662 + } + }, + { + "ph": "f", "id": 50662, "pid": 435, "tid": 435, "ts": 1776649791317259, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317260, "dur": 0, + "args": { + "External id": 50663, + "cbid": 273, "correlation": 50663 + } + }, + { + "ph": "f", "id": 50663, "pid": 435, "tid": 435, "ts": 1776649791317260, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791317269, "dur": 15, + "args": { + "External id": 50664, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50664, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50664, "pid": 0, "tid": 7, "ts": 1776649791317269, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317261, "dur": 4, + "args": { + "External id": 50664, + "cbid": 211, "correlation": 50664 + } + }, + { + "ph": "s", "id": 50664, "pid": 435, "tid": 435, "ts": 1776649791317261, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791317313, "dur": 1, + "args": { + "External id": 6659, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50682, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50682, "pid": 0, "tid": 7, "ts": 1776649791317313, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317304, "dur": 6, + "args": { + "External id": 6659, + "cbid": 211, "correlation": 50682 + } + }, + { + "ph": "s", "id": 50682, "pid": 435, "tid": 435, "ts": 1776649791317304, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317316, "dur": 0, + "args": { + "External id": 50688, + "cbid": 200, "correlation": 50688 + } + }, + { + "ph": "f", "id": 50688, "pid": 435, "tid": 435, "ts": 1776649791317316, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317317, "dur": 0, + "args": { + "External id": 50689, + "cbid": 273, "correlation": 50689 + } + }, + { + "ph": "f", "id": 50689, "pid": 435, "tid": 435, "ts": 1776649791317317, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791317326, "dur": 15, + "args": { + "External id": 50690, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50690, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50690, "pid": 0, "tid": 7, "ts": 1776649791317326, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317318, "dur": 4, + "args": { + "External id": 50690, + "cbid": 211, "correlation": 50690 + } + }, + { + "ph": "s", "id": 50690, "pid": 435, "tid": 435, "ts": 1776649791317318, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791317362, "dur": 1, + "args": { + "External id": 6664, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50708, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50708, "pid": 0, "tid": 7, "ts": 1776649791317362, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317352, "dur": 6, + "args": { + "External id": 6664, + "cbid": 211, "correlation": 50708 + } + }, + { + "ph": "s", "id": 50708, "pid": 435, "tid": 435, "ts": 1776649791317352, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317365, "dur": 0, + "args": { + "External id": 50714, + "cbid": 200, "correlation": 50714 + } + }, + { + "ph": "f", "id": 50714, "pid": 435, "tid": 435, "ts": 1776649791317365, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317365, "dur": 0, + "args": { + "External id": 50715, + "cbid": 273, "correlation": 50715 + } + }, + { + "ph": "f", "id": 50715, "pid": 435, "tid": 435, "ts": 1776649791317365, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791317374, "dur": 15, + "args": { + "External id": 50716, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50716, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50716, "pid": 0, "tid": 7, "ts": 1776649791317374, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317366, "dur": 4, + "args": { + "External id": 50716, + "cbid": 211, "correlation": 50716 + } + }, + { + "ph": "s", "id": 50716, "pid": 435, "tid": 435, "ts": 1776649791317366, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791317410, "dur": 1, + "args": { + "External id": 6669, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50734, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50734, "pid": 0, "tid": 7, "ts": 1776649791317410, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317401, "dur": 6, + "args": { + "External id": 6669, + "cbid": 211, "correlation": 50734 + } + }, + { + "ph": "s", "id": 50734, "pid": 435, "tid": 435, "ts": 1776649791317401, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317413, "dur": 0, + "args": { + "External id": 50740, + "cbid": 200, "correlation": 50740 + } + }, + { + "ph": "f", "id": 50740, "pid": 435, "tid": 435, "ts": 1776649791317413, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317414, "dur": 0, + "args": { + "External id": 50741, + "cbid": 273, "correlation": 50741 + } + }, + { + "ph": "f", "id": 50741, "pid": 435, "tid": 435, "ts": 1776649791317414, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791317423, "dur": 15, + "args": { + "External id": 50742, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50742, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50742, "pid": 0, "tid": 7, "ts": 1776649791317423, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317415, "dur": 4, + "args": { + "External id": 50742, + "cbid": 211, "correlation": 50742 + } + }, + { + "ph": "s", "id": 50742, "pid": 435, "tid": 435, "ts": 1776649791317415, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791317459, "dur": 1, + "args": { + "External id": 6674, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50760, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50760, "pid": 0, "tid": 7, "ts": 1776649791317459, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317450, "dur": 6, + "args": { + "External id": 6674, + "cbid": 211, "correlation": 50760 + } + }, + { + "ph": "s", "id": 50760, "pid": 435, "tid": 435, "ts": 1776649791317450, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317462, "dur": 0, + "args": { + "External id": 50766, + "cbid": 200, "correlation": 50766 + } + }, + { + "ph": "f", "id": 50766, "pid": 435, "tid": 435, "ts": 1776649791317462, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317462, "dur": 0, + "args": { + "External id": 50767, + "cbid": 273, "correlation": 50767 + } + }, + { + "ph": "f", "id": 50767, "pid": 435, "tid": 435, "ts": 1776649791317462, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791317472, "dur": 15, + "args": { + "External id": 50768, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50768, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50768, "pid": 0, "tid": 7, "ts": 1776649791317472, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317463, "dur": 4, + "args": { + "External id": 50768, + "cbid": 211, "correlation": 50768 + } + }, + { + "ph": "s", "id": 50768, "pid": 435, "tid": 435, "ts": 1776649791317463, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791317508, "dur": 1, + "args": { + "External id": 6679, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50786, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50786, "pid": 0, "tid": 7, "ts": 1776649791317508, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317499, "dur": 6, + "args": { + "External id": 6679, + "cbid": 211, "correlation": 50786 + } + }, + { + "ph": "s", "id": 50786, "pid": 435, "tid": 435, "ts": 1776649791317499, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317511, "dur": 0, + "args": { + "External id": 50792, + "cbid": 200, "correlation": 50792 + } + }, + { + "ph": "f", "id": 50792, "pid": 435, "tid": 435, "ts": 1776649791317511, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317512, "dur": 0, + "args": { + "External id": 50793, + "cbid": 273, "correlation": 50793 + } + }, + { + "ph": "f", "id": 50793, "pid": 435, "tid": 435, "ts": 1776649791317512, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791317521, "dur": 15, + "args": { + "External id": 50794, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50794, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50794, "pid": 0, "tid": 7, "ts": 1776649791317521, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317513, "dur": 4, + "args": { + "External id": 50794, + "cbid": 211, "correlation": 50794 + } + }, + { + "ph": "s", "id": 50794, "pid": 435, "tid": 435, "ts": 1776649791317513, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791317556, "dur": 1, + "args": { + "External id": 6684, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50812, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50812, "pid": 0, "tid": 7, "ts": 1776649791317556, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317547, "dur": 6, + "args": { + "External id": 6684, + "cbid": 211, "correlation": 50812 + } + }, + { + "ph": "s", "id": 50812, "pid": 435, "tid": 435, "ts": 1776649791317547, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317559, "dur": 0, + "args": { + "External id": 50818, + "cbid": 200, "correlation": 50818 + } + }, + { + "ph": "f", "id": 50818, "pid": 435, "tid": 435, "ts": 1776649791317559, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317560, "dur": 0, + "args": { + "External id": 50819, + "cbid": 273, "correlation": 50819 + } + }, + { + "ph": "f", "id": 50819, "pid": 435, "tid": 435, "ts": 1776649791317560, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791317569, "dur": 15, + "args": { + "External id": 50820, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50820, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50820, "pid": 0, "tid": 7, "ts": 1776649791317569, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317561, "dur": 5, + "args": { + "External id": 50820, + "cbid": 211, "correlation": 50820 + } + }, + { + "ph": "s", "id": 50820, "pid": 435, "tid": 435, "ts": 1776649791317561, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791317604, "dur": 1, + "args": { + "External id": 6689, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50838, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50838, "pid": 0, "tid": 7, "ts": 1776649791317604, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317595, "dur": 6, + "args": { + "External id": 6689, + "cbid": 211, "correlation": 50838 + } + }, + { + "ph": "s", "id": 50838, "pid": 435, "tid": 435, "ts": 1776649791317595, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317608, "dur": 0, + "args": { + "External id": 50844, + "cbid": 200, "correlation": 50844 + } + }, + { + "ph": "f", "id": 50844, "pid": 435, "tid": 435, "ts": 1776649791317608, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317608, "dur": 0, + "args": { + "External id": 50845, + "cbid": 273, "correlation": 50845 + } + }, + { + "ph": "f", "id": 50845, "pid": 435, "tid": 435, "ts": 1776649791317608, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791317617, "dur": 16, + "args": { + "External id": 50846, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50846, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50846, "pid": 0, "tid": 7, "ts": 1776649791317617, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317609, "dur": 4, + "args": { + "External id": 50846, + "cbid": 211, "correlation": 50846 + } + }, + { + "ph": "s", "id": 50846, "pid": 435, "tid": 435, "ts": 1776649791317609, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791317653, "dur": 1, + "args": { + "External id": 6694, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50864, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50864, "pid": 0, "tid": 7, "ts": 1776649791317653, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317643, "dur": 6, + "args": { + "External id": 6694, + "cbid": 211, "correlation": 50864 + } + }, + { + "ph": "s", "id": 50864, "pid": 435, "tid": 435, "ts": 1776649791317643, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317656, "dur": 0, + "args": { + "External id": 50870, + "cbid": 200, "correlation": 50870 + } + }, + { + "ph": "f", "id": 50870, "pid": 435, "tid": 435, "ts": 1776649791317656, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317657, "dur": 0, + "args": { + "External id": 50871, + "cbid": 273, "correlation": 50871 + } + }, + { + "ph": "f", "id": 50871, "pid": 435, "tid": 435, "ts": 1776649791317657, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791317666, "dur": 15, + "args": { + "External id": 50872, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50872, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50872, "pid": 0, "tid": 7, "ts": 1776649791317666, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317658, "dur": 5, + "args": { + "External id": 50872, + "cbid": 211, "correlation": 50872 + } + }, + { + "ph": "s", "id": 50872, "pid": 435, "tid": 435, "ts": 1776649791317658, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791317701, "dur": 1, + "args": { + "External id": 6699, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50890, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50890, "pid": 0, "tid": 7, "ts": 1776649791317701, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317692, "dur": 6, + "args": { + "External id": 6699, + "cbid": 211, "correlation": 50890 + } + }, + { + "ph": "s", "id": 50890, "pid": 435, "tid": 435, "ts": 1776649791317692, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317704, "dur": 0, + "args": { + "External id": 50896, + "cbid": 200, "correlation": 50896 + } + }, + { + "ph": "f", "id": 50896, "pid": 435, "tid": 435, "ts": 1776649791317704, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317705, "dur": 0, + "args": { + "External id": 50897, + "cbid": 273, "correlation": 50897 + } + }, + { + "ph": "f", "id": 50897, "pid": 435, "tid": 435, "ts": 1776649791317705, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791317714, "dur": 15, + "args": { + "External id": 50898, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50898, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50898, "pid": 0, "tid": 7, "ts": 1776649791317714, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317706, "dur": 4, + "args": { + "External id": 50898, + "cbid": 211, "correlation": 50898 + } + }, + { + "ph": "s", "id": 50898, "pid": 435, "tid": 435, "ts": 1776649791317706, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791317749, "dur": 1, + "args": { + "External id": 6704, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50916, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50916, "pid": 0, "tid": 7, "ts": 1776649791317749, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317740, "dur": 7, + "args": { + "External id": 6704, + "cbid": 211, "correlation": 50916 + } + }, + { + "ph": "s", "id": 50916, "pid": 435, "tid": 435, "ts": 1776649791317740, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317753, "dur": 0, + "args": { + "External id": 50922, + "cbid": 200, "correlation": 50922 + } + }, + { + "ph": "f", "id": 50922, "pid": 435, "tid": 435, "ts": 1776649791317753, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317753, "dur": 0, + "args": { + "External id": 50923, + "cbid": 273, "correlation": 50923 + } + }, + { + "ph": "f", "id": 50923, "pid": 435, "tid": 435, "ts": 1776649791317753, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791317763, "dur": 15, + "args": { + "External id": 50924, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50924, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50924, "pid": 0, "tid": 7, "ts": 1776649791317763, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317755, "dur": 5, + "args": { + "External id": 50924, + "cbid": 211, "correlation": 50924 + } + }, + { + "ph": "s", "id": 50924, "pid": 435, "tid": 435, "ts": 1776649791317755, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791317798, "dur": 1, + "args": { + "External id": 6709, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50942, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50942, "pid": 0, "tid": 7, "ts": 1776649791317798, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317788, "dur": 6, + "args": { + "External id": 6709, + "cbid": 211, "correlation": 50942 + } + }, + { + "ph": "s", "id": 50942, "pid": 435, "tid": 435, "ts": 1776649791317788, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317801, "dur": 0, + "args": { + "External id": 50948, + "cbid": 200, "correlation": 50948 + } + }, + { + "ph": "f", "id": 50948, "pid": 435, "tid": 435, "ts": 1776649791317801, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317801, "dur": 0, + "args": { + "External id": 50949, + "cbid": 273, "correlation": 50949 + } + }, + { + "ph": "f", "id": 50949, "pid": 435, "tid": 435, "ts": 1776649791317801, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791317810, "dur": 16, + "args": { + "External id": 50950, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50950, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50950, "pid": 0, "tid": 7, "ts": 1776649791317810, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317802, "dur": 4, + "args": { + "External id": 50950, + "cbid": 211, "correlation": 50950 + } + }, + { + "ph": "s", "id": 50950, "pid": 435, "tid": 435, "ts": 1776649791317802, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791317845, "dur": 1, + "args": { + "External id": 6714, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50968, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50968, "pid": 0, "tid": 7, "ts": 1776649791317845, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317836, "dur": 6, + "args": { + "External id": 6714, + "cbid": 211, "correlation": 50968 + } + }, + { + "ph": "s", "id": 50968, "pid": 435, "tid": 435, "ts": 1776649791317836, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317849, "dur": 0, + "args": { + "External id": 50974, + "cbid": 200, "correlation": 50974 + } + }, + { + "ph": "f", "id": 50974, "pid": 435, "tid": 435, "ts": 1776649791317849, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317849, "dur": 0, + "args": { + "External id": 50975, + "cbid": 273, "correlation": 50975 + } + }, + { + "ph": "f", "id": 50975, "pid": 435, "tid": 435, "ts": 1776649791317849, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791317859, "dur": 15, + "args": { + "External id": 50976, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50976, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50976, "pid": 0, "tid": 7, "ts": 1776649791317859, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317851, "dur": 5, + "args": { + "External id": 50976, + "cbid": 211, "correlation": 50976 + } + }, + { + "ph": "s", "id": 50976, "pid": 435, "tid": 435, "ts": 1776649791317851, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791317894, "dur": 1, + "args": { + "External id": 6719, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 50994, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 50994, "pid": 0, "tid": 7, "ts": 1776649791317894, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317885, "dur": 6, + "args": { + "External id": 6719, + "cbid": 211, "correlation": 50994 + } + }, + { + "ph": "s", "id": 50994, "pid": 435, "tid": 435, "ts": 1776649791317885, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317897, "dur": 0, + "args": { + "External id": 51000, + "cbid": 200, "correlation": 51000 + } + }, + { + "ph": "f", "id": 51000, "pid": 435, "tid": 435, "ts": 1776649791317897, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317898, "dur": 0, + "args": { + "External id": 51001, + "cbid": 273, "correlation": 51001 + } + }, + { + "ph": "f", "id": 51001, "pid": 435, "tid": 435, "ts": 1776649791317898, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791317907, "dur": 15, + "args": { + "External id": 51002, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51002, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51002, "pid": 0, "tid": 7, "ts": 1776649791317907, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317899, "dur": 4, + "args": { + "External id": 51002, + "cbid": 211, "correlation": 51002 + } + }, + { + "ph": "s", "id": 51002, "pid": 435, "tid": 435, "ts": 1776649791317899, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791317943, "dur": 1, + "args": { + "External id": 6724, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51020, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51020, "pid": 0, "tid": 7, "ts": 1776649791317943, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317933, "dur": 6, + "args": { + "External id": 6724, + "cbid": 211, "correlation": 51020 + } + }, + { + "ph": "s", "id": 51020, "pid": 435, "tid": 435, "ts": 1776649791317933, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317946, "dur": 0, + "args": { + "External id": 51026, + "cbid": 200, "correlation": 51026 + } + }, + { + "ph": "f", "id": 51026, "pid": 435, "tid": 435, "ts": 1776649791317946, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317947, "dur": 0, + "args": { + "External id": 51027, + "cbid": 273, "correlation": 51027 + } + }, + { + "ph": "f", "id": 51027, "pid": 435, "tid": 435, "ts": 1776649791317947, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791317956, "dur": 15, + "args": { + "External id": 51028, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51028, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51028, "pid": 0, "tid": 7, "ts": 1776649791317956, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317948, "dur": 5, + "args": { + "External id": 51028, + "cbid": 211, "correlation": 51028 + } + }, + { + "ph": "s", "id": 51028, "pid": 435, "tid": 435, "ts": 1776649791317948, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791317991, "dur": 1, + "args": { + "External id": 6729, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51046, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51046, "pid": 0, "tid": 7, "ts": 1776649791317991, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317982, "dur": 6, + "args": { + "External id": 6729, + "cbid": 211, "correlation": 51046 + } + }, + { + "ph": "s", "id": 51046, "pid": 435, "tid": 435, "ts": 1776649791317982, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317994, "dur": 0, + "args": { + "External id": 51052, + "cbid": 200, "correlation": 51052 + } + }, + { + "ph": "f", "id": 51052, "pid": 435, "tid": 435, "ts": 1776649791317994, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791317995, "dur": 0, + "args": { + "External id": 51053, + "cbid": 273, "correlation": 51053 + } + }, + { + "ph": "f", "id": 51053, "pid": 435, "tid": 435, "ts": 1776649791317995, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791318004, "dur": 16, + "args": { + "External id": 51054, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51054, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51054, "pid": 0, "tid": 7, "ts": 1776649791318004, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791317996, "dur": 5, + "args": { + "External id": 51054, + "cbid": 211, "correlation": 51054 + } + }, + { + "ph": "s", "id": 51054, "pid": 435, "tid": 435, "ts": 1776649791317996, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791318039, "dur": 1, + "args": { + "External id": 6734, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51072, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51072, "pid": 0, "tid": 7, "ts": 1776649791318039, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318030, "dur": 6, + "args": { + "External id": 6734, + "cbid": 211, "correlation": 51072 + } + }, + { + "ph": "s", "id": 51072, "pid": 435, "tid": 435, "ts": 1776649791318030, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318042, "dur": 0, + "args": { + "External id": 51078, + "cbid": 200, "correlation": 51078 + } + }, + { + "ph": "f", "id": 51078, "pid": 435, "tid": 435, "ts": 1776649791318042, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318043, "dur": 0, + "args": { + "External id": 51079, + "cbid": 273, "correlation": 51079 + } + }, + { + "ph": "f", "id": 51079, "pid": 435, "tid": 435, "ts": 1776649791318043, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791318052, "dur": 16, + "args": { + "External id": 51080, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51080, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51080, "pid": 0, "tid": 7, "ts": 1776649791318052, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318044, "dur": 4, + "args": { + "External id": 51080, + "cbid": 211, "correlation": 51080 + } + }, + { + "ph": "s", "id": 51080, "pid": 435, "tid": 435, "ts": 1776649791318044, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791318088, "dur": 1, + "args": { + "External id": 6739, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51098, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51098, "pid": 0, "tid": 7, "ts": 1776649791318088, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318078, "dur": 6, + "args": { + "External id": 6739, + "cbid": 211, "correlation": 51098 + } + }, + { + "ph": "s", "id": 51098, "pid": 435, "tid": 435, "ts": 1776649791318078, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318090, "dur": 0, + "args": { + "External id": 51104, + "cbid": 200, "correlation": 51104 + } + }, + { + "ph": "f", "id": 51104, "pid": 435, "tid": 435, "ts": 1776649791318090, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318091, "dur": 0, + "args": { + "External id": 51105, + "cbid": 273, "correlation": 51105 + } + }, + { + "ph": "f", "id": 51105, "pid": 435, "tid": 435, "ts": 1776649791318091, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791318100, "dur": 15, + "args": { + "External id": 51106, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51106, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51106, "pid": 0, "tid": 7, "ts": 1776649791318100, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318092, "dur": 4, + "args": { + "External id": 51106, + "cbid": 211, "correlation": 51106 + } + }, + { + "ph": "s", "id": 51106, "pid": 435, "tid": 435, "ts": 1776649791318092, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791318135, "dur": 1, + "args": { + "External id": 6744, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51124, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51124, "pid": 0, "tid": 7, "ts": 1776649791318135, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318126, "dur": 6, + "args": { + "External id": 6744, + "cbid": 211, "correlation": 51124 + } + }, + { + "ph": "s", "id": 51124, "pid": 435, "tid": 435, "ts": 1776649791318126, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318138, "dur": 0, + "args": { + "External id": 51130, + "cbid": 200, "correlation": 51130 + } + }, + { + "ph": "f", "id": 51130, "pid": 435, "tid": 435, "ts": 1776649791318138, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318139, "dur": 0, + "args": { + "External id": 51131, + "cbid": 273, "correlation": 51131 + } + }, + { + "ph": "f", "id": 51131, "pid": 435, "tid": 435, "ts": 1776649791318139, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791318148, "dur": 15, + "args": { + "External id": 51132, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51132, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51132, "pid": 0, "tid": 7, "ts": 1776649791318148, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318140, "dur": 4, + "args": { + "External id": 51132, + "cbid": 211, "correlation": 51132 + } + }, + { + "ph": "s", "id": 51132, "pid": 435, "tid": 435, "ts": 1776649791318140, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791318183, "dur": 1, + "args": { + "External id": 6749, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51150, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51150, "pid": 0, "tid": 7, "ts": 1776649791318183, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318174, "dur": 6, + "args": { + "External id": 6749, + "cbid": 211, "correlation": 51150 + } + }, + { + "ph": "s", "id": 51150, "pid": 435, "tid": 435, "ts": 1776649791318174, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318186, "dur": 0, + "args": { + "External id": 51156, + "cbid": 200, "correlation": 51156 + } + }, + { + "ph": "f", "id": 51156, "pid": 435, "tid": 435, "ts": 1776649791318186, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318187, "dur": 0, + "args": { + "External id": 51157, + "cbid": 273, "correlation": 51157 + } + }, + { + "ph": "f", "id": 51157, "pid": 435, "tid": 435, "ts": 1776649791318187, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791318196, "dur": 15, + "args": { + "External id": 51158, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51158, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51158, "pid": 0, "tid": 7, "ts": 1776649791318196, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318188, "dur": 4, + "args": { + "External id": 51158, + "cbid": 211, "correlation": 51158 + } + }, + { + "ph": "s", "id": 51158, "pid": 435, "tid": 435, "ts": 1776649791318188, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791318231, "dur": 1, + "args": { + "External id": 6754, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51176, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51176, "pid": 0, "tid": 7, "ts": 1776649791318231, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318222, "dur": 6, + "args": { + "External id": 6754, + "cbid": 211, "correlation": 51176 + } + }, + { + "ph": "s", "id": 51176, "pid": 435, "tid": 435, "ts": 1776649791318222, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318235, "dur": 0, + "args": { + "External id": 51182, + "cbid": 200, "correlation": 51182 + } + }, + { + "ph": "f", "id": 51182, "pid": 435, "tid": 435, "ts": 1776649791318235, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318235, "dur": 0, + "args": { + "External id": 51183, + "cbid": 273, "correlation": 51183 + } + }, + { + "ph": "f", "id": 51183, "pid": 435, "tid": 435, "ts": 1776649791318235, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791318244, "dur": 15, + "args": { + "External id": 51184, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51184, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51184, "pid": 0, "tid": 7, "ts": 1776649791318244, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318237, "dur": 4, + "args": { + "External id": 51184, + "cbid": 211, "correlation": 51184 + } + }, + { + "ph": "s", "id": 51184, "pid": 435, "tid": 435, "ts": 1776649791318237, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791318280, "dur": 1, + "args": { + "External id": 6759, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51202, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51202, "pid": 0, "tid": 7, "ts": 1776649791318280, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318271, "dur": 6, + "args": { + "External id": 6759, + "cbid": 211, "correlation": 51202 + } + }, + { + "ph": "s", "id": 51202, "pid": 435, "tid": 435, "ts": 1776649791318271, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318283, "dur": 0, + "args": { + "External id": 51208, + "cbid": 200, "correlation": 51208 + } + }, + { + "ph": "f", "id": 51208, "pid": 435, "tid": 435, "ts": 1776649791318283, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318284, "dur": 0, + "args": { + "External id": 51209, + "cbid": 273, "correlation": 51209 + } + }, + { + "ph": "f", "id": 51209, "pid": 435, "tid": 435, "ts": 1776649791318284, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791318293, "dur": 15, + "args": { + "External id": 51210, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51210, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51210, "pid": 0, "tid": 7, "ts": 1776649791318293, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318285, "dur": 5, + "args": { + "External id": 51210, + "cbid": 211, "correlation": 51210 + } + }, + { + "ph": "s", "id": 51210, "pid": 435, "tid": 435, "ts": 1776649791318285, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791318328, "dur": 1, + "args": { + "External id": 6764, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51228, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51228, "pid": 0, "tid": 7, "ts": 1776649791318328, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318319, "dur": 6, + "args": { + "External id": 6764, + "cbid": 211, "correlation": 51228 + } + }, + { + "ph": "s", "id": 51228, "pid": 435, "tid": 435, "ts": 1776649791318319, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318331, "dur": 0, + "args": { + "External id": 51234, + "cbid": 200, "correlation": 51234 + } + }, + { + "ph": "f", "id": 51234, "pid": 435, "tid": 435, "ts": 1776649791318331, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318332, "dur": 0, + "args": { + "External id": 51235, + "cbid": 273, "correlation": 51235 + } + }, + { + "ph": "f", "id": 51235, "pid": 435, "tid": 435, "ts": 1776649791318332, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791318341, "dur": 15, + "args": { + "External id": 51236, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51236, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51236, "pid": 0, "tid": 7, "ts": 1776649791318341, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318333, "dur": 4, + "args": { + "External id": 51236, + "cbid": 211, "correlation": 51236 + } + }, + { + "ph": "s", "id": 51236, "pid": 435, "tid": 435, "ts": 1776649791318333, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791318376, "dur": 1, + "args": { + "External id": 6769, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51254, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51254, "pid": 0, "tid": 7, "ts": 1776649791318376, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318367, "dur": 6, + "args": { + "External id": 6769, + "cbid": 211, "correlation": 51254 + } + }, + { + "ph": "s", "id": 51254, "pid": 435, "tid": 435, "ts": 1776649791318367, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318380, "dur": 0, + "args": { + "External id": 51260, + "cbid": 200, "correlation": 51260 + } + }, + { + "ph": "f", "id": 51260, "pid": 435, "tid": 435, "ts": 1776649791318380, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318381, "dur": 0, + "args": { + "External id": 51261, + "cbid": 273, "correlation": 51261 + } + }, + { + "ph": "f", "id": 51261, "pid": 435, "tid": 435, "ts": 1776649791318381, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791318390, "dur": 15, + "args": { + "External id": 51262, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51262, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51262, "pid": 0, "tid": 7, "ts": 1776649791318390, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318382, "dur": 4, + "args": { + "External id": 51262, + "cbid": 211, "correlation": 51262 + } + }, + { + "ph": "s", "id": 51262, "pid": 435, "tid": 435, "ts": 1776649791318382, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791318425, "dur": 1, + "args": { + "External id": 6774, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51280, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51280, "pid": 0, "tid": 7, "ts": 1776649791318425, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318416, "dur": 6, + "args": { + "External id": 6774, + "cbid": 211, "correlation": 51280 + } + }, + { + "ph": "s", "id": 51280, "pid": 435, "tid": 435, "ts": 1776649791318416, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318428, "dur": 0, + "args": { + "External id": 51286, + "cbid": 200, "correlation": 51286 + } + }, + { + "ph": "f", "id": 51286, "pid": 435, "tid": 435, "ts": 1776649791318428, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318429, "dur": 0, + "args": { + "External id": 51287, + "cbid": 273, "correlation": 51287 + } + }, + { + "ph": "f", "id": 51287, "pid": 435, "tid": 435, "ts": 1776649791318429, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791318438, "dur": 16, + "args": { + "External id": 51288, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51288, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51288, "pid": 0, "tid": 7, "ts": 1776649791318438, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318430, "dur": 4, + "args": { + "External id": 51288, + "cbid": 211, "correlation": 51288 + } + }, + { + "ph": "s", "id": 51288, "pid": 435, "tid": 435, "ts": 1776649791318430, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791318474, "dur": 1, + "args": { + "External id": 6779, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51306, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51306, "pid": 0, "tid": 7, "ts": 1776649791318474, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318464, "dur": 6, + "args": { + "External id": 6779, + "cbid": 211, "correlation": 51306 + } + }, + { + "ph": "s", "id": 51306, "pid": 435, "tid": 435, "ts": 1776649791318464, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318477, "dur": 0, + "args": { + "External id": 51312, + "cbid": 200, "correlation": 51312 + } + }, + { + "ph": "f", "id": 51312, "pid": 435, "tid": 435, "ts": 1776649791318477, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318477, "dur": 0, + "args": { + "External id": 51313, + "cbid": 273, "correlation": 51313 + } + }, + { + "ph": "f", "id": 51313, "pid": 435, "tid": 435, "ts": 1776649791318477, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791318486, "dur": 17, + "args": { + "External id": 51314, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51314, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51314, "pid": 0, "tid": 7, "ts": 1776649791318486, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318478, "dur": 4, + "args": { + "External id": 51314, + "cbid": 211, "correlation": 51314 + } + }, + { + "ph": "s", "id": 51314, "pid": 435, "tid": 435, "ts": 1776649791318478, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791318522, "dur": 1, + "args": { + "External id": 6784, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51332, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51332, "pid": 0, "tid": 7, "ts": 1776649791318522, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318512, "dur": 6, + "args": { + "External id": 6784, + "cbid": 211, "correlation": 51332 + } + }, + { + "ph": "s", "id": 51332, "pid": 435, "tid": 435, "ts": 1776649791318512, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318525, "dur": 0, + "args": { + "External id": 51338, + "cbid": 200, "correlation": 51338 + } + }, + { + "ph": "f", "id": 51338, "pid": 435, "tid": 435, "ts": 1776649791318525, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318526, "dur": 0, + "args": { + "External id": 51339, + "cbid": 273, "correlation": 51339 + } + }, + { + "ph": "f", "id": 51339, "pid": 435, "tid": 435, "ts": 1776649791318526, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791318535, "dur": 15, + "args": { + "External id": 51340, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51340, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51340, "pid": 0, "tid": 7, "ts": 1776649791318535, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318527, "dur": 5, + "args": { + "External id": 51340, + "cbid": 211, "correlation": 51340 + } + }, + { + "ph": "s", "id": 51340, "pid": 435, "tid": 435, "ts": 1776649791318527, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791318570, "dur": 1, + "args": { + "External id": 6789, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51358, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51358, "pid": 0, "tid": 7, "ts": 1776649791318570, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318561, "dur": 6, + "args": { + "External id": 6789, + "cbid": 211, "correlation": 51358 + } + }, + { + "ph": "s", "id": 51358, "pid": 435, "tid": 435, "ts": 1776649791318561, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318574, "dur": 0, + "args": { + "External id": 51364, + "cbid": 200, "correlation": 51364 + } + }, + { + "ph": "f", "id": 51364, "pid": 435, "tid": 435, "ts": 1776649791318574, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318574, "dur": 0, + "args": { + "External id": 51365, + "cbid": 273, "correlation": 51365 + } + }, + { + "ph": "f", "id": 51365, "pid": 435, "tid": 435, "ts": 1776649791318574, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791318583, "dur": 15, + "args": { + "External id": 51366, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51366, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51366, "pid": 0, "tid": 7, "ts": 1776649791318583, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318575, "dur": 5, + "args": { + "External id": 51366, + "cbid": 211, "correlation": 51366 + } + }, + { + "ph": "s", "id": 51366, "pid": 435, "tid": 435, "ts": 1776649791318575, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791318619, "dur": 1, + "args": { + "External id": 6794, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51384, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51384, "pid": 0, "tid": 7, "ts": 1776649791318619, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318609, "dur": 6, + "args": { + "External id": 6794, + "cbid": 211, "correlation": 51384 + } + }, + { + "ph": "s", "id": 51384, "pid": 435, "tid": 435, "ts": 1776649791318609, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318622, "dur": 0, + "args": { + "External id": 51390, + "cbid": 200, "correlation": 51390 + } + }, + { + "ph": "f", "id": 51390, "pid": 435, "tid": 435, "ts": 1776649791318622, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318622, "dur": 0, + "args": { + "External id": 51391, + "cbid": 273, "correlation": 51391 + } + }, + { + "ph": "f", "id": 51391, "pid": 435, "tid": 435, "ts": 1776649791318622, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791318632, "dur": 15, + "args": { + "External id": 51392, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51392, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51392, "pid": 0, "tid": 7, "ts": 1776649791318632, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318623, "dur": 4, + "args": { + "External id": 51392, + "cbid": 211, "correlation": 51392 + } + }, + { + "ph": "s", "id": 51392, "pid": 435, "tid": 435, "ts": 1776649791318623, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791318667, "dur": 1, + "args": { + "External id": 6799, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51410, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51410, "pid": 0, "tid": 7, "ts": 1776649791318667, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318658, "dur": 6, + "args": { + "External id": 6799, + "cbid": 211, "correlation": 51410 + } + }, + { + "ph": "s", "id": 51410, "pid": 435, "tid": 435, "ts": 1776649791318658, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318670, "dur": 0, + "args": { + "External id": 51416, + "cbid": 200, "correlation": 51416 + } + }, + { + "ph": "f", "id": 51416, "pid": 435, "tid": 435, "ts": 1776649791318670, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318671, "dur": 0, + "args": { + "External id": 51417, + "cbid": 273, "correlation": 51417 + } + }, + { + "ph": "f", "id": 51417, "pid": 435, "tid": 435, "ts": 1776649791318671, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791318680, "dur": 15, + "args": { + "External id": 51418, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51418, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51418, "pid": 0, "tid": 7, "ts": 1776649791318680, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318672, "dur": 4, + "args": { + "External id": 51418, + "cbid": 211, "correlation": 51418 + } + }, + { + "ph": "s", "id": 51418, "pid": 435, "tid": 435, "ts": 1776649791318672, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791318715, "dur": 1, + "args": { + "External id": 6804, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51436, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51436, "pid": 0, "tid": 7, "ts": 1776649791318715, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318706, "dur": 6, + "args": { + "External id": 6804, + "cbid": 211, "correlation": 51436 + } + }, + { + "ph": "s", "id": 51436, "pid": 435, "tid": 435, "ts": 1776649791318706, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318718, "dur": 0, + "args": { + "External id": 51442, + "cbid": 200, "correlation": 51442 + } + }, + { + "ph": "f", "id": 51442, "pid": 435, "tid": 435, "ts": 1776649791318718, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318719, "dur": 0, + "args": { + "External id": 51443, + "cbid": 273, "correlation": 51443 + } + }, + { + "ph": "f", "id": 51443, "pid": 435, "tid": 435, "ts": 1776649791318719, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791318728, "dur": 15, + "args": { + "External id": 51444, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51444, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51444, "pid": 0, "tid": 7, "ts": 1776649791318728, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318720, "dur": 5, + "args": { + "External id": 51444, + "cbid": 211, "correlation": 51444 + } + }, + { + "ph": "s", "id": 51444, "pid": 435, "tid": 435, "ts": 1776649791318720, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791318764, "dur": 1, + "args": { + "External id": 6809, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51462, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51462, "pid": 0, "tid": 7, "ts": 1776649791318764, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318754, "dur": 6, + "args": { + "External id": 6809, + "cbid": 211, "correlation": 51462 + } + }, + { + "ph": "s", "id": 51462, "pid": 435, "tid": 435, "ts": 1776649791318754, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318767, "dur": 0, + "args": { + "External id": 51468, + "cbid": 200, "correlation": 51468 + } + }, + { + "ph": "f", "id": 51468, "pid": 435, "tid": 435, "ts": 1776649791318767, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318767, "dur": 0, + "args": { + "External id": 51469, + "cbid": 273, "correlation": 51469 + } + }, + { + "ph": "f", "id": 51469, "pid": 435, "tid": 435, "ts": 1776649791318767, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791318776, "dur": 15, + "args": { + "External id": 51470, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51470, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51470, "pid": 0, "tid": 7, "ts": 1776649791318776, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318768, "dur": 4, + "args": { + "External id": 51470, + "cbid": 211, "correlation": 51470 + } + }, + { + "ph": "s", "id": 51470, "pid": 435, "tid": 435, "ts": 1776649791318768, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791318812, "dur": 1, + "args": { + "External id": 6814, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51488, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51488, "pid": 0, "tid": 7, "ts": 1776649791318812, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318803, "dur": 6, + "args": { + "External id": 6814, + "cbid": 211, "correlation": 51488 + } + }, + { + "ph": "s", "id": 51488, "pid": 435, "tid": 435, "ts": 1776649791318803, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318815, "dur": 0, + "args": { + "External id": 51494, + "cbid": 200, "correlation": 51494 + } + }, + { + "ph": "f", "id": 51494, "pid": 435, "tid": 435, "ts": 1776649791318815, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318816, "dur": 0, + "args": { + "External id": 51495, + "cbid": 273, "correlation": 51495 + } + }, + { + "ph": "f", "id": 51495, "pid": 435, "tid": 435, "ts": 1776649791318816, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791318825, "dur": 15, + "args": { + "External id": 51496, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51496, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51496, "pid": 0, "tid": 7, "ts": 1776649791318825, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318817, "dur": 4, + "args": { + "External id": 51496, + "cbid": 211, "correlation": 51496 + } + }, + { + "ph": "s", "id": 51496, "pid": 435, "tid": 435, "ts": 1776649791318817, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791318863, "dur": 1, + "args": { + "External id": 6819, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51514, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51514, "pid": 0, "tid": 7, "ts": 1776649791318863, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318851, "dur": 8, + "args": { + "External id": 6819, + "cbid": 211, "correlation": 51514 + } + }, + { + "ph": "s", "id": 51514, "pid": 435, "tid": 435, "ts": 1776649791318851, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318866, "dur": 0, + "args": { + "External id": 51520, + "cbid": 200, "correlation": 51520 + } + }, + { + "ph": "f", "id": 51520, "pid": 435, "tid": 435, "ts": 1776649791318866, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318867, "dur": 0, + "args": { + "External id": 51521, + "cbid": 273, "correlation": 51521 + } + }, + { + "ph": "f", "id": 51521, "pid": 435, "tid": 435, "ts": 1776649791318867, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791318876, "dur": 17, + "args": { + "External id": 51522, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51522, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51522, "pid": 0, "tid": 7, "ts": 1776649791318876, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318868, "dur": 5, + "args": { + "External id": 51522, + "cbid": 211, "correlation": 51522 + } + }, + { + "ph": "s", "id": 51522, "pid": 435, "tid": 435, "ts": 1776649791318868, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791318913, "dur": 1, + "args": { + "External id": 6824, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51540, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51540, "pid": 0, "tid": 7, "ts": 1776649791318913, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318903, "dur": 7, + "args": { + "External id": 6824, + "cbid": 211, "correlation": 51540 + } + }, + { + "ph": "s", "id": 51540, "pid": 435, "tid": 435, "ts": 1776649791318903, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318916, "dur": 0, + "args": { + "External id": 51546, + "cbid": 200, "correlation": 51546 + } + }, + { + "ph": "f", "id": 51546, "pid": 435, "tid": 435, "ts": 1776649791318916, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318917, "dur": 0, + "args": { + "External id": 51547, + "cbid": 273, "correlation": 51547 + } + }, + { + "ph": "f", "id": 51547, "pid": 435, "tid": 435, "ts": 1776649791318917, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791318926, "dur": 15, + "args": { + "External id": 51548, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51548, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51548, "pid": 0, "tid": 7, "ts": 1776649791318926, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318918, "dur": 5, + "args": { + "External id": 51548, + "cbid": 211, "correlation": 51548 + } + }, + { + "ph": "s", "id": 51548, "pid": 435, "tid": 435, "ts": 1776649791318918, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791318964, "dur": 1, + "args": { + "External id": 6829, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51566, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51566, "pid": 0, "tid": 7, "ts": 1776649791318964, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318955, "dur": 6, + "args": { + "External id": 6829, + "cbid": 211, "correlation": 51566 + } + }, + { + "ph": "s", "id": 51566, "pid": 435, "tid": 435, "ts": 1776649791318955, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318967, "dur": 0, + "args": { + "External id": 51572, + "cbid": 200, "correlation": 51572 + } + }, + { + "ph": "f", "id": 51572, "pid": 435, "tid": 435, "ts": 1776649791318967, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791318968, "dur": 0, + "args": { + "External id": 51573, + "cbid": 273, "correlation": 51573 + } + }, + { + "ph": "f", "id": 51573, "pid": 435, "tid": 435, "ts": 1776649791318968, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791318977, "dur": 15, + "args": { + "External id": 51574, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51574, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51574, "pid": 0, "tid": 7, "ts": 1776649791318977, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791318969, "dur": 4, + "args": { + "External id": 51574, + "cbid": 211, "correlation": 51574 + } + }, + { + "ph": "s", "id": 51574, "pid": 435, "tid": 435, "ts": 1776649791318969, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791318999, "dur": 3, + "args": { + "External id": 51582, + "cbid": 135, "correlation": 51582 + } + }, + { + "ph": "f", "id": 51582, "pid": 435, "tid": 435, "ts": 1776649791318999, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791319014, "dur": 6, + "args": { + "External id": 51587, + "cbid": 165, "correlation": 51587 + } + }, + { + "ph": "s", "id": 51587, "pid": 435, "tid": 435, "ts": 1776649791319014, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649791319023, "dur": 4, + "args": { + "External id": 51590, + "cbid": 139, "correlation": 51590 + } + }, + { + "ph": "f", "id": 51590, "pid": 435, "tid": 435, "ts": 1776649791319023, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791319078, "dur": 1, + "args": { + "External id": 6834, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51618, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51618, "pid": 0, "tid": 7, "ts": 1776649791319078, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319068, "dur": 8, + "args": { + "External id": 6834, + "cbid": 211, "correlation": 51618 + } + }, + { + "ph": "s", "id": 51618, "pid": 435, "tid": 435, "ts": 1776649791319068, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319083, "dur": 0, + "args": { + "External id": 51624, + "cbid": 200, "correlation": 51624 + } + }, + { + "ph": "f", "id": 51624, "pid": 435, "tid": 435, "ts": 1776649791319083, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319083, "dur": 0, + "args": { + "External id": 51625, + "cbid": 273, "correlation": 51625 + } + }, + { + "ph": "f", "id": 51625, "pid": 435, "tid": 435, "ts": 1776649791319083, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791319093, "dur": 15, + "args": { + "External id": 51626, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51626, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51626, "pid": 0, "tid": 7, "ts": 1776649791319093, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319085, "dur": 5, + "args": { + "External id": 51626, + "cbid": 211, "correlation": 51626 + } + }, + { + "ph": "s", "id": 51626, "pid": 435, "tid": 435, "ts": 1776649791319085, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791319131, "dur": 1, + "args": { + "External id": 6839, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51644, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51644, "pid": 0, "tid": 7, "ts": 1776649791319131, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319121, "dur": 7, + "args": { + "External id": 6839, + "cbid": 211, "correlation": 51644 + } + }, + { + "ph": "s", "id": 51644, "pid": 435, "tid": 435, "ts": 1776649791319121, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319134, "dur": 0, + "args": { + "External id": 51650, + "cbid": 200, "correlation": 51650 + } + }, + { + "ph": "f", "id": 51650, "pid": 435, "tid": 435, "ts": 1776649791319134, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319135, "dur": 0, + "args": { + "External id": 51651, + "cbid": 273, "correlation": 51651 + } + }, + { + "ph": "f", "id": 51651, "pid": 435, "tid": 435, "ts": 1776649791319135, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791319144, "dur": 15, + "args": { + "External id": 51652, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51652, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51652, "pid": 0, "tid": 7, "ts": 1776649791319144, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319136, "dur": 5, + "args": { + "External id": 51652, + "cbid": 211, "correlation": 51652 + } + }, + { + "ph": "s", "id": 51652, "pid": 435, "tid": 435, "ts": 1776649791319136, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791319180, "dur": 1, + "args": { + "External id": 6844, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51670, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51670, "pid": 0, "tid": 7, "ts": 1776649791319180, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319171, "dur": 6, + "args": { + "External id": 6844, + "cbid": 211, "correlation": 51670 + } + }, + { + "ph": "s", "id": 51670, "pid": 435, "tid": 435, "ts": 1776649791319171, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319183, "dur": 0, + "args": { + "External id": 51676, + "cbid": 200, "correlation": 51676 + } + }, + { + "ph": "f", "id": 51676, "pid": 435, "tid": 435, "ts": 1776649791319183, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319184, "dur": 0, + "args": { + "External id": 51677, + "cbid": 273, "correlation": 51677 + } + }, + { + "ph": "f", "id": 51677, "pid": 435, "tid": 435, "ts": 1776649791319184, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791319193, "dur": 15, + "args": { + "External id": 51678, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51678, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51678, "pid": 0, "tid": 7, "ts": 1776649791319193, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319185, "dur": 4, + "args": { + "External id": 51678, + "cbid": 211, "correlation": 51678 + } + }, + { + "ph": "s", "id": 51678, "pid": 435, "tid": 435, "ts": 1776649791319185, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791319229, "dur": 1, + "args": { + "External id": 6849, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51696, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51696, "pid": 0, "tid": 7, "ts": 1776649791319229, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319220, "dur": 7, + "args": { + "External id": 6849, + "cbid": 211, "correlation": 51696 + } + }, + { + "ph": "s", "id": 51696, "pid": 435, "tid": 435, "ts": 1776649791319220, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319232, "dur": 0, + "args": { + "External id": 51702, + "cbid": 200, "correlation": 51702 + } + }, + { + "ph": "f", "id": 51702, "pid": 435, "tid": 435, "ts": 1776649791319232, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319233, "dur": 0, + "args": { + "External id": 51703, + "cbid": 273, "correlation": 51703 + } + }, + { + "ph": "f", "id": 51703, "pid": 435, "tid": 435, "ts": 1776649791319233, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791319242, "dur": 15, + "args": { + "External id": 51704, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51704, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51704, "pid": 0, "tid": 7, "ts": 1776649791319242, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319234, "dur": 5, + "args": { + "External id": 51704, + "cbid": 211, "correlation": 51704 + } + }, + { + "ph": "s", "id": 51704, "pid": 435, "tid": 435, "ts": 1776649791319234, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791319278, "dur": 1, + "args": { + "External id": 6854, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51722, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51722, "pid": 0, "tid": 7, "ts": 1776649791319278, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319269, "dur": 6, + "args": { + "External id": 6854, + "cbid": 211, "correlation": 51722 + } + }, + { + "ph": "s", "id": 51722, "pid": 435, "tid": 435, "ts": 1776649791319269, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319281, "dur": 0, + "args": { + "External id": 51728, + "cbid": 200, "correlation": 51728 + } + }, + { + "ph": "f", "id": 51728, "pid": 435, "tid": 435, "ts": 1776649791319281, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319282, "dur": 0, + "args": { + "External id": 51729, + "cbid": 273, "correlation": 51729 + } + }, + { + "ph": "f", "id": 51729, "pid": 435, "tid": 435, "ts": 1776649791319282, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791319291, "dur": 15, + "args": { + "External id": 51730, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51730, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51730, "pid": 0, "tid": 7, "ts": 1776649791319291, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319283, "dur": 5, + "args": { + "External id": 51730, + "cbid": 211, "correlation": 51730 + } + }, + { + "ph": "s", "id": 51730, "pid": 435, "tid": 435, "ts": 1776649791319283, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791319326, "dur": 1, + "args": { + "External id": 6859, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51748, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51748, "pid": 0, "tid": 7, "ts": 1776649791319326, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319317, "dur": 6, + "args": { + "External id": 6859, + "cbid": 211, "correlation": 51748 + } + }, + { + "ph": "s", "id": 51748, "pid": 435, "tid": 435, "ts": 1776649791319317, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319329, "dur": 0, + "args": { + "External id": 51754, + "cbid": 200, "correlation": 51754 + } + }, + { + "ph": "f", "id": 51754, "pid": 435, "tid": 435, "ts": 1776649791319329, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319330, "dur": 0, + "args": { + "External id": 51755, + "cbid": 273, "correlation": 51755 + } + }, + { + "ph": "f", "id": 51755, "pid": 435, "tid": 435, "ts": 1776649791319330, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791319339, "dur": 16, + "args": { + "External id": 51756, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51756, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51756, "pid": 0, "tid": 7, "ts": 1776649791319339, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319331, "dur": 5, + "args": { + "External id": 51756, + "cbid": 211, "correlation": 51756 + } + }, + { + "ph": "s", "id": 51756, "pid": 435, "tid": 435, "ts": 1776649791319331, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791319375, "dur": 1, + "args": { + "External id": 6864, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51774, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51774, "pid": 0, "tid": 7, "ts": 1776649791319375, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319365, "dur": 6, + "args": { + "External id": 6864, + "cbid": 211, "correlation": 51774 + } + }, + { + "ph": "s", "id": 51774, "pid": 435, "tid": 435, "ts": 1776649791319365, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319378, "dur": 0, + "args": { + "External id": 51780, + "cbid": 200, "correlation": 51780 + } + }, + { + "ph": "f", "id": 51780, "pid": 435, "tid": 435, "ts": 1776649791319378, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319378, "dur": 0, + "args": { + "External id": 51781, + "cbid": 273, "correlation": 51781 + } + }, + { + "ph": "f", "id": 51781, "pid": 435, "tid": 435, "ts": 1776649791319378, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791319388, "dur": 16, + "args": { + "External id": 51782, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51782, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51782, "pid": 0, "tid": 7, "ts": 1776649791319388, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319380, "dur": 4, + "args": { + "External id": 51782, + "cbid": 211, "correlation": 51782 + } + }, + { + "ph": "s", "id": 51782, "pid": 435, "tid": 435, "ts": 1776649791319380, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791319423, "dur": 1, + "args": { + "External id": 6869, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51800, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51800, "pid": 0, "tid": 7, "ts": 1776649791319423, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319414, "dur": 6, + "args": { + "External id": 6869, + "cbid": 211, "correlation": 51800 + } + }, + { + "ph": "s", "id": 51800, "pid": 435, "tid": 435, "ts": 1776649791319414, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319426, "dur": 0, + "args": { + "External id": 51806, + "cbid": 200, "correlation": 51806 + } + }, + { + "ph": "f", "id": 51806, "pid": 435, "tid": 435, "ts": 1776649791319426, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319426, "dur": 0, + "args": { + "External id": 51807, + "cbid": 273, "correlation": 51807 + } + }, + { + "ph": "f", "id": 51807, "pid": 435, "tid": 435, "ts": 1776649791319426, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791319435, "dur": 15, + "args": { + "External id": 51808, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51808, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51808, "pid": 0, "tid": 7, "ts": 1776649791319435, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319428, "dur": 4, + "args": { + "External id": 51808, + "cbid": 211, "correlation": 51808 + } + }, + { + "ph": "s", "id": 51808, "pid": 435, "tid": 435, "ts": 1776649791319428, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791319471, "dur": 1, + "args": { + "External id": 6874, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51826, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51826, "pid": 0, "tid": 7, "ts": 1776649791319471, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319462, "dur": 6, + "args": { + "External id": 6874, + "cbid": 211, "correlation": 51826 + } + }, + { + "ph": "s", "id": 51826, "pid": 435, "tid": 435, "ts": 1776649791319462, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319474, "dur": 0, + "args": { + "External id": 51832, + "cbid": 200, "correlation": 51832 + } + }, + { + "ph": "f", "id": 51832, "pid": 435, "tid": 435, "ts": 1776649791319474, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319474, "dur": 0, + "args": { + "External id": 51833, + "cbid": 273, "correlation": 51833 + } + }, + { + "ph": "f", "id": 51833, "pid": 435, "tid": 435, "ts": 1776649791319474, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791319483, "dur": 15, + "args": { + "External id": 51834, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51834, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51834, "pid": 0, "tid": 7, "ts": 1776649791319483, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319476, "dur": 4, + "args": { + "External id": 51834, + "cbid": 211, "correlation": 51834 + } + }, + { + "ph": "s", "id": 51834, "pid": 435, "tid": 435, "ts": 1776649791319476, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791319519, "dur": 1, + "args": { + "External id": 6879, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51852, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51852, "pid": 0, "tid": 7, "ts": 1776649791319519, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319510, "dur": 6, + "args": { + "External id": 6879, + "cbid": 211, "correlation": 51852 + } + }, + { + "ph": "s", "id": 51852, "pid": 435, "tid": 435, "ts": 1776649791319510, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319522, "dur": 0, + "args": { + "External id": 51858, + "cbid": 200, "correlation": 51858 + } + }, + { + "ph": "f", "id": 51858, "pid": 435, "tid": 435, "ts": 1776649791319522, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319522, "dur": 0, + "args": { + "External id": 51859, + "cbid": 273, "correlation": 51859 + } + }, + { + "ph": "f", "id": 51859, "pid": 435, "tid": 435, "ts": 1776649791319522, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791319531, "dur": 15, + "args": { + "External id": 51860, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51860, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51860, "pid": 0, "tid": 7, "ts": 1776649791319531, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319524, "dur": 4, + "args": { + "External id": 51860, + "cbid": 211, "correlation": 51860 + } + }, + { + "ph": "s", "id": 51860, "pid": 435, "tid": 435, "ts": 1776649791319524, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791319546, "dur": 6, + "args": { + "External id": 51863, + "cbid": 165, "correlation": 51863 + } + }, + { + "ph": "s", "id": 51863, "pid": 435, "tid": 435, "ts": 1776649791319546, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791319573, "dur": 3, + "args": { + "External id": 51872, + "cbid": 135, "correlation": 51872 + } + }, + { + "ph": "f", "id": 51872, "pid": 435, "tid": 435, "ts": 1776649791319573, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791319622, "dur": 1, + "args": { + "External id": 6884, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51892, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51892, "pid": 0, "tid": 7, "ts": 1776649791319622, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319611, "dur": 8, + "args": { + "External id": 6884, + "cbid": 211, "correlation": 51892 + } + }, + { + "ph": "s", "id": 51892, "pid": 435, "tid": 435, "ts": 1776649791319611, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319626, "dur": 0, + "args": { + "External id": 51898, + "cbid": 200, "correlation": 51898 + } + }, + { + "ph": "f", "id": 51898, "pid": 435, "tid": 435, "ts": 1776649791319626, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319627, "dur": 1, + "args": { + "External id": 51899, + "cbid": 273, "correlation": 51899 + } + }, + { + "ph": "f", "id": 51899, "pid": 435, "tid": 435, "ts": 1776649791319627, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791319636, "dur": 15, + "args": { + "External id": 51900, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51900, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51900, "pid": 0, "tid": 7, "ts": 1776649791319636, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319628, "dur": 5, + "args": { + "External id": 51900, + "cbid": 211, "correlation": 51900 + } + }, + { + "ph": "s", "id": 51900, "pid": 435, "tid": 435, "ts": 1776649791319628, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791319673, "dur": 1, + "args": { + "External id": 6889, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51918, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51918, "pid": 0, "tid": 7, "ts": 1776649791319673, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319664, "dur": 6, + "args": { + "External id": 6889, + "cbid": 211, "correlation": 51918 + } + }, + { + "ph": "s", "id": 51918, "pid": 435, "tid": 435, "ts": 1776649791319664, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319677, "dur": 0, + "args": { + "External id": 51924, + "cbid": 200, "correlation": 51924 + } + }, + { + "ph": "f", "id": 51924, "pid": 435, "tid": 435, "ts": 1776649791319677, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319677, "dur": 0, + "args": { + "External id": 51925, + "cbid": 273, "correlation": 51925 + } + }, + { + "ph": "f", "id": 51925, "pid": 435, "tid": 435, "ts": 1776649791319677, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791319686, "dur": 15, + "args": { + "External id": 51926, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51926, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51926, "pid": 0, "tid": 7, "ts": 1776649791319686, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319678, "dur": 4, + "args": { + "External id": 51926, + "cbid": 211, "correlation": 51926 + } + }, + { + "ph": "s", "id": 51926, "pid": 435, "tid": 435, "ts": 1776649791319678, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791319722, "dur": 1, + "args": { + "External id": 6894, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51944, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51944, "pid": 0, "tid": 7, "ts": 1776649791319722, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319713, "dur": 6, + "args": { + "External id": 6894, + "cbid": 211, "correlation": 51944 + } + }, + { + "ph": "s", "id": 51944, "pid": 435, "tid": 435, "ts": 1776649791319713, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319725, "dur": 0, + "args": { + "External id": 51950, + "cbid": 200, "correlation": 51950 + } + }, + { + "ph": "f", "id": 51950, "pid": 435, "tid": 435, "ts": 1776649791319725, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319725, "dur": 0, + "args": { + "External id": 51951, + "cbid": 273, "correlation": 51951 + } + }, + { + "ph": "f", "id": 51951, "pid": 435, "tid": 435, "ts": 1776649791319725, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791319734, "dur": 15, + "args": { + "External id": 51952, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51952, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51952, "pid": 0, "tid": 7, "ts": 1776649791319734, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319727, "dur": 4, + "args": { + "External id": 51952, + "cbid": 211, "correlation": 51952 + } + }, + { + "ph": "s", "id": 51952, "pid": 435, "tid": 435, "ts": 1776649791319727, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791319771, "dur": 1, + "args": { + "External id": 6899, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51970, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51970, "pid": 0, "tid": 7, "ts": 1776649791319771, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319761, "dur": 6, + "args": { + "External id": 6899, + "cbid": 211, "correlation": 51970 + } + }, + { + "ph": "s", "id": 51970, "pid": 435, "tid": 435, "ts": 1776649791319761, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319774, "dur": 0, + "args": { + "External id": 51976, + "cbid": 200, "correlation": 51976 + } + }, + { + "ph": "f", "id": 51976, "pid": 435, "tid": 435, "ts": 1776649791319774, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319775, "dur": 0, + "args": { + "External id": 51977, + "cbid": 273, "correlation": 51977 + } + }, + { + "ph": "f", "id": 51977, "pid": 435, "tid": 435, "ts": 1776649791319775, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791319784, "dur": 15, + "args": { + "External id": 51978, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51978, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51978, "pid": 0, "tid": 7, "ts": 1776649791319784, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319776, "dur": 4, + "args": { + "External id": 51978, + "cbid": 211, "correlation": 51978 + } + }, + { + "ph": "s", "id": 51978, "pid": 435, "tid": 435, "ts": 1776649791319776, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791319819, "dur": 1, + "args": { + "External id": 6904, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 51996, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 51996, "pid": 0, "tid": 7, "ts": 1776649791319819, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319810, "dur": 6, + "args": { + "External id": 6904, + "cbid": 211, "correlation": 51996 + } + }, + { + "ph": "s", "id": 51996, "pid": 435, "tid": 435, "ts": 1776649791319810, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319822, "dur": 0, + "args": { + "External id": 52002, + "cbid": 200, "correlation": 52002 + } + }, + { + "ph": "f", "id": 52002, "pid": 435, "tid": 435, "ts": 1776649791319822, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319823, "dur": 0, + "args": { + "External id": 52003, + "cbid": 273, "correlation": 52003 + } + }, + { + "ph": "f", "id": 52003, "pid": 435, "tid": 435, "ts": 1776649791319823, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791319832, "dur": 16, + "args": { + "External id": 52004, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52004, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52004, "pid": 0, "tid": 7, "ts": 1776649791319832, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319824, "dur": 5, + "args": { + "External id": 52004, + "cbid": 211, "correlation": 52004 + } + }, + { + "ph": "s", "id": 52004, "pid": 435, "tid": 435, "ts": 1776649791319824, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791319873, "dur": 1, + "args": { + "External id": 6909, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52022, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52022, "pid": 0, "tid": 7, "ts": 1776649791319873, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319864, "dur": 6, + "args": { + "External id": 6909, + "cbid": 211, "correlation": 52022 + } + }, + { + "ph": "s", "id": 52022, "pid": 435, "tid": 435, "ts": 1776649791319864, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319876, "dur": 0, + "args": { + "External id": 52028, + "cbid": 200, "correlation": 52028 + } + }, + { + "ph": "f", "id": 52028, "pid": 435, "tid": 435, "ts": 1776649791319876, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319877, "dur": 0, + "args": { + "External id": 52029, + "cbid": 273, "correlation": 52029 + } + }, + { + "ph": "f", "id": 52029, "pid": 435, "tid": 435, "ts": 1776649791319877, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791319886, "dur": 15, + "args": { + "External id": 52030, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52030, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52030, "pid": 0, "tid": 7, "ts": 1776649791319886, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319878, "dur": 4, + "args": { + "External id": 52030, + "cbid": 211, "correlation": 52030 + } + }, + { + "ph": "s", "id": 52030, "pid": 435, "tid": 435, "ts": 1776649791319878, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791319921, "dur": 1, + "args": { + "External id": 6914, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52048, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52048, "pid": 0, "tid": 7, "ts": 1776649791319921, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319912, "dur": 6, + "args": { + "External id": 6914, + "cbid": 211, "correlation": 52048 + } + }, + { + "ph": "s", "id": 52048, "pid": 435, "tid": 435, "ts": 1776649791319912, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319924, "dur": 0, + "args": { + "External id": 52054, + "cbid": 200, "correlation": 52054 + } + }, + { + "ph": "f", "id": 52054, "pid": 435, "tid": 435, "ts": 1776649791319924, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319925, "dur": 0, + "args": { + "External id": 52055, + "cbid": 273, "correlation": 52055 + } + }, + { + "ph": "f", "id": 52055, "pid": 435, "tid": 435, "ts": 1776649791319925, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791319934, "dur": 15, + "args": { + "External id": 52056, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52056, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52056, "pid": 0, "tid": 7, "ts": 1776649791319934, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319926, "dur": 4, + "args": { + "External id": 52056, + "cbid": 211, "correlation": 52056 + } + }, + { + "ph": "s", "id": 52056, "pid": 435, "tid": 435, "ts": 1776649791319926, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791319969, "dur": 1, + "args": { + "External id": 6919, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52074, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52074, "pid": 0, "tid": 7, "ts": 1776649791319969, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319960, "dur": 6, + "args": { + "External id": 6919, + "cbid": 211, "correlation": 52074 + } + }, + { + "ph": "s", "id": 52074, "pid": 435, "tid": 435, "ts": 1776649791319960, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319973, "dur": 0, + "args": { + "External id": 52080, + "cbid": 200, "correlation": 52080 + } + }, + { + "ph": "f", "id": 52080, "pid": 435, "tid": 435, "ts": 1776649791319973, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791319973, "dur": 0, + "args": { + "External id": 52081, + "cbid": 273, "correlation": 52081 + } + }, + { + "ph": "f", "id": 52081, "pid": 435, "tid": 435, "ts": 1776649791319973, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791319982, "dur": 15, + "args": { + "External id": 52082, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52082, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52082, "pid": 0, "tid": 7, "ts": 1776649791319982, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791319974, "dur": 4, + "args": { + "External id": 52082, + "cbid": 211, "correlation": 52082 + } + }, + { + "ph": "s", "id": 52082, "pid": 435, "tid": 435, "ts": 1776649791319974, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791320018, "dur": 1, + "args": { + "External id": 6924, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52100, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52100, "pid": 0, "tid": 7, "ts": 1776649791320018, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320008, "dur": 6, + "args": { + "External id": 6924, + "cbid": 211, "correlation": 52100 + } + }, + { + "ph": "s", "id": 52100, "pid": 435, "tid": 435, "ts": 1776649791320008, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320021, "dur": 0, + "args": { + "External id": 52106, + "cbid": 200, "correlation": 52106 + } + }, + { + "ph": "f", "id": 52106, "pid": 435, "tid": 435, "ts": 1776649791320021, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320021, "dur": 0, + "args": { + "External id": 52107, + "cbid": 273, "correlation": 52107 + } + }, + { + "ph": "f", "id": 52107, "pid": 435, "tid": 435, "ts": 1776649791320021, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791320030, "dur": 15, + "args": { + "External id": 52108, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52108, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52108, "pid": 0, "tid": 7, "ts": 1776649791320030, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320023, "dur": 4, + "args": { + "External id": 52108, + "cbid": 211, "correlation": 52108 + } + }, + { + "ph": "s", "id": 52108, "pid": 435, "tid": 435, "ts": 1776649791320023, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791320066, "dur": 1, + "args": { + "External id": 6929, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52126, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52126, "pid": 0, "tid": 7, "ts": 1776649791320066, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320056, "dur": 6, + "args": { + "External id": 6929, + "cbid": 211, "correlation": 52126 + } + }, + { + "ph": "s", "id": 52126, "pid": 435, "tid": 435, "ts": 1776649791320056, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320069, "dur": 0, + "args": { + "External id": 52132, + "cbid": 200, "correlation": 52132 + } + }, + { + "ph": "f", "id": 52132, "pid": 435, "tid": 435, "ts": 1776649791320069, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320069, "dur": 0, + "args": { + "External id": 52133, + "cbid": 273, "correlation": 52133 + } + }, + { + "ph": "f", "id": 52133, "pid": 435, "tid": 435, "ts": 1776649791320069, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791320078, "dur": 15, + "args": { + "External id": 52134, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52134, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52134, "pid": 0, "tid": 7, "ts": 1776649791320078, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320070, "dur": 5, + "args": { + "External id": 52134, + "cbid": 211, "correlation": 52134 + } + }, + { + "ph": "s", "id": 52134, "pid": 435, "tid": 435, "ts": 1776649791320070, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791320116, "dur": 1, + "args": { + "External id": 6934, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52152, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52152, "pid": 0, "tid": 7, "ts": 1776649791320116, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320106, "dur": 6, + "args": { + "External id": 6934, + "cbid": 211, "correlation": 52152 + } + }, + { + "ph": "s", "id": 52152, "pid": 435, "tid": 435, "ts": 1776649791320106, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320119, "dur": 0, + "args": { + "External id": 52158, + "cbid": 200, "correlation": 52158 + } + }, + { + "ph": "f", "id": 52158, "pid": 435, "tid": 435, "ts": 1776649791320119, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320120, "dur": 0, + "args": { + "External id": 52159, + "cbid": 273, "correlation": 52159 + } + }, + { + "ph": "f", "id": 52159, "pid": 435, "tid": 435, "ts": 1776649791320120, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791320129, "dur": 15, + "args": { + "External id": 52160, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52160, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52160, "pid": 0, "tid": 7, "ts": 1776649791320129, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320121, "dur": 5, + "args": { + "External id": 52160, + "cbid": 211, "correlation": 52160 + } + }, + { + "ph": "s", "id": 52160, "pid": 435, "tid": 435, "ts": 1776649791320121, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791320164, "dur": 1, + "args": { + "External id": 6939, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52178, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52178, "pid": 0, "tid": 7, "ts": 1776649791320164, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320155, "dur": 6, + "args": { + "External id": 6939, + "cbid": 211, "correlation": 52178 + } + }, + { + "ph": "s", "id": 52178, "pid": 435, "tid": 435, "ts": 1776649791320155, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320168, "dur": 0, + "args": { + "External id": 52184, + "cbid": 200, "correlation": 52184 + } + }, + { + "ph": "f", "id": 52184, "pid": 435, "tid": 435, "ts": 1776649791320168, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320168, "dur": 0, + "args": { + "External id": 52185, + "cbid": 273, "correlation": 52185 + } + }, + { + "ph": "f", "id": 52185, "pid": 435, "tid": 435, "ts": 1776649791320168, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791320177, "dur": 15, + "args": { + "External id": 52186, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52186, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52186, "pid": 0, "tid": 7, "ts": 1776649791320177, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320169, "dur": 4, + "args": { + "External id": 52186, + "cbid": 211, "correlation": 52186 + } + }, + { + "ph": "s", "id": 52186, "pid": 435, "tid": 435, "ts": 1776649791320169, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791320213, "dur": 1, + "args": { + "External id": 6944, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52204, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52204, "pid": 0, "tid": 7, "ts": 1776649791320213, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320204, "dur": 6, + "args": { + "External id": 6944, + "cbid": 211, "correlation": 52204 + } + }, + { + "ph": "s", "id": 52204, "pid": 435, "tid": 435, "ts": 1776649791320204, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320236, "dur": 0, + "args": { + "External id": 52210, + "cbid": 200, "correlation": 52210 + } + }, + { + "ph": "f", "id": 52210, "pid": 435, "tid": 435, "ts": 1776649791320236, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320236, "dur": 1, + "args": { + "External id": 52211, + "cbid": 273, "correlation": 52211 + } + }, + { + "ph": "f", "id": 52211, "pid": 435, "tid": 435, "ts": 1776649791320236, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791320247, "dur": 17, + "args": { + "External id": 52212, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52212, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52212, "pid": 0, "tid": 7, "ts": 1776649791320247, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320238, "dur": 6, + "args": { + "External id": 52212, + "cbid": 211, "correlation": 52212 + } + }, + { + "ph": "s", "id": 52212, "pid": 435, "tid": 435, "ts": 1776649791320238, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791320287, "dur": 1, + "args": { + "External id": 6949, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52230, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52230, "pid": 0, "tid": 7, "ts": 1776649791320287, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320278, "dur": 6, + "args": { + "External id": 6949, + "cbid": 211, "correlation": 52230 + } + }, + { + "ph": "s", "id": 52230, "pid": 435, "tid": 435, "ts": 1776649791320278, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320291, "dur": 0, + "args": { + "External id": 52236, + "cbid": 200, "correlation": 52236 + } + }, + { + "ph": "f", "id": 52236, "pid": 435, "tid": 435, "ts": 1776649791320291, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320291, "dur": 0, + "args": { + "External id": 52237, + "cbid": 273, "correlation": 52237 + } + }, + { + "ph": "f", "id": 52237, "pid": 435, "tid": 435, "ts": 1776649791320291, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791320301, "dur": 16, + "args": { + "External id": 52238, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52238, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52238, "pid": 0, "tid": 7, "ts": 1776649791320301, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320293, "dur": 5, + "args": { + "External id": 52238, + "cbid": 211, "correlation": 52238 + } + }, + { + "ph": "s", "id": 52238, "pid": 435, "tid": 435, "ts": 1776649791320293, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791320337, "dur": 1, + "args": { + "External id": 6954, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52256, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52256, "pid": 0, "tid": 7, "ts": 1776649791320337, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320327, "dur": 6, + "args": { + "External id": 6954, + "cbid": 211, "correlation": 52256 + } + }, + { + "ph": "s", "id": 52256, "pid": 435, "tid": 435, "ts": 1776649791320327, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320340, "dur": 0, + "args": { + "External id": 52262, + "cbid": 200, "correlation": 52262 + } + }, + { + "ph": "f", "id": 52262, "pid": 435, "tid": 435, "ts": 1776649791320340, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320340, "dur": 0, + "args": { + "External id": 52263, + "cbid": 273, "correlation": 52263 + } + }, + { + "ph": "f", "id": 52263, "pid": 435, "tid": 435, "ts": 1776649791320340, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791320349, "dur": 15, + "args": { + "External id": 52264, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52264, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52264, "pid": 0, "tid": 7, "ts": 1776649791320349, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320341, "dur": 4, + "args": { + "External id": 52264, + "cbid": 211, "correlation": 52264 + } + }, + { + "ph": "s", "id": 52264, "pid": 435, "tid": 435, "ts": 1776649791320341, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791320384, "dur": 1, + "args": { + "External id": 6959, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52282, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52282, "pid": 0, "tid": 7, "ts": 1776649791320384, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320375, "dur": 6, + "args": { + "External id": 6959, + "cbid": 211, "correlation": 52282 + } + }, + { + "ph": "s", "id": 52282, "pid": 435, "tid": 435, "ts": 1776649791320375, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320387, "dur": 0, + "args": { + "External id": 52288, + "cbid": 200, "correlation": 52288 + } + }, + { + "ph": "f", "id": 52288, "pid": 435, "tid": 435, "ts": 1776649791320387, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320388, "dur": 0, + "args": { + "External id": 52289, + "cbid": 273, "correlation": 52289 + } + }, + { + "ph": "f", "id": 52289, "pid": 435, "tid": 435, "ts": 1776649791320388, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791320397, "dur": 15, + "args": { + "External id": 52290, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52290, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52290, "pid": 0, "tid": 7, "ts": 1776649791320397, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320389, "dur": 4, + "args": { + "External id": 52290, + "cbid": 211, "correlation": 52290 + } + }, + { + "ph": "s", "id": 52290, "pid": 435, "tid": 435, "ts": 1776649791320389, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791320432, "dur": 1, + "args": { + "External id": 6964, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52308, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52308, "pid": 0, "tid": 7, "ts": 1776649791320432, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320423, "dur": 6, + "args": { + "External id": 6964, + "cbid": 211, "correlation": 52308 + } + }, + { + "ph": "s", "id": 52308, "pid": 435, "tid": 435, "ts": 1776649791320423, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320435, "dur": 0, + "args": { + "External id": 52314, + "cbid": 200, "correlation": 52314 + } + }, + { + "ph": "f", "id": 52314, "pid": 435, "tid": 435, "ts": 1776649791320435, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320436, "dur": 0, + "args": { + "External id": 52315, + "cbid": 273, "correlation": 52315 + } + }, + { + "ph": "f", "id": 52315, "pid": 435, "tid": 435, "ts": 1776649791320436, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791320445, "dur": 15, + "args": { + "External id": 52316, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52316, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52316, "pid": 0, "tid": 7, "ts": 1776649791320445, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320437, "dur": 4, + "args": { + "External id": 52316, + "cbid": 211, "correlation": 52316 + } + }, + { + "ph": "s", "id": 52316, "pid": 435, "tid": 435, "ts": 1776649791320437, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791320480, "dur": 1, + "args": { + "External id": 6969, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52334, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52334, "pid": 0, "tid": 7, "ts": 1776649791320480, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320471, "dur": 6, + "args": { + "External id": 6969, + "cbid": 211, "correlation": 52334 + } + }, + { + "ph": "s", "id": 52334, "pid": 435, "tid": 435, "ts": 1776649791320471, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320483, "dur": 0, + "args": { + "External id": 52340, + "cbid": 200, "correlation": 52340 + } + }, + { + "ph": "f", "id": 52340, "pid": 435, "tid": 435, "ts": 1776649791320483, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320484, "dur": 0, + "args": { + "External id": 52341, + "cbid": 273, "correlation": 52341 + } + }, + { + "ph": "f", "id": 52341, "pid": 435, "tid": 435, "ts": 1776649791320484, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791320493, "dur": 15, + "args": { + "External id": 52342, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52342, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52342, "pid": 0, "tid": 7, "ts": 1776649791320493, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320485, "dur": 4, + "args": { + "External id": 52342, + "cbid": 211, "correlation": 52342 + } + }, + { + "ph": "s", "id": 52342, "pid": 435, "tid": 435, "ts": 1776649791320485, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791320528, "dur": 1, + "args": { + "External id": 6974, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52360, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52360, "pid": 0, "tid": 7, "ts": 1776649791320528, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320519, "dur": 6, + "args": { + "External id": 6974, + "cbid": 211, "correlation": 52360 + } + }, + { + "ph": "s", "id": 52360, "pid": 435, "tid": 435, "ts": 1776649791320519, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320531, "dur": 0, + "args": { + "External id": 52366, + "cbid": 200, "correlation": 52366 + } + }, + { + "ph": "f", "id": 52366, "pid": 435, "tid": 435, "ts": 1776649791320531, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320532, "dur": 0, + "args": { + "External id": 52367, + "cbid": 273, "correlation": 52367 + } + }, + { + "ph": "f", "id": 52367, "pid": 435, "tid": 435, "ts": 1776649791320532, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791320541, "dur": 15, + "args": { + "External id": 52368, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52368, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52368, "pid": 0, "tid": 7, "ts": 1776649791320541, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320533, "dur": 4, + "args": { + "External id": 52368, + "cbid": 211, "correlation": 52368 + } + }, + { + "ph": "s", "id": 52368, "pid": 435, "tid": 435, "ts": 1776649791320533, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791320576, "dur": 1, + "args": { + "External id": 6979, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52386, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52386, "pid": 0, "tid": 7, "ts": 1776649791320576, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320567, "dur": 6, + "args": { + "External id": 6979, + "cbid": 211, "correlation": 52386 + } + }, + { + "ph": "s", "id": 52386, "pid": 435, "tid": 435, "ts": 1776649791320567, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320579, "dur": 0, + "args": { + "External id": 52392, + "cbid": 200, "correlation": 52392 + } + }, + { + "ph": "f", "id": 52392, "pid": 435, "tid": 435, "ts": 1776649791320579, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320580, "dur": 0, + "args": { + "External id": 52393, + "cbid": 273, "correlation": 52393 + } + }, + { + "ph": "f", "id": 52393, "pid": 435, "tid": 435, "ts": 1776649791320580, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791320589, "dur": 15, + "args": { + "External id": 52394, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52394, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52394, "pid": 0, "tid": 7, "ts": 1776649791320589, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320581, "dur": 4, + "args": { + "External id": 52394, + "cbid": 211, "correlation": 52394 + } + }, + { + "ph": "s", "id": 52394, "pid": 435, "tid": 435, "ts": 1776649791320581, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791320624, "dur": 1, + "args": { + "External id": 6984, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52412, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52412, "pid": 0, "tid": 7, "ts": 1776649791320624, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320615, "dur": 6, + "args": { + "External id": 6984, + "cbid": 211, "correlation": 52412 + } + }, + { + "ph": "s", "id": 52412, "pid": 435, "tid": 435, "ts": 1776649791320615, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320627, "dur": 0, + "args": { + "External id": 52418, + "cbid": 200, "correlation": 52418 + } + }, + { + "ph": "f", "id": 52418, "pid": 435, "tid": 435, "ts": 1776649791320627, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320627, "dur": 0, + "args": { + "External id": 52419, + "cbid": 273, "correlation": 52419 + } + }, + { + "ph": "f", "id": 52419, "pid": 435, "tid": 435, "ts": 1776649791320627, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791320637, "dur": 15, + "args": { + "External id": 52420, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52420, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52420, "pid": 0, "tid": 7, "ts": 1776649791320637, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320629, "dur": 5, + "args": { + "External id": 52420, + "cbid": 211, "correlation": 52420 + } + }, + { + "ph": "s", "id": 52420, "pid": 435, "tid": 435, "ts": 1776649791320629, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791320672, "dur": 1, + "args": { + "External id": 6989, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52438, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52438, "pid": 0, "tid": 7, "ts": 1776649791320672, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320663, "dur": 6, + "args": { + "External id": 6989, + "cbid": 211, "correlation": 52438 + } + }, + { + "ph": "s", "id": 52438, "pid": 435, "tid": 435, "ts": 1776649791320663, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320675, "dur": 0, + "args": { + "External id": 52444, + "cbid": 200, "correlation": 52444 + } + }, + { + "ph": "f", "id": 52444, "pid": 435, "tid": 435, "ts": 1776649791320675, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320676, "dur": 0, + "args": { + "External id": 52445, + "cbid": 273, "correlation": 52445 + } + }, + { + "ph": "f", "id": 52445, "pid": 435, "tid": 435, "ts": 1776649791320676, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791320685, "dur": 16, + "args": { + "External id": 52446, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52446, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52446, "pid": 0, "tid": 7, "ts": 1776649791320685, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320677, "dur": 4, + "args": { + "External id": 52446, + "cbid": 211, "correlation": 52446 + } + }, + { + "ph": "s", "id": 52446, "pid": 435, "tid": 435, "ts": 1776649791320677, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791320720, "dur": 1, + "args": { + "External id": 6994, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52464, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52464, "pid": 0, "tid": 7, "ts": 1776649791320720, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320711, "dur": 6, + "args": { + "External id": 6994, + "cbid": 211, "correlation": 52464 + } + }, + { + "ph": "s", "id": 52464, "pid": 435, "tid": 435, "ts": 1776649791320711, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320723, "dur": 0, + "args": { + "External id": 52470, + "cbid": 200, "correlation": 52470 + } + }, + { + "ph": "f", "id": 52470, "pid": 435, "tid": 435, "ts": 1776649791320723, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320723, "dur": 0, + "args": { + "External id": 52471, + "cbid": 273, "correlation": 52471 + } + }, + { + "ph": "f", "id": 52471, "pid": 435, "tid": 435, "ts": 1776649791320723, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791320732, "dur": 16, + "args": { + "External id": 52472, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52472, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52472, "pid": 0, "tid": 7, "ts": 1776649791320732, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320725, "dur": 4, + "args": { + "External id": 52472, + "cbid": 211, "correlation": 52472 + } + }, + { + "ph": "s", "id": 52472, "pid": 435, "tid": 435, "ts": 1776649791320725, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791320768, "dur": 1, + "args": { + "External id": 6999, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52490, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52490, "pid": 0, "tid": 7, "ts": 1776649791320768, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320759, "dur": 6, + "args": { + "External id": 6999, + "cbid": 211, "correlation": 52490 + } + }, + { + "ph": "s", "id": 52490, "pid": 435, "tid": 435, "ts": 1776649791320759, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320771, "dur": 0, + "args": { + "External id": 52496, + "cbid": 200, "correlation": 52496 + } + }, + { + "ph": "f", "id": 52496, "pid": 435, "tid": 435, "ts": 1776649791320771, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320771, "dur": 0, + "args": { + "External id": 52497, + "cbid": 273, "correlation": 52497 + } + }, + { + "ph": "f", "id": 52497, "pid": 435, "tid": 435, "ts": 1776649791320771, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791320780, "dur": 16, + "args": { + "External id": 52498, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52498, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52498, "pid": 0, "tid": 7, "ts": 1776649791320780, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320773, "dur": 4, + "args": { + "External id": 52498, + "cbid": 211, "correlation": 52498 + } + }, + { + "ph": "s", "id": 52498, "pid": 435, "tid": 435, "ts": 1776649791320773, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791320816, "dur": 1, + "args": { + "External id": 7004, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52516, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52516, "pid": 0, "tid": 7, "ts": 1776649791320816, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320806, "dur": 6, + "args": { + "External id": 7004, + "cbid": 211, "correlation": 52516 + } + }, + { + "ph": "s", "id": 52516, "pid": 435, "tid": 435, "ts": 1776649791320806, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320819, "dur": 0, + "args": { + "External id": 52522, + "cbid": 200, "correlation": 52522 + } + }, + { + "ph": "f", "id": 52522, "pid": 435, "tid": 435, "ts": 1776649791320819, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320819, "dur": 0, + "args": { + "External id": 52523, + "cbid": 273, "correlation": 52523 + } + }, + { + "ph": "f", "id": 52523, "pid": 435, "tid": 435, "ts": 1776649791320819, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791320828, "dur": 15, + "args": { + "External id": 52524, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52524, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52524, "pid": 0, "tid": 7, "ts": 1776649791320828, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320820, "dur": 4, + "args": { + "External id": 52524, + "cbid": 211, "correlation": 52524 + } + }, + { + "ph": "s", "id": 52524, "pid": 435, "tid": 435, "ts": 1776649791320820, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791320864, "dur": 1, + "args": { + "External id": 7009, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52542, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52542, "pid": 0, "tid": 7, "ts": 1776649791320864, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320854, "dur": 6, + "args": { + "External id": 7009, + "cbid": 211, "correlation": 52542 + } + }, + { + "ph": "s", "id": 52542, "pid": 435, "tid": 435, "ts": 1776649791320854, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320867, "dur": 0, + "args": { + "External id": 52548, + "cbid": 200, "correlation": 52548 + } + }, + { + "ph": "f", "id": 52548, "pid": 435, "tid": 435, "ts": 1776649791320867, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320867, "dur": 0, + "args": { + "External id": 52549, + "cbid": 273, "correlation": 52549 + } + }, + { + "ph": "f", "id": 52549, "pid": 435, "tid": 435, "ts": 1776649791320867, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791320877, "dur": 15, + "args": { + "External id": 52550, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52550, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52550, "pid": 0, "tid": 7, "ts": 1776649791320877, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320868, "dur": 4, + "args": { + "External id": 52550, + "cbid": 211, "correlation": 52550 + } + }, + { + "ph": "s", "id": 52550, "pid": 435, "tid": 435, "ts": 1776649791320868, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791320912, "dur": 1, + "args": { + "External id": 7014, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52568, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52568, "pid": 0, "tid": 7, "ts": 1776649791320912, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320903, "dur": 6, + "args": { + "External id": 7014, + "cbid": 211, "correlation": 52568 + } + }, + { + "ph": "s", "id": 52568, "pid": 435, "tid": 435, "ts": 1776649791320903, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320915, "dur": 0, + "args": { + "External id": 52574, + "cbid": 200, "correlation": 52574 + } + }, + { + "ph": "f", "id": 52574, "pid": 435, "tid": 435, "ts": 1776649791320915, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320916, "dur": 0, + "args": { + "External id": 52575, + "cbid": 273, "correlation": 52575 + } + }, + { + "ph": "f", "id": 52575, "pid": 435, "tid": 435, "ts": 1776649791320916, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791320925, "dur": 15, + "args": { + "External id": 52576, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52576, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52576, "pid": 0, "tid": 7, "ts": 1776649791320925, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320917, "dur": 5, + "args": { + "External id": 52576, + "cbid": 211, "correlation": 52576 + } + }, + { + "ph": "s", "id": 52576, "pid": 435, "tid": 435, "ts": 1776649791320917, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791320960, "dur": 1, + "args": { + "External id": 7019, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52594, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52594, "pid": 0, "tid": 7, "ts": 1776649791320960, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320951, "dur": 6, + "args": { + "External id": 7019, + "cbid": 211, "correlation": 52594 + } + }, + { + "ph": "s", "id": 52594, "pid": 435, "tid": 435, "ts": 1776649791320951, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320963, "dur": 0, + "args": { + "External id": 52600, + "cbid": 200, "correlation": 52600 + } + }, + { + "ph": "f", "id": 52600, "pid": 435, "tid": 435, "ts": 1776649791320963, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791320964, "dur": 0, + "args": { + "External id": 52601, + "cbid": 273, "correlation": 52601 + } + }, + { + "ph": "f", "id": 52601, "pid": 435, "tid": 435, "ts": 1776649791320964, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791320973, "dur": 15, + "args": { + "External id": 52602, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52602, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52602, "pid": 0, "tid": 7, "ts": 1776649791320973, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320965, "dur": 5, + "args": { + "External id": 52602, + "cbid": 211, "correlation": 52602 + } + }, + { + "ph": "s", "id": 52602, "pid": 435, "tid": 435, "ts": 1776649791320965, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791321008, "dur": 1, + "args": { + "External id": 7024, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52620, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52620, "pid": 0, "tid": 7, "ts": 1776649791321008, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791320999, "dur": 6, + "args": { + "External id": 7024, + "cbid": 211, "correlation": 52620 + } + }, + { + "ph": "s", "id": 52620, "pid": 435, "tid": 435, "ts": 1776649791320999, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321011, "dur": 0, + "args": { + "External id": 52626, + "cbid": 200, "correlation": 52626 + } + }, + { + "ph": "f", "id": 52626, "pid": 435, "tid": 435, "ts": 1776649791321011, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321012, "dur": 0, + "args": { + "External id": 52627, + "cbid": 273, "correlation": 52627 + } + }, + { + "ph": "f", "id": 52627, "pid": 435, "tid": 435, "ts": 1776649791321012, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791321021, "dur": 15, + "args": { + "External id": 52628, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52628, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52628, "pid": 0, "tid": 7, "ts": 1776649791321021, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321013, "dur": 4, + "args": { + "External id": 52628, + "cbid": 211, "correlation": 52628 + } + }, + { + "ph": "s", "id": 52628, "pid": 435, "tid": 435, "ts": 1776649791321013, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791321056, "dur": 1, + "args": { + "External id": 7029, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52646, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52646, "pid": 0, "tid": 7, "ts": 1776649791321056, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321047, "dur": 6, + "args": { + "External id": 7029, + "cbid": 211, "correlation": 52646 + } + }, + { + "ph": "s", "id": 52646, "pid": 435, "tid": 435, "ts": 1776649791321047, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321060, "dur": 0, + "args": { + "External id": 52652, + "cbid": 200, "correlation": 52652 + } + }, + { + "ph": "f", "id": 52652, "pid": 435, "tid": 435, "ts": 1776649791321060, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321060, "dur": 0, + "args": { + "External id": 52653, + "cbid": 273, "correlation": 52653 + } + }, + { + "ph": "f", "id": 52653, "pid": 435, "tid": 435, "ts": 1776649791321060, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791321069, "dur": 15, + "args": { + "External id": 52654, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52654, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52654, "pid": 0, "tid": 7, "ts": 1776649791321069, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321061, "dur": 4, + "args": { + "External id": 52654, + "cbid": 211, "correlation": 52654 + } + }, + { + "ph": "s", "id": 52654, "pid": 435, "tid": 435, "ts": 1776649791321061, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791321105, "dur": 1, + "args": { + "External id": 7034, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52672, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52672, "pid": 0, "tid": 7, "ts": 1776649791321105, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321095, "dur": 6, + "args": { + "External id": 7034, + "cbid": 211, "correlation": 52672 + } + }, + { + "ph": "s", "id": 52672, "pid": 435, "tid": 435, "ts": 1776649791321095, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321108, "dur": 0, + "args": { + "External id": 52678, + "cbid": 200, "correlation": 52678 + } + }, + { + "ph": "f", "id": 52678, "pid": 435, "tid": 435, "ts": 1776649791321108, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321108, "dur": 0, + "args": { + "External id": 52679, + "cbid": 273, "correlation": 52679 + } + }, + { + "ph": "f", "id": 52679, "pid": 435, "tid": 435, "ts": 1776649791321108, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791321117, "dur": 16, + "args": { + "External id": 52680, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52680, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52680, "pid": 0, "tid": 7, "ts": 1776649791321117, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321110, "dur": 4, + "args": { + "External id": 52680, + "cbid": 211, "correlation": 52680 + } + }, + { + "ph": "s", "id": 52680, "pid": 435, "tid": 435, "ts": 1776649791321110, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791321152, "dur": 1, + "args": { + "External id": 7039, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52698, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52698, "pid": 0, "tid": 7, "ts": 1776649791321152, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321143, "dur": 6, + "args": { + "External id": 7039, + "cbid": 211, "correlation": 52698 + } + }, + { + "ph": "s", "id": 52698, "pid": 435, "tid": 435, "ts": 1776649791321143, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321155, "dur": 0, + "args": { + "External id": 52704, + "cbid": 200, "correlation": 52704 + } + }, + { + "ph": "f", "id": 52704, "pid": 435, "tid": 435, "ts": 1776649791321155, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321156, "dur": 0, + "args": { + "External id": 52705, + "cbid": 273, "correlation": 52705 + } + }, + { + "ph": "f", "id": 52705, "pid": 435, "tid": 435, "ts": 1776649791321156, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791321165, "dur": 15, + "args": { + "External id": 52706, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52706, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52706, "pid": 0, "tid": 7, "ts": 1776649791321165, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321157, "dur": 4, + "args": { + "External id": 52706, + "cbid": 211, "correlation": 52706 + } + }, + { + "ph": "s", "id": 52706, "pid": 435, "tid": 435, "ts": 1776649791321157, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791321200, "dur": 1, + "args": { + "External id": 7044, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52724, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52724, "pid": 0, "tid": 7, "ts": 1776649791321200, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321191, "dur": 6, + "args": { + "External id": 7044, + "cbid": 211, "correlation": 52724 + } + }, + { + "ph": "s", "id": 52724, "pid": 435, "tid": 435, "ts": 1776649791321191, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321203, "dur": 0, + "args": { + "External id": 52730, + "cbid": 200, "correlation": 52730 + } + }, + { + "ph": "f", "id": 52730, "pid": 435, "tid": 435, "ts": 1776649791321203, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321204, "dur": 0, + "args": { + "External id": 52731, + "cbid": 273, "correlation": 52731 + } + }, + { + "ph": "f", "id": 52731, "pid": 435, "tid": 435, "ts": 1776649791321204, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791321213, "dur": 15, + "args": { + "External id": 52732, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52732, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52732, "pid": 0, "tid": 7, "ts": 1776649791321213, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321205, "dur": 4, + "args": { + "External id": 52732, + "cbid": 211, "correlation": 52732 + } + }, + { + "ph": "s", "id": 52732, "pid": 435, "tid": 435, "ts": 1776649791321205, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791321248, "dur": 1, + "args": { + "External id": 7049, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52750, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52750, "pid": 0, "tid": 7, "ts": 1776649791321248, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321239, "dur": 6, + "args": { + "External id": 7049, + "cbid": 211, "correlation": 52750 + } + }, + { + "ph": "s", "id": 52750, "pid": 435, "tid": 435, "ts": 1776649791321239, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321252, "dur": 0, + "args": { + "External id": 52756, + "cbid": 200, "correlation": 52756 + } + }, + { + "ph": "f", "id": 52756, "pid": 435, "tid": 435, "ts": 1776649791321252, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321252, "dur": 0, + "args": { + "External id": 52757, + "cbid": 273, "correlation": 52757 + } + }, + { + "ph": "f", "id": 52757, "pid": 435, "tid": 435, "ts": 1776649791321252, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791321261, "dur": 16, + "args": { + "External id": 52758, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52758, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52758, "pid": 0, "tid": 7, "ts": 1776649791321261, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321253, "dur": 5, + "args": { + "External id": 52758, + "cbid": 211, "correlation": 52758 + } + }, + { + "ph": "s", "id": 52758, "pid": 435, "tid": 435, "ts": 1776649791321253, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791321297, "dur": 1, + "args": { + "External id": 7054, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52776, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52776, "pid": 0, "tid": 7, "ts": 1776649791321297, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321288, "dur": 6, + "args": { + "External id": 7054, + "cbid": 211, "correlation": 52776 + } + }, + { + "ph": "s", "id": 52776, "pid": 435, "tid": 435, "ts": 1776649791321288, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321300, "dur": 0, + "args": { + "External id": 52782, + "cbid": 200, "correlation": 52782 + } + }, + { + "ph": "f", "id": 52782, "pid": 435, "tid": 435, "ts": 1776649791321300, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321301, "dur": 0, + "args": { + "External id": 52783, + "cbid": 273, "correlation": 52783 + } + }, + { + "ph": "f", "id": 52783, "pid": 435, "tid": 435, "ts": 1776649791321301, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791321310, "dur": 15, + "args": { + "External id": 52784, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52784, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52784, "pid": 0, "tid": 7, "ts": 1776649791321310, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321302, "dur": 4, + "args": { + "External id": 52784, + "cbid": 211, "correlation": 52784 + } + }, + { + "ph": "s", "id": 52784, "pid": 435, "tid": 435, "ts": 1776649791321302, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791321345, "dur": 1, + "args": { + "External id": 7059, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52802, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52802, "pid": 0, "tid": 7, "ts": 1776649791321345, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321336, "dur": 6, + "args": { + "External id": 7059, + "cbid": 211, "correlation": 52802 + } + }, + { + "ph": "s", "id": 52802, "pid": 435, "tid": 435, "ts": 1776649791321336, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321348, "dur": 0, + "args": { + "External id": 52808, + "cbid": 200, "correlation": 52808 + } + }, + { + "ph": "f", "id": 52808, "pid": 435, "tid": 435, "ts": 1776649791321348, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321349, "dur": 0, + "args": { + "External id": 52809, + "cbid": 273, "correlation": 52809 + } + }, + { + "ph": "f", "id": 52809, "pid": 435, "tid": 435, "ts": 1776649791321349, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791321358, "dur": 15, + "args": { + "External id": 52810, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52810, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52810, "pid": 0, "tid": 7, "ts": 1776649791321358, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321350, "dur": 4, + "args": { + "External id": 52810, + "cbid": 211, "correlation": 52810 + } + }, + { + "ph": "s", "id": 52810, "pid": 435, "tid": 435, "ts": 1776649791321350, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791321393, "dur": 1, + "args": { + "External id": 7064, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52828, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52828, "pid": 0, "tid": 7, "ts": 1776649791321393, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321384, "dur": 6, + "args": { + "External id": 7064, + "cbid": 211, "correlation": 52828 + } + }, + { + "ph": "s", "id": 52828, "pid": 435, "tid": 435, "ts": 1776649791321384, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321396, "dur": 0, + "args": { + "External id": 52834, + "cbid": 200, "correlation": 52834 + } + }, + { + "ph": "f", "id": 52834, "pid": 435, "tid": 435, "ts": 1776649791321396, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321396, "dur": 0, + "args": { + "External id": 52835, + "cbid": 273, "correlation": 52835 + } + }, + { + "ph": "f", "id": 52835, "pid": 435, "tid": 435, "ts": 1776649791321396, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791321405, "dur": 15, + "args": { + "External id": 52836, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52836, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52836, "pid": 0, "tid": 7, "ts": 1776649791321405, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321397, "dur": 4, + "args": { + "External id": 52836, + "cbid": 211, "correlation": 52836 + } + }, + { + "ph": "s", "id": 52836, "pid": 435, "tid": 435, "ts": 1776649791321397, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791321440, "dur": 1, + "args": { + "External id": 7069, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52854, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52854, "pid": 0, "tid": 7, "ts": 1776649791321440, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321431, "dur": 6, + "args": { + "External id": 7069, + "cbid": 211, "correlation": 52854 + } + }, + { + "ph": "s", "id": 52854, "pid": 435, "tid": 435, "ts": 1776649791321431, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321443, "dur": 0, + "args": { + "External id": 52860, + "cbid": 200, "correlation": 52860 + } + }, + { + "ph": "f", "id": 52860, "pid": 435, "tid": 435, "ts": 1776649791321443, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321444, "dur": 0, + "args": { + "External id": 52861, + "cbid": 273, "correlation": 52861 + } + }, + { + "ph": "f", "id": 52861, "pid": 435, "tid": 435, "ts": 1776649791321444, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791321453, "dur": 15, + "args": { + "External id": 52862, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52862, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52862, "pid": 0, "tid": 7, "ts": 1776649791321453, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321445, "dur": 4, + "args": { + "External id": 52862, + "cbid": 211, "correlation": 52862 + } + }, + { + "ph": "s", "id": 52862, "pid": 435, "tid": 435, "ts": 1776649791321445, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791321489, "dur": 1, + "args": { + "External id": 7074, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52880, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52880, "pid": 0, "tid": 7, "ts": 1776649791321489, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321479, "dur": 6, + "args": { + "External id": 7074, + "cbid": 211, "correlation": 52880 + } + }, + { + "ph": "s", "id": 52880, "pid": 435, "tid": 435, "ts": 1776649791321479, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321492, "dur": 0, + "args": { + "External id": 52886, + "cbid": 200, "correlation": 52886 + } + }, + { + "ph": "f", "id": 52886, "pid": 435, "tid": 435, "ts": 1776649791321492, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321492, "dur": 0, + "args": { + "External id": 52887, + "cbid": 273, "correlation": 52887 + } + }, + { + "ph": "f", "id": 52887, "pid": 435, "tid": 435, "ts": 1776649791321492, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791321501, "dur": 15, + "args": { + "External id": 52888, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52888, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52888, "pid": 0, "tid": 7, "ts": 1776649791321501, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321494, "dur": 4, + "args": { + "External id": 52888, + "cbid": 211, "correlation": 52888 + } + }, + { + "ph": "s", "id": 52888, "pid": 435, "tid": 435, "ts": 1776649791321494, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791321537, "dur": 1, + "args": { + "External id": 7079, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52906, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52906, "pid": 0, "tid": 7, "ts": 1776649791321537, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321528, "dur": 6, + "args": { + "External id": 7079, + "cbid": 211, "correlation": 52906 + } + }, + { + "ph": "s", "id": 52906, "pid": 435, "tid": 435, "ts": 1776649791321528, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321540, "dur": 0, + "args": { + "External id": 52912, + "cbid": 200, "correlation": 52912 + } + }, + { + "ph": "f", "id": 52912, "pid": 435, "tid": 435, "ts": 1776649791321540, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321541, "dur": 0, + "args": { + "External id": 52913, + "cbid": 273, "correlation": 52913 + } + }, + { + "ph": "f", "id": 52913, "pid": 435, "tid": 435, "ts": 1776649791321541, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791321550, "dur": 16, + "args": { + "External id": 52914, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52914, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52914, "pid": 0, "tid": 7, "ts": 1776649791321550, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321542, "dur": 5, + "args": { + "External id": 52914, + "cbid": 211, "correlation": 52914 + } + }, + { + "ph": "s", "id": 52914, "pid": 435, "tid": 435, "ts": 1776649791321542, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791321585, "dur": 1, + "args": { + "External id": 7084, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52932, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52932, "pid": 0, "tid": 7, "ts": 1776649791321585, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321576, "dur": 6, + "args": { + "External id": 7084, + "cbid": 211, "correlation": 52932 + } + }, + { + "ph": "s", "id": 52932, "pid": 435, "tid": 435, "ts": 1776649791321576, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321588, "dur": 0, + "args": { + "External id": 52938, + "cbid": 200, "correlation": 52938 + } + }, + { + "ph": "f", "id": 52938, "pid": 435, "tid": 435, "ts": 1776649791321588, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321589, "dur": 0, + "args": { + "External id": 52939, + "cbid": 273, "correlation": 52939 + } + }, + { + "ph": "f", "id": 52939, "pid": 435, "tid": 435, "ts": 1776649791321589, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791321598, "dur": 15, + "args": { + "External id": 52940, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52940, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52940, "pid": 0, "tid": 7, "ts": 1776649791321598, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321590, "dur": 5, + "args": { + "External id": 52940, + "cbid": 211, "correlation": 52940 + } + }, + { + "ph": "s", "id": 52940, "pid": 435, "tid": 435, "ts": 1776649791321590, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791321634, "dur": 1, + "args": { + "External id": 7089, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52958, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52958, "pid": 0, "tid": 7, "ts": 1776649791321634, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321625, "dur": 6, + "args": { + "External id": 7089, + "cbid": 211, "correlation": 52958 + } + }, + { + "ph": "s", "id": 52958, "pid": 435, "tid": 435, "ts": 1776649791321625, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321637, "dur": 0, + "args": { + "External id": 52964, + "cbid": 200, "correlation": 52964 + } + }, + { + "ph": "f", "id": 52964, "pid": 435, "tid": 435, "ts": 1776649791321637, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321638, "dur": 0, + "args": { + "External id": 52965, + "cbid": 273, "correlation": 52965 + } + }, + { + "ph": "f", "id": 52965, "pid": 435, "tid": 435, "ts": 1776649791321638, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791321647, "dur": 16, + "args": { + "External id": 52966, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52966, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52966, "pid": 0, "tid": 7, "ts": 1776649791321647, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321639, "dur": 4, + "args": { + "External id": 52966, + "cbid": 211, "correlation": 52966 + } + }, + { + "ph": "s", "id": 52966, "pid": 435, "tid": 435, "ts": 1776649791321639, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791321683, "dur": 1, + "args": { + "External id": 7094, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52984, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52984, "pid": 0, "tid": 7, "ts": 1776649791321683, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321673, "dur": 6, + "args": { + "External id": 7094, + "cbid": 211, "correlation": 52984 + } + }, + { + "ph": "s", "id": 52984, "pid": 435, "tid": 435, "ts": 1776649791321673, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321686, "dur": 0, + "args": { + "External id": 52990, + "cbid": 200, "correlation": 52990 + } + }, + { + "ph": "f", "id": 52990, "pid": 435, "tid": 435, "ts": 1776649791321686, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321686, "dur": 0, + "args": { + "External id": 52991, + "cbid": 273, "correlation": 52991 + } + }, + { + "ph": "f", "id": 52991, "pid": 435, "tid": 435, "ts": 1776649791321686, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791321695, "dur": 15, + "args": { + "External id": 52992, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 52992, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 52992, "pid": 0, "tid": 7, "ts": 1776649791321695, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321688, "dur": 4, + "args": { + "External id": 52992, + "cbid": 211, "correlation": 52992 + } + }, + { + "ph": "s", "id": 52992, "pid": 435, "tid": 435, "ts": 1776649791321688, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791321731, "dur": 1, + "args": { + "External id": 7099, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53010, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53010, "pid": 0, "tid": 7, "ts": 1776649791321731, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321722, "dur": 6, + "args": { + "External id": 7099, + "cbid": 211, "correlation": 53010 + } + }, + { + "ph": "s", "id": 53010, "pid": 435, "tid": 435, "ts": 1776649791321722, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321734, "dur": 0, + "args": { + "External id": 53016, + "cbid": 200, "correlation": 53016 + } + }, + { + "ph": "f", "id": 53016, "pid": 435, "tid": 435, "ts": 1776649791321734, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321735, "dur": 0, + "args": { + "External id": 53017, + "cbid": 273, "correlation": 53017 + } + }, + { + "ph": "f", "id": 53017, "pid": 435, "tid": 435, "ts": 1776649791321735, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791321744, "dur": 15, + "args": { + "External id": 53018, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53018, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53018, "pid": 0, "tid": 7, "ts": 1776649791321744, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321736, "dur": 4, + "args": { + "External id": 53018, + "cbid": 211, "correlation": 53018 + } + }, + { + "ph": "s", "id": 53018, "pid": 435, "tid": 435, "ts": 1776649791321736, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791321779, "dur": 1, + "args": { + "External id": 7104, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53036, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53036, "pid": 0, "tid": 7, "ts": 1776649791321779, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321770, "dur": 6, + "args": { + "External id": 7104, + "cbid": 211, "correlation": 53036 + } + }, + { + "ph": "s", "id": 53036, "pid": 435, "tid": 435, "ts": 1776649791321770, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321783, "dur": 0, + "args": { + "External id": 53042, + "cbid": 200, "correlation": 53042 + } + }, + { + "ph": "f", "id": 53042, "pid": 435, "tid": 435, "ts": 1776649791321783, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321783, "dur": 0, + "args": { + "External id": 53043, + "cbid": 273, "correlation": 53043 + } + }, + { + "ph": "f", "id": 53043, "pid": 435, "tid": 435, "ts": 1776649791321783, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791321792, "dur": 15, + "args": { + "External id": 53044, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53044, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53044, "pid": 0, "tid": 7, "ts": 1776649791321792, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321784, "dur": 4, + "args": { + "External id": 53044, + "cbid": 211, "correlation": 53044 + } + }, + { + "ph": "s", "id": 53044, "pid": 435, "tid": 435, "ts": 1776649791321784, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791321827, "dur": 1, + "args": { + "External id": 7109, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53062, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53062, "pid": 0, "tid": 7, "ts": 1776649791321827, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321818, "dur": 6, + "args": { + "External id": 7109, + "cbid": 211, "correlation": 53062 + } + }, + { + "ph": "s", "id": 53062, "pid": 435, "tid": 435, "ts": 1776649791321818, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321830, "dur": 0, + "args": { + "External id": 53068, + "cbid": 200, "correlation": 53068 + } + }, + { + "ph": "f", "id": 53068, "pid": 435, "tid": 435, "ts": 1776649791321830, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321831, "dur": 0, + "args": { + "External id": 53069, + "cbid": 273, "correlation": 53069 + } + }, + { + "ph": "f", "id": 53069, "pid": 435, "tid": 435, "ts": 1776649791321831, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791321840, "dur": 15, + "args": { + "External id": 53070, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53070, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53070, "pid": 0, "tid": 7, "ts": 1776649791321840, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321832, "dur": 5, + "args": { + "External id": 53070, + "cbid": 211, "correlation": 53070 + } + }, + { + "ph": "s", "id": 53070, "pid": 435, "tid": 435, "ts": 1776649791321832, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791321875, "dur": 1, + "args": { + "External id": 7114, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53088, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53088, "pid": 0, "tid": 7, "ts": 1776649791321875, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321866, "dur": 6, + "args": { + "External id": 7114, + "cbid": 211, "correlation": 53088 + } + }, + { + "ph": "s", "id": 53088, "pid": 435, "tid": 435, "ts": 1776649791321866, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321878, "dur": 0, + "args": { + "External id": 53094, + "cbid": 200, "correlation": 53094 + } + }, + { + "ph": "f", "id": 53094, "pid": 435, "tid": 435, "ts": 1776649791321878, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321878, "dur": 0, + "args": { + "External id": 53095, + "cbid": 273, "correlation": 53095 + } + }, + { + "ph": "f", "id": 53095, "pid": 435, "tid": 435, "ts": 1776649791321878, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791321887, "dur": 15, + "args": { + "External id": 53096, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53096, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53096, "pid": 0, "tid": 7, "ts": 1776649791321887, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321880, "dur": 4, + "args": { + "External id": 53096, + "cbid": 211, "correlation": 53096 + } + }, + { + "ph": "s", "id": 53096, "pid": 435, "tid": 435, "ts": 1776649791321880, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791321922, "dur": 1, + "args": { + "External id": 7119, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53114, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53114, "pid": 0, "tid": 7, "ts": 1776649791321922, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321913, "dur": 6, + "args": { + "External id": 7119, + "cbid": 211, "correlation": 53114 + } + }, + { + "ph": "s", "id": 53114, "pid": 435, "tid": 435, "ts": 1776649791321913, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321926, "dur": 0, + "args": { + "External id": 53120, + "cbid": 200, "correlation": 53120 + } + }, + { + "ph": "f", "id": 53120, "pid": 435, "tid": 435, "ts": 1776649791321926, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321926, "dur": 0, + "args": { + "External id": 53121, + "cbid": 273, "correlation": 53121 + } + }, + { + "ph": "f", "id": 53121, "pid": 435, "tid": 435, "ts": 1776649791321926, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791321935, "dur": 16, + "args": { + "External id": 53122, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53122, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53122, "pid": 0, "tid": 7, "ts": 1776649791321935, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321927, "dur": 4, + "args": { + "External id": 53122, + "cbid": 211, "correlation": 53122 + } + }, + { + "ph": "s", "id": 53122, "pid": 435, "tid": 435, "ts": 1776649791321927, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791321971, "dur": 1, + "args": { + "External id": 7124, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53140, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53140, "pid": 0, "tid": 7, "ts": 1776649791321971, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321961, "dur": 6, + "args": { + "External id": 7124, + "cbid": 211, "correlation": 53140 + } + }, + { + "ph": "s", "id": 53140, "pid": 435, "tid": 435, "ts": 1776649791321961, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321974, "dur": 0, + "args": { + "External id": 53146, + "cbid": 200, "correlation": 53146 + } + }, + { + "ph": "f", "id": 53146, "pid": 435, "tid": 435, "ts": 1776649791321974, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791321974, "dur": 0, + "args": { + "External id": 53147, + "cbid": 273, "correlation": 53147 + } + }, + { + "ph": "f", "id": 53147, "pid": 435, "tid": 435, "ts": 1776649791321974, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791321983, "dur": 16, + "args": { + "External id": 53148, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53148, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53148, "pid": 0, "tid": 7, "ts": 1776649791321983, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791321975, "dur": 4, + "args": { + "External id": 53148, + "cbid": 211, "correlation": 53148 + } + }, + { + "ph": "s", "id": 53148, "pid": 435, "tid": 435, "ts": 1776649791321975, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::FillFunctor, at::detail::Array >(int, at::native::FillFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791322019, "dur": 1, + "args": { + "External id": 7129, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53166, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 0.009259, + "warps per SM": 0.037037, + "grid": [1, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53166, "pid": 0, "tid": 7, "ts": 1776649791322019, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791322009, "dur": 6, + "args": { + "External id": 7129, + "cbid": 211, "correlation": 53166 + } + }, + { + "ph": "s", "id": 53166, "pid": 435, "tid": 435, "ts": 1776649791322009, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceGetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791322022, "dur": 0, + "args": { + "External id": 53172, + "cbid": 200, "correlation": 53172 + } + }, + { + "ph": "f", "id": 53172, "pid": 435, "tid": 435, "ts": 1776649791322022, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaFuncSetAttribute", "pid": 435, "tid": 435, + "ts": 1776649791322022, "dur": 0, + "args": { + "External id": 53173, + "cbid": 273, "correlation": 53173 + } + }, + { + "ph": "f", "id": 53173, "pid": 435, "tid": 435, "ts": 1776649791322022, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void Marlin<256, 1, 8, 8, 4, 8>(int4 const*, int4 const*, int4*, int4 const*, int, int, int, int*)", "pid": 0, "tid": 7, + "ts": 1776649791322032, "dur": 15, + "args": { + "External id": 53174, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53174, + "registers per thread": 128, + "shared memory": 98304, + "blocks per SM": 1.000000, + "warps per SM": 8.000000, + "grid": [108, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53174, "pid": 0, "tid": 7, "ts": 1776649791322032, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791322024, "dur": 5, + "args": { + "External id": 53174, + "cbid": 211, "correlation": 53174 + } + }, + { + "ph": "s", "id": 53174, "pid": 435, "tid": 435, "ts": 1776649791322024, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791322054, "dur": 4, + "args": { + "External id": 53182, + "cbid": 135, "correlation": 53182 + } + }, + { + "ph": "f", "id": 53182, "pid": 435, "tid": 435, "ts": 1776649791322054, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791322069, "dur": 5, + "args": { + "External id": 53187, + "cbid": 165, "correlation": 53187 + } + }, + { + "ph": "s", "id": 53187, "pid": 435, "tid": 435, "ts": 1776649791322069, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649791322077, "dur": 4, + "args": { + "External id": 53190, + "cbid": 139, "correlation": 53190 + } + }, + { + "ph": "f", "id": 53190, "pid": 435, "tid": 435, "ts": 1776649791322077, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649791322254, "dur": 2, + "args": { + "External id": 7132, + "cbid": 317, "correlation": 53215 + } + }, + { + "ph": "f", "id": 53215, "pid": 435, "tid": 435, "ts": 1776649791322254, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649791322258, "dur": 0, + "args": { + "External id": 7132, + "cbid": 317, "correlation": 53219 + } + }, + { + "ph": "f", "id": 53219, "pid": 435, "tid": 435, "ts": 1776649791322258, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::distribution_elementwise_grid_stride_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::(anonymous namespace)::distribution_nullary_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2} const&, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(int, float)#1}>(int, at::PhiloxCudaState, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::(anonymous namespace)::distribution_nullary_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2} const&, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(int, float)#1})", "pid": 0, "tid": 7, + "ts": 1776649791322275, "dur": 3, + "args": { + "External id": 7132, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53221, + "registers per thread": 38, + "shared memory": 0, + "blocks per SM": 0.148148, + "warps per SM": 1.185185, + "grid": [16, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 2 + } + }, + { + "ph": "f", "id": 53221, "pid": 0, "tid": 7, "ts": 1776649791322275, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791322261, "dur": 13, + "args": { + "External id": 7132, + "cbid": 211, "correlation": 53221 + } + }, + { + "ph": "s", "id": 53221, "pid": 435, "tid": 435, "ts": 1776649791322261, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649791322288, "dur": 0, + "args": { + "External id": 53228, + "cbid": 317, "correlation": 53228 + } + }, + { + "ph": "f", "id": 53228, "pid": 435, "tid": 435, "ts": 1776649791322288, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649791322294, "dur": 0, + "args": { + "External id": 53232, + "cbid": 317, "correlation": 53232 + } + }, + { + "ph": "f", "id": 53232, "pid": 435, "tid": 435, "ts": 1776649791322294, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649791322319, "dur": 0, + "args": { + "External id": 7135, + "cbid": 317, "correlation": 53247 + } + }, + { + "ph": "f", "id": 53247, "pid": 435, "tid": 435, "ts": 1776649791322319, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649791322321, "dur": 0, + "args": { + "External id": 7135, + "cbid": 317, "correlation": 53251 + } + }, + { + "ph": "f", "id": 53251, "pid": 435, "tid": 435, "ts": 1776649791322321, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::distribution_elementwise_grid_stride_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::(anonymous namespace)::distribution_nullary_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2} const&, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(int, float)#1}>(int, at::PhiloxCudaState, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::(anonymous namespace)::distribution_nullary_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2}, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_and_transform(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1}>(at::TensorIteratorBase&, at::CUDAGeneratorImpl*, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(curandStatePhilox4_32_10*)#2} const&, at::native::templates::cuda::normal_kernel(at::TensorBase const&, double, double, at::CUDAGeneratorImpl*)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(float)#1})::{lambda(int, float)#1})", "pid": 0, "tid": 7, + "ts": 1776649791322333, "dur": 165, + "args": { + "External id": 7135, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53253, + "registers per thread": 38, + "shared memory": 0, + "blocks per SM": 8.000000, + "warps per SM": 64.000000, + "grid": [864, 1, 1], + "block": [256, 1, 1], + "est. achieved occupancy %": 75 + } + }, + { + "ph": "f", "id": 53253, "pid": 0, "tid": 7, "ts": 1776649791322333, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791322323, "dur": 8, + "args": { + "External id": 7135, + "cbid": 211, "correlation": 53253 + } + }, + { + "ph": "s", "id": 53253, "pid": 435, "tid": 435, "ts": 1776649791322323, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor >, at::detail::Array >(int, at::native::AUnaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791322499, "dur": 78, + "args": { + "External id": 7136, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53263, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 53263, "pid": 0, "tid": 7, "ts": 1776649791322499, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791322391, "dur": 10, + "args": { + "External id": 7136, + "cbid": 211, "correlation": 53263 + } + }, + { + "ph": "s", "id": 53263, "pid": 435, "tid": 435, "ts": 1776649791322391, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AbsFunctor, at::detail::Array >(int, at::native::AbsFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791322580, "dur": 77, + "args": { + "External id": 7141, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53283, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 53283, "pid": 0, "tid": 7, "ts": 1776649791322580, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791322494, "dur": 10, + "args": { + "External id": 7141, + "cbid": 211, "correlation": 53283 + } + }, + { + "ph": "s", "id": 53283, "pid": 435, "tid": 435, "ts": 1776649791322494, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::reduce_kernel<128, 4, at::native::ReduceOp >, unsigned int, c10::Half, 4> >(at::native::ReduceOp >, unsigned int, c10::Half, 4>)", "pid": 0, "tid": 7, + "ts": 1776649791322658, "dur": 72, + "args": { + "External id": 7143, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53296, + "registers per thread": 49, + "shared memory": 1040, + "blocks per SM": 18.962963, + "warps per SM": 75.851852, + "grid": [2048, 1, 1], + "block": [32, 4, 1], + "est. achieved occupancy %": 56 + } + }, + { + "ph": "f", "id": 53296, "pid": 0, "tid": 7, "ts": 1776649791322658, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791322559, "dur": 11, + "args": { + "External id": 7143, + "cbid": 211, "correlation": 53296 + } + }, + { + "ph": "s", "id": 53296, "pid": 435, "tid": 435, "ts": 1776649791322559, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BUnaryFunctor >, at::detail::Array >(int, at::native::BUnaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791322732, "dur": 3, + "args": { + "External id": 7145, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53306, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 4.740741, + "warps per SM": 18.962963, + "grid": [512, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 30 + } + }, + { + "ph": "f", "id": 53306, "pid": 0, "tid": 7, "ts": 1776649791322732, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791322599, "dur": 9, + "args": { + "External id": 7145, + "cbid": 211, "correlation": 53306 + } + }, + { + "ph": "s", "id": 53306, "pid": 435, "tid": 435, "ts": 1776649791322599, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#8}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#8}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791322737, "dur": 3, + "args": { + "External id": 7146, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53316, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 4.740741, + "warps per SM": 18.962963, + "grid": [512, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 30 + } + }, + { + "ph": "f", "id": 53316, "pid": 0, "tid": 7, "ts": 1776649791322737, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791322641, "dur": 9, + "args": { + "External id": 7146, + "cbid": 211, "correlation": 53316 + } + }, + { + "ph": "s", "id": 53316, "pid": 435, "tid": 435, "ts": 1776649791322641, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649791322742, "dur": 193, + "args": { + "External id": 7151, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53326, + "registers per thread": 20, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 53326, "pid": 0, "tid": 7, "ts": 1776649791322742, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791322721, "dur": 9, + "args": { + "External id": 7151, + "cbid": 211, "correlation": 53326 + } + }, + { + "ph": "s", "id": 53326, "pid": 435, "tid": 435, "ts": 1776649791322721, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaStreamIsCapturing", "pid": 435, "tid": 435, + "ts": 1776649791322750, "dur": 0, + "args": { + "External id": 7152, + "cbid": 317, "correlation": 53335 + } + }, + { + "ph": "f", "id": 53335, "pid": 435, "tid": 435, "ts": 1776649791322750, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaMalloc", "pid": 435, "tid": 435, + "ts": 1776649791322751, "dur": 508, + "args": { + "External id": 7152, + "cbid": 20, "correlation": 53336 + } + }, + { + "ph": "f", "id": 53336, "pid": 435, "tid": 435, "ts": 1776649791322751, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array >(int, at::native::round_kernel_cuda(at::TensorIteratorBase&)::{lambda()#1}::operator()() const::{lambda()#3}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791323293, "dur": 77, + "args": { + "External id": 7152, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53340, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 53340, "pid": 0, "tid": 7, "ts": 1776649791323293, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791323275, "dur": 14, + "args": { + "External id": 7152, + "cbid": 211, "correlation": 53340 + } + }, + { + "ph": "s", "id": 53340, "pid": 435, "tid": 435, "ts": 1776649791323275, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#8}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array >(int, at::native::(anonymous namespace)::launch_clamp_scalar(at::TensorIteratorBase&, c10::Scalar, c10::Scalar, at::native::detail::ClampLimits)::{lambda()#1}::operator()() const::{lambda()#8}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791323372, "dur": 80, + "args": { + "External id": 7153, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53350, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 53350, "pid": 0, "tid": 7, "ts": 1776649791323372, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791323323, "dur": 7, + "args": { + "External id": 7153, + "cbid": 211, "correlation": 53350 + } + }, + { + "ph": "s", "id": 53350, "pid": 435, "tid": 435, "ts": 1776649791323323, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649791323453, "dur": 183, + "args": { + "External id": 7158, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53363, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 53363, "pid": 0, "tid": 7, "ts": 1776649791323453, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791323373, "dur": 8, + "args": { + "External id": 7158, + "cbid": 211, "correlation": 53363 + } + }, + { + "ph": "s", "id": 53363, "pid": 435, "tid": 435, "ts": 1776649791323373, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, at::detail::Array >(int, at::native::CUDAFunctorOnSelf_add, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791323638, "dur": 63, + "args": { + "External id": 7159, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53374, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 53374, "pid": 0, "tid": 7, "ts": 1776649791323638, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791323420, "dur": 9, + "args": { + "External id": 7159, + "cbid": 211, "correlation": 53374 + } + }, + { + "ph": "s", "id": 53374, "pid": 435, "tid": 435, "ts": 1776649791323420, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#1}::operator()() const::{lambda(unsigned char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649791323703, "dur": 111, + "args": { + "External id": 7163, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53387, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 53387, "pid": 0, "tid": 7, "ts": 1776649791323703, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791323452, "dur": 8, + "args": { + "External id": 7163, + "cbid": 211, "correlation": 53387 + } + }, + { + "ph": "s", "id": 53387, "pid": 435, "tid": 435, "ts": 1776649791323452, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast >(at::TensorIteratorBase&, at::native::BUnaryFunctor const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast >(at::TensorIteratorBase&, at::native::BUnaryFunctor const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649791323816, "dur": 77, + "args": { + "External id": 7174, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53407, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 53407, "pid": 0, "tid": 7, "ts": 1776649791323816, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791323569, "dur": 11, + "args": { + "External id": 7174, + "cbid": 211, "correlation": 53407 + } + }, + { + "ph": "s", "id": 53407, "pid": 435, "tid": 435, "ts": 1776649791323569, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast > >(at::TensorIteratorBase&, at::native::BinaryFunctor > const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649791323895, "dur": 83, + "args": { + "External id": 7177, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53417, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 53417, "pid": 0, "tid": 7, "ts": 1776649791323895, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791323610, "dur": 8, + "args": { + "External id": 7177, + "cbid": 211, "correlation": 53417 + } + }, + { + "ph": "s", "id": 53417, "pid": 435, "tid": 435, "ts": 1776649791323610, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::AUnaryFunctor >, at::detail::Array >(int, at::native::AUnaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791323979, "dur": 33, + "args": { + "External id": 7181, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53427, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 53427, "pid": 0, "tid": 7, "ts": 1776649791323979, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791323667, "dur": 8, + "args": { + "External id": 7181, + "cbid": 211, "correlation": 53427 + } + }, + { + "ph": "s", "id": 53427, "pid": 435, "tid": 435, "ts": 1776649791323667, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649791324014, "dur": 57, + "args": { + "External id": 7185, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53440, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 53440, "pid": 0, "tid": 7, "ts": 1776649791324014, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791323700, "dur": 7, + "args": { + "External id": 7185, + "cbid": 211, "correlation": 53440 + } + }, + { + "ph": "s", "id": 53440, "pid": 435, "tid": 435, "ts": 1776649791323700, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BUnaryFunctor, at::detail::Array >(int, at::native::BUnaryFunctor, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791324073, "dur": 33, + "args": { + "External id": 7186, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53454, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 53454, "pid": 0, "tid": 7, "ts": 1776649791324073, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791323735, "dur": 8, + "args": { + "External id": 7186, + "cbid": 211, "correlation": 53454 + } + }, + { + "ph": "s", "id": 53454, "pid": 435, "tid": 435, "ts": 1776649791323735, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#2}::operator()() const::{lambda(signed char)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649791324107, "dur": 57, + "args": { + "External id": 7191, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53467, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 303.407410, + "warps per SM": 1213.629639, + "grid": [32768, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 53467, "pid": 0, "tid": 7, "ts": 1776649791324107, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791323762, "dur": 6, + "args": { + "External id": 7191, + "cbid": 211, "correlation": 53467 + } + }, + { + "ph": "s", "id": 53467, "pid": 435, "tid": 435, "ts": 1776649791323762, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::CatArrayBatchedCopy_contig, unsigned int, 3, 128, 1>(at::native::(anonymous namespace)::OpaqueType<1u>*, at::native::(anonymous namespace)::CatArrInputTensorMetadata, unsigned int, 128, 1>, at::native::(anonymous namespace)::TensorSizeStride, int, unsigned int)", "pid": 0, "tid": 7, + "ts": 1776649791324166, "dur": 125, + "args": { + "External id": 7197, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53479, + "registers per thread": 21, + "shared memory": 0, + "blocks per SM": 4.000000, + "warps per SM": 64.000000, + "grid": [216, 2, 1], + "block": [512, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 53479, "pid": 0, "tid": 7, "ts": 1776649791324166, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791323844, "dur": 11, + "args": { + "External id": 7197, + "cbid": 211, "correlation": 53479 + } + }, + { + "ph": "s", "id": 53479, "pid": 435, "tid": 435, "ts": 1776649791323844, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 4, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast(at::TensorIteratorBase&, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1} const&)::{lambda(int)#1})", "pid": 0, "tid": 7, + "ts": 1776649791324293, "dur": 136, + "args": { + "External id": 7208, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53492, + "registers per thread": 16, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 53492, "pid": 0, "tid": 7, "ts": 1776649791324293, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791323924, "dur": 10, + "args": { + "External id": 7208, + "cbid": 211, "correlation": 53492 + } + }, + { + "ph": "s", "id": 53492, "pid": 435, "tid": 435, "ts": 1776649791323924, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::CUDAFunctorOnSelf_add, at::detail::Array >(int, at::native::CUDAFunctorOnSelf_add, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791324431, "dur": 63, + "args": { + "External id": 7211, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53503, + "registers per thread": 18, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 53503, "pid": 0, "tid": 7, "ts": 1776649791324431, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791323967, "dur": 7, + "args": { + "External id": 7211, + "cbid": 211, "correlation": 53503 + } + }, + { + "ph": "s", "id": 53503, "pid": 435, "tid": 435, "ts": 1776649791323967, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::unrolled_elementwise_kernel, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1> >(int, at::native::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#10}::operator()() const::{lambda(c10::Half)#1}, at::detail::Array, TrivialOffsetCalculator<1, unsigned int>, TrivialOffsetCalculator<1, unsigned int>, at::native::memory::LoadWithCast<1>, at::native::memory::StoreWithCast<1>)", "pid": 0, "tid": 7, + "ts": 1776649791324496, "dur": 179, + "args": { + "External id": 7215, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53516, + "registers per thread": 32, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 53516, "pid": 0, "tid": 7, "ts": 1776649791324496, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791323997, "dur": 8, + "args": { + "External id": 7215, + "cbid": 211, "correlation": 53516 + } + }, + { + "ph": "s", "id": 53516, "pid": 435, "tid": 435, "ts": 1776649791323997, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "kernel", "name": "void at::native::vectorized_elementwise_kernel<4, at::native::BinaryFunctor >, at::detail::Array >(int, at::native::BinaryFunctor >, at::detail::Array)", "pid": 0, "tid": 7, + "ts": 1776649791324677, "dur": 114, + "args": { + "External id": 7216, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53527, + "registers per thread": 22, + "shared memory": 0, + "blocks per SM": 606.814819, + "warps per SM": 2427.259277, + "grid": [65536, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 100 + } + }, + { + "ph": "f", "id": 53527, "pid": 0, "tid": 7, "ts": 1776649791324677, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791324025, "dur": 7, + "args": { + "External id": 7216, + "cbid": 211, "correlation": 53527 + } + }, + { + "ph": "s", "id": 53527, "pid": 435, "tid": 435, "ts": 1776649791324025, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791324444, "dur": 8, + "args": { + "External id": 7218, + "cbid": 210, "correlation": 53554 + } + }, + { + "ph": "f", "id": 53554, "pid": 435, "tid": 435, "ts": 1776649791324444, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791324463, "dur": 0, + "args": { + "External id": 7218, + "cbid": 210, "correlation": 53555 + } + }, + { + "ph": "f", "id": 53555, "pid": 435, "tid": 435, "ts": 1776649791324463, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791324793, "dur": 50, + "args": { + "External id": 7218, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53556, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53556, "pid": 0, "tid": 7, "ts": 1776649791324793, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791324487, "dur": 16, + "args": { + "External id": 7218, + "cbid": 211, "correlation": 53556 + } + }, + { + "ph": "s", "id": 53556, "pid": 435, "tid": 435, "ts": 1776649791324487, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791324544, "dur": 0, + "args": { + "External id": 7220, + "cbid": 210, "correlation": 53583 + } + }, + { + "ph": "f", "id": 53583, "pid": 435, "tid": 435, "ts": 1776649791324544, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791324844, "dur": 50, + "args": { + "External id": 7220, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53584, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53584, "pid": 0, "tid": 7, "ts": 1776649791324844, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791324545, "dur": 6, + "args": { + "External id": 7220, + "cbid": 211, "correlation": 53584 + } + }, + { + "ph": "s", "id": 53584, "pid": 435, "tid": 435, "ts": 1776649791324545, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791324577, "dur": 0, + "args": { + "External id": 7222, + "cbid": 210, "correlation": 53611 + } + }, + { + "ph": "f", "id": 53611, "pid": 435, "tid": 435, "ts": 1776649791324577, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791324896, "dur": 49, + "args": { + "External id": 7222, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53612, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53612, "pid": 0, "tid": 7, "ts": 1776649791324896, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791324578, "dur": 6, + "args": { + "External id": 7222, + "cbid": 211, "correlation": 53612 + } + }, + { + "ph": "s", "id": 53612, "pid": 435, "tid": 435, "ts": 1776649791324578, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791324608, "dur": 0, + "args": { + "External id": 7224, + "cbid": 210, "correlation": 53639 + } + }, + { + "ph": "f", "id": 53639, "pid": 435, "tid": 435, "ts": 1776649791324608, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791324947, "dur": 50, + "args": { + "External id": 7224, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53640, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53640, "pid": 0, "tid": 7, "ts": 1776649791324947, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791324610, "dur": 5, + "args": { + "External id": 7224, + "cbid": 211, "correlation": 53640 + } + }, + { + "ph": "s", "id": 53640, "pid": 435, "tid": 435, "ts": 1776649791324610, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791324638, "dur": 0, + "args": { + "External id": 7226, + "cbid": 210, "correlation": 53667 + } + }, + { + "ph": "f", "id": 53667, "pid": 435, "tid": 435, "ts": 1776649791324638, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791324999, "dur": 51, + "args": { + "External id": 7226, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53668, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53668, "pid": 0, "tid": 7, "ts": 1776649791324999, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791324639, "dur": 6, + "args": { + "External id": 7226, + "cbid": 211, "correlation": 53668 + } + }, + { + "ph": "s", "id": 53668, "pid": 435, "tid": 435, "ts": 1776649791324639, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791324669, "dur": 0, + "args": { + "External id": 7228, + "cbid": 210, "correlation": 53695 + } + }, + { + "ph": "f", "id": 53695, "pid": 435, "tid": 435, "ts": 1776649791324669, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791325052, "dur": 49, + "args": { + "External id": 7228, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53696, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53696, "pid": 0, "tid": 7, "ts": 1776649791325052, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791324670, "dur": 6, + "args": { + "External id": 7228, + "cbid": 211, "correlation": 53696 + } + }, + { + "ph": "s", "id": 53696, "pid": 435, "tid": 435, "ts": 1776649791324670, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791324699, "dur": 0, + "args": { + "External id": 7230, + "cbid": 210, "correlation": 53723 + } + }, + { + "ph": "f", "id": 53723, "pid": 435, "tid": 435, "ts": 1776649791324699, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791325103, "dur": 50, + "args": { + "External id": 7230, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53724, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53724, "pid": 0, "tid": 7, "ts": 1776649791325103, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791324700, "dur": 5, + "args": { + "External id": 7230, + "cbid": 211, "correlation": 53724 + } + }, + { + "ph": "s", "id": 53724, "pid": 435, "tid": 435, "ts": 1776649791324700, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791324729, "dur": 0, + "args": { + "External id": 7232, + "cbid": 210, "correlation": 53751 + } + }, + { + "ph": "f", "id": 53751, "pid": 435, "tid": 435, "ts": 1776649791324729, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791325155, "dur": 52, + "args": { + "External id": 7232, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53752, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53752, "pid": 0, "tid": 7, "ts": 1776649791325155, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791324730, "dur": 5, + "args": { + "External id": 7232, + "cbid": 211, "correlation": 53752 + } + }, + { + "ph": "s", "id": 53752, "pid": 435, "tid": 435, "ts": 1776649791324730, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791324758, "dur": 0, + "args": { + "External id": 7234, + "cbid": 210, "correlation": 53779 + } + }, + { + "ph": "f", "id": 53779, "pid": 435, "tid": 435, "ts": 1776649791324758, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791325208, "dur": 51, + "args": { + "External id": 7234, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53780, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53780, "pid": 0, "tid": 7, "ts": 1776649791325208, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791324760, "dur": 5, + "args": { + "External id": 7234, + "cbid": 211, "correlation": 53780 + } + }, + { + "ph": "s", "id": 53780, "pid": 435, "tid": 435, "ts": 1776649791324760, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791324788, "dur": 0, + "args": { + "External id": 7236, + "cbid": 210, "correlation": 53807 + } + }, + { + "ph": "f", "id": 53807, "pid": 435, "tid": 435, "ts": 1776649791324788, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791325262, "dur": 41, + "args": { + "External id": 7236, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53808, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53808, "pid": 0, "tid": 7, "ts": 1776649791325262, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791324789, "dur": 6, + "args": { + "External id": 7236, + "cbid": 211, "correlation": 53808 + } + }, + { + "ph": "s", "id": 53808, "pid": 435, "tid": 435, "ts": 1776649791324789, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791324821, "dur": 487, + "args": { + "External id": 53814, + "cbid": 165, "correlation": 53814 + } + }, + { + "ph": "s", "id": 53814, "pid": 435, "tid": 435, "ts": 1776649791324821, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791325337, "dur": 4, + "args": { + "External id": 53823, + "cbid": 135, "correlation": 53823 + } + }, + { + "ph": "f", "id": 53823, "pid": 435, "tid": 435, "ts": 1776649791325337, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791325372, "dur": 1, + "args": { + "External id": 7238, + "cbid": 210, "correlation": 53849 + } + }, + { + "ph": "f", "id": 53849, "pid": 435, "tid": 435, "ts": 1776649791325372, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791325384, "dur": 41, + "args": { + "External id": 7238, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53850, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53850, "pid": 0, "tid": 7, "ts": 1776649791325384, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791325374, "dur": 8, + "args": { + "External id": 7238, + "cbid": 211, "correlation": 53850 + } + }, + { + "ph": "s", "id": 53850, "pid": 435, "tid": 435, "ts": 1776649791325374, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791325408, "dur": 0, + "args": { + "External id": 7240, + "cbid": 210, "correlation": 53877 + } + }, + { + "ph": "f", "id": 53877, "pid": 435, "tid": 435, "ts": 1776649791325408, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791325427, "dur": 41, + "args": { + "External id": 7240, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53878, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53878, "pid": 0, "tid": 7, "ts": 1776649791325427, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791325409, "dur": 6, + "args": { + "External id": 7240, + "cbid": 211, "correlation": 53878 + } + }, + { + "ph": "s", "id": 53878, "pid": 435, "tid": 435, "ts": 1776649791325409, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791325439, "dur": 0, + "args": { + "External id": 7242, + "cbid": 210, "correlation": 53905 + } + }, + { + "ph": "f", "id": 53905, "pid": 435, "tid": 435, "ts": 1776649791325439, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791325470, "dur": 41, + "args": { + "External id": 7242, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53906, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53906, "pid": 0, "tid": 7, "ts": 1776649791325470, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791325441, "dur": 6, + "args": { + "External id": 7242, + "cbid": 211, "correlation": 53906 + } + }, + { + "ph": "s", "id": 53906, "pid": 435, "tid": 435, "ts": 1776649791325441, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791325470, "dur": 0, + "args": { + "External id": 7244, + "cbid": 210, "correlation": 53933 + } + }, + { + "ph": "f", "id": 53933, "pid": 435, "tid": 435, "ts": 1776649791325470, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791325512, "dur": 41, + "args": { + "External id": 7244, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53934, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53934, "pid": 0, "tid": 7, "ts": 1776649791325512, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791325471, "dur": 6, + "args": { + "External id": 7244, + "cbid": 211, "correlation": 53934 + } + }, + { + "ph": "s", "id": 53934, "pid": 435, "tid": 435, "ts": 1776649791325471, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791325500, "dur": 0, + "args": { + "External id": 7246, + "cbid": 210, "correlation": 53961 + } + }, + { + "ph": "f", "id": 53961, "pid": 435, "tid": 435, "ts": 1776649791325500, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791325555, "dur": 46, + "args": { + "External id": 7246, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53962, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53962, "pid": 0, "tid": 7, "ts": 1776649791325555, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791325502, "dur": 6, + "args": { + "External id": 7246, + "cbid": 211, "correlation": 53962 + } + }, + { + "ph": "s", "id": 53962, "pid": 435, "tid": 435, "ts": 1776649791325502, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791325531, "dur": 0, + "args": { + "External id": 7248, + "cbid": 210, "correlation": 53989 + } + }, + { + "ph": "f", "id": 53989, "pid": 435, "tid": 435, "ts": 1776649791325531, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791325603, "dur": 45, + "args": { + "External id": 7248, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 53990, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 53990, "pid": 0, "tid": 7, "ts": 1776649791325603, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791325532, "dur": 6, + "args": { + "External id": 7248, + "cbid": 211, "correlation": 53990 + } + }, + { + "ph": "s", "id": 53990, "pid": 435, "tid": 435, "ts": 1776649791325532, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791325561, "dur": 0, + "args": { + "External id": 7250, + "cbid": 210, "correlation": 54017 + } + }, + { + "ph": "f", "id": 54017, "pid": 435, "tid": 435, "ts": 1776649791325561, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791325649, "dur": 50, + "args": { + "External id": 7250, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54018, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54018, "pid": 0, "tid": 7, "ts": 1776649791325649, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791325562, "dur": 5, + "args": { + "External id": 7250, + "cbid": 211, "correlation": 54018 + } + }, + { + "ph": "s", "id": 54018, "pid": 435, "tid": 435, "ts": 1776649791325562, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791325591, "dur": 0, + "args": { + "External id": 7252, + "cbid": 210, "correlation": 54045 + } + }, + { + "ph": "f", "id": 54045, "pid": 435, "tid": 435, "ts": 1776649791325591, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791325701, "dur": 44, + "args": { + "External id": 7252, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54046, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54046, "pid": 0, "tid": 7, "ts": 1776649791325701, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791325592, "dur": 5, + "args": { + "External id": 7252, + "cbid": 211, "correlation": 54046 + } + }, + { + "ph": "s", "id": 54046, "pid": 435, "tid": 435, "ts": 1776649791325592, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791325620, "dur": 0, + "args": { + "External id": 7254, + "cbid": 210, "correlation": 54073 + } + }, + { + "ph": "f", "id": 54073, "pid": 435, "tid": 435, "ts": 1776649791325620, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791325747, "dur": 44, + "args": { + "External id": 7254, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54074, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54074, "pid": 0, "tid": 7, "ts": 1776649791325747, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791325622, "dur": 5, + "args": { + "External id": 7254, + "cbid": 211, "correlation": 54074 + } + }, + { + "ph": "s", "id": 54074, "pid": 435, "tid": 435, "ts": 1776649791325622, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791325651, "dur": 0, + "args": { + "External id": 7256, + "cbid": 210, "correlation": 54101 + } + }, + { + "ph": "f", "id": 54101, "pid": 435, "tid": 435, "ts": 1776649791325651, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791325792, "dur": 43, + "args": { + "External id": 7256, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54102, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54102, "pid": 0, "tid": 7, "ts": 1776649791325792, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791325652, "dur": 6, + "args": { + "External id": 7256, + "cbid": 211, "correlation": 54102 + } + }, + { + "ph": "s", "id": 54102, "pid": 435, "tid": 435, "ts": 1776649791325652, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791325680, "dur": 0, + "args": { + "External id": 7258, + "cbid": 210, "correlation": 54129 + } + }, + { + "ph": "f", "id": 54129, "pid": 435, "tid": 435, "ts": 1776649791325680, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791325837, "dur": 49, + "args": { + "External id": 7258, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54130, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54130, "pid": 0, "tid": 7, "ts": 1776649791325837, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791325682, "dur": 6, + "args": { + "External id": 7258, + "cbid": 211, "correlation": 54130 + } + }, + { + "ph": "s", "id": 54130, "pid": 435, "tid": 435, "ts": 1776649791325682, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791325711, "dur": 0, + "args": { + "External id": 7260, + "cbid": 210, "correlation": 54157 + } + }, + { + "ph": "f", "id": 54157, "pid": 435, "tid": 435, "ts": 1776649791325711, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791325888, "dur": 52, + "args": { + "External id": 7260, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54158, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54158, "pid": 0, "tid": 7, "ts": 1776649791325888, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791325712, "dur": 5, + "args": { + "External id": 7260, + "cbid": 211, "correlation": 54158 + } + }, + { + "ph": "s", "id": 54158, "pid": 435, "tid": 435, "ts": 1776649791325712, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791325740, "dur": 0, + "args": { + "External id": 7262, + "cbid": 210, "correlation": 54185 + } + }, + { + "ph": "f", "id": 54185, "pid": 435, "tid": 435, "ts": 1776649791325740, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791325941, "dur": 49, + "args": { + "External id": 7262, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54186, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54186, "pid": 0, "tid": 7, "ts": 1776649791325941, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791325742, "dur": 6, + "args": { + "External id": 7262, + "cbid": 211, "correlation": 54186 + } + }, + { + "ph": "s", "id": 54186, "pid": 435, "tid": 435, "ts": 1776649791325742, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791325771, "dur": 0, + "args": { + "External id": 7264, + "cbid": 210, "correlation": 54213 + } + }, + { + "ph": "f", "id": 54213, "pid": 435, "tid": 435, "ts": 1776649791325771, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791325992, "dur": 51, + "args": { + "External id": 7264, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54214, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54214, "pid": 0, "tid": 7, "ts": 1776649791325992, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791325772, "dur": 6, + "args": { + "External id": 7264, + "cbid": 211, "correlation": 54214 + } + }, + { + "ph": "s", "id": 54214, "pid": 435, "tid": 435, "ts": 1776649791325772, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791325801, "dur": 0, + "args": { + "External id": 7266, + "cbid": 210, "correlation": 54241 + } + }, + { + "ph": "f", "id": 54241, "pid": 435, "tid": 435, "ts": 1776649791325801, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791326045, "dur": 49, + "args": { + "External id": 7266, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54242, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54242, "pid": 0, "tid": 7, "ts": 1776649791326045, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791325802, "dur": 6, + "args": { + "External id": 7266, + "cbid": 211, "correlation": 54242 + } + }, + { + "ph": "s", "id": 54242, "pid": 435, "tid": 435, "ts": 1776649791325802, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791325831, "dur": 0, + "args": { + "External id": 7268, + "cbid": 210, "correlation": 54269 + } + }, + { + "ph": "f", "id": 54269, "pid": 435, "tid": 435, "ts": 1776649791325831, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791326096, "dur": 52, + "args": { + "External id": 7268, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54270, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54270, "pid": 0, "tid": 7, "ts": 1776649791326096, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791325832, "dur": 5, + "args": { + "External id": 7268, + "cbid": 211, "correlation": 54270 + } + }, + { + "ph": "s", "id": 54270, "pid": 435, "tid": 435, "ts": 1776649791325832, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791325861, "dur": 0, + "args": { + "External id": 7270, + "cbid": 210, "correlation": 54297 + } + }, + { + "ph": "f", "id": 54297, "pid": 435, "tid": 435, "ts": 1776649791325861, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791326150, "dur": 41, + "args": { + "External id": 7270, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54298, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54298, "pid": 0, "tid": 7, "ts": 1776649791326150, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791325863, "dur": 5, + "args": { + "External id": 7270, + "cbid": 211, "correlation": 54298 + } + }, + { + "ph": "s", "id": 54298, "pid": 435, "tid": 435, "ts": 1776649791325863, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791325891, "dur": 0, + "args": { + "External id": 7272, + "cbid": 210, "correlation": 54325 + } + }, + { + "ph": "f", "id": 54325, "pid": 435, "tid": 435, "ts": 1776649791325891, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791326193, "dur": 46, + "args": { + "External id": 7272, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54326, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54326, "pid": 0, "tid": 7, "ts": 1776649791326193, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791325893, "dur": 6, + "args": { + "External id": 7272, + "cbid": 211, "correlation": 54326 + } + }, + { + "ph": "s", "id": 54326, "pid": 435, "tid": 435, "ts": 1776649791325893, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791325922, "dur": 0, + "args": { + "External id": 7274, + "cbid": 210, "correlation": 54353 + } + }, + { + "ph": "f", "id": 54353, "pid": 435, "tid": 435, "ts": 1776649791325922, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791326240, "dur": 43, + "args": { + "External id": 7274, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54354, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54354, "pid": 0, "tid": 7, "ts": 1776649791326240, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791325923, "dur": 5, + "args": { + "External id": 7274, + "cbid": 211, "correlation": 54354 + } + }, + { + "ph": "s", "id": 54354, "pid": 435, "tid": 435, "ts": 1776649791325923, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791325951, "dur": 0, + "args": { + "External id": 7276, + "cbid": 210, "correlation": 54381 + } + }, + { + "ph": "f", "id": 54381, "pid": 435, "tid": 435, "ts": 1776649791325951, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791326285, "dur": 43, + "args": { + "External id": 7276, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54382, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54382, "pid": 0, "tid": 7, "ts": 1776649791326285, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791325953, "dur": 6, + "args": { + "External id": 7276, + "cbid": 211, "correlation": 54382 + } + }, + { + "ph": "s", "id": 54382, "pid": 435, "tid": 435, "ts": 1776649791325953, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791325981, "dur": 0, + "args": { + "External id": 7278, + "cbid": 210, "correlation": 54409 + } + }, + { + "ph": "f", "id": 54409, "pid": 435, "tid": 435, "ts": 1776649791325981, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791326331, "dur": 47, + "args": { + "External id": 7278, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54410, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54410, "pid": 0, "tid": 7, "ts": 1776649791326331, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791325983, "dur": 5, + "args": { + "External id": 7278, + "cbid": 211, "correlation": 54410 + } + }, + { + "ph": "s", "id": 54410, "pid": 435, "tid": 435, "ts": 1776649791325983, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326011, "dur": 0, + "args": { + "External id": 7280, + "cbid": 210, "correlation": 54437 + } + }, + { + "ph": "f", "id": 54437, "pid": 435, "tid": 435, "ts": 1776649791326011, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791326379, "dur": 47, + "args": { + "External id": 7280, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54438, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54438, "pid": 0, "tid": 7, "ts": 1776649791326379, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326012, "dur": 6, + "args": { + "External id": 7280, + "cbid": 211, "correlation": 54438 + } + }, + { + "ph": "s", "id": 54438, "pid": 435, "tid": 435, "ts": 1776649791326012, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326041, "dur": 0, + "args": { + "External id": 7282, + "cbid": 210, "correlation": 54465 + } + }, + { + "ph": "f", "id": 54465, "pid": 435, "tid": 435, "ts": 1776649791326041, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791326428, "dur": 42, + "args": { + "External id": 7282, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54466, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54466, "pid": 0, "tid": 7, "ts": 1776649791326428, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326042, "dur": 6, + "args": { + "External id": 7282, + "cbid": 211, "correlation": 54466 + } + }, + { + "ph": "s", "id": 54466, "pid": 435, "tid": 435, "ts": 1776649791326042, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326072, "dur": 0, + "args": { + "External id": 7284, + "cbid": 210, "correlation": 54493 + } + }, + { + "ph": "f", "id": 54493, "pid": 435, "tid": 435, "ts": 1776649791326072, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791326471, "dur": 47, + "args": { + "External id": 7284, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54494, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54494, "pid": 0, "tid": 7, "ts": 1776649791326471, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326073, "dur": 6, + "args": { + "External id": 7284, + "cbid": 211, "correlation": 54494 + } + }, + { + "ph": "s", "id": 54494, "pid": 435, "tid": 435, "ts": 1776649791326073, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326102, "dur": 0, + "args": { + "External id": 7286, + "cbid": 210, "correlation": 54521 + } + }, + { + "ph": "f", "id": 54521, "pid": 435, "tid": 435, "ts": 1776649791326102, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791326520, "dur": 52, + "args": { + "External id": 7286, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54522, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54522, "pid": 0, "tid": 7, "ts": 1776649791326520, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326103, "dur": 6, + "args": { + "External id": 7286, + "cbid": 211, "correlation": 54522 + } + }, + { + "ph": "s", "id": 54522, "pid": 435, "tid": 435, "ts": 1776649791326103, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326132, "dur": 0, + "args": { + "External id": 7288, + "cbid": 210, "correlation": 54549 + } + }, + { + "ph": "f", "id": 54549, "pid": 435, "tid": 435, "ts": 1776649791326132, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791326573, "dur": 44, + "args": { + "External id": 7288, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54550, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54550, "pid": 0, "tid": 7, "ts": 1776649791326573, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326133, "dur": 6, + "args": { + "External id": 7288, + "cbid": 211, "correlation": 54550 + } + }, + { + "ph": "s", "id": 54550, "pid": 435, "tid": 435, "ts": 1776649791326133, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326162, "dur": 0, + "args": { + "External id": 7290, + "cbid": 210, "correlation": 54577 + } + }, + { + "ph": "f", "id": 54577, "pid": 435, "tid": 435, "ts": 1776649791326162, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791326618, "dur": 49, + "args": { + "External id": 7290, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54578, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54578, "pid": 0, "tid": 7, "ts": 1776649791326618, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326163, "dur": 6, + "args": { + "External id": 7290, + "cbid": 211, "correlation": 54578 + } + }, + { + "ph": "s", "id": 54578, "pid": 435, "tid": 435, "ts": 1776649791326163, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326193, "dur": 0, + "args": { + "External id": 7292, + "cbid": 210, "correlation": 54605 + } + }, + { + "ph": "f", "id": 54605, "pid": 435, "tid": 435, "ts": 1776649791326193, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791326669, "dur": 49, + "args": { + "External id": 7292, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54606, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54606, "pid": 0, "tid": 7, "ts": 1776649791326669, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326194, "dur": 6, + "args": { + "External id": 7292, + "cbid": 211, "correlation": 54606 + } + }, + { + "ph": "s", "id": 54606, "pid": 435, "tid": 435, "ts": 1776649791326194, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326223, "dur": 0, + "args": { + "External id": 7294, + "cbid": 210, "correlation": 54633 + } + }, + { + "ph": "f", "id": 54633, "pid": 435, "tid": 435, "ts": 1776649791326223, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791326721, "dur": 49, + "args": { + "External id": 7294, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54634, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54634, "pid": 0, "tid": 7, "ts": 1776649791326721, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326224, "dur": 6, + "args": { + "External id": 7294, + "cbid": 211, "correlation": 54634 + } + }, + { + "ph": "s", "id": 54634, "pid": 435, "tid": 435, "ts": 1776649791326224, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326253, "dur": 0, + "args": { + "External id": 7296, + "cbid": 210, "correlation": 54661 + } + }, + { + "ph": "f", "id": 54661, "pid": 435, "tid": 435, "ts": 1776649791326253, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791326771, "dur": 53, + "args": { + "External id": 7296, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54662, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54662, "pid": 0, "tid": 7, "ts": 1776649791326771, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326254, "dur": 6, + "args": { + "External id": 7296, + "cbid": 211, "correlation": 54662 + } + }, + { + "ph": "s", "id": 54662, "pid": 435, "tid": 435, "ts": 1776649791326254, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326283, "dur": 0, + "args": { + "External id": 7298, + "cbid": 210, "correlation": 54689 + } + }, + { + "ph": "f", "id": 54689, "pid": 435, "tid": 435, "ts": 1776649791326283, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791326826, "dur": 44, + "args": { + "External id": 7298, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54690, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54690, "pid": 0, "tid": 7, "ts": 1776649791326826, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326285, "dur": 6, + "args": { + "External id": 7298, + "cbid": 211, "correlation": 54690 + } + }, + { + "ph": "s", "id": 54690, "pid": 435, "tid": 435, "ts": 1776649791326285, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326313, "dur": 0, + "args": { + "External id": 7300, + "cbid": 210, "correlation": 54717 + } + }, + { + "ph": "f", "id": 54717, "pid": 435, "tid": 435, "ts": 1776649791326313, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791326871, "dur": 49, + "args": { + "External id": 7300, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54718, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54718, "pid": 0, "tid": 7, "ts": 1776649791326871, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326315, "dur": 6, + "args": { + "External id": 7300, + "cbid": 211, "correlation": 54718 + } + }, + { + "ph": "s", "id": 54718, "pid": 435, "tid": 435, "ts": 1776649791326315, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326343, "dur": 0, + "args": { + "External id": 7302, + "cbid": 210, "correlation": 54745 + } + }, + { + "ph": "f", "id": 54745, "pid": 435, "tid": 435, "ts": 1776649791326343, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791326922, "dur": 47, + "args": { + "External id": 7302, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54746, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54746, "pid": 0, "tid": 7, "ts": 1776649791326922, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326345, "dur": 6, + "args": { + "External id": 7302, + "cbid": 211, "correlation": 54746 + } + }, + { + "ph": "s", "id": 54746, "pid": 435, "tid": 435, "ts": 1776649791326345, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326374, "dur": 0, + "args": { + "External id": 7304, + "cbid": 210, "correlation": 54773 + } + }, + { + "ph": "f", "id": 54773, "pid": 435, "tid": 435, "ts": 1776649791326374, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791326970, "dur": 48, + "args": { + "External id": 7304, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54774, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54774, "pid": 0, "tid": 7, "ts": 1776649791326970, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326375, "dur": 5, + "args": { + "External id": 7304, + "cbid": 211, "correlation": 54774 + } + }, + { + "ph": "s", "id": 54774, "pid": 435, "tid": 435, "ts": 1776649791326375, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326404, "dur": 0, + "args": { + "External id": 7306, + "cbid": 210, "correlation": 54801 + } + }, + { + "ph": "f", "id": 54801, "pid": 435, "tid": 435, "ts": 1776649791326404, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791327021, "dur": 47, + "args": { + "External id": 7306, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54802, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54802, "pid": 0, "tid": 7, "ts": 1776649791327021, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326405, "dur": 6, + "args": { + "External id": 7306, + "cbid": 211, "correlation": 54802 + } + }, + { + "ph": "s", "id": 54802, "pid": 435, "tid": 435, "ts": 1776649791326405, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326434, "dur": 0, + "args": { + "External id": 7308, + "cbid": 210, "correlation": 54829 + } + }, + { + "ph": "f", "id": 54829, "pid": 435, "tid": 435, "ts": 1776649791326434, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791327069, "dur": 44, + "args": { + "External id": 7308, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54830, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54830, "pid": 0, "tid": 7, "ts": 1776649791327069, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326435, "dur": 6, + "args": { + "External id": 7308, + "cbid": 211, "correlation": 54830 + } + }, + { + "ph": "s", "id": 54830, "pid": 435, "tid": 435, "ts": 1776649791326435, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326464, "dur": 0, + "args": { + "External id": 7310, + "cbid": 210, "correlation": 54857 + } + }, + { + "ph": "f", "id": 54857, "pid": 435, "tid": 435, "ts": 1776649791326464, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791327114, "dur": 44, + "args": { + "External id": 7310, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54858, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54858, "pid": 0, "tid": 7, "ts": 1776649791327114, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326465, "dur": 6, + "args": { + "External id": 7310, + "cbid": 211, "correlation": 54858 + } + }, + { + "ph": "s", "id": 54858, "pid": 435, "tid": 435, "ts": 1776649791326465, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326493, "dur": 0, + "args": { + "External id": 7312, + "cbid": 210, "correlation": 54885 + } + }, + { + "ph": "f", "id": 54885, "pid": 435, "tid": 435, "ts": 1776649791326493, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791327160, "dur": 41, + "args": { + "External id": 7312, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54886, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54886, "pid": 0, "tid": 7, "ts": 1776649791327160, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326495, "dur": 6, + "args": { + "External id": 7312, + "cbid": 211, "correlation": 54886 + } + }, + { + "ph": "s", "id": 54886, "pid": 435, "tid": 435, "ts": 1776649791326495, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326523, "dur": 0, + "args": { + "External id": 7314, + "cbid": 210, "correlation": 54913 + } + }, + { + "ph": "f", "id": 54913, "pid": 435, "tid": 435, "ts": 1776649791326523, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791327202, "dur": 50, + "args": { + "External id": 7314, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54914, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54914, "pid": 0, "tid": 7, "ts": 1776649791327202, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326525, "dur": 6, + "args": { + "External id": 7314, + "cbid": 211, "correlation": 54914 + } + }, + { + "ph": "s", "id": 54914, "pid": 435, "tid": 435, "ts": 1776649791326525, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326554, "dur": 0, + "args": { + "External id": 7316, + "cbid": 210, "correlation": 54941 + } + }, + { + "ph": "f", "id": 54941, "pid": 435, "tid": 435, "ts": 1776649791326554, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791327254, "dur": 51, + "args": { + "External id": 7316, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54942, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54942, "pid": 0, "tid": 7, "ts": 1776649791327254, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326555, "dur": 5, + "args": { + "External id": 7316, + "cbid": 211, "correlation": 54942 + } + }, + { + "ph": "s", "id": 54942, "pid": 435, "tid": 435, "ts": 1776649791326555, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326583, "dur": 0, + "args": { + "External id": 7318, + "cbid": 210, "correlation": 54969 + } + }, + { + "ph": "f", "id": 54969, "pid": 435, "tid": 435, "ts": 1776649791326583, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791327306, "dur": 45, + "args": { + "External id": 7318, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54970, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54970, "pid": 0, "tid": 7, "ts": 1776649791327306, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326585, "dur": 5, + "args": { + "External id": 7318, + "cbid": 211, "correlation": 54970 + } + }, + { + "ph": "s", "id": 54970, "pid": 435, "tid": 435, "ts": 1776649791326585, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326613, "dur": 0, + "args": { + "External id": 7320, + "cbid": 210, "correlation": 54997 + } + }, + { + "ph": "f", "id": 54997, "pid": 435, "tid": 435, "ts": 1776649791326613, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791327353, "dur": 44, + "args": { + "External id": 7320, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 54998, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 54998, "pid": 0, "tid": 7, "ts": 1776649791327353, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326614, "dur": 5, + "args": { + "External id": 7320, + "cbid": 211, "correlation": 54998 + } + }, + { + "ph": "s", "id": 54998, "pid": 435, "tid": 435, "ts": 1776649791326614, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326643, "dur": 0, + "args": { + "External id": 7322, + "cbid": 210, "correlation": 55025 + } + }, + { + "ph": "f", "id": 55025, "pid": 435, "tid": 435, "ts": 1776649791326643, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791327398, "dur": 52, + "args": { + "External id": 7322, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55026, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55026, "pid": 0, "tid": 7, "ts": 1776649791327398, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326644, "dur": 6, + "args": { + "External id": 7322, + "cbid": 211, "correlation": 55026 + } + }, + { + "ph": "s", "id": 55026, "pid": 435, "tid": 435, "ts": 1776649791326644, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326673, "dur": 0, + "args": { + "External id": 7324, + "cbid": 210, "correlation": 55053 + } + }, + { + "ph": "f", "id": 55053, "pid": 435, "tid": 435, "ts": 1776649791326673, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791327452, "dur": 48, + "args": { + "External id": 7324, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55054, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55054, "pid": 0, "tid": 7, "ts": 1776649791327452, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326674, "dur": 6, + "args": { + "External id": 7324, + "cbid": 211, "correlation": 55054 + } + }, + { + "ph": "s", "id": 55054, "pid": 435, "tid": 435, "ts": 1776649791326674, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326703, "dur": 0, + "args": { + "External id": 7326, + "cbid": 210, "correlation": 55081 + } + }, + { + "ph": "f", "id": 55081, "pid": 435, "tid": 435, "ts": 1776649791326703, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791327502, "dur": 49, + "args": { + "External id": 7326, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55082, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55082, "pid": 0, "tid": 7, "ts": 1776649791327502, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326704, "dur": 5, + "args": { + "External id": 7326, + "cbid": 211, "correlation": 55082 + } + }, + { + "ph": "s", "id": 55082, "pid": 435, "tid": 435, "ts": 1776649791326704, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326733, "dur": 0, + "args": { + "External id": 7328, + "cbid": 210, "correlation": 55109 + } + }, + { + "ph": "f", "id": 55109, "pid": 435, "tid": 435, "ts": 1776649791326733, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791327553, "dur": 46, + "args": { + "External id": 7328, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55110, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55110, "pid": 0, "tid": 7, "ts": 1776649791327553, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326734, "dur": 5, + "args": { + "External id": 7328, + "cbid": 211, "correlation": 55110 + } + }, + { + "ph": "s", "id": 55110, "pid": 435, "tid": 435, "ts": 1776649791326734, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326762, "dur": 0, + "args": { + "External id": 7330, + "cbid": 210, "correlation": 55137 + } + }, + { + "ph": "f", "id": 55137, "pid": 435, "tid": 435, "ts": 1776649791326762, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791327602, "dur": 46, + "args": { + "External id": 7330, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55138, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55138, "pid": 0, "tid": 7, "ts": 1776649791327602, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326764, "dur": 6, + "args": { + "External id": 7330, + "cbid": 211, "correlation": 55138 + } + }, + { + "ph": "s", "id": 55138, "pid": 435, "tid": 435, "ts": 1776649791326764, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326792, "dur": 0, + "args": { + "External id": 7332, + "cbid": 210, "correlation": 55165 + } + }, + { + "ph": "f", "id": 55165, "pid": 435, "tid": 435, "ts": 1776649791326792, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791327650, "dur": 48, + "args": { + "External id": 7332, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55166, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55166, "pid": 0, "tid": 7, "ts": 1776649791327650, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326793, "dur": 6, + "args": { + "External id": 7332, + "cbid": 211, "correlation": 55166 + } + }, + { + "ph": "s", "id": 55166, "pid": 435, "tid": 435, "ts": 1776649791326793, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326822, "dur": 0, + "args": { + "External id": 7334, + "cbid": 210, "correlation": 55193 + } + }, + { + "ph": "f", "id": 55193, "pid": 435, "tid": 435, "ts": 1776649791326822, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791327700, "dur": 49, + "args": { + "External id": 7334, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55194, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55194, "pid": 0, "tid": 7, "ts": 1776649791327700, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326823, "dur": 6, + "args": { + "External id": 7334, + "cbid": 211, "correlation": 55194 + } + }, + { + "ph": "s", "id": 55194, "pid": 435, "tid": 435, "ts": 1776649791326823, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791326852, "dur": 0, + "args": { + "External id": 7336, + "cbid": 210, "correlation": 55221 + } + }, + { + "ph": "f", "id": 55221, "pid": 435, "tid": 435, "ts": 1776649791326852, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791327751, "dur": 43, + "args": { + "External id": 7336, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55222, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55222, "pid": 0, "tid": 7, "ts": 1776649791327751, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791326853, "dur": 6, + "args": { + "External id": 7336, + "cbid": 211, "correlation": 55222 + } + }, + { + "ph": "s", "id": 55222, "pid": 435, "tid": 435, "ts": 1776649791326853, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791326888, "dur": 4, + "args": { + "External id": 55233, + "cbid": 135, "correlation": 55233 + } + }, + { + "ph": "f", "id": 55233, "pid": 435, "tid": 435, "ts": 1776649791326888, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791326904, "dur": 900, + "args": { + "External id": 55238, + "cbid": 165, "correlation": 55238 + } + }, + { + "ph": "s", "id": 55238, "pid": 435, "tid": 435, "ts": 1776649791326904, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649791327808, "dur": 5, + "args": { + "External id": 55241, + "cbid": 139, "correlation": 55241 + } + }, + { + "ph": "f", "id": 55241, "pid": 435, "tid": 435, "ts": 1776649791327808, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791327847, "dur": 1, + "args": { + "External id": 7338, + "cbid": 210, "correlation": 55275 + } + }, + { + "ph": "f", "id": 55275, "pid": 435, "tid": 435, "ts": 1776649791327847, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791327859, "dur": 41, + "args": { + "External id": 7338, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55276, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55276, "pid": 0, "tid": 7, "ts": 1776649791327859, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791327849, "dur": 8, + "args": { + "External id": 7338, + "cbid": 211, "correlation": 55276 + } + }, + { + "ph": "s", "id": 55276, "pid": 435, "tid": 435, "ts": 1776649791327849, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791327883, "dur": 0, + "args": { + "External id": 7340, + "cbid": 210, "correlation": 55303 + } + }, + { + "ph": "f", "id": 55303, "pid": 435, "tid": 435, "ts": 1776649791327883, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791327902, "dur": 44, + "args": { + "External id": 7340, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55304, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55304, "pid": 0, "tid": 7, "ts": 1776649791327902, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791327884, "dur": 6, + "args": { + "External id": 7340, + "cbid": 211, "correlation": 55304 + } + }, + { + "ph": "s", "id": 55304, "pid": 435, "tid": 435, "ts": 1776649791327884, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791327914, "dur": 0, + "args": { + "External id": 7342, + "cbid": 210, "correlation": 55331 + } + }, + { + "ph": "f", "id": 55331, "pid": 435, "tid": 435, "ts": 1776649791327914, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791327947, "dur": 48, + "args": { + "External id": 7342, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55332, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55332, "pid": 0, "tid": 7, "ts": 1776649791327947, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791327915, "dur": 6, + "args": { + "External id": 7342, + "cbid": 211, "correlation": 55332 + } + }, + { + "ph": "s", "id": 55332, "pid": 435, "tid": 435, "ts": 1776649791327915, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791327944, "dur": 0, + "args": { + "External id": 7344, + "cbid": 210, "correlation": 55359 + } + }, + { + "ph": "f", "id": 55359, "pid": 435, "tid": 435, "ts": 1776649791327944, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791327997, "dur": 46, + "args": { + "External id": 7344, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55360, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55360, "pid": 0, "tid": 7, "ts": 1776649791327997, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791327945, "dur": 6, + "args": { + "External id": 7344, + "cbid": 211, "correlation": 55360 + } + }, + { + "ph": "s", "id": 55360, "pid": 435, "tid": 435, "ts": 1776649791327945, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791327974, "dur": 0, + "args": { + "External id": 7346, + "cbid": 210, "correlation": 55387 + } + }, + { + "ph": "f", "id": 55387, "pid": 435, "tid": 435, "ts": 1776649791327974, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791328045, "dur": 46, + "args": { + "External id": 7346, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55388, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55388, "pid": 0, "tid": 7, "ts": 1776649791328045, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791327976, "dur": 6, + "args": { + "External id": 7346, + "cbid": 211, "correlation": 55388 + } + }, + { + "ph": "s", "id": 55388, "pid": 435, "tid": 435, "ts": 1776649791327976, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791328005, "dur": 0, + "args": { + "External id": 7348, + "cbid": 210, "correlation": 55415 + } + }, + { + "ph": "f", "id": 55415, "pid": 435, "tid": 435, "ts": 1776649791328005, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791328093, "dur": 41, + "args": { + "External id": 7348, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55416, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55416, "pid": 0, "tid": 7, "ts": 1776649791328093, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791328006, "dur": 6, + "args": { + "External id": 7348, + "cbid": 211, "correlation": 55416 + } + }, + { + "ph": "s", "id": 55416, "pid": 435, "tid": 435, "ts": 1776649791328006, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791328035, "dur": 0, + "args": { + "External id": 7350, + "cbid": 210, "correlation": 55443 + } + }, + { + "ph": "f", "id": 55443, "pid": 435, "tid": 435, "ts": 1776649791328035, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791328136, "dur": 46, + "args": { + "External id": 7350, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55444, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55444, "pid": 0, "tid": 7, "ts": 1776649791328136, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791328037, "dur": 6, + "args": { + "External id": 7350, + "cbid": 211, "correlation": 55444 + } + }, + { + "ph": "s", "id": 55444, "pid": 435, "tid": 435, "ts": 1776649791328037, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791328065, "dur": 0, + "args": { + "External id": 7352, + "cbid": 210, "correlation": 55471 + } + }, + { + "ph": "f", "id": 55471, "pid": 435, "tid": 435, "ts": 1776649791328065, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791328184, "dur": 49, + "args": { + "External id": 7352, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55472, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55472, "pid": 0, "tid": 7, "ts": 1776649791328184, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791328066, "dur": 6, + "args": { + "External id": 7352, + "cbid": 211, "correlation": 55472 + } + }, + { + "ph": "s", "id": 55472, "pid": 435, "tid": 435, "ts": 1776649791328066, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791328095, "dur": 0, + "args": { + "External id": 7354, + "cbid": 210, "correlation": 55499 + } + }, + { + "ph": "f", "id": 55499, "pid": 435, "tid": 435, "ts": 1776649791328095, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791328234, "dur": 51, + "args": { + "External id": 7354, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55500, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55500, "pid": 0, "tid": 7, "ts": 1776649791328234, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791328097, "dur": 6, + "args": { + "External id": 7354, + "cbid": 211, "correlation": 55500 + } + }, + { + "ph": "s", "id": 55500, "pid": 435, "tid": 435, "ts": 1776649791328097, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791328125, "dur": 0, + "args": { + "External id": 7356, + "cbid": 210, "correlation": 55527 + } + }, + { + "ph": "f", "id": 55527, "pid": 435, "tid": 435, "ts": 1776649791328125, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791328287, "dur": 47, + "args": { + "External id": 7356, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55528, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55528, "pid": 0, "tid": 7, "ts": 1776649791328287, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791328127, "dur": 6, + "args": { + "External id": 7356, + "cbid": 211, "correlation": 55528 + } + }, + { + "ph": "s", "id": 55528, "pid": 435, "tid": 435, "ts": 1776649791328127, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791328152, "dur": 187, + "args": { + "External id": 55534, + "cbid": 165, "correlation": 55534 + } + }, + { + "ph": "s", "id": 55534, "pid": 435, "tid": 435, "ts": 1776649791328152, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791328365, "dur": 4, + "args": { + "External id": 55543, + "cbid": 135, "correlation": 55543 + } + }, + { + "ph": "f", "id": 55543, "pid": 435, "tid": 435, "ts": 1776649791328365, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791328400, "dur": 1, + "args": { + "External id": 7358, + "cbid": 210, "correlation": 55569 + } + }, + { + "ph": "f", "id": 55569, "pid": 435, "tid": 435, "ts": 1776649791328400, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791328414, "dur": 40, + "args": { + "External id": 7358, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55570, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55570, "pid": 0, "tid": 7, "ts": 1776649791328414, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791328403, "dur": 9, + "args": { + "External id": 7358, + "cbid": 211, "correlation": 55570 + } + }, + { + "ph": "s", "id": 55570, "pid": 435, "tid": 435, "ts": 1776649791328403, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791328439, "dur": 0, + "args": { + "External id": 7360, + "cbid": 210, "correlation": 55597 + } + }, + { + "ph": "f", "id": 55597, "pid": 435, "tid": 435, "ts": 1776649791328439, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791328456, "dur": 43, + "args": { + "External id": 7360, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55598, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55598, "pid": 0, "tid": 7, "ts": 1776649791328456, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791328440, "dur": 6, + "args": { + "External id": 7360, + "cbid": 211, "correlation": 55598 + } + }, + { + "ph": "s", "id": 55598, "pid": 435, "tid": 435, "ts": 1776649791328440, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791328470, "dur": 0, + "args": { + "External id": 7362, + "cbid": 210, "correlation": 55625 + } + }, + { + "ph": "f", "id": 55625, "pid": 435, "tid": 435, "ts": 1776649791328470, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791328501, "dur": 46, + "args": { + "External id": 7362, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55626, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55626, "pid": 0, "tid": 7, "ts": 1776649791328501, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791328472, "dur": 6, + "args": { + "External id": 7362, + "cbid": 211, "correlation": 55626 + } + }, + { + "ph": "s", "id": 55626, "pid": 435, "tid": 435, "ts": 1776649791328472, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791328500, "dur": 0, + "args": { + "External id": 7364, + "cbid": 210, "correlation": 55653 + } + }, + { + "ph": "f", "id": 55653, "pid": 435, "tid": 435, "ts": 1776649791328500, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791328548, "dur": 47, + "args": { + "External id": 7364, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55654, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55654, "pid": 0, "tid": 7, "ts": 1776649791328548, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791328502, "dur": 5, + "args": { + "External id": 7364, + "cbid": 211, "correlation": 55654 + } + }, + { + "ph": "s", "id": 55654, "pid": 435, "tid": 435, "ts": 1776649791328502, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791328531, "dur": 0, + "args": { + "External id": 7366, + "cbid": 210, "correlation": 55681 + } + }, + { + "ph": "f", "id": 55681, "pid": 435, "tid": 435, "ts": 1776649791328531, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791328597, "dur": 47, + "args": { + "External id": 7366, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55682, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55682, "pid": 0, "tid": 7, "ts": 1776649791328597, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791328532, "dur": 5, + "args": { + "External id": 7366, + "cbid": 211, "correlation": 55682 + } + }, + { + "ph": "s", "id": 55682, "pid": 435, "tid": 435, "ts": 1776649791328532, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791328560, "dur": 0, + "args": { + "External id": 7368, + "cbid": 210, "correlation": 55709 + } + }, + { + "ph": "f", "id": 55709, "pid": 435, "tid": 435, "ts": 1776649791328560, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791328646, "dur": 49, + "args": { + "External id": 7368, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55710, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55710, "pid": 0, "tid": 7, "ts": 1776649791328646, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791328562, "dur": 6, + "args": { + "External id": 7368, + "cbid": 211, "correlation": 55710 + } + }, + { + "ph": "s", "id": 55710, "pid": 435, "tid": 435, "ts": 1776649791328562, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791328591, "dur": 0, + "args": { + "External id": 7370, + "cbid": 210, "correlation": 55737 + } + }, + { + "ph": "f", "id": 55737, "pid": 435, "tid": 435, "ts": 1776649791328591, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791328696, "dur": 44, + "args": { + "External id": 7370, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55738, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55738, "pid": 0, "tid": 7, "ts": 1776649791328696, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791328592, "dur": 6, + "args": { + "External id": 7370, + "cbid": 211, "correlation": 55738 + } + }, + { + "ph": "s", "id": 55738, "pid": 435, "tid": 435, "ts": 1776649791328592, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791328621, "dur": 0, + "args": { + "External id": 7372, + "cbid": 210, "correlation": 55765 + } + }, + { + "ph": "f", "id": 55765, "pid": 435, "tid": 435, "ts": 1776649791328621, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791328742, "dur": 45, + "args": { + "External id": 7372, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55766, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55766, "pid": 0, "tid": 7, "ts": 1776649791328742, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791328623, "dur": 6, + "args": { + "External id": 7372, + "cbid": 211, "correlation": 55766 + } + }, + { + "ph": "s", "id": 55766, "pid": 435, "tid": 435, "ts": 1776649791328623, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791328651, "dur": 0, + "args": { + "External id": 7374, + "cbid": 210, "correlation": 55793 + } + }, + { + "ph": "f", "id": 55793, "pid": 435, "tid": 435, "ts": 1776649791328651, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791328789, "dur": 46, + "args": { + "External id": 7374, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55794, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55794, "pid": 0, "tid": 7, "ts": 1776649791328789, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791328653, "dur": 5, + "args": { + "External id": 7374, + "cbid": 211, "correlation": 55794 + } + }, + { + "ph": "s", "id": 55794, "pid": 435, "tid": 435, "ts": 1776649791328653, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791328682, "dur": 0, + "args": { + "External id": 7376, + "cbid": 210, "correlation": 55821 + } + }, + { + "ph": "f", "id": 55821, "pid": 435, "tid": 435, "ts": 1776649791328682, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791328837, "dur": 49, + "args": { + "External id": 7376, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55822, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55822, "pid": 0, "tid": 7, "ts": 1776649791328837, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791328683, "dur": 6, + "args": { + "External id": 7376, + "cbid": 211, "correlation": 55822 + } + }, + { + "ph": "s", "id": 55822, "pid": 435, "tid": 435, "ts": 1776649791328683, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791328712, "dur": 0, + "args": { + "External id": 7378, + "cbid": 210, "correlation": 55849 + } + }, + { + "ph": "f", "id": 55849, "pid": 435, "tid": 435, "ts": 1776649791328712, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791328888, "dur": 44, + "args": { + "External id": 7378, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55850, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55850, "pid": 0, "tid": 7, "ts": 1776649791328888, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791328713, "dur": 6, + "args": { + "External id": 7378, + "cbid": 211, "correlation": 55850 + } + }, + { + "ph": "s", "id": 55850, "pid": 435, "tid": 435, "ts": 1776649791328713, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791328742, "dur": 0, + "args": { + "External id": 7380, + "cbid": 210, "correlation": 55877 + } + }, + { + "ph": "f", "id": 55877, "pid": 435, "tid": 435, "ts": 1776649791328742, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791328933, "dur": 49, + "args": { + "External id": 7380, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55878, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55878, "pid": 0, "tid": 7, "ts": 1776649791328933, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791328744, "dur": 5, + "args": { + "External id": 7380, + "cbid": 211, "correlation": 55878 + } + }, + { + "ph": "s", "id": 55878, "pid": 435, "tid": 435, "ts": 1776649791328744, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791328773, "dur": 0, + "args": { + "External id": 7382, + "cbid": 210, "correlation": 55905 + } + }, + { + "ph": "f", "id": 55905, "pid": 435, "tid": 435, "ts": 1776649791328773, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791328984, "dur": 46, + "args": { + "External id": 7382, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55906, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55906, "pid": 0, "tid": 7, "ts": 1776649791328984, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791328774, "dur": 5, + "args": { + "External id": 7382, + "cbid": 211, "correlation": 55906 + } + }, + { + "ph": "s", "id": 55906, "pid": 435, "tid": 435, "ts": 1776649791328774, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791328802, "dur": 0, + "args": { + "External id": 7384, + "cbid": 210, "correlation": 55933 + } + }, + { + "ph": "f", "id": 55933, "pid": 435, "tid": 435, "ts": 1776649791328802, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791329032, "dur": 48, + "args": { + "External id": 7384, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55934, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55934, "pid": 0, "tid": 7, "ts": 1776649791329032, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791328804, "dur": 6, + "args": { + "External id": 7384, + "cbid": 211, "correlation": 55934 + } + }, + { + "ph": "s", "id": 55934, "pid": 435, "tid": 435, "ts": 1776649791328804, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791328833, "dur": 0, + "args": { + "External id": 7386, + "cbid": 210, "correlation": 55961 + } + }, + { + "ph": "f", "id": 55961, "pid": 435, "tid": 435, "ts": 1776649791328833, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791329082, "dur": 49, + "args": { + "External id": 7386, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55962, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55962, "pid": 0, "tid": 7, "ts": 1776649791329082, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791328834, "dur": 6, + "args": { + "External id": 7386, + "cbid": 211, "correlation": 55962 + } + }, + { + "ph": "s", "id": 55962, "pid": 435, "tid": 435, "ts": 1776649791328834, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791328863, "dur": 0, + "args": { + "External id": 7388, + "cbid": 210, "correlation": 55989 + } + }, + { + "ph": "f", "id": 55989, "pid": 435, "tid": 435, "ts": 1776649791328863, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791329132, "dur": 46, + "args": { + "External id": 7388, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 55990, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 55990, "pid": 0, "tid": 7, "ts": 1776649791329132, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791328865, "dur": 6, + "args": { + "External id": 7388, + "cbid": 211, "correlation": 55990 + } + }, + { + "ph": "s", "id": 55990, "pid": 435, "tid": 435, "ts": 1776649791328865, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791328893, "dur": 0, + "args": { + "External id": 7390, + "cbid": 210, "correlation": 56017 + } + }, + { + "ph": "f", "id": 56017, "pid": 435, "tid": 435, "ts": 1776649791328893, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791329180, "dur": 49, + "args": { + "External id": 7390, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56018, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56018, "pid": 0, "tid": 7, "ts": 1776649791329180, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791328895, "dur": 6, + "args": { + "External id": 7390, + "cbid": 211, "correlation": 56018 + } + }, + { + "ph": "s", "id": 56018, "pid": 435, "tid": 435, "ts": 1776649791328895, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791328923, "dur": 0, + "args": { + "External id": 7392, + "cbid": 210, "correlation": 56045 + } + }, + { + "ph": "f", "id": 56045, "pid": 435, "tid": 435, "ts": 1776649791328923, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791329231, "dur": 46, + "args": { + "External id": 7392, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56046, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56046, "pid": 0, "tid": 7, "ts": 1776649791329231, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791328925, "dur": 5, + "args": { + "External id": 7392, + "cbid": 211, "correlation": 56046 + } + }, + { + "ph": "s", "id": 56046, "pid": 435, "tid": 435, "ts": 1776649791328925, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791328953, "dur": 0, + "args": { + "External id": 7394, + "cbid": 210, "correlation": 56073 + } + }, + { + "ph": "f", "id": 56073, "pid": 435, "tid": 435, "ts": 1776649791328953, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791329279, "dur": 43, + "args": { + "External id": 7394, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56074, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56074, "pid": 0, "tid": 7, "ts": 1776649791329279, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791328955, "dur": 6, + "args": { + "External id": 7394, + "cbid": 211, "correlation": 56074 + } + }, + { + "ph": "s", "id": 56074, "pid": 435, "tid": 435, "ts": 1776649791328955, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791328983, "dur": 0, + "args": { + "External id": 7396, + "cbid": 210, "correlation": 56101 + } + }, + { + "ph": "f", "id": 56101, "pid": 435, "tid": 435, "ts": 1776649791328983, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791329324, "dur": 49, + "args": { + "External id": 7396, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56102, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56102, "pid": 0, "tid": 7, "ts": 1776649791329324, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791328985, "dur": 5, + "args": { + "External id": 7396, + "cbid": 211, "correlation": 56102 + } + }, + { + "ph": "s", "id": 56102, "pid": 435, "tid": 435, "ts": 1776649791328985, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329014, "dur": 0, + "args": { + "External id": 7398, + "cbid": 210, "correlation": 56129 + } + }, + { + "ph": "f", "id": 56129, "pid": 435, "tid": 435, "ts": 1776649791329014, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791329375, "dur": 46, + "args": { + "External id": 7398, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56130, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56130, "pid": 0, "tid": 7, "ts": 1776649791329375, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329015, "dur": 6, + "args": { + "External id": 7398, + "cbid": 211, "correlation": 56130 + } + }, + { + "ph": "s", "id": 56130, "pid": 435, "tid": 435, "ts": 1776649791329015, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329044, "dur": 0, + "args": { + "External id": 7400, + "cbid": 210, "correlation": 56157 + } + }, + { + "ph": "f", "id": 56157, "pid": 435, "tid": 435, "ts": 1776649791329044, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791329423, "dur": 50, + "args": { + "External id": 7400, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56158, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56158, "pid": 0, "tid": 7, "ts": 1776649791329423, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329045, "dur": 6, + "args": { + "External id": 7400, + "cbid": 211, "correlation": 56158 + } + }, + { + "ph": "s", "id": 56158, "pid": 435, "tid": 435, "ts": 1776649791329045, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329074, "dur": 0, + "args": { + "External id": 7402, + "cbid": 210, "correlation": 56185 + } + }, + { + "ph": "f", "id": 56185, "pid": 435, "tid": 435, "ts": 1776649791329074, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791329475, "dur": 46, + "args": { + "External id": 7402, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56186, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56186, "pid": 0, "tid": 7, "ts": 1776649791329475, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329075, "dur": 5, + "args": { + "External id": 7402, + "cbid": 211, "correlation": 56186 + } + }, + { + "ph": "s", "id": 56186, "pid": 435, "tid": 435, "ts": 1776649791329075, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329104, "dur": 0, + "args": { + "External id": 7404, + "cbid": 210, "correlation": 56213 + } + }, + { + "ph": "f", "id": 56213, "pid": 435, "tid": 435, "ts": 1776649791329104, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791329522, "dur": 44, + "args": { + "External id": 7404, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56214, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56214, "pid": 0, "tid": 7, "ts": 1776649791329522, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329105, "dur": 6, + "args": { + "External id": 7404, + "cbid": 211, "correlation": 56214 + } + }, + { + "ph": "s", "id": 56214, "pid": 435, "tid": 435, "ts": 1776649791329105, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329134, "dur": 0, + "args": { + "External id": 7406, + "cbid": 210, "correlation": 56241 + } + }, + { + "ph": "f", "id": 56241, "pid": 435, "tid": 435, "ts": 1776649791329134, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791329568, "dur": 46, + "args": { + "External id": 7406, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56242, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56242, "pid": 0, "tid": 7, "ts": 1776649791329568, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329135, "dur": 5, + "args": { + "External id": 7406, + "cbid": 211, "correlation": 56242 + } + }, + { + "ph": "s", "id": 56242, "pid": 435, "tid": 435, "ts": 1776649791329135, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329163, "dur": 0, + "args": { + "External id": 7408, + "cbid": 210, "correlation": 56269 + } + }, + { + "ph": "f", "id": 56269, "pid": 435, "tid": 435, "ts": 1776649791329163, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791329616, "dur": 44, + "args": { + "External id": 7408, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56270, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56270, "pid": 0, "tid": 7, "ts": 1776649791329616, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329165, "dur": 6, + "args": { + "External id": 7408, + "cbid": 211, "correlation": 56270 + } + }, + { + "ph": "s", "id": 56270, "pid": 435, "tid": 435, "ts": 1776649791329165, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329193, "dur": 0, + "args": { + "External id": 7410, + "cbid": 210, "correlation": 56297 + } + }, + { + "ph": "f", "id": 56297, "pid": 435, "tid": 435, "ts": 1776649791329193, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791329662, "dur": 49, + "args": { + "External id": 7410, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56298, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56298, "pid": 0, "tid": 7, "ts": 1776649791329662, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329195, "dur": 6, + "args": { + "External id": 7410, + "cbid": 211, "correlation": 56298 + } + }, + { + "ph": "s", "id": 56298, "pid": 435, "tid": 435, "ts": 1776649791329195, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329223, "dur": 0, + "args": { + "External id": 7412, + "cbid": 210, "correlation": 56325 + } + }, + { + "ph": "f", "id": 56325, "pid": 435, "tid": 435, "ts": 1776649791329223, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791329712, "dur": 53, + "args": { + "External id": 7412, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56326, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56326, "pid": 0, "tid": 7, "ts": 1776649791329712, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329225, "dur": 6, + "args": { + "External id": 7412, + "cbid": 211, "correlation": 56326 + } + }, + { + "ph": "s", "id": 56326, "pid": 435, "tid": 435, "ts": 1776649791329225, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329253, "dur": 0, + "args": { + "External id": 7414, + "cbid": 210, "correlation": 56353 + } + }, + { + "ph": "f", "id": 56353, "pid": 435, "tid": 435, "ts": 1776649791329253, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791329768, "dur": 46, + "args": { + "External id": 7414, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56354, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56354, "pid": 0, "tid": 7, "ts": 1776649791329768, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329255, "dur": 6, + "args": { + "External id": 7414, + "cbid": 211, "correlation": 56354 + } + }, + { + "ph": "s", "id": 56354, "pid": 435, "tid": 435, "ts": 1776649791329255, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329283, "dur": 0, + "args": { + "External id": 7416, + "cbid": 210, "correlation": 56381 + } + }, + { + "ph": "f", "id": 56381, "pid": 435, "tid": 435, "ts": 1776649791329283, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791329815, "dur": 48, + "args": { + "External id": 7416, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56382, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56382, "pid": 0, "tid": 7, "ts": 1776649791329815, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329285, "dur": 6, + "args": { + "External id": 7416, + "cbid": 211, "correlation": 56382 + } + }, + { + "ph": "s", "id": 56382, "pid": 435, "tid": 435, "ts": 1776649791329285, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329314, "dur": 0, + "args": { + "External id": 7418, + "cbid": 210, "correlation": 56409 + } + }, + { + "ph": "f", "id": 56409, "pid": 435, "tid": 435, "ts": 1776649791329314, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791329865, "dur": 49, + "args": { + "External id": 7418, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56410, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56410, "pid": 0, "tid": 7, "ts": 1776649791329865, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329315, "dur": 6, + "args": { + "External id": 7418, + "cbid": 211, "correlation": 56410 + } + }, + { + "ph": "s", "id": 56410, "pid": 435, "tid": 435, "ts": 1776649791329315, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329343, "dur": 0, + "args": { + "External id": 7420, + "cbid": 210, "correlation": 56437 + } + }, + { + "ph": "f", "id": 56437, "pid": 435, "tid": 435, "ts": 1776649791329343, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791329916, "dur": 51, + "args": { + "External id": 7420, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56438, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56438, "pid": 0, "tid": 7, "ts": 1776649791329916, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329345, "dur": 6, + "args": { + "External id": 7420, + "cbid": 211, "correlation": 56438 + } + }, + { + "ph": "s", "id": 56438, "pid": 435, "tid": 435, "ts": 1776649791329345, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329374, "dur": 0, + "args": { + "External id": 7422, + "cbid": 210, "correlation": 56465 + } + }, + { + "ph": "f", "id": 56465, "pid": 435, "tid": 435, "ts": 1776649791329374, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791329969, "dur": 41, + "args": { + "External id": 7422, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56466, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56466, "pid": 0, "tid": 7, "ts": 1776649791329969, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329375, "dur": 5, + "args": { + "External id": 7422, + "cbid": 211, "correlation": 56466 + } + }, + { + "ph": "s", "id": 56466, "pid": 435, "tid": 435, "ts": 1776649791329375, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329403, "dur": 0, + "args": { + "External id": 7424, + "cbid": 210, "correlation": 56493 + } + }, + { + "ph": "f", "id": 56493, "pid": 435, "tid": 435, "ts": 1776649791329403, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791330011, "dur": 44, + "args": { + "External id": 7424, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56494, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56494, "pid": 0, "tid": 7, "ts": 1776649791330011, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329405, "dur": 6, + "args": { + "External id": 7424, + "cbid": 211, "correlation": 56494 + } + }, + { + "ph": "s", "id": 56494, "pid": 435, "tid": 435, "ts": 1776649791329405, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329433, "dur": 0, + "args": { + "External id": 7426, + "cbid": 210, "correlation": 56521 + } + }, + { + "ph": "f", "id": 56521, "pid": 435, "tid": 435, "ts": 1776649791329433, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791330057, "dur": 45, + "args": { + "External id": 7426, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56522, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56522, "pid": 0, "tid": 7, "ts": 1776649791330057, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329435, "dur": 6, + "args": { + "External id": 7426, + "cbid": 211, "correlation": 56522 + } + }, + { + "ph": "s", "id": 56522, "pid": 435, "tid": 435, "ts": 1776649791329435, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329464, "dur": 0, + "args": { + "External id": 7428, + "cbid": 210, "correlation": 56549 + } + }, + { + "ph": "f", "id": 56549, "pid": 435, "tid": 435, "ts": 1776649791329464, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791330104, "dur": 44, + "args": { + "External id": 7428, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56550, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56550, "pid": 0, "tid": 7, "ts": 1776649791330104, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329465, "dur": 6, + "args": { + "External id": 7428, + "cbid": 211, "correlation": 56550 + } + }, + { + "ph": "s", "id": 56550, "pid": 435, "tid": 435, "ts": 1776649791329465, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329494, "dur": 0, + "args": { + "External id": 7430, + "cbid": 210, "correlation": 56577 + } + }, + { + "ph": "f", "id": 56577, "pid": 435, "tid": 435, "ts": 1776649791329494, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791330150, "dur": 44, + "args": { + "External id": 7430, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56578, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56578, "pid": 0, "tid": 7, "ts": 1776649791330150, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329495, "dur": 6, + "args": { + "External id": 7430, + "cbid": 211, "correlation": 56578 + } + }, + { + "ph": "s", "id": 56578, "pid": 435, "tid": 435, "ts": 1776649791329495, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329523, "dur": 0, + "args": { + "External id": 7432, + "cbid": 210, "correlation": 56605 + } + }, + { + "ph": "f", "id": 56605, "pid": 435, "tid": 435, "ts": 1776649791329523, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791330195, "dur": 46, + "args": { + "External id": 7432, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56606, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56606, "pid": 0, "tid": 7, "ts": 1776649791330195, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329525, "dur": 6, + "args": { + "External id": 7432, + "cbid": 211, "correlation": 56606 + } + }, + { + "ph": "s", "id": 56606, "pid": 435, "tid": 435, "ts": 1776649791329525, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329553, "dur": 0, + "args": { + "External id": 7434, + "cbid": 210, "correlation": 56633 + } + }, + { + "ph": "f", "id": 56633, "pid": 435, "tid": 435, "ts": 1776649791329553, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791330243, "dur": 47, + "args": { + "External id": 7434, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56634, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56634, "pid": 0, "tid": 7, "ts": 1776649791330243, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329555, "dur": 6, + "args": { + "External id": 7434, + "cbid": 211, "correlation": 56634 + } + }, + { + "ph": "s", "id": 56634, "pid": 435, "tid": 435, "ts": 1776649791329555, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329583, "dur": 0, + "args": { + "External id": 7436, + "cbid": 210, "correlation": 56661 + } + }, + { + "ph": "f", "id": 56661, "pid": 435, "tid": 435, "ts": 1776649791329583, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791330292, "dur": 51, + "args": { + "External id": 7436, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56662, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56662, "pid": 0, "tid": 7, "ts": 1776649791330292, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329584, "dur": 6, + "args": { + "External id": 7436, + "cbid": 211, "correlation": 56662 + } + }, + { + "ph": "s", "id": 56662, "pid": 435, "tid": 435, "ts": 1776649791329584, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329613, "dur": 0, + "args": { + "External id": 7438, + "cbid": 210, "correlation": 56689 + } + }, + { + "ph": "f", "id": 56689, "pid": 435, "tid": 435, "ts": 1776649791329613, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791330345, "dur": 43, + "args": { + "External id": 7438, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56690, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56690, "pid": 0, "tid": 7, "ts": 1776649791330345, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329614, "dur": 6, + "args": { + "External id": 7438, + "cbid": 211, "correlation": 56690 + } + }, + { + "ph": "s", "id": 56690, "pid": 435, "tid": 435, "ts": 1776649791329614, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329643, "dur": 0, + "args": { + "External id": 7440, + "cbid": 210, "correlation": 56717 + } + }, + { + "ph": "f", "id": 56717, "pid": 435, "tid": 435, "ts": 1776649791329643, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791330390, "dur": 50, + "args": { + "External id": 7440, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56718, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56718, "pid": 0, "tid": 7, "ts": 1776649791330390, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329644, "dur": 5, + "args": { + "External id": 7440, + "cbid": 211, "correlation": 56718 + } + }, + { + "ph": "s", "id": 56718, "pid": 435, "tid": 435, "ts": 1776649791329644, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329673, "dur": 0, + "args": { + "External id": 7442, + "cbid": 210, "correlation": 56745 + } + }, + { + "ph": "f", "id": 56745, "pid": 435, "tid": 435, "ts": 1776649791329673, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791330442, "dur": 48, + "args": { + "External id": 7442, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56746, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56746, "pid": 0, "tid": 7, "ts": 1776649791330442, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329674, "dur": 5, + "args": { + "External id": 7442, + "cbid": 211, "correlation": 56746 + } + }, + { + "ph": "s", "id": 56746, "pid": 435, "tid": 435, "ts": 1776649791329674, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329703, "dur": 0, + "args": { + "External id": 7444, + "cbid": 210, "correlation": 56773 + } + }, + { + "ph": "f", "id": 56773, "pid": 435, "tid": 435, "ts": 1776649791329703, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791330492, "dur": 52, + "args": { + "External id": 7444, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56774, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56774, "pid": 0, "tid": 7, "ts": 1776649791330492, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329704, "dur": 6, + "args": { + "External id": 7444, + "cbid": 211, "correlation": 56774 + } + }, + { + "ph": "s", "id": 56774, "pid": 435, "tid": 435, "ts": 1776649791329704, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329733, "dur": 0, + "args": { + "External id": 7446, + "cbid": 210, "correlation": 56801 + } + }, + { + "ph": "f", "id": 56801, "pid": 435, "tid": 435, "ts": 1776649791329733, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791330546, "dur": 44, + "args": { + "External id": 7446, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56802, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56802, "pid": 0, "tid": 7, "ts": 1776649791330546, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329734, "dur": 6, + "args": { + "External id": 7446, + "cbid": 211, "correlation": 56802 + } + }, + { + "ph": "s", "id": 56802, "pid": 435, "tid": 435, "ts": 1776649791329734, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329763, "dur": 0, + "args": { + "External id": 7448, + "cbid": 210, "correlation": 56829 + } + }, + { + "ph": "f", "id": 56829, "pid": 435, "tid": 435, "ts": 1776649791329763, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791330592, "dur": 46, + "args": { + "External id": 7448, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56830, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56830, "pid": 0, "tid": 7, "ts": 1776649791330592, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329764, "dur": 6, + "args": { + "External id": 7448, + "cbid": 211, "correlation": 56830 + } + }, + { + "ph": "s", "id": 56830, "pid": 435, "tid": 435, "ts": 1776649791329764, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329793, "dur": 0, + "args": { + "External id": 7450, + "cbid": 210, "correlation": 56857 + } + }, + { + "ph": "f", "id": 56857, "pid": 435, "tid": 435, "ts": 1776649791329793, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791330640, "dur": 48, + "args": { + "External id": 7450, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56858, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56858, "pid": 0, "tid": 7, "ts": 1776649791330640, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329794, "dur": 6, + "args": { + "External id": 7450, + "cbid": 211, "correlation": 56858 + } + }, + { + "ph": "s", "id": 56858, "pid": 435, "tid": 435, "ts": 1776649791329794, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329823, "dur": 0, + "args": { + "External id": 7452, + "cbid": 210, "correlation": 56885 + } + }, + { + "ph": "f", "id": 56885, "pid": 435, "tid": 435, "ts": 1776649791329823, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791330690, "dur": 48, + "args": { + "External id": 7452, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56886, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56886, "pid": 0, "tid": 7, "ts": 1776649791330690, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329824, "dur": 6, + "args": { + "External id": 7452, + "cbid": 211, "correlation": 56886 + } + }, + { + "ph": "s", "id": 56886, "pid": 435, "tid": 435, "ts": 1776649791329824, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329854, "dur": 0, + "args": { + "External id": 7454, + "cbid": 210, "correlation": 56913 + } + }, + { + "ph": "f", "id": 56913, "pid": 435, "tid": 435, "ts": 1776649791329854, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791330740, "dur": 46, + "args": { + "External id": 7454, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56914, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56914, "pid": 0, "tid": 7, "ts": 1776649791330740, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329856, "dur": 6, + "args": { + "External id": 7454, + "cbid": 211, "correlation": 56914 + } + }, + { + "ph": "s", "id": 56914, "pid": 435, "tid": 435, "ts": 1776649791329856, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791329888, "dur": 0, + "args": { + "External id": 7456, + "cbid": 210, "correlation": 56941 + } + }, + { + "ph": "f", "id": 56941, "pid": 435, "tid": 435, "ts": 1776649791329888, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791330788, "dur": 43, + "args": { + "External id": 7456, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56942, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56942, "pid": 0, "tid": 7, "ts": 1776649791330788, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791329889, "dur": 5, + "args": { + "External id": 7456, + "cbid": 211, "correlation": 56942 + } + }, + { + "ph": "s", "id": 56942, "pid": 435, "tid": 435, "ts": 1776649791329889, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791329923, "dur": 4, + "args": { + "External id": 56953, + "cbid": 135, "correlation": 56953 + } + }, + { + "ph": "f", "id": 56953, "pid": 435, "tid": 435, "ts": 1776649791329923, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791329938, "dur": 902, + "args": { + "External id": 56958, + "cbid": 165, "correlation": 56958 + } + }, + { + "ph": "s", "id": 56958, "pid": 435, "tid": 435, "ts": 1776649791329938, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventElapsedTime", "pid": 435, "tid": 435, + "ts": 1776649791330844, "dur": 5, + "args": { + "External id": 56961, + "cbid": 139, "correlation": 56961 + } + }, + { + "ph": "f", "id": 56961, "pid": 435, "tid": 435, "ts": 1776649791330844, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791330882, "dur": 1, + "args": { + "External id": 7458, + "cbid": 210, "correlation": 56995 + } + }, + { + "ph": "f", "id": 56995, "pid": 435, "tid": 435, "ts": 1776649791330882, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791330895, "dur": 40, + "args": { + "External id": 7458, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 56996, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 56996, "pid": 0, "tid": 7, "ts": 1776649791330895, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791330884, "dur": 8, + "args": { + "External id": 7458, + "cbid": 211, "correlation": 56996 + } + }, + { + "ph": "s", "id": 56996, "pid": 435, "tid": 435, "ts": 1776649791330884, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791330918, "dur": 0, + "args": { + "External id": 7460, + "cbid": 210, "correlation": 57023 + } + }, + { + "ph": "f", "id": 57023, "pid": 435, "tid": 435, "ts": 1776649791330918, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791330937, "dur": 40, + "args": { + "External id": 7460, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57024, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57024, "pid": 0, "tid": 7, "ts": 1776649791330937, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791330920, "dur": 6, + "args": { + "External id": 7460, + "cbid": 211, "correlation": 57024 + } + }, + { + "ph": "s", "id": 57024, "pid": 435, "tid": 435, "ts": 1776649791330920, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791330949, "dur": 0, + "args": { + "External id": 7462, + "cbid": 210, "correlation": 57051 + } + }, + { + "ph": "f", "id": 57051, "pid": 435, "tid": 435, "ts": 1776649791330949, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791330980, "dur": 43, + "args": { + "External id": 7462, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57052, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57052, "pid": 0, "tid": 7, "ts": 1776649791330980, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791330951, "dur": 6, + "args": { + "External id": 7462, + "cbid": 211, "correlation": 57052 + } + }, + { + "ph": "s", "id": 57052, "pid": 435, "tid": 435, "ts": 1776649791330951, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791330980, "dur": 0, + "args": { + "External id": 7464, + "cbid": 210, "correlation": 57079 + } + }, + { + "ph": "f", "id": 57079, "pid": 435, "tid": 435, "ts": 1776649791330980, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791331024, "dur": 50, + "args": { + "External id": 7464, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57080, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57080, "pid": 0, "tid": 7, "ts": 1776649791331024, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791330982, "dur": 6, + "args": { + "External id": 7464, + "cbid": 211, "correlation": 57080 + } + }, + { + "ph": "s", "id": 57080, "pid": 435, "tid": 435, "ts": 1776649791330982, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791331011, "dur": 0, + "args": { + "External id": 7466, + "cbid": 210, "correlation": 57107 + } + }, + { + "ph": "f", "id": 57107, "pid": 435, "tid": 435, "ts": 1776649791331011, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791331076, "dur": 46, + "args": { + "External id": 7466, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57108, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57108, "pid": 0, "tid": 7, "ts": 1776649791331076, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791331012, "dur": 6, + "args": { + "External id": 7466, + "cbid": 211, "correlation": 57108 + } + }, + { + "ph": "s", "id": 57108, "pid": 435, "tid": 435, "ts": 1776649791331012, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791331041, "dur": 0, + "args": { + "External id": 7468, + "cbid": 210, "correlation": 57135 + } + }, + { + "ph": "f", "id": 57135, "pid": 435, "tid": 435, "ts": 1776649791331041, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791331124, "dur": 46, + "args": { + "External id": 7468, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57136, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57136, "pid": 0, "tid": 7, "ts": 1776649791331124, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791331042, "dur": 6, + "args": { + "External id": 7468, + "cbid": 211, "correlation": 57136 + } + }, + { + "ph": "s", "id": 57136, "pid": 435, "tid": 435, "ts": 1776649791331042, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791331070, "dur": 0, + "args": { + "External id": 7470, + "cbid": 210, "correlation": 57163 + } + }, + { + "ph": "f", "id": 57163, "pid": 435, "tid": 435, "ts": 1776649791331070, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791331172, "dur": 46, + "args": { + "External id": 7470, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57164, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57164, "pid": 0, "tid": 7, "ts": 1776649791331172, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791331072, "dur": 5, + "args": { + "External id": 7470, + "cbid": 211, "correlation": 57164 + } + }, + { + "ph": "s", "id": 57164, "pid": 435, "tid": 435, "ts": 1776649791331072, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791331100, "dur": 0, + "args": { + "External id": 7472, + "cbid": 210, "correlation": 57191 + } + }, + { + "ph": "f", "id": 57191, "pid": 435, "tid": 435, "ts": 1776649791331100, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791331220, "dur": 49, + "args": { + "External id": 7472, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57192, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57192, "pid": 0, "tid": 7, "ts": 1776649791331220, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791331101, "dur": 6, + "args": { + "External id": 7472, + "cbid": 211, "correlation": 57192 + } + }, + { + "ph": "s", "id": 57192, "pid": 435, "tid": 435, "ts": 1776649791331101, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791331130, "dur": 0, + "args": { + "External id": 7474, + "cbid": 210, "correlation": 57219 + } + }, + { + "ph": "f", "id": 57219, "pid": 435, "tid": 435, "ts": 1776649791331130, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791331271, "dur": 54, + "args": { + "External id": 7474, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57220, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57220, "pid": 0, "tid": 7, "ts": 1776649791331271, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791331131, "dur": 5, + "args": { + "External id": 7474, + "cbid": 211, "correlation": 57220 + } + }, + { + "ph": "s", "id": 57220, "pid": 435, "tid": 435, "ts": 1776649791331131, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791331160, "dur": 0, + "args": { + "External id": 7476, + "cbid": 210, "correlation": 57247 + } + }, + { + "ph": "f", "id": 57247, "pid": 435, "tid": 435, "ts": 1776649791331160, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791331327, "dur": 46, + "args": { + "External id": 7476, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57248, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57248, "pid": 0, "tid": 7, "ts": 1776649791331327, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791331161, "dur": 5, + "args": { + "External id": 7476, + "cbid": 211, "correlation": 57248 + } + }, + { + "ph": "s", "id": 57248, "pid": 435, "tid": 435, "ts": 1776649791331161, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaDeviceSynchronize", "pid": 435, "tid": 435, + "ts": 1776649791331186, "dur": 191, + "args": { + "External id": 57254, + "cbid": 165, "correlation": 57254 + } + }, + { + "ph": "s", "id": 57254, "pid": 435, "tid": 435, "ts": 1776649791331186, + "cat": "ac2g", "name": "ac2g" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaEventRecord", "pid": 435, "tid": 435, + "ts": 1776649791331399, "dur": 3, + "args": { + "External id": 57263, + "cbid": 135, "correlation": 57263 + } + }, + { + "ph": "f", "id": 57263, "pid": 435, "tid": 435, "ts": 1776649791331399, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaOccupancyMaxActiveBlocksPerMultiprocessor", "pid": 435, "tid": 435, + "ts": 1776649791331428, "dur": 1, + "args": { + "External id": 7478, + "cbid": 210, "correlation": 57289 + } + }, + { + "ph": "f", "id": 57289, "pid": 435, "tid": 435, "ts": 1776649791331428, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "kernel", "name": "ampere_fp16_s16816gemm_fp16_128x64_ldg8_f2f_stages_64x4_nn", "pid": 0, "tid": 7, + "ts": 1776649791331440, "dur": 40, + "args": { + "External id": 7478, + "queued": 0, "device": 0, "context": 1, + "stream": 7, "correlation": 57290, + "registers per thread": 150, + "shared memory": 98304, + "blocks per SM": 0.592593, + "warps per SM": 2.370370, + "grid": [64, 1, 1], + "block": [128, 1, 1], + "est. achieved occupancy %": 0 + } + }, + { + "ph": "f", "id": 57290, "pid": 0, "tid": 7, "ts": 1776649791331440, + "cat": "ac2g", "name": "ac2g", "bp": "e" + }, + { + "ph": "X", "cat": "cuda_runtime", "name": "cudaLaunchKernel", "pid": 435, "tid": 435, + "ts": 1776649791331430, "dur": 8, + "args": { + "External id": 7478, + "cbid": 211, "correlation": 57290 + } + }, + { + "ph": "s", "id": 57290, "pid": 435, "tid": 435, "ts": 1776649791331430, + "cat": "ac2g", "name": "ac2g" + }, + { + "name": "process_name", "ph": "M", "ts": 1776649789501904, "pid": 435, "tid": 0, + "args": { + "name": "python" + } + }, + { + "name": "process_labels", "ph": "M", "ts": 1776649789501904, "pid": 435, "tid": 0, + "args": { + "labels": "CPU" + } + }, + { + "name": "process_sort_index", "ph": "M", "ts": 1776649789501904, "pid": 435, "tid": 0, + "args": { + "sort_index": 435 + } + }, + { + "name": "process_name", "ph": "M", "ts": 1776649789501904, "pid": 0, "tid": 0, + "args": { + "name": "python" + } + }, + { + "name": "process_labels", "ph": "M", "ts": 1776649789501904, "pid": 0, "tid": 0, + "args": { + "labels": "GPU 0" + } + }, + { + "name": "process_sort_index", "ph": "M", "ts": 1776649789501904, "pid": 0, "tid": 0, + "args": { + "sort_index": 16777216 + } + }, + { + "name": "process_name", "ph": "M", "ts": 1776649789501904, "pid": 1, "tid": 0, + "args": { + "name": "python" + } + }, + { + "name": "process_labels", "ph": "M", "ts": 1776649789501904, "pid": 1, "tid": 0, + "args": { + "labels": "GPU 1" + } + }, + { + "name": "process_sort_index", "ph": "M", "ts": 1776649789501904, "pid": 1, "tid": 0, + "args": { + "sort_index": 16777217 + } + }, + { + "name": "process_name", "ph": "M", "ts": 1776649789501904, "pid": 2, "tid": 0, + "args": { + "name": "python" + } + }, + { + "name": "process_labels", "ph": "M", "ts": 1776649789501904, "pid": 2, "tid": 0, + "args": { + "labels": "GPU 2" + } + }, + { + "name": "process_sort_index", "ph": "M", "ts": 1776649789501904, "pid": 2, "tid": 0, + "args": { + "sort_index": 16777218 + } + }, + { + "name": "process_name", "ph": "M", "ts": 1776649789501904, "pid": 3, "tid": 0, + "args": { + "name": "python" + } + }, + { + "name": "process_labels", "ph": "M", "ts": 1776649789501904, "pid": 3, "tid": 0, + "args": { + "labels": "GPU 3" + } + }, + { + "name": "process_sort_index", "ph": "M", "ts": 1776649789501904, "pid": 3, "tid": 0, + "args": { + "sort_index": 16777219 + } + }, + { + "name": "process_name", "ph": "M", "ts": 1776649789501904, "pid": 4, "tid": 0, + "args": { + "name": "python" + } + }, + { + "name": "process_labels", "ph": "M", "ts": 1776649789501904, "pid": 4, "tid": 0, + "args": { + "labels": "GPU 4" + } + }, + { + "name": "process_sort_index", "ph": "M", "ts": 1776649789501904, "pid": 4, "tid": 0, + "args": { + "sort_index": 16777220 + } + }, + { + "name": "process_name", "ph": "M", "ts": 1776649789501904, "pid": 5, "tid": 0, + "args": { + "name": "python" + } + }, + { + "name": "process_labels", "ph": "M", "ts": 1776649789501904, "pid": 5, "tid": 0, + "args": { + "labels": "GPU 5" + } + }, + { + "name": "process_sort_index", "ph": "M", "ts": 1776649789501904, "pid": 5, "tid": 0, + "args": { + "sort_index": 16777221 + } + }, + { + "name": "process_name", "ph": "M", "ts": 1776649789501904, "pid": 6, "tid": 0, + "args": { + "name": "python" + } + }, + { + "name": "process_labels", "ph": "M", "ts": 1776649789501904, "pid": 6, "tid": 0, + "args": { + "labels": "GPU 6" + } + }, + { + "name": "process_sort_index", "ph": "M", "ts": 1776649789501904, "pid": 6, "tid": 0, + "args": { + "sort_index": 16777222 + } + }, + { + "name": "process_name", "ph": "M", "ts": 1776649789501904, "pid": 7, "tid": 0, + "args": { + "name": "python" + } + }, + { + "name": "process_labels", "ph": "M", "ts": 1776649789501904, "pid": 7, "tid": 0, + "args": { + "labels": "GPU 7" + } + }, + { + "name": "process_sort_index", "ph": "M", "ts": 1776649789501904, "pid": 7, "tid": 0, + "args": { + "sort_index": 16777223 + } + }, + { + "name": "thread_name", "ph": "M", "ts": 1776649789501904, "pid": 0, "tid": 7, + "args": { + "name": "stream 7 " + } + }, + { + "name": "thread_sort_index", "ph": "M", "ts": 1776649789501904, "pid": 0, "tid": 7, + "args": { + "sort_index": 7 + } + }, + { + "name": "thread_name", "ph": "M", "ts": 1776649789501904, "pid": 435, "tid": 435, + "args": { + "name": "thread 435 (pt_main_thread)" + } + }, + { + "name": "thread_sort_index", "ph": "M", "ts": 1776649789501904, "pid": 435, "tid": 435, + "args": { + "sort_index": 435 + } + }, + { + "name": "thread_name", "ph": "M", "ts": 1776649789501904, "pid": 435, "tid": 435, + "args": { + "name": "thread 435 (pt_main_thread)" + } + }, + { + "name": "thread_sort_index", "ph": "M", "ts": 1776649789501904, "pid": 435, "tid": 435, + "args": { + "sort_index": 435 + } + }, + { + "ph": "X", "cat": "Trace", "ts": 1776649789501790, "dur": 3843429, + "pid": "Spans", "tid": "PyTorch Profiler", + "name": "PyTorch Profiler (0)", + "args": { + "Op count": 0 + } + }, + { + "name": "process_sort_index", "ph": "M", "ts": 1776649789501790, + "pid": "Spans", "tid": 0, + "args": { + "sort_index": 536870912 + } + }, + { + "name": "Iteration Start: PyTorch Profiler", "ph": "i", "s": "g", + "pid": "Traces", "tid": "Trace PyTorch Profiler", "ts": 1776649789501790 + }, + { + "name": "Record Window End", "ph": "i", "s": "g", + "pid": "", "tid": "", "ts": 1776649793364821 + } + ], + "traceName": "docs/profiling/marlin/20260420T014943Z/trace.json" +} \ No newline at end of file From 8bdb85e9d34bb0ce97d0dc69903ceeab79cc85fe Mon Sep 17 00:00:00 2001 From: swordfish-autoresearch Date: Mon, 20 Apr 2026 01:50:20 +0000 Subject: [PATCH 2/2] autoresearch: link PR in INDEX.md --- docs/profiling/INDEX.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/docs/profiling/INDEX.md b/docs/profiling/INDEX.md index 732fc21..e1bcc00 100644 --- a/docs/profiling/INDEX.md +++ b/docs/profiling/INDEX.md @@ -5,4 +5,4 @@ Newest first. PR column links to the draft PR carrying the artifacts. | timestamp (UTC) | source SHA | shapes | impls | GPU | 8b-b1 marlin TFLOPS | run dir | PR | |---|---|---|---|---|---|---|---| -| 20260420T014943Z | `20ab7f3` | voice | fp16,marlin | NVIDIA A100-SXM4-80GB | 0.7 | [`20260420T014943Z/`](./marlin/20260420T014943Z/) | _pending_ | +| 20260420T014943Z | `20ab7f3` | voice | fp16,marlin | NVIDIA A100-SXM4-80GB | 0.7 | [`20260420T014943Z/`](./marlin/20260420T014943Z/) | [link](https://github.com/chokevin/swordfish/pull/2) |